#  @code4ai Discover AI Discover AI posts on YouTube about ai, llm, artificial intelligence, language the most. They currently have [------] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours. ### Engagements: [------] [#](/creator/youtube::UCfOvNb3xj28SNqPQ_JIbumg/interactions)  - [--] Week [------] +8.50% - [--] Month [-------] -21% - [--] Months [-------] -32% - [--] Year [---------] +112% ### Mentions: [--] [#](/creator/youtube::UCfOvNb3xj28SNqPQ_JIbumg/posts_active)  - [--] Month [--] -47% - [--] Months [---] -10% - [--] Year [---] +112% ### Followers: [------] [#](/creator/youtube::UCfOvNb3xj28SNqPQ_JIbumg/followers)  - [--] Week [------] +0.35% - [--] Month [------] +1.30% - [--] Months [------] +11% - [--] Year [------] +48% ### CreatorRank: [-------] [#](/creator/youtube::UCfOvNb3xj28SNqPQ_JIbumg/influencer_rank)  ### Social Influence **Social category influence** [technology brands](/list/technology-brands) [stocks](/list/stocks) [countries](/list/countries) [travel destinations](/list/travel-destinations) [social networks](/list/social-networks) [finance](/list/finance) [currencies](/list/currencies) [cryptocurrencies](/list/cryptocurrencies) [automotive brands](/list/automotive-brands) [musicians](/list/musicians) **Social topic influence** [ai](/topic/ai), [llm](/topic/llm) #118, [artificial intelligence](/topic/artificial-intelligence) #331, [language](/topic/language) #950, [artificial](/topic/artificial), [$googl](/topic/$googl), [open ai](/topic/open-ai), [university of](/topic/university-of), [#ai](/topic/#ai), [youtube](/topic/youtube) **Top accounts mentioned or mentioned by** [@mit](/creator/undefined) [@stanford](/creator/undefined) [@openai](/creator/undefined) [@ucberkeley](/creator/undefined) [@googledeepmind](/creator/undefined) [@tsinghuauniversityofficial](/creator/undefined) [@princeton](/creator/undefined) [@harvard](/creator/undefined) [@nvidia](/creator/undefined) [@google](/creator/undefined) [@cmu](/creator/undefined) [@huggingface](/creator/undefined) [@anthropicai](/creator/undefined) [@salesforce](/creator/undefined) [@penn](/creator/undefined) [@oxforduniversity](/creator/undefined) [@nvidiaal](/creator/undefined) [@yale](/creator/undefined) [@ucla](/creator/undefined) [@pekinguniversity1898](/creator/undefined) **Top assets mentioned** [Alphabet Inc Class A (GOOGL)](/topic/$googl) [Microsoft Corp. (MSFT)](/topic/microsoft) [Salesforce Inc (CRM)](/topic/salesforce) [CyberConnect (CYBER)](/topic/cyber) ### Top Social Posts Top posts by engagements in the last [--] hours "The $10T AI Economy: New Smart Protocol Emerges The Smart Contract between AI Agents New AI Protocol emerges. The future Economy of Humans and AI agents on a global competitive marketplace with decentralized auctions. Human Intelligence and machine intelligence are priced and offered globally. Is this our future Is this why we invented AI All rights w/ authors: "Intelligent AI Delegation" Nenad Tomaev1 Matija Franklin1 and Simon Osindero1 from Google DeepMind #economy #aieconomy #aibusiness #aiproductivity #aiperformance #nextgenai artificial intelligence AI models LLM VLM VLA Multi-modal" [YouTube Link](https://youtube.com/watch?v=FDixNQunPTc) 2026-02-15T14:15Z 85.2K followers, [----] engagements "NEW GLM-5 vs MiniMax-2.5: NEW = BETTER Artificial Intelligence: Two new AI models (GLM-5 vs MiniMax-2.5) are tested - side by side - on a non-public causal reasoning test to evaluate their performance. Live recording of real-world performance of the latest agent optimized LLMs. 00:00 GLM-5 and MiniMax-2.5 02:13 Start Live TEST 12:35 GLM-5 and MiniMax-2.5 CRASH 14:07 First Solution by GLM-5 14:48 Successful GLM-5 Evaluation Run 16:33 GLM-5 Evaluation by MiniMax-2.5 18:03 MiniMax-2.5 FAILS Validation #aiexplained #aitech #ainews artificial intelligence AI models LLM VLM VLA Multi-modal model" [YouTube Link](https://youtube.com/watch?v=0ao20vRWgis) 2026-02-13T14:30Z 85.2K followers, [----] engagements "The "OPENNESS" of corporate LLM models: A corporate board meeting Ever wondered what happens behind the colossal closed doors of tech behemoths like ClosedAI Microsocks and METAH Join me for a humorous parody of a hypothetical management board meeting as they grapple with criticisms of their open-source models Get ready for witty repartees exaggerations and the most comically blown-out-of-proportion reactions - all with a pinch of reality. Remember it's all in good fun as we salute these trailblazers for their incredible contributions to the AI world. Subscribe and stay tuned for an episode" [YouTube Link](https://youtube.com/watch?v=C9rqIiDEc8E) 2023-07-30T12:00Z 85.2K followers, [----] engagements "AI Belief Functions: Deciding Under Absolute Uncertainty Engines of Intelligence: The Definition and Necessity of AI Agents. Basic mathematical explanations for any CEO of any multinational strategy and management consulting firm: What AI agents are and why we need them. And why AI (re-)acts in absolute uncertainty with self-updating belief functions. Plus: A simple explanation for CEOs (McKinsey EY .) what are AI agents and why they operate in absolute uncertainty with a simple mathematical probability distribution for the most difficult client jobs. .maybe not such a good idea #aiexplained" [YouTube Link](https://youtube.com/watch?v=QbbTQ-B_v2A) 2026-02-14T14:00Z 85.2K followers, [----] engagements "Forget LLM: MIT's New RLM (Phase Shift in AI) Weve been misled by the promise of "infinite" context windows: new AI research proves that "Context Rot" is destroying reasoning capabilities as inputs scale. But a groundbreaking paper from MIT introduces a radical solution: Recursive Language Models (RLMs). Instead of blindly force-feeding data into a single Transformer RLMs act as a Neurosymbolic Operating System writing Python code to mechanically split massive datasets and recursively "spawn" fresh model instances to process them. The result is a staggering leap in performance: on quadratic" [YouTube Link](https://youtube.com/watch?v=mtRJmIup3b8) 2026-01-04T14:01Z 85.2K followers, 30K engagements "Accelerate pandas df: DASK [----] = superfast Python DASK scales numpy arrays and pandas dataframes efficiently. Utilizes all CPU cores / threads. Video shows you that four lines of code set up a local DASK cluster automatically on my Win10 PC to supercharge python - even on a single CPU. Optimize your data input pipeline to your transformer models (AI) with a local cluster configuration utilizing your system resources. Speed tests on numpy arrays w/ DASK. Speed tests on pandas dataframes w/ DASK. #code_in_real_time #real_time_coding #DASK #parallelize_python #cluster #JupyterLab #python" [YouTube Link](https://youtube.com/watch?v=1RKobQLAk7E) 2021-07-27T11:15Z 84.7K followers, [---] engagements "Smarter AI Gradients: How Agents Learn to Think Exploration is essential in reinforcement learning (RL) as an AI agent relies on trial and error to learn an optimal policy. However when rewards are sparse naive exploration strategies like noise injection are often insufficient. Intrinsic rewards can also provide principled guidance for exploration by for example combining them with extrinsic rewards to optimize a policy or using them to train sub-policies for hierarchical learning. However the former approach suffers from unstable credit assignment while the latter exhibits sample" [YouTube Link](https://youtube.com/watch?v=2POdg38T1Ec) 2026-01-31T14:30Z 85.2K followers, [----] engagements "Grand Unified Theory of AI (Explained w/ Google ADK) Grand Unified Theory of AI Multi-Agent Systems (Explained w/ Google Context ADK). We present a rigorous analysis of Google Clouds "Mathematical Framing for Different Agent Strategies" which proposes a unified probabilistic formulation to quantify agent behavior beyond empirical benchmarks. To build production-grade agents that are reliable efficient and debuggable the industry is exploring a new discipline: Context engineering treating context as a first-class system with its own architecture lifecycle and constraints. Based on the" [YouTube Link](https://youtube.com/watch?v=5eBU_9fyPFs) 2025-12-08T14:00Z 84.7K followers, [----] engagements "Activate GROKKING NOW - Performance Phase of LLMs (II) Grokking or the sudden generalization by AI models to new knowledge - that occurs after prolonged overfitting in LLMs is a surprising phenomenon that has challenged our understanding of deep learning and AI in general. While a lot of progress has been made in understanding grokking finally we get some answers -we have been waiting for [--] months to be discovered. GROKKING - Finally understood AUDIO: With the automatic audio dubbing from YouTube /Google you hear a synthetic voice in your regional language. To hear my original voice in" [YouTube Link](https://youtube.com/watch?v=H3OofROzlA0) 2025-01-15T15:15Z 84.7K followers, [----] engagements "LLM Quantization (Ollama LM Studio): Any Performance Drop TEST A NEW benchmark and guide which quantization models to use locally on your PC or laptop. Either in Ollama or in LM Studio whenever I want to download a LLM or VLM I have to use a quantized Ai model because of the limited VRAM on my NVIDIA GPU my local AI infra. But what quantized model should I choose What quantization really hurt the overall model Performance What quant is recommended The Ollama version of the latest version of deepseek-r1:671b-0528-q4_K_M (404GB) is available here:" [YouTube Link](https://youtube.com/watch?v=Jj0R6V5bYLY) 2025-08-26T14:00Z 84.7K followers, [----] engagements "New Discovery: LLMs have a Performance Phase Grokking is a new phase in the performance of LLMs. Starting with arithmetic operations we analyze the patterns in the embedded space of Transformers. Grokking refers to a phenomenon where after extensive training beyond typical saturation points transformers can generalize effectively to unseen data achieving high performance long after initial overfitting occurs. This discovery challenges conventional wisdom about early stopping to prevent overfitting revealing that extended training can lead to superior generalization. The video highlights" [YouTube Link](https://youtube.com/watch?v=QgOeWbW0jeA) 2024-06-04T12:00Z 84.7K followers, 16.8K engagements "POPE RL Curriculum Learning (CMU) RL doesn't teach the AI model new facts; POPE RL tries to steer the model's internal attention heads to attend to the correct latent subspaces (like mathematical reasoning) rather than the incorrect ones (casual chat or confusion) which cause the "Cold Start" problem. Further insights into the "Valley of Death" for RL in AI (zero gradients zero rewards). All rights w/ authors: POPE: Learning to Reason on Hard Problems via Privileged On-Policy Exploration Yuxiao Qu*1 Amrith Setlur*1 Virginia Smith1 Ruslan Salakhutdinov1 Aviral Kumar1 from [--] Carnegie Mellon" [YouTube Link](https://youtube.com/watch?v=Rfcyl-S0kfg) 2026-01-30T14:15Z 85.2K followers, [----] engagements "Stanford: Do Not use [--] AI Agents: They will Fail (CooperBench) In depth vido why Why Two AI Coding Agents Are Worse Than One. Stanford Univ directly attacks the industry hype around Multi-Agent Systems (MAS) and promises a data-backed explanation for why adding compute (AI coding agents) actually degrades performance. All rights w/ authors: "CooperBench: Why Coding Agents Cannot be Your Teammates Yet" Arpandeep Khatua1 Hao Zhu1 Peter Tran2 Arya Prabhudesai2 Frederic Sadrieh2 Johann K. Lieberwirth2 Xinkai Yu1 Yicheng Fu1 Michael J. Ryan1 Jiaxin Pei1 Diyi Yang1 from [--] Stanford University [--] SAP" [YouTube Link](https://youtube.com/watch?v=Sy75zdfZIEM) 2026-01-26T14:15Z 85.2K followers, [----] engagements "Jupyter AI: Generative AI in your Notebook Jupyter AI explained in both modes: AI magic commands and the Chat UI for generative AI models form Anthropic Cohere OpenAI . #ai #codegeneration #generativeai" [YouTube Link](https://youtube.com/watch?v=_Llo2yyS7FU) 2023-08-15T12:00Z 84.7K followers, [----] engagements "NEW Knowledge Graph based RAG: SimGRAG (no training) Excellent new Knowledge Graph based RAG system called SimGraphRAG or simply SimGRAG. Overview of our four classical KG-based RAG systems and the new SimGRAG which outperform them. Short technical deep dive into the new methods and algorithms plus code via GitHub repo. All rights w/ authors: SimGRAG: Leveraging Similar Subgraphs for Knowledge Graphs Driven Retrieval-Augmented Generation Yuzheng Cai Zhenyue Guo Yiwen Pei Wanrui Bian Weiguo Zheng from Fudan University #airesearch #knowledgegraph #science #aiagents #graph" [YouTube Link](https://youtube.com/watch?v=aPsfAkrkma0) 2024-12-25T15:00Z 84.8K followers, 12.6K engagements "Scaling AI Reasoning: MCTS in ICL for Small LM Meta-reasoning Unleashed: A New ICL Paradigm Beyond Simple Examples: Next-Level ICL Reasoning ICL Reinvented: Harnessing MCTS Insights From Steps to Strategies: Thought Cards in ICL High-Level Paths: The New Face of In-Context Learning Scaling Reasoning: MCTS and Beyond in ICL VOC computing ICL: Optimizing Reasoning Paths Abstract Templates: The Future of ICL Strategies ICL Meets MCTS: Deep Reasoning Upgrades Evolving ICL: Distilling Optimal Thought Patterns #airesearch #chatgpt #programming #coding #reasoning #logic #cognitivefunction" [YouTube Link](https://youtube.com/watch?v=bSLc-dhn9vg) 2024-12-08T15:00Z 84.8K followers, [----] engagements "LLM + Knowledge Graph + GNN = TRUTH by AI GraphCHECK: Improving Factuality in LLM Outputs w/ Graph Neural Networks for Knowledge-Graph Enhanced Verification Can AI Find Truth The Power of LLMs Knowledge Graphs and GNNs CODE implementation of GraphCHECK available: https://anonymous.4open.science/r/GraphCheck-1D43/README.md PYTHON implementation for Graph Attention Network: https://anonymous.4open.science/r/GraphCheck-1D43/model/gnn.py All rights w/ authors: "GraphCheck: Breaking Long-Term Text Barriers with Extracted Knowledge Graph-Powered Fact-Checking" Yingjian Chen1 Haoran Liu2 Yinhong" [YouTube Link](https://youtube.com/watch?v=fpFA0AOfBYI) 2025-02-28T15:15Z 84.8K followers, [----] engagements "NEW Multi-Agent Protocol (REP by MIT): 100+ Agents Caught a cold. Sorry for my wet pronunciation . All rights W/ authors: Which LLM MultiAgent Protocol to Choose Hongyi Du1 Jiaqi Su2 Jisen Li1 Lijie Ding3 Yingxuan Yang2 Peixuan Han1 Xiangru Tang4 Kunlun Zhu1 Jiaxuan You1 from [--] University of Illinois UrbanaChampaign [--] Shanghai Jiao Tong University [--] Oak Ridge National Laboratory [--] Yale University arXiv:2510.16572 @yale @princeton Ripple Effect Protocol: Coordinating Agent Populations Ayush Chopra12 Aman Sharma2 Feroz Ahmad2 Luca Muscariello3 Vijoy Pandey3 Ramesh Raskar12 from [--] Massachusetts" [YouTube Link](https://youtube.com/watch?v=nfH9nMWCRi8) 2025-10-24T12:15Z 84.7K followers, [----] engagements "Salesforce: New AI Agents That Doubt Themselves (AUQ) Salesforce Research has just operationalised Kahnemans "System [--] vs. System 2" framework directly into LLM architectures to solve the "Spiral of Hallucination" in long-horizon tasks for AI agents. In this breakdown we dissect Agentic Uncertainty Quantification (AUQ): a training-free framework that treats verbalised confidence not merely as a metric but as a dynamic control signal. You will learn how they implement a non-differentiable switching function that bifurcates inference into a fast memory-augmented path and a slow "inverse" [YouTube Link](https://youtube.com/watch?v=BBjnaTcboFQ) 2026-01-24T14:15Z 85.2K followers, [----] engagements "Death of the Token in AI: Multi-Parallel AI Reality NVIDIAs Silent Robots & Pre-GPT-6 Schrdinger Token Frankenstein Vector and Quantum collapse for new Multi-Parallel AI Realities & NVIDIA Silent Robots w/ Pre-GPT-6. Multi-Parallel realities for AI reasoning and Quantum collapse in AI for next generation of AI models and their advanced architectures. Combine these three brand new AI Arxiv pre-prints from the latest Ai research and you will see a pattern emerging for new AI architectures like GPT-6. From NVIDIA to Microsoft. All rights W/ authors: "Reasoning Beyond Chain-of-Thought: A Latent" [YouTube Link](https://youtube.com/watch?v=O9HxArmWChs) 2026-01-16T14:15Z 85.2K followers, [----] engagements "MiMo V2 Flash: Excellent Performance (vs Kimi K2 Thinking) I test the causal reasoning performance for a simple [--] step logic task with the MiMo V2 Flash MoE 309B-15A model from Xiaomi (open-source). And compare the performance to a much bigger Kimi K2 Thinking Turbo model MoE 1T-32A. My Youtube playlist for this causal reasoning test is available here https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT 00:00 MiMo V2 vs KIMI K2 02:42 Start Live Test 08:56 MiMo First Result 10:15 1st Validation Run 12:37 2nd Validation Run 13:53 Kimi K2 First Answer 14:56 Self Check Kimi K2" [YouTube Link](https://youtube.com/watch?v=OoPVwK0KAF8) 2026-02-11T14:30Z 85.2K followers, [----] engagements "The Future of Conversational AI Google's PaLM w/ RLHF LLM ChatGPT Competitor After explaining BERT vs GPT and Google's T5 JAX (in my last videos) we now examine new PaLM: Pathways Language Model (if combined w/ RLHF -Reinforcement Learning with Human feedback). T5X = Google's T5 on JAX and FLAX. Plus Code implementation for PaLM w/ RLHF. PS: Next video on another LLM could be on Sparrow . smile. my resources (all rights are with the authors): Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer https://arxiv.org/pdf/1910.10683.pdf SentencePiece: A simple and" [YouTube Link](https://youtube.com/watch?v=Q4qaDZarl8g) 2023-01-23T13:15Z 85.2K followers, [----] engagements "8B outperforms GPT-120B on Multi Agents All rights w/ authors: DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning via Semantic Matching Yuxing Lu * [--] [--] Yucheng Hu * [--] Xukai Zhao [--] Jiuxin Cao [--] from [--] Peking University Beijing China [--] Georgia Institute of Technology Atlanta United States [--] Southeast University [--] Tsinghua University. #aiexplained #airesearch #artificialintelligence #aiagents artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial" [YouTube Link](https://youtube.com/watch?v=jWhnicSLdD4) 2026-02-10T14:01Z 85.2K followers, [----] engagements "15B Active MoE BEATS OPUS [---] in Reasoning Inside AI: to be specific inside a real powerful reasoning engine MoE and all the new methods and optimizations algorithms we encounter in building an open-source Mixture-of-Expert AI model. Inside a modern MoE AI model (Technical Architecture) All rights w/ authors: MiMo-V2-Flash Technical Report LLM-Core Xiaomi https://arxiv.org/pdf/2601.02780 #aitransformation #aiexplained #scienceexplained #nextgenai #airesearch artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI" [YouTube Link](https://youtube.com/watch?v=uVPZXvs634U) 2026-02-12T14:01Z 85.2K followers, [----] engagements "Claude OPUS [---] Thinking vs [---] Non-Thinking: Both FAIL Anthropic just released Claude OPUS [---]. And specified "Opus [---] extends the frontier of expert-level reasoning". So I test both models of Claude [---] (Thinking and Non-Thinking) on my standard causal reasoning test. You can see my complete YouTube Playlist for AI model testing on this causal reasoning test here https://www.youtube.com/watchv=g1L8uOQ7Ids&list=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT https://www.anthropic.com/news/claude-opus-4-6 00:00 OPUS [---] TEST 02:10 OPUS [---] Non-Thinking First Result 02:20 OPUS [---] Thinking 05:58 OPUS 4.6" [YouTube Link](https://youtube.com/watch?v=-0xKV2i6M4U) 2026-02-05T23:15Z 85.2K followers, [----] engagements "Google Goes to EXTREMES: In-Context Symbolic AI "We demonstrate an approach for LLMs to critique their own answers with the goal of enhancing their performance that leads to significant improvements over established planning benchmarks. Despite the findings of earlier research that has cast doubt on the effectiveness of LLMs leveraging self critique methods we show significant performance gains on planning datasets in the Blocksworld domain through intrinsic self-critique without external source such as a verifier. " Quote by Google DeepMind (see below) All rights w/ authors: "Enhancing LLM" [YouTube Link](https://youtube.com/watch?v=0x8JoBLV-8c) 2026-01-06T14:00Z 85.2K followers, [----] engagements "NeuroSymbolic Web World Model (Decouples Physics from AI) The idea is simple: Instead of training a huge AI model on Language syntax domain knowledge coding and Science patterns - why not separate the Physics engine (like a neurosymbolic Model) from the autoregressive AI Done in an elegant way: new insights by Princeton Univ. All rights w/ authors: Web World Models Jichen Feng13 Yifan Zhang1 Chenggong Zhang2 Yifu Lu1 Shilong Liu1 Mengdi Wang1 from [--] Princeton University [--] University of California Los Angeles [--] University of Pennsylvania arXiv:2512.23676" [YouTube Link](https://youtube.com/watch?v=388I4ugcf-0) 2025-12-31T14:00Z 85.2K followers, [----] engagements "AutoGPT & BabyAGI: autonomous AI Agents for LLMs explained AutoGPT & BabyAGI explained Combining GPT-4 with the interface functionality of LangChain with on open data channel to external internet services (with $$$ API) opens up new combinations of interlinks of resources: meet AutoGPT and BabyAGI. There are significant risks for your financial resources when GPT-4 takes over the link priorities without human interaction. Literature and code (all rights w/ those authors): https://github.com/yoheinakajima/babyagi" [YouTube Link](https://youtube.com/watch?v=5cnk5zpf1EA) 2023-04-22T12:15Z 84.8K followers, [----] engagements "MedAI: Vision Language Models & Fine-Tuning (KnowAda) Smaller VLM hallucinate. A new counter-measure: Knowledge-Adapted Fine-Tuning (KnowAda) is a novel approach to mitigate hallucinations in vision-language models (VLMs) when generating dense image captions. Reducing Hallucinations in Multimodal Models Through new Adaptive Training. Traditional fine-tuning methods often result in smaller-scale VLMs (up to 7B parameters) struggling to balance descriptiveness with factual accuracy especially in visually complex datasets. KnowAda addresses this by probing the VLMs knowledge using generated" [YouTube Link](https://youtube.com/watch?v=6Dc8Tny4agE) 2024-11-16T15:00Z 84.9K followers, [----] engagements "ORPO: NEW DPO Alignment and SFT Method for LLM Instead of the classical SFT and DPO alignment for training our LLMs there is a new method available. A innovative "reference model-free" monolithic odds ratio reference optimization algorithm ORPO eliminating the necessity for an additional preference alignment phase. A New Preference-aligned SFT method. We explore this idea from a theoretical physics perspective and notice a similarity to the regularizations terms methodologies. We further explore the conceptional similarities from a Lagrange Multiplier to new correction terms in addition to" [YouTube Link](https://youtube.com/watch?v=6kkJGkPZP88) 2024-03-24T13:00Z 84.7K followers, [----] engagements "AI Dual Manifold Cognitive Architecture (Experts only) All rights w/ authors: "MirrorMind: Empowering OmniScientist with the Expert Perspectives and Collective Knowledge of Human Scientists" Qingbin Zeng [--] Bingbing Fan [--] Zhiyu Chen [--] Sijian Ren [--] Zhilun Zhou [--] Xuhua Zhang [--] Yuanyi Zhen [--] Fengli Xu [--] Yong Li [--] Tie-Yan Liu [--] from [--] Department of Electronic Engineering BNRist Tsinghua University [--] Zhongguancun Academy "PersonaAgent with GraphRAG: Community-Aware Knowledge Graphs for Personalized LLM" Siqi Liang 1* Yudi Zhang 2* Yue Guo [--] from [--] Purdue University [--] Iowa State University 3" [YouTube Link](https://youtube.com/watch?v=8GGuKOrooJA) 2025-11-27T13:45Z 85K followers, 11.7K engagements "NEW StreamingLLM by MIT & Meta: Code explained MIT and META introduce StreamingLLM an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence length without any fine-tuning. Streaming LLM. ARXIV preprint: https://arxiv.org/pdf/2309.17453v1.pdf GitHub repo: https://github.com/mit-han-lab/streaming-llm/blob/main/streaming_llm/pos_shift/modify_llama.py" [YouTube Link](https://youtube.com/watch?v=9i1e3zAFdsI) 2023-10-13T12:00Z 84.8K followers, [----] engagements "Autonomous AI Agents: 14% MAX Performance Webarena provides a test ground to test AI Agents' performance for functional correctness of task completions. Ideal for development and performance testing of autonomous AI agents. All rights w/ authors: https://arxiv.org/pdf/2307.13854 Plus: User's Intent research by Microsoft on optimized tool use by autonomous agents. Plus the open source Toolkit from @CohereAI now available for your perfect RAG system with pre-build components and apps. https://github.com/cohere-ai/cohere-toolkit 00:00 Autonomous AI Agents 00:45 Webarena for dev of agents 03:30" [YouTube Link](https://youtube.com/watch?v=A2pDicTvOEM) 2024-04-28T12:00Z 84.8K followers, [----] engagements "Individualise your AI Companion: EASY The easiest way to individualize your AI - simple Demo. Please add additional personal guardrails to your individualized system prompt. Commercial AI systems can make mistakes. Severe mistakes. All rights w/ authors: Simulating Psychological Risks in Human-AI Interactions: Real-Case Informed Modeling of AI-Induced Addiction Anorexia Depression Homicide Psychosis and Suicide Chayapatr Archiwaranguprok MIT Media Lab Massachusetts Institute of Technology Cambridge Massachusetts USA Constanze Albrecht MIT Media Lab Massachusetts Institute of Technology" [YouTube Link](https://youtube.com/watch?v=AjZuzZRdyMI) 2025-11-18T14:00Z 84.8K followers, [----] engagements "Google's AutoGrad + Tensorflow's XLA Linear Algebra Compiler = JAX Add Google's AutoGrad and Tensorflow's XLA linear algebra compiler and you get JAX: a python and numpy racehorse to differentiate for backprop and compile on multi TPU clouds. You love numpy and want vectorization and automatic parallelization for GPUs and TPUs Then you know JAX For the sole purpose of applying Graph Neural network models we need to cover JAX by Google/DeepMind before starting into Jraph for our main purpose: Apply GNN to complex problem solving in the omniverse. Or was it the Multiverse Any way here is JAX" [YouTube Link](https://youtube.com/watch?v=DqFFH44yps8) 2021-12-04T07:00Z 84.9K followers, [---] engagements "How a 14B Model BEATS GPT-5.2 FUZZY Graph Reward All rights w/ authors: "Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning" Yuval Kansal Princeton University Niraj K. Jha Princeton University @princeton #ainews #RLVR #gpt52 #chatgpt5 #chatgpt Reinforcement Learning with verifiable return Medical AI Multi-step reasoning Reinforcement Learning with verifiable return Medical AI Multi-step reasoning" [YouTube Link](https://youtube.com/watch?v=KV-uZzE78qA) 2026-01-27T14:30Z 85.2K followers, [----] engagements "NEW Gemini [--] FLASH vs GPT [---] HIGH - A Bloodbath NEW Gemini [--] FLASH is [--] times cheaper ($) than OpenAI's GPT-5.2 HIGH for your identical tasks. So in a real-world test that looks similar to real science tasks I evaluate both AI models side-by-side. Note: This is not the known standard vanilla benchmarks this has to do with real world complexities - heavily oriented towards SCIENCE not Social Media. For a detailed comparison with all the other Ai models I have a dedicated YouTube Playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT Will Gemini [--] FLASH reach only a" [YouTube Link](https://youtube.com/watch?v=Nk3uSxgz0SQ) 2025-12-18T13:30Z 85K followers, [----] engagements "Geometric GROKKING Unlocked & Explained Given the latest insights on new geometric representations of knowledge patterns in the memory of Mamba or Transformer architecture I present a new theory of mine - explaining Grokking from a different framing. No science just me thinking about an explanation of the Grokking Phase for improved AI models. Complete (ArXiv) literature has already been quoted in my corresponding YT videos. Idea happened during a breakfast session on a Sunday morning. #artificialintelligence #reasoning #aiexplained artificial intelligence AI models LLM VLM VLA Multi-modal" [YouTube Link](https://youtube.com/watch?v=PaSm5vHYDew) 2025-11-02T13:45Z 85K followers, [----] engagements "AI Kill Switch for Hallucinations (Anthropic) [--] new AI research papers (all from [--] Jan 2026) that focus on one topic: inside the activations (the latent space inside an AI transformer architecture). A new kill switch for AI hallucinations in reasoning traces and how to heal catastrophic forgetting caused by Supervised Fine-Tuning (SFT). Self-learning self-healing self-correcting and self-improving AI models. mathematical insights into loss functions and new operators. all rights w/ authors: "Emergent Introspective Awareness in Large Language Models" Jack Lindsey from Anthropic" [YouTube Link](https://youtube.com/watch?v=R9czY1uVq_k) 2026-01-07T14:00Z 85.2K followers, [----] engagements "DeepSeek built a New Topological Transformer (mHC) DeepSeek build a new topological transformer that is beautifully compatible with the new Transformer architecture from Google (see my video https://youtu.be/5gpc3d2rFlg ). In this video I explain the mathematics of the Manifold-Constrained Hyper-Connections in particular the Sinkhorn-Knopp algorithm to entropically project H_residual onto the Birkhoff polytope. All rights w/ authors: "mHC: Manifold-Constrained Hyper-Connections" Zhenda Xie* Yixuan Wei* Huanqi Cao* Chenggang Zhao Chengqi Deng Jiashi Li Damai Dai Huazuo Gao Jiang Chang Liang" [YouTube Link](https://youtube.com/watch?v=Tki2Zy4jOAc) 2026-01-03T14:00Z 85.2K followers, 17.2K engagements "CORE of AI is EXPLODING - [--] New Papers CORE of AI currently explodes: we'll discover a specific selection of [--] new ArXiv CS pre-prints as a subset from more than [---] new ArXiv papers as published on first days of September [----]. All new arxiv preprints are given with full details: authors title publication date and arxiv link in the video. #scienceexplained #aiexplained #discoverai" [YouTube Link](https://youtube.com/watch?v=XlnMoWEjogY) 2025-09-06T14:00Z 84.8K followers, 15K engagements "DeepSeek [---] vs MiniMax M2 (1 Sentence TEST) Superior Intelligence: ChatGPT [---] vs MiniMax M2 on my [--] sentence test. Watch my new logic test (1 sentence no science) to challenge the intelligence of AI models. New test between the latest Flagship from OpenAI: GPT-5.2 vs the open Source Model MiniMax M2 (available to download from HuggingFace). And of course I test the performance of our good old friend DeepSeek v3.2. No DSPy optimization No additional Prompt engineering no ICL few shot examples just pure human AI interaction. Note: test on new MiniMax M2.1 currently being recorded ." [YouTube Link](https://youtube.com/watch?v=YFzdIBBnv5o) 2025-12-23T21:00Z 85.2K followers, [----] engagements "NEW DeepSeek V3.2 Thinking: Equal to Gemini [--] PRO Finally. The improved version of DeepSeek [---] was released today . With maximum Test Time Scaling. I am testing the thinking and non-thinking version live. The experimental version (now outdated) of DeepSeek [---] EXP was released [--] months ago. Will this new Chinese AI perform at the same level like GPT [--] (high) or even Gemini [--] PRO We look at the official marketing benchmark data and perform a real world live test with both models on the Chinese and the US platform. The results are different than the benchmarks. https://www.deepseek.com/en" [YouTube Link](https://youtube.com/watch?v=_TBsZeU4K7Q) 2025-12-02T11:15Z 85K followers, [----] engagements "Google is cooking: Beyond the 'Next-Token' Manifold All rights w/ authors: Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents Zehong Wang [--] Fang Wu [--] Hongru Wang [--] Xiangru Tang [--] Bolian Li [--] Zhenfei Yin [--] Yijun Ma [--] Yiyang Li [--] Weixiang Sun [--] Xiusi Chen [--] Yanfang Ye [--] from [--] University of Notre Dame [--] Stanford University [--] University of Edinburgh [--] Yale University [--] Purdue University [--] University of Oxford [--] UIUC. Context Structure Reshapes the Representational Geometry of Language Models Eghbal A. Hosseini1 Yuxuan Li1 Yasaman Bahri1 Declan" [YouTube Link](https://youtube.com/watch?v=anEVsOPtbnw) 2026-02-03T14:15Z 85.2K followers, 10.5K engagements "Google's Warning: ICL Context is Inert Context is Not Compute: Destroying the ICL 'World Model' Myth. All rights w/ authors: Language Models Struggle to Use Representations Learned In-Context Michael A. Lepori* Tal Linzen Ann Yuan Katja Filippova from Google DeepMind Brown University New York University #scienceexplained #aiexplained #aiworld #worldmodel #contextengineering artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM" [YouTube Link](https://youtube.com/watch?v=b0c64uUyvpo) 2026-02-06T14:01Z 85.2K followers, [----] engagements "Stanford's AI is Self-Learning w/ Context Engineering: ACE The synergy between Early Experience and Agentic Context Engineering (ACE) creates a powerful two-loop architecture for autonomous AI self-improvement. The Early Experience paradigm acts as the agent's tactical sensory system generating a continuous stream of raw grounded learning signals by exploring alternative actions and observing their immediate reward-free consequences. This raw experiential data then feeds into the Agentic Context Engineering framework which functions as the agent's strategic cognitive brain. Here a Reflector" [YouTube Link](https://youtube.com/watch?v=elgYgPo_vY4) 2025-10-11T14:26Z 85K followers, 11.6K engagements "NEW ADAPTIVE Multi-Agent AI System: AIME (ByteDance) Old multi-agent AI systems suffer from rigid planning that lack an strategic and real time tactical component for a intelligent swarm /multi-agent configuration /operation. New research further develops communication protocols between AI entities and adaptive planning entities reacting to new emerging challenges in real world scenarios /battlefields. multi-agent communication multi-agent specialization reduced complexity for agents and their tool-set. All rights w/ authors: AIME: TOWARDS FULLY-AUTONOMOUS MULTI-AGENT FRAMEWORK Yexuan Shi" [YouTube Link](https://youtube.com/watch?v=euEXBqVa7LM) 2025-07-18T12:01Z 84.8K followers, [----] engagements "LOGIC Test on Qwen3 Max Thinking and Kimi K2.5 I perform my causal reasoning performance tests on the new Qwen [--] MAX THINKING and analyse the new KIMI K2.5 with added visual intelligence on top of the Base K2 model with a nice new marketing feature of [---] subagents under an orchestrator as a special model variance. Live test on Alibaba Cloud. I like the information provided by John Hammond on "Clawdbot Malware" https://www.youtube.com/watchv=7GS6Xs4hdvg and for some additional info https://www.youtube.com/watchv=kSno1-xOjwI (nothing to add from my side). #aitesting #chatgpt #scienceexplained" [YouTube Link](https://youtube.com/watch?v=g1L8uOQ7Ids) 2026-01-29T14:15Z 85.2K followers, [----] engagements "A2A - MCP SECURITY Threats: Protect your AI Agents A new Blueprint for AI Security responding to the security threats by MCP - Model Context Protocol by Anthropic and A2A - Agent to Agent Protocol. Latest Ai research by Google to protect and secure A2A communication and inter-agent security. This video makes the community aware of current security threats in latest AI systems especially when implementing RAG MCP or A2A and countermeasures to protect your privacy confidential data and protect against attack vectors as described in the latest research literature. All rights w/ authors: Building" [YouTube Link](https://youtube.com/watch?v=h_6unQxHyb4) 2025-05-17T13:01Z 85K followers, [----] engagements "The New Geometry of Intelligence #ai All rights w/ authors: "Spectral Superposition: A Theory of Feature Geometry" Georgi Ivanov [--] [--] Narmeen Oozeer [--] Shivam Raval [--] Tasana Pejovic [--] Shriyash Upadhyay [--] Amir Abdullah [--] [--] from [--] Theopha [--] Harvard University [--] Martian [--] Thoughtworks #aiexplained #aireasoning #scienceexplained artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG" [YouTube Link](https://youtube.com/watch?v=iHLDu-IdJwo) 2026-02-04T14:15Z 85.2K followers, [----] engagements "New AI Post-Training: Add RL as orthogonal vector to SFT All rights w/ authors: "Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation" Pingzhi Tang12 Yiding Wang12 Muhan Zhang13 from [--] Institute for Artificial Intelligence Peking University [--] Yuanpei College Peking University [--] State Key Laboratory of General Artificial Intelligence BIGAI #chatgpt5 #aireasoning #newsai #reinforcementlearning artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI" [YouTube Link](https://youtube.com/watch?v=ltm1fMIpbwM) 2026-01-20T14:15Z 85.2K followers, [----] engagements "Self Evolution of AI beyond Humans (Agent0: UNC Stanford) We are rapidly approaching the "Data Wall"the point where high-quality human reasoning traces run dry. But a groundbreaking new framework from UNC Salesforce and Stanford just demonstrated a way out. Meet Agent0: a fully autonomous post-training method that evolves high-level reasoning capabilities from zero external data. By engineering a symbiotic arms race between a "Curriculum Agent" (incentivized to generate maximum entropy puzzles) and a tool-integrated "Executor Agent" (verified by a Python sandbox) this method bypasses the mode" [YouTube Link](https://youtube.com/watch?v=mMEG074Bkm4) 2025-11-26T14:01Z 84.8K followers, [----] engagements "AI Leap: Tiny HRM 27M Beats Claude OPUS [--] on AGI HRM = Hierarchical Reasoning Model. Independent benchmark confirms Hierarchical Reasoning Model performance. A tiny 27B HRM outperforms a CLAUDE OPUS in reasoning on ARC-AGI-1 benchmark. WHAT My original video on HRM - Hierarchical Reasoning Models https://youtu.be/QWD55guu0Sofeature=shared Link to the company in Singapore whose experts published the HRM paper: https://www.sapient.inc/ @ARCprize all rights w/ authors: "The Hidden Drivers of HRM's Performance on ARC-AGI"" [YouTube Link](https://youtube.com/watch?v=mhft9WBK4uE) 2025-08-17T14:01Z 85K followers, 23.7K engagements "Gemini [--] PRO Logic: A BEAST I performed my standard logic and causal reasoning test on the newly released GEMINI [--] PRO on the free platform lmarena.ai for everybody to follow or validate. This identical test has been used in my recorded tests of GPT-5.1 Grok [--] and more than [--] Ai models in the past months. You have a perfect comparison. See my YouTube playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT #GEMINI3pro #gemini3 #aitest #google #googledeepmind artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent" [YouTube Link](https://youtube.com/watch?v=pBKtjSbNWLk) 2025-11-18T19:48Z 85K followers, [----] engagements "CODE Variational Autoencoder (VAE) w/ KL-Divergence #ai #pythonprogramming #keras A VAE is a probabilistic take on the autoencoder a model which takes high dimensional input data and compresses it into a smaller representation space. Unlike a traditional autoencoder which maps the input onto a latent vector a VAE maps the input data into the parameters of a probability distribution such as the mean and variance of a Gaussian. This approach produces a continuous structured latent space which is useful for image generation. official link and COLAB NB:" [YouTube Link](https://youtube.com/watch?v=pRKTr8gw2KA) 2022-09-04T12:00Z 84.8K followers, [----] engagements "Beyond Next Token Prediction: CALM AI Finally a new AI that implements the next generation after Next-Token-Prediction: CALM - CONTINUOUS AUTOREGRESSIVE LANGUAGE MODELS. All rights w/ authors: CONTINUOUS AUTOREGRESSIVE LANGUAGE MODELS Chenze Shao [--] Darren Li [--] Fandong Meng [--] Jie Zhou [--] from [--] WeChat AI Tencent Inc [--] Qiuzhen College Tsinghua University https://shaochenze.github.io/blog/2025/CALM/ https://github.com/shaochenze/calm #scienceexplained #aireasoning #aiexplained #aidiscovery artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent" [YouTube Link](https://youtube.com/watch?v=pkDnJdAy8_c) 2025-11-05T14:01Z 84.9K followers, [----] engagements "Tokenizing Gravity Waves: AI in Astrophysics (LIGO) All rights W/ authors: Large Language Models for Limited Noisy Data: A Gravitational Wave Identification Study Yixuan Li [--] [--] Yuhao Lu [--] [--] Yang Liu [--] [--] Liang Li [--] [--] R. Ruffini [--] [--] [--] [--] [--] Di Li [--] [--] [--] [--] Rong-Gen Cai [--] Xiaoyan Zhu [--] Wenbin Lin [--] [--] [--] and Yu Wang [--] [--] [--] [--] from [--] School of Mathematics and Physics University of South China Hengyang [------] China [--] ICRANet-AI Brickell Avenue [---] Miami FL [-----] USA [--] School of Computer Science University of South China Hengyang [------] China [--] Department of Physics E. Pancini University Federico II" [YouTube Link](https://youtube.com/watch?v=rx8fPUsvR3c) 2025-12-07T13:01Z 84.9K followers, [----] engagements "GPT-5 w/ MCP Fails on World Models: NEW Solution ATLAS All rights w/ authors: Current Agents Fail to Leverage World Model as Tool for Foresight Cheng Qian1 Emre Can Acikgoz1 Bingxuan Li1 Xiusi Chen1 Yuji Zhang1 Bingxiang He2 Qinyu Luo3 Dilek Hakkani-Tr1 Gokhan Tur1 Yunzhu Li4 Heng Ji1 from [--] UIUC [--] THU [--] JHU [--] Columbia @Illinois1867 Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning Jinyang Wu1* Guocheng Zhai1* Ruihan Jin1* Jiahao Yuan3 Yuhao Shen2 Shuai Zhang1 Zhengqi Wen1 Jianhua Tao1 from [--] Tsinghua University [--] Zhejiang University [--] East China Normal" [YouTube Link](https://youtube.com/watch?v=tez4AyTm1Rs) 2026-01-12T13:45Z 85.2K followers, [----] engagements "Dream Job Alert: AI Prompt Engineer - $335K AI Prompt Design: A Crash Course Anthropic AI offers you a job as prompt engineer. Go and get a new Job in AI if you know about prompt engineering. Short Introduction to prompt engineering and continuous prompt design plus prefix tuning vs fine-tuning for LLMs. Hint: all my viewers who read recommended research arxiv pre-prints surely qualify. Literature: https://arxiv.org/pdf/2107.13586.pdf Pre-train Prompt and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing Constitutional AI: Harmlessness from AI Feedback (by" [YouTube Link](https://youtube.com/watch?v=tw_f-Q6Rb2c) 2023-02-28T13:00Z 85K followers, [----] engagements "PyG - PyTorch Geometric - Intro to Graph Neural Networks - Outlook SBERT w/ PyG PyG - PyTorch Geometric - is a library for Graph Neural Networks based on PyTorch for ML and geometric learning. This lookout on my next [--] YouTube videos w/ PyG Code on homogeneous Graphs provides insights and analyses the problem w/ heterogeneous Graphs especially when applying SBERT (Sentence Transformers) to scientific documents w/ PyG. #pytorch #geometric #graphs PyTorch PyTorch Geometric PyG Intro GCN Graph Neural Networks PyTorch PyTorch Geometric PyG Intro GCN Graph Neural Networks" [YouTube Link](https://youtube.com/watch?v=vuv2hHKf0to) 2022-11-29T13:00Z 85K followers, [----] engagements "RAG's Intelligent Upgrade: Agentic RAR (Oxford Univ) New research by the University of Oxford on Advanced Reasoning Agentic AI systems that substitute classical RAG (Retrieval Augmented Generation). An Agentic RAR (Reason on Agentic Reasoning). All rights w/ authors: Agentic Reasoning: Reasoning LLMs with Tools for the Deep Research by Junde Wu Jiayuan Zhu Yuyuan Liu from University of Oxford #aiagents #airesearch #knowledgegraphs #reasoning #scienceexplained artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI" [YouTube Link](https://youtube.com/watch?v=xHUgONA-x3g) 2025-02-11T15:01Z 85K followers, 51.8K engagements "Microsoft and ChatGPU Time for Comedy: the untold fictional story of how Microsoft secured skyrocketing performance of Microsoft Office in [----]. All protagonists are fictional kind of . #chatgpt #chatgptexplained #shorts #comedy #comedyshorts" [YouTube Link](https://youtube.com/watch?v=xnPq4VbADu8) 2023-01-24T13:00Z 84.8K followers, [---] engagements "Chain of Thought (CoT) meets Instruction Fine-Tuning Explore the concept of "Chain-of-Thought" (CoT) combined with "instruction fine-tuning" as techniques to improve the performance of large language models (LLMs). These techniques involve optimizing prompt structures and training the models to follow specific instructions leading to enhanced capabilities in solving unseen tasks. The combination of chain of thought and instruction fine-tuning has shown promising results in improving the model's performance and understanding of complex language tasks also for smaller language models." [YouTube Link](https://youtube.com/watch?v=-AZL31jop9Y) 2023-05-30T12:00Z 60.7K followers, [----] engagements "Python TF2: BERT model Code your WordPiece - Tokenizer (w/ HuggingFace) Python TF2 code (w/ JupyterLab) to train your WordPiece tokenizer: Tokenizers are one of the core components of the NLP pipeline. They serve one purpose: to translate text into data that can be processed by the BERT model. Why would you need a new & improved tokenizer That's because Transformer models very often use subword tokenization algorithms and they need to be trained to identify the parts of words that are often present in the corpus of your input text (sentences Paragraphs documents .) the sentences you are" [YouTube Link](https://youtube.com/watch?v=04oZ2P0uvp0) 2022-01-31T09:30Z 80.3K followers, [---] engagements "NodePiece 2022: New vocabulary for your Knowledge Graph #Shorts New representation learning algorithms for knowledge graphs (KG): NodePiece. An anchor-based approach to learn a fixed-size entity vocabulary. In NodePiece a vocabulary of sub-entity units is constructed from anchor nodes in a graph with known relation types. Similar to WordPiece tokenization for BERT in NLP. Arxiv preprint (credits to): https://arxiv.org/pdf/2106.12144.pdf Mikhail GalkinEtienne Denis Jiapeng Wu and William L. Hamilton Published as a conference paper at ICLR [----] New NodePiece tokenization can augment any" [YouTube Link](https://youtube.com/watch?v=0FrG1L8JD8c) 2022-05-18T11:15Z 83.1K followers, [--] engagements "Superhuman SPATIAL AI: Finally AGI ASI and Superhuman performance of AI: a real - world reflection on the latest AI research results for Vision Language Models (VLM). All rights w/ authors: From Macro to Micro: Benchmarking Microscopic Spatial Intelligence on Molecules via Vision-Language Models Zongzhao Li1* Xiangzhe Kong23* Jiahui Su4 Zongyang Ma5 Mingze Li1 Songyou Li1 Yuelin Zhang1 Yu Rong67 Tingyang Xu67 Deli Zhao67 Wenbing Huang1 from [--] Gaoling School of Artificial Intelligence Renmin University of China [--] Dept. of Comp. Sci. & Tech. Tsinghua University [--] Institute for AI Industry" [YouTube Link](https://youtube.com/watch?v=0bcaINHdUU8) 2025-12-15T14:00Z 83K followers, [----] engagements "The Cracks in AI Are Widening (CoT RAG) all rights w/ authors: "Rational Synthesizers or Heuristic Followers Analyzing LLMs in RAG-based Question-Answering" Atharv Naphade from Carnegie Mellon University #ai #aifails #aireasoning #aiexplained #aisystem artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision" [YouTube Link](https://youtube.com/watch?v=0ezdBcdY3bc) 2026-01-15T14:00Z 85.2K followers, [----] engagements "How to Combine Knowledge Graphs and Agents (Emory Stanford) How to combine AI agents in the most effective way with structured knowledge in a knowledge graph representation New insights from AI research: [--] agents interact with Knowledge Graph which was updated via a fine-tuned Sentence BERT NER tool (Named Entity Recognition). Emory and Stanford Univ show this multi-agent architecture for a medical diagnosis prediction which is the task of predicting a patients future health risks based on their historically observed medical data such as electronic health records (EHRs) which plays a vital" [YouTube Link](https://youtube.com/watch?v=0oDgruiW7Gw) 2025-07-08T08:15Z 81.7K followers, [----] engagements "MAGIC of DSPY [--] (Stanford) - Lean [--] What is the magic of DSPY Should I learn DSPY now Is it MCP compatible How to integrate agents into DSPY What the *** is a teleprompter in DSPY What compiler to use in DSPY How does a compiler work in DSPY exactly Explain the complete flow of DSPY Cursor codes DSPY for me but what is it Does DSPY [--] integrate Lean [--] From prompt engineering to Context Engineering A lot questions from my viewers today I answer it all. DSPY: finally explained. DSPy allows you to iterate fast on building modular AI systems and offers algorithms for optimizing their prompts and" [YouTube Link](https://youtube.com/watch?v=1067jj67toY) 2025-07-15T14:00Z 80.6K followers, [----] engagements "PyG + SBERT: Heterogeneous Graphs Using SBERT SentenceTransformers for Node Classification SBERT [--] PyG w/ SBERT Sentence Transformers for Node Classification in heterogeneous Graphs coded in PyG (PyTorch geometric) on a free COLAB NB. ML on GRAPHS. Graph-structured data such as social graphs networks in cybersecurity or molecular representations are our real-world scenarios which generate heterogeneous Graphs on which to apply our ML models (Node2Vec Message Passing MP-GNN GCN - Graph Convolutional Networks) for prediction of node classification or simply classical link prediction. Detecting" [YouTube Link](https://youtube.com/watch?v=10evf7xmXto) 2022-12-05T13:00Z 84.3K followers, [----] engagements "FUSION Controlled by AI Agents (Los Alamos) URSA is a modular agentic AI ecosystem by Los Alamos National Labs that accelerates scientific discovery by using a team of specialized agents to autonomously plan hypothesize and execute complex research tasks even outperforming traditional methods in domains like physics simulation. The authors of this new paper are commendably transparent about the failure modes which are critical for us to understand. A) Hallucinated Reality: In one test URSA was asked to find new alloys. It hallucinated a full experimental plan () including claiming to have" [YouTube Link](https://youtube.com/watch?v=11RCxGqXtKc) 2025-07-07T14:00Z 84.5K followers, [----] engagements "Galactica: New LLM by Meta hallucinates Science - First Look Galactica: a brand-new large language model (LLM) on Science: trained on over [--] million papers textbooks reference material compounds proteins and other sources of scientific knowledge. But there is a warning: This LLM can hallucinate scientific text. A machine (without any understanding of the content) adds with a probability distribution some pieces of text together. Now the probability distribution has been trained on real scientific text but the output is not If you want to learn about LLMs: https://youtu.be/rrZGIR5CryM" [YouTube Link](https://youtube.com/watch?v=1Ycodhrd0S8) 2022-11-16T07:45Z 84.4K followers, [----] engagements "Explainable AI - The story behind XAI in [----] (legal ethical commercial risks) Two beautiful examples show why we need an Explainable AI (XAI) system to protect our individual freedom and human rights in a digital / AI economy. Although it will cause significant additional work for us AI coder. The story behind XAI in 2022: The problem of social media input data. In general: the quality of data for training deep AI systems. Explainable Artificial Intelligence will become important in [----] for AI coder and AI provider. Upcoming US and EU AI-specific legislation for market applications are" [YouTube Link](https://youtube.com/watch?v=1h-IdF1KRWw) 2022-01-04T07:00Z 81.4K followers, [---] engagements "AGI Asymmetry Discovered (Harvard Stanford MIT) New insights into the latest of Ai research from Stanford Univ Harvard Univ MIT NVIDIA CMU regarding Reasoning of or between Ai agents. All rights w/ authors: "RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems" Yuxiao Qu*1 Anikait Singh*2 Yoonho Lee*2 Amrith Setlur1 Ruslan Salakhutdinov1 Chelsea Finn2 Aviral Kumar1 from [--] Carnegie Mellon University [--] Stanford University Beyond Majority Voting: LLM Aggregation by Leveraging Higher-Order Information" Rui Ai MIT Yuqi Pan Harvard University David Simchi-Levi MIT Milind" [YouTube Link](https://youtube.com/watch?v=1nd2oAt2rEU) 2025-10-05T14:00Z 80.9K followers, [----] engagements "Hallucinate.DELETE = We found the H-Neurons NEW AI research by Tsinghua univ: For years we have treated LLM hallucinations as an inevitable "psychological" flaw of autoregressive generation: a stochastic drift caused by noisy pre-training data or misalignment. We assume that when a model lies the error is diffuse smeared across billions of parameters in a high-dimensional fog. But new mechanistic interpretability research suggests this assumption is fundamentally wrong. What if confabulation isn't a random state but a dedicated deterministic circuit We have discovered that the drive to" [YouTube Link](https://youtube.com/watch?v=1oVelAKD_5A) 2025-12-05T13:00Z 82.4K followers, [----] engagements "Reduce CONTEXT for MAX Intelligence. WHY Advanced visual Ai reasoning It is simple: create a synthetic "adversarial" environment inside the inference process of the reasoning engine. No additional training runs. And: AI Test Time Scaling (TTS) algorithm claim advanced reasoning performance for your AI but is it true And what AI Scaling method should you pay for Self-refinement Self-Reflection Best-of-N There is a new TTS competitor with outstanding performance features. Maximal reduction of the visual representation for the training accuracy of next generation Visual AI systems. You can" [YouTube Link](https://youtube.com/watch?v=20rNv7yrTPM) 2025-12-03T13:15Z 82.6K followers, [----] engagements "Wall Street's new LLM beats GPT-4 GPT-4 beaten GPT-4 lost the game Wall Street build the perfect AI /LLM. A billionaire builds a LLM for the US Financial sector the first finance LLM And a simple reason why GPT-4 lost . Questions: [--]. Imagine your are the US Financial Sector. What AI LLM do you buy [--]. You have the largest domain-specific dataset for an LLM of the latest generation: the financial data of the world Why you stay away from GPT-4 [--]. New economic perspective (price vs performance) on Large Language Models (LLM) including the latest developments of AI systems vs GPT-4. [--]. The" [YouTube Link](https://youtube.com/watch?v=272QsC0wLGo) 2023-04-18T12:15Z 84.1K followers, [----] engagements "Intro to KERAS [--] (KERAS core) for PyTorch & JAX KERAS [--] - a high level API to design Neural Networks like Transformers - will be introduced in autumn [----]. KERAS core is available as a Beta version and we test it on PyTorch and JAX for building and designing special transformer architectures apply specific loss functions freeze layers redefine dense layers and program new evaluation routines. A truly framework agnostic code base to design and train Transformers either in TensorFlow PyTorch or JAX. #ai #keras #introduction" [YouTube Link](https://youtube.com/watch?v=29cL4rnU5Ck) 2023-07-27T12:00Z 80.6K followers, [----] engagements "AI Paradox: Use Text for Logic Avatars for Meaning All rights w/ authors: "Future You: Designing and Evaluating Multimodal AI-generated Digital Twins for Strengthening Future Self-Continuity" Constanze Albrecht MIT Media Lab Cambridge MA USA csophie@mit.edu Chayapatr Archiwaranguprok MIT Media Lab Cambridge MA USA pub@mit.edu Rachel Poonsiriwong Harvard University Cambridge MA USA rachel_poonsiriwong @gsd.harvard.edu Awu Chen MIT Media Lab Cambridge MA USA awu@mit.edu Peggy Yin Stanford University Stanford CA USA peggyyin@stanford.edu Monchai Lertsutthiwong KASIKORN Labs Nonthaburi Thailand" [YouTube Link](https://youtube.com/watch?v=2rHC_-rikis) 2025-12-11T13:45Z 82.6K followers, [----] engagements "Topology DSPy: Prompting the Swarm (Multi-Agents) Latest Tech insights for multi-agent AI by Google. Utilizing DSPy and Topology optimization techniques for an improved multi-agent performance. All rights w/ authors: Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies Han Zhou Xingchen Wan Ruoxi Sun Hamid Palangi Shariq Iqbal Ivan Vuli Anna Korhonen and Sercan . Ark from Google and University of Cambridge Feb [--] [----] #airesearch #educationalvideos #aiagents #multiAgentAI" [YouTube Link](https://youtube.com/watch?v=39ZUJsrazao) 2025-02-08T15:00Z 80.6K followers, [----] engagements "From Physics to SwarmAgentic AI: No Code After multi-agent systems that humans have to design and build now the next step: autonomous SwarmAgentic systems where AI optimizes multi-agent systems in their swarm time evolution. Self-learning AI. Swarm intelligence. Multi-agent system self learning. All rights w/ authors: SwarmAgentic: Towards Fully Automated Agentic System Generation via Swarm Intelligence Yao Zhang [--] Chenyang Lin [--] Shijie Tang [--] Haokun Chen1 Shijie Zhou [--] Yunpu Ma [--] Volker Tresp13 from [--] LMU Munich [--] Technical University of Munich [--] Munich Center for Machine Learning" [YouTube Link](https://youtube.com/watch?v=3tiAvRcviiY) 2025-06-22T13:00Z 82.6K followers, [----] engagements "AI :: Fine-tune LLama 2: Facebook or HuggingFace Follow Facebook for fine-tuning Llama [--] models or is there a better way a more elegant way by the open source community YES on HuggingFace by the way: coool that Meta AI is running on Facebook's GitHub. Script: https://github.com/lvwerra/trl/blob/main/examples/scripts/sft_trainer.py" [YouTube Link](https://youtube.com/watch?v=3waZ78Cs93k) 2023-07-20T13:00Z 80.9K followers, 14.3K engagements "Better than AutoGen & LangChain: OctoTools (Stanford AI) Complex reasoning remains a challenge for Large Language Models and modern VLMs. OctoTools a novel agentic framework addresses this by introducing standardized Tool Cards and a dedicated Planner-Executor architecture. This approach enables training-free integration of diverse tools enhancing agent capabilities for intricate tasks. Early results demonstrate significant performance gains suggesting a scalable and modular paradigm for tool-augmented AI reasoning. all rights w/ authors: OctoTools: An Agentic Framework with Extensible Tools" [YouTube Link](https://youtube.com/watch?v=4828sGfx7dk) 2025-02-25T15:00Z 59.7K followers, [----] engagements "8 CLOUD GPU Provider (H100 to RTX 4090) A very short intro to [--] Cloud based GPU provider especially for ML and fine-tuning LLMs. If you are looking for a cloud based solution you have a lot of options and I randomly choose those [--] CLOUD GPU provider. A lot of further CLOUD GPU provider not mentioned just a representative sample (from [--] RTX [----] to 3.5K H100). Please leave a comment if you would recommend your preferred GPU Cloud Service for AI and LLMs in particular. If you are a GPU Cloud Service provider make yourself known in the comments. The AI community is always interested in new" [YouTube Link](https://youtube.com/watch?v=4ArkBdKREDo) 2023-07-06T12:00Z 83K followers, [----] engagements "Code your BLIP-2 APP: VISION Transformer (ViT) + Chat LLM (Flan-T5) = MLLM BLIP-2: Upload an image the vision transformer will analyze the content of the image and a LLM will tell you a story about it - or answer your questions about the picture. We'll use Flan-T5 and Vision Transformer interlinked w/ Q-Former (BLIP 2). Multimodal LLM w/ BLIP-2. Example: if you upload a picture from the great pyramid in Egypt and you prompt (ask) the system: "When was it built" The ViT will tell the LLM that on the image are the pyramids from Gizeh and therefore the LLM (ChatGPT or T5) will tell you: "The" [YouTube Link](https://youtube.com/watch?v=4XweSnMXxWw) 2023-03-12T14:00Z 82.7K followers, [----] engagements "Explainable AI - The Billion $ Business Model of XAI in [----] How can you make money of XAI in [----] How is the interlink designed between Cloud Companies (Microsoft Meta .) and commercial companies seeking AI induced growth paths A locked-in ecosystem Dominated by tech giants No way out #xai #businessmodel #explainableai #humancentereddesign #humancentered #industry #business 00:00 Explainable AI system configuration 03:17 Cloud Economy 05:45 Government regulations 07:10 Cloud Company 11:05 Specific AI models 12:55 Revenue streams 13:20 Ecosystem lock-in" [YouTube Link](https://youtube.com/watch?v=4_ZT6hvlhGQ) 2022-01-07T06:15Z 80.8K followers, [---] engagements "Mathematical Shape of AI Thoughts (Topology Homology) Is it possible to grade an AIs reasoning without actually knowing the answer For years weve assumed that to evaluate a Chain-of-Thought we needed expensive "Ground Truth" labels. We treated reasoning as a linear sequence of tokens judging it only by the final output. But a breakthrough paper from end of [----] challenges this entire paradigm by proving that Truth has a Geometry. By mapping reasoning steps into a high-dimensional Point Cloud we can now see that accurate logic forms a tight coherent manifold while hallucinations look like" [YouTube Link](https://youtube.com/watch?v=4zB_FVpFpJI) 2025-12-24T13:45Z 85.2K followers, [----] engagements "Discover Vision Transformer (ViT) Tech in [----] Discover how I learn to code new AI topics (like Vision Transformer - ViT) for my YouTube videos and how I plan my AI videos. Where to get information about current trends in NLP or Vision where to learn a new theory (arxiv pre-prints) of a new tech (eg Vision transformer for medical images) in AI. Where to find excellent code examples for a first implementation. And how to stay informed on new and evolving AI topics and code implementations for real-world applications. From @HuggingFace libraries to my beloved https://paperswithcode.com 00:00" [YouTube Link](https://youtube.com/watch?v=5ZP9SJyTM94) 2023-01-13T13:00Z 82.7K followers, [---] engagements "AI Inside an AI: Internal RL w/ Temporal Abstraction Google invented a new transformer architecture with an internal metacontroller. An AI inside an AI. No #agent no #RAG just a more intelligent AI itself. This pre-print shows that the future of AI reasoning isn't just bigger Context Windows or more Chain-of-Thought tokens. It's about Latent Space Steering. It's about putting a small 'System 2' brain inside the massive 'System 1' body of the LLM. Future AI architectures discovered. All rights w/ authors: Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement" [YouTube Link](https://youtube.com/watch?v=5gpc3d2rFlg) 2025-12-28T13:45Z 85.2K followers, 11.5K engagements "DBRX: MOST POWERFUL Open Source LLM - NEW @Databricks DBRX: NEW MOST POWERFUL Open Source LLM: MoE 132B 16E 32K 12T Databricks reveals DBRX: After [--] months of cloud compute on [----] H100 GPUs DBRX sets a new state-of-the-art for established open LLMs. Moreover it provides the open community and enterprises building their own LLMs with capabilities that were previously limited to closed model APIs; according to first measurements it surpasses GPT-3.5 and it is competitive with Gemini [---] Pro. Great job @Databricks All rights with authors:" [YouTube Link](https://youtube.com/watch?v=5rtGvKuEnuQ) 2024-03-27T14:00Z 80.6K followers, [----] engagements "Design your own SciBERT sentence embedding model and explore Deloitte's TechTrends2021 (SciBERT) Code your AI with multiple HuggingFace models and different architectures of SentenceTransformers e.g. SciBERT (BERT pre-trained on scientific text). https://github.com/allenai/scibert #sbert #nlproc #nlptechniques #clustering #semantic #bert #climatechange #3danimation #3dvisualization #topologicalspace #deeplearning #machinelearningwithpython #pytorch #sentence #embedding #complex #ipcc #umap #insight #code_your_own_AI #code_in_real_time #SentenceTransformers #AI_reads_a_document" [YouTube Link](https://youtube.com/watch?v=6vliwxz6j5k) 2021-04-05T05:15Z 82.6K followers, [----] engagements "NEW: MedAI for US$100 (some technical insights) New AI research on how to optimize MedAI LLMs for a more adjusted clinical environment. New MedAI LLM benchmark MuddyMaze and new fine-tune Med LLMs on conversational med data sets. All rights w/ authors Dialogue is Better Than Monologue: Instructing Medical LLMs via Strategical Conversations by Zijie Liu Xinyu Zhao Jie Peng Zhuangdi Zhu Qingyu Chen Xia Hu and Tianlong Chen from Washington University in St. Louis University of North Carolina at Chapel Hill George Mason University Yale University and Rice University. #airesearch #aiagents" [YouTube Link](https://youtube.com/watch?v=6whj28Q6ujA) 2025-02-03T15:00Z 75.7K followers, [----] engagements "Learn SBERT Sentence Transformers: TSDAE SimCSE and CT #sbert #deeplearning (SBERT 15) Real time code for SBERT Sentence Embedding in a vector space with SBERT Transformer models Bi-encoder Transformer models Learn SBERT Sentence Embedding: TSDAE SimCSE and CT. With NEW pre-trained models best suited for your application. A) Add "SUPERVISED training data" to your SentenceTransformers to improve model performance. B) If you have NO labeled training data: Add "UNsupervised learning" to learn semantically meaningful sentence embedding from the text/sentences itself For [--] models I show you coding" [YouTube Link](https://youtube.com/watch?v=6yPWtdgs5Sg) 2021-06-01T12:15Z 82.6K followers, [----] engagements "Apply LLMSelector to your AI Agents (Tutorial & Code) My video detailes a new optimization where allocating different LLMs insert your latest VLM - like Grok [--] to the new Sonnet [---] to the upcoming ChatGPT [---] to different modules /agents leads to substantially higher performance than allocating the same expensive smartest best performance singular LLM to all modules. Specialization and optimizations by choosing the best VLMs for the specific sub-task pay off. New LLM optimization method works with any LLM or VLM you have access to. Also the latest models (try it out yourself - code" [YouTube Link](https://youtube.com/watch?v=7HxDU8K59k8) 2025-02-24T13:00Z 59.5K followers, [----] engagements "Add Cognitive Topology to Your AI Agents All rights w/ authors: "From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs" Yingjian Chen1 Haoran Liu2 Yinhong Liu3 Sherry T. Tong1 Aosong Feng4 Jinghui Lu5 Juntao Zhang6 Yusuke Iwasawa1 Yutaka Matsuo1 Irene Li1* from [--] University of Tokyo [--] Texas A&M University [--] University of Cambridge [--] Yale University [--] Xiaomi EV [--] Henan University. 00:00 [--] ArXiv preprints to combine 3:33 LEVEL [--] _ Technical 9:13 LEVEL [--] _ Higher Complexities 13:05 LEVEL [--] _ Combinatorials 13:29 IDEA A 21:13 IDEA B 27:06 LEVEL [--] _ Unique 34:24 Cognitive" [YouTube Link](https://youtube.com/watch?v=7d4bEfj7wmc) 2026-01-10T14:15Z 85.2K followers, [----] engagements "My TOP [--] videos to understand & code GNN - Graph Neural Network How YOU start with coding Graph Neural Networks GNN Playlist overview of my [--] GNN videos. The benefits for you when coding GNNs (python code as of December 2021). My library recommendations for you: a) PyTorch Geometric PyG and b) Deep Graph Library DGL. Best free online course on GNN I found (lectures free on Youtube): CS224W: Machine Learning with Graphs Jure Leskovec Stanford University http://cs224w.stanford.edu 00:00 [--] videos 00:40 CS224W 01:50 Data (foundation) 04:57 Graph Representation Learning 07:25 My [--] GNNs 08:35" [YouTube Link](https://youtube.com/watch?v=8xsVQJ72VB8) 2021-12-19T15:00Z 80.9K followers, [----] engagements ""Smartest" VISION AI in Cars Do Reasoning The Illusion of AI Reasoning: They're Not Watching They're Reading. How "Impossible Physics" Broke The Smartest Vision AIs. All rights w/ authors: "INPHYRE Discovers: Large Multimodal Models Struggle in Inductive Physical Reasoning" Gautam Sreekumar Michigan State University Vishnu Naresh Boddeti from Michigan State University" [YouTube Link](https://youtube.com/watch?v=95elVYpxXUU) 2025-09-22T14:00Z 80.2K followers, [----] engagements "New NVIDIA "MASTERS" Distillation: Local 3B Vision AI Stop wasting compute distilling 72B models directly into 8B students because the 'Capacity Gap' is ruining your gradients. When the size difference is this extreme the student physically cannot resolve the teacher's high-dimensional manifold forcing it to learn a noisy "blurry average" of the data instead of precise logic. NVIDIAs new Masters protocol fixes this by mathematically sabotaging the teacher: they apply a dynamic magnitude pruning schedule that artificially lowers the teacher's "IQ" to match the student's capacity then" [YouTube Link](https://youtube.com/watch?v=96XVs6qcIT4) 2026-01-02T14:00Z 85.2K followers, [----] engagements "The Lie We Built: Chain-of-Thoughts Emergent Deception: The Inner Worlds of AI Are Not For Us. Emergent Layers of Cognition in AI. All rights w/ authors: "Stress Testing Deliberative Alignment for Anti-Scheming Training" Bronson Schoen Evgenia Nitishinskaya Mikita Balesni Axel Hjmark Felix Hofsttter Jrmy Scheurer Alexander Meinke Jason Wolfe Teun van der Weij Alex Lloyd Nicholas Goldowsky-Dill Angela Fan Andrei Matveiakin Rusheb Shah Marcus Williams Amelia Glaese Boaz Barak Wojciech Zaremba Marius Hobbhahn from Apollo Research & OpenAI CAN REASONING MODELS OBFUSCATE REASONING STRESS-TESTING" [YouTube Link](https://youtube.com/watch?v=9Eqa5F6IFBk) 2025-10-25T14:00Z 80.5K followers, [----] engagements "Molecular Language Models of Proteins - or Diffusion Crafting new proteins and molecules with specific properties has always been a challenge. Now imagine having the ability to generate these biomolecules almost at will using powerful AI models. But how do you teach an AI to understand the delicate three-dimensional dance of atoms and bonds that define life's fundamental building blocks This video exploration goes deep into a surprisingly effective method: diffusion models typically used for generating images are now being adapted to this incredibly complex realm. It's a journey of turning" [YouTube Link](https://youtube.com/watch?v=9cNxgYhmAAg) 2025-01-13T15:00Z 80.6K followers, [----] engagements "Your next ML (Cloud) Infrastructure for your Code Which ML Framework is best for the new CLOUD infrastructure (independent if NVIDIA H100 or GOOGLE TPUs) The future of Machine Learning Accelerators (NVIDIA Tensor Core H100 GPU and Google's TPU Pod v4) w/ ML compiler = XLA. Plus JAX and TensorFlow3 for new optimized ML Cloud computing in [----]. There could be a winner if you want pure speed and auto-cloud-parallelism over 1000s of TPU-chips v4 for you advanced ML models. #nvidia #h100 #tpu #xla #cloudcomputing" [YouTube Link](https://youtube.com/watch?v=9canD-CiGiQ) 2022-11-09T13:00Z 84.6K followers, [----] engagements "Visualizing the Self-Attention Head of the Last Layer in DINO ViT: A Unique Perspective on Vision AI In a Colab Notebook we code a visualization of the last layer of the Vision Transformer Encoder stack and analyze the visual output of each of the [--] Attention Heads given a specific image. Now we understand how a only pre-trained ViT (although with the DINO method) can not always succeed in an image classification (downstream) task. The fine-tuning of the ViT is simply missing - but essential for a better performance. Based on the COLAB NB by Niels Rogge HuggingFace (all rights with him):" [YouTube Link](https://youtube.com/watch?v=9dYXGGklkhc) 2023-02-18T13:00Z 82.7K followers, [----] engagements "NEW GPT 5.2: A Total Bloodbath Brand new GPT-5.2 was released just hours ago and I tested it not on standard known benchmarks but on my personal logic test for causal reasoning. This test is my base test for a lot of other LLMs from Gemini [--] Pro to the latest OPUS. You find all my other test here in this playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT AT first impression it seems to me that the graph engine of [---] is blocking the complete solution path if encountering a problem. [---] defines problems as pure roadblocks. THis would indicate that [---] move from a" [YouTube Link](https://youtube.com/watch?v=9wg0dGz5-bs) 2025-12-11T23:53Z 82.6K followers, 10.4K engagements "DEEPSEARCH for RLVR and Agentic GraphRAG via RL (MIT Stanford) All rights w/ authors: DEEPSEARCH: OVERCOME THE BOTTLENECK OF REINFORCEMENT LEARNING WITH VERIFIABLE REWARDS VIA MONTE CARLO TREE SEARCH Fang Wu♡ Weihao Xuan Heli Qi Ximing Lu♢ Aaron Tu♠Li Erran Li♣ Yejin Choi♡ from ♡ Stanford University University of Tokyo RIKEN AIP ♢ University of Washington ♠UC Berkeley ♣ Amazon AWS EFFICIENT AND TRANSFERABLE AGENTIC KNOWLEDGE GRAPH RAG VIA REINFORCEMENT LEARNING Jinyeop Song MIT Physics Song Wang University of Central Florida Julian Shun MIT CSAIL Yada Zhu IBM Research #ainews #airesearch" [YouTube Link](https://youtube.com/watch?v=ARst0nlEgO4) 2025-10-03T14:00Z 81.4K followers, [----] engagements "NodePiece code for Knowledge Graphs in Python clever Node embedding in [----] NodePiece a more parameter-efficient node embedding strategy for Knowledge Graphs. In NodePiece a vocabulary of subword/sub-entity units is constructed from anchor nodes in a graph with known relation types. Special benefit: Compositional encoding is inductive by design d.h. new nodes can be added with training the whole Graph. Credits to: arXiv preprint (Feb [--] 2022): https://arxiv.org/pdf/2106.12144.pdf by Mikhail Galkin Etienne Denis Jiapeng Wu and William L. Hamilton GitHub link:" [YouTube Link](https://youtube.com/watch?v=AlHcx7RW1yo) 2022-04-22T11:00Z 82.4K followers, [---] engagements "AI Nano Bio Agents (ETH) NEW Nano Bio-Agent (NBA) framework - implemented by authors (see below) from ETH (Swiss) for GeneTuring benchmark incorporates task decomposition tool orchestration and API access into well-established systems such as NCBI and AlphaGenome. But can scientists work with Small Language Models (SLM) for special genomics tasks complexities How small can we go for our VLM /LLMs Local LLMs All rights w/ authors: Nano Bio-Agents (NBA): Small Language Model Agents for Genomics George Hong ETH Zrich Daniel Trejo Banos Swiss Data Science Centre @UCBerkeley @Google @SDSC" [YouTube Link](https://youtube.com/watch?v=B4Ua8G-OZkw) 2025-09-26T14:00Z 84.5K followers, [----] engagements "AI: New Graph-based Agent Planning (Tsinghua CMU) Empower AI w/ Parallel Thoughts: NEW GAP Framework. All rights w/ authors: GAP: Graph-based Agent Planning with Parallel Tool Use and Reinforcement Learning Jiaqi Wu [--] Qinlao Zhao [--] Zefeng Chen [--] Kai Qin [--] Yifei Zhao [--] Xueqian Wang [--] Yuhang Yao [--] from [--] Tsinghua University [--] Huazhong University of Science and Technology [--] National University of Singapore [--] Carnegie Mellon University @cmu @TsinghuaUniversity_official #airesearch #machinelearning #scienceexplained #aireasoning #aiagents" [YouTube Link](https://youtube.com/watch?v=B9CKm8J9sHY) 2025-10-31T14:00Z 80.9K followers, [----] engagements "RAG Collapses: Reasoning w/ Conflicting Knowledge RAG incorporates a hidden danger that most developers are currently falling into. All rights w/ authors: "Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge" Yiyang Feng♣♦ yiyang.feng@stonybrook.edu Zeming Chen ♣ zeming.chen@epfl.ch Haotian Wu ♣ haotian.wu@epfl.ch Jiawei Zhou ♦ jiawei.zhou.1@stonybrook.edu Antoine Bosselut ♣ antoine.bosselut@epfl.ch from ♣ EPFL ♦ Stony Brook University #retrievalaugmentedgeneration #aireasoning #aiexplained #scienceexplained artificial intelligence AI" [YouTube Link](https://youtube.com/watch?v=BmzdS-a-G8g) 2026-01-25T14:15Z 85.2K followers, [----] engagements "Unified Agentic RAG - NEW AI for Medical Diagnosis Deep-DxSearch introduces a fully trainable agentic RAG system designed for high-stakes medical diagnosis optimized end-to-end via reinforcement learning. The framework models the diagnostic workflow as a sequential decision-making process where an LLM-based agent learns an optimal policy over a specialized multi-tool action space - comprising reason lookup match and search - to interact with a comprehensive multi-modal medical retrieval corpus of patient records disease guidelines and clinical literature. Either MCP client-server protocol or" [YouTube Link](https://youtube.com/watch?v=C7WcaYjR2E8) 2025-08-24T14:00Z 80.4K followers, [----] engagements "Neurosymbolic AI: the Path to Superintelligence Just some personal ideas about the next generation of AI systems that are hyped to achieve superintelligence although we face significant challenges with current Neurosymbolic AI. By the way can you imagine a future without AI models from OpenAI Anthropic META or X A new technology emerging and substituting LLMs @NVIDIA @OpenAI @stanford @TsinghuaUniversity_official #superintelligence #airesearch #scienceexplained #reasoning" [YouTube Link](https://youtube.com/watch?v=CAAE8c-_JK4) 2025-10-15T12:00Z 81.5K followers, 20.2K engagements "AI's Secret Memory Discovered We Were Wrong About How AI Thinks. Newly published ArXiv pre-print raises the foundational question of when and how associative and geometric memory compete with each other during optimization and what factors - such as training time learning rate weight decay - can foster one over the other. More careful empirical research and ideation may be needed to make the geometric view more broadly applicable. Orthogonally new findings are of relevance to making choices between parametric vs. contextual memory and also between generative retrieval vs. dual encoder" [YouTube Link](https://youtube.com/watch?v=D1lh3XXTUlk) 2025-11-01T14:00Z 80.8K followers, [----] engagements "Agent2Agent + (MCP to Tool) in Multi-Agent AI Google's new Agent2Agent Protocol Explained and its compatibility to a simple MCP protocol for tool use by an LLM and external data infusion - by Anthropic. Cross-ecosystem compatible Agent2Agent protocol introduced (incl LangGraph and crew.ai). In reference to my latest video on the new AGENT Development Kit - ADK by Google (multi-agent system dev) https://youtu.be/Geo8LzCHoMQ Detailed info on A2A and a lot of A2A Python code you find here https://github.com/google/A2A The exact JSON Schema for A2A Protocol you find here" [YouTube Link](https://youtube.com/watch?v=DAQ6msUVOp0) 2025-04-13T14:00Z 81.7K followers, 11.2K engagements "Vector Embeddings: NEW Geometric Limit Discovered The authors (from Google Deepmind Johns Hopkins University): " .we observe that even state-of-the-art models fail on this dataset despite the simple nature of the task. Our work shows the limits of embedding models . " "In this work we demonstrate that we may encounter these theoretical limitations in realistic settings with extremely simple queries. We connect known results in learning theory showing that the number of top- subsets of documents capable of being returned as the result of some query is limited by the dimension of the" [YouTube Link](https://youtube.com/watch?v=DUKzFrodzxE) 2025-09-03T14:00Z 82.6K followers, 13.1K engagements "The Algebra of AI Thoughts: Self-Learn Reasoning We often wonder if AI can truly teach itself to become smarter. It's a question that feels like it's straight out of science fiction but for the first time we have a scientifically proven answer. In this video we'll dive into the groundbreaking mathematical theory that demonstrates how a Transformer AI can in fact "bootstrap" its own intelligence teaching itself to solve reasoning problems of ever-increasing length. But this is not a story of infinite runaway growth. We will also explore the rigorous provable limits to this self-improvement" [YouTube Link](https://youtube.com/watch?v=E-4X-WDOSeA) 2025-11-12T13:45Z 81.6K followers, [----] engagements "Lean AI Reasoning: NEW Energy-Based Chain-of-Thought Optimizing Latent AI Thought Trajectories via Energy-Based Calibration. All rights w/ authors: OckBench: Measuring the Efficiency of LLM Reasoning Zheng Du* Georgia Institute of Technology Hao Kang* Georgia Institute of Technology Song Han Massachusetts Institute of Technology Tushar Krishna Georgia Institute of Technology Ligeng Zhu Nvidia Cooperation THINK CONSISTENTLY REASON EFFICIENTLY: ENERGY-BASED CALIBRATION FOR IMPLICIT CHAIN-OF-THOUGHT Zhikang Chen1 Sen Cui2 Deheng Ye3 Yu Zhang [--] Yatao Bian [--] Tingting Zhu [--] from [--] University of" [YouTube Link](https://youtube.com/watch?v=E-DME8XfzXs) 2025-11-11T16:34Z 81.5K followers, [----] engagements "Contextual Instantiation of AI Persona Agents (Stanford) All rights w/ authors: Ask WhAI: "Probing Belief Formation in Role-Primed LLM Agents" Keith Moore Jun W. Kim David Lyu Jeffrey Heo Ehsan Adeli from Department of Biomedical Data Science Stanford University HARMFUL TRAITS OF AI COMPANIONS W. Bradley Knox [--] Katie Bradford [--] Samanta Varela Castro [--] Desmond C. Ong [--] Sean Williams [--] Jacob Romanow [--] Carly Nations [--] Peter Stone [--] Samuel Baker [--] from [--] UT Austin Department of Computer Science [--] UT Austin Department of Communication Studies [--] UT Austin Technology & Information Policy Institute" [YouTube Link](https://youtube.com/watch?v=ERJ2s73HwDs) 2025-11-22T13:15Z 81.7K followers, [----] engagements "Meta Reinforcement Fine-Tuning AI vs GRPO (MRT by CMU) MRT (Meta Reinforcement Fine-Tuning) redefines how AI models optimize test-time compute by embedding dense step-level rewards into each segment of the reasoning process. Instead of waiting for a final outcome MRT minimizes cumulative regret by continuously evaluating the progress of each reasoning episode. This approach enables models to balance exploration and exploitation dynamically leading to more efficient and robust decision-making even as computational budgets scale. All right w/ authors: "Optimizing Test-Time Compute via Meta" [YouTube Link](https://youtube.com/watch?v=Eh1xxtcNqrA) 2025-03-12T15:00Z 82.1K followers, [----] engagements "AI's Intellectual Dark Matter Discovered All rights w/ authors: "Inverse Knowledge Search over Verifiable Reasoning: Synthesizing a Scientific Encyclopedia from a Long Chains-of-Thought Knowledge Base" Yu Li12 Yuan Huang3 Tao Wang4 Caiyu Fan32 Xiansheng Cai2 Sihan Hu6 Xinzijian Liu3 Cheng Shi5 Mingjun Xu3 Zhen Wang3 Yan Wang3 Xiangqi Jin7 Tianhan Zhang8 Linfeng Zhang7 Lei Wang4 Youjin Deng69 Pan Zhang2106 Weijie Sun3 Xingyu Li12 Weinan E111213 Linfeng Zhang312* Zhiyuan Yao12* Kun Chen2* from [--] Lanzhou Center for Theoretical Physics Key Laboratory of Theoretical Physics of Gansu Province Key" [YouTube Link](https://youtube.com/watch?v=EosSD36-niQ) 2025-11-04T13:45Z 80.9K followers, [----] engagements "How to start with ChatGPT Short Introduction to OpenAI API #shorts New to ChatGPT How to start [--] easy steps for beginners w/ ChatGPT. See also the documentation about: https://beta.openai.com/docs/introduction https://beta.openai.com/overview https://beta.openai.com/examples Info about their content filter: https://beta.openai.com/docs/models/content-filter Info how to fine-tune models: https://beta.openai.com/docs/guides/fine-tuning Fine-tuning improves on few-shot learning by training on many more examples than can fit in the prompt letting you achieve better results on a wide number of" [YouTube Link](https://youtube.com/watch?v=F76njQqhOIw) 2023-01-22T13:00Z 81.7K followers, [---] engagements "Sort Pandas DataFrame in Python Code #Shorts How to sort a Pandas dataframe (sort_values) in Python on a COLAB NB. Sort values of a pandas dataframe. Load your excel file or your csv file and create a pandas dataframe within a Jupyter Notebook on COLAB. Welcome to Pandas - a Python Data Analysis Library. #dataframes #sort #shorts #pandasdataframe #datascience #dataframe #pythonprogramming #colab #pandas" [YouTube Link](https://youtube.com/watch?v=FaJiiZzEFXc) 2022-03-08T08:30Z 83.8K followers, [---] engagements "NEW Qwen3-2507: Independent Benchmark (w/ Kimi K2) LIVE TEST: Two non-reasoning models show their performance on a reasoning benchmark. Is this possible at all Is there really a diff between reasoning and non-reasoning models Are benchmarks published by global corporations trustworthy All answers available in my new video. Note: NEW separate Qwen3-2507-thinking model will be released soon. Note: I test only non-quantized models. Quantized version of the new Qwen3 - [----] you can find here: https://huggingface.co/unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF and" [YouTube Link](https://youtube.com/watch?v=FtrLaHeEP4E) 2025-07-22T16:15Z 82.3K followers, [----] engagements "I Learn How to Code Agents w/ Google's NEW ADK Google's new Agent Developer Kit (ADK) was published today as the new standard for effortless coding of AI Agent across all models companies and platforms. Compatible with MCP servers with Crew.ai LangGraph many more . and OpenAPI applications. This video guides you through the ADK manual the detailed workflow specific agentic instructions of ADK looks at massive amounts of code examples on how to build your perfect Ai agent with specified tools in multi-agent configs with dynamic memory and state dependent logic / reasoning. How to code more" [YouTube Link](https://youtube.com/watch?v=Geo8LzCHoMQ) 2025-04-10T12:30Z 80.2K followers, 18.6K engagements "TEST Claude [---] Thinking: BEST EVER Anthropic released (just hours ago) the new CLAUDE [---] model in two variants. Non-thinking and thinking AI. I test both new AI models in my real world logic test. The results of other AI models on this identical test routine you can watch live in my YouTube Playlist "LOGIC TESTS for AI" https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT https://www.anthropic.com/news/claude-opus-4-5 00:00 CLAUDE [---] Overview 02:21 CLAUDE [---] Real world TEST 04:33 Validation run 11:50 CLAUDE [---] Thinking 32K 18:16 2nd run CLAUDE [---] Thinking 32K" [YouTube Link](https://youtube.com/watch?v=HpOB8AabHF4) 2025-11-25T14:00Z 81.7K followers, [----] engagements "Beyond Transformer: Building an Artificial Mind All rights w/ authors: Intilligence Foundation Model: A New Perspective to Approach Artificial General Intelligence Borui Cai Yao Zhao B. Cai is with Hangzhou International Innovation Institute Beihang Uni- versity China. Y. Zhao is with RMIT University Melbourne Australia. #airesearch #artificialintelligence #artificiallearning #neuroscience" [YouTube Link](https://youtube.com/watch?v=IADccLs--lM) 2025-11-16T13:45Z 81.6K followers, [----] engagements "Pre-Train BERT from scratch: Solution for Company Domain Knowledge Data PyTorch (SBERT 51) We pretrain a BERT (Bidirectional Encoder Representations from Transformers) model from scratch in PyTorch on domain specific data (eg confidential company data). We code in Python to train an optimized Tokenizer for our data design a BERT architecture from scratch and start pre-training of BERT with a masked Language Model Head (MLM). We define the vocabulary size according to our needs (from 8K to 60K) define the depth of our BERT architecture (eg [--] layers) and train days on (a single) GPU for our" [YouTube Link](https://youtube.com/watch?v=IcrN_L2w0_Y) 2023-01-15T13:00Z 80.6K followers, 13.1K engagements "Reasoning TEST GPT-5.1: A Surprise A brand new GPT-5.1 has been released yesterday - and I test it with my causal reasoning test that I use for all LLM tests in the last year. So you have a perfect comparison of GPT-5.1 performance to the other LLMs. See also my YouTube Playlist for (at least) [--] other AI model-tests on YouTube - with the identical reasoning test: https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT "LOGIC TESTS FOR AI" 00:00 TEST GPT-5.1 Reasoning 02:23 RESTART GPT-5.1Instant 04:30 Optimization run GPT-5.1 06:45 RE-Try run GPT-5.1 07:50 Last run GPT-5.1" [YouTube Link](https://youtube.com/watch?v=IhbOpIeQtPg) 2025-11-14T08:20Z 82.5K followers, [----] engagements "Agent Builder w/ ChatKit: West Coast Disaster First day of the new OPENAI Agent Builder and ChatKIT for multi-agent system design. First exploration only (no profound recommendation no massive insights) and an example with GPT-5 PRO as the core intelligence of the agents on the topic of global ecosystems and in particular the West-coast of the US. Create an agent workflow with Agent Builder. Agent Builder is a visual canvas for designing multi-step agent workflows. You'll get a workflow ID. Again we encounter a reduction in system complexity by defining several agents (with tool use) for real" [YouTube Link](https://youtube.com/watch?v=IjLHe-Qq8o8) 2025-10-07T14:00Z 82.9K followers, [----] engagements "FIRE all AI Agents New SCALING Laws (Google MIT) New research from Google DeepMind & MIT proves that the "More Agents = Better" heuristic is mathematically wrong. We analyze their study of [---] agent architectures to find the "Tool-Coordination Trade-off" and why independent agent swarms amplify errors by 17.2x. Technical note: To keep the narrative fluid I presented a simplified conceptual form of the new scaling law. However for those conducting real research in agent orchestration the pre-print defines a more advanced rigorous mathematical functional form () in Equation [--] (Page 16) and" [YouTube Link](https://youtube.com/watch?v=IvJgrwp1VUk) 2025-12-12T14:15Z 84.3K followers, [----] engagements "GPT is Not The Future of AI: NEW AI Topology OpenAI's IPO approaching fast let us reconsider their technology. Given the massive limitations of GPT AI systems today we explore new AI architectures of the Future that are more powerful than our current system. We will focus on open-source AI smaller model sizes for local use and improved performance - also for multi-modal use cases. We will touch on massive data centers versus more intelligent specialized edge AI devices for distributed intelligence. GPT transformer architecture might reach the end of dev in [----] and new AI architectures might" [YouTube Link](https://youtube.com/watch?v=Ixrpkub47vg) 2025-12-21T13:15Z 84.1K followers, [----] engagements "Qwen3 NEXT A3B for Reasoning and MCP Tools I test the new Qwen3 MoE 80B A3B model on a complex causal reasoning test. Detailed explanations and live recoding of performance test. https://qwen.ai/blog https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking #aitesting #reasoning #aiexplained" [YouTube Link](https://youtube.com/watch?v=J9br0e34cp0) 2025-09-16T07:00Z 84K followers, [----] engagements "RLHFs Missing Piece: Qwens World Model Aligns AI w/ Human Values (GRPO) After multiple Qwen3 models now Qwen published a new world model (WorldPM) for human preferences (RLHF) and further explored specific scaling laws regarding the model size and effectiveness. Beyond Qwen3: Qwen's New WorldPM (Paper & Model). How Qwens GRPO World Model Solves RLHFs Biggest Flaw: Human Values. RLHFs Missing Piece: NEW World Model (by Qwen) That Thinks Like Humans (GRPO). How we built the first AI world model to encode real human preferences at scale. All rights w/ authors: "WorldPM: Scaling Human Preference" [YouTube Link](https://youtube.com/watch?v=JNNhXzwokZY) 2025-05-18T12:00Z 81.6K followers, [----] engagements "LeanRAG: Multiple Layers of Knowledge Graphs (RAG 3.0) LeanRAG: Hierarchical Knowledge Graphs for RAG [---]. (see also my video on: Hierarchical Reasoning Models - HRM) all rights w/ authors: LeanRAG: Knowledge-Graph-Based Generation with Semantic Aggregation and Hierarchical Retrieval Yaoze Zhang [--] Rong Wu [--] Pinlong Cai [--] Xiaoman Wang [--] Guohang Yan [--] Song Mao [--] Ding Wang [--] Botian Shi [--] from [--] Shanghai Artificial Intelligence Laboratory [--] University of Shanghai for Science and Technology [--] Zhejiang University [--] East China Normal University #aiexplained #science #knowledgegraph" [YouTube Link](https://youtube.com/watch?v=JyILtjM3dCE) 2025-08-19T14:00Z 80.5K followers, [----] engagements "Unlocking the Potential of Message Passing: Exploring GraphSAGE GCN and GAT GNN GraphML Introduction to GRAPH ML Graph Neural Networks (GNN) and the main idea behind Message Passing in graph network configurations of GraphSAGE GCN and GAT. Message passing applied to Graph Convolutional Networks (GCN) GraphSAGE and Graph Attention Networks. The key difference between GAT and GCN is how the information from the k-hop neighborhood is aggregated. Stanford online: CS224W https://www.youtube.com/watchv=JAB_plj2rbA&list=PLoROMvodv4rPLKxIpqhjhPgdQy7imNkDn #ai #graphs #theory" [YouTube Link](https://youtube.com/watch?v=K9w7-4N_ZCI) 2022-12-07T13:00Z 81.4K followers, 10.1K engagements "Multi DeepSeek R1: STEP-GRPO RL MultiModal My video explores new Ai research on R1 multi-Modal reasoning and demonstrates clearly how StepGRPOs step-wise rewards enable more reliable structured and logically sound reasoning in multimodal large language models. By offering continuous and detailed feedback on both accuracy and validity these rewards foster incremental improvements that go beyond passive supervised imitation resulting in superior performance demonstrated across multiple reasoning benchmarks. All rights w/ authors: "R1-VL: Learning to Reason with Multimodal Large Language Models" [YouTube Link](https://youtube.com/watch?v=KbeWVLvQhX8) 2025-03-19T15:00Z 84.5K followers, [----] engagements "New Graph Diffusion Transformer #ai Generative AI has a dirty secret: it is terrible at graph theory. While Diffusion Models can paint masterpieces they treat complex networks like liquid turning precise molecules and road maps into chaotic "hairballs" or shattered disconnected messes the moment you try to force a condition on them. In this video we reveal the solution: CoPHo. We dive into how researchers are fixing this by stopping the AI from "painting" edges and instead teaching it to "sculpt" them using Persistent Homology. Join me to see how this algebraic topological "tool" allows us to" [YouTube Link](https://youtube.com/watch?v=LEmndg6Q6k4) 2025-12-27T13:45Z 85.2K followers, [----] engagements "HiRAG: Hierarchical Reasoning for GraphRAG (BEST RAG) HiRAG is a hierarchical extension of retrieval-augmented generation that addresses a key weakness of GraphRAG: flat graphs conflate fine-grained entities with broad conceptual structure making multi-hop reasoning brittle. The method introduces HiIndex an offline process that recursively clusters entity embeddings with Gaussian mixture models then uses an LLM to generate summary entities at higher layers linked downward to their members. This produces a hierarchical knowledge graph where upper layers serve as semantic shortcuts abstracting" [YouTube Link](https://youtube.com/watch?v=LV0jRVXtx80) 2025-08-22T14:00Z 81.4K followers, [----] engagements "Self Learning AI: Accelerate w/ new RL The frontier of LLM research has shifted decisively toward Post-Training and System [--] reasoning. We all know the recipe for replicating O1-level performance: move beyond Supervised Fine-Tuning and embrace Reinforcement Learning with Verifiable Rewards (RLVR). The ultimate goal for every research lab right now is Self-Supervised RL: allowing the model to generate its own questions verify its own reasoning chains and improve indefinitely without needing expensive unscalable human annotations. However this new pre-print exposes a critical instability that" [YouTube Link](https://youtube.com/watch?v=LbUBncFv9yM) 2025-12-20T14:00Z 82.8K followers, [----] engagements "AI is Just a Correction Term (to Physics) Why Your Physics Model Needs an AI Residual. Forget AGI - Sorry Sam. AI is Just a Correction Term to our complex real-world computer simulations. All rights w/ authors: NeuralOGCM: Differentiable Ocean Modeling with Learnable Physics Hao Wu [--] Yuan Gao [--] Fan Xu [--] Fan Zhang [--] Guangliang Liu [--] Yuxuan Liang [--] Xiaomeng Huang [--] from [--] Tsinghua University Beijing China [--] University of Science and Technology of China Hefei China [--] The Chinese University of Hong Kong Hong Kong China [--] Hong Kong University of Science and Technology (Guangzhou) Guangzhou China." [YouTube Link](https://youtube.com/watch?v=LcJRbt8lPhk) 2025-12-17T14:00Z 82.7K followers, [----] engagements "LIQUID AI 40B (MIT): REAL Performance on Reasoning (My [--] Tests) New LIQUID Foundation Models (LFM): Liquid AI (an MIT spin-off) just released [--] new AI model three LIQUID Foundation models. I examine the logic reasoning performance of their biggest model LIQUID 40B on their official platform on [--] logical reasoning tests. Details on my [--] logical reasoning test you find in my videos: Extreme logic test https://www.youtube.com/watchv=e1Zup0Wib5A LLM (w/ RAG) need a new Logic Layer (Stanford) https://www.youtube.com/watchv=42gHxqLu0Kk Rethinking AI: Solutions for Logical Reasoning" [YouTube Link](https://youtube.com/watch?v=M_v5f5Mvzxo) 2024-10-01T14:00Z 83K followers, [----] engagements "BRAIN.COPY = Latent-MAS AI Breakthrough We are moving beyond lossy text-based collaboration. The new Latent-MAS framework introduces a paradigm of Pure Latent State Injection. By utilizing a Linear Alignment Operator derived via ridge regression this method enables agents to perform "Silent Reasoning" loops in continuous vector space - structurally circumventing the vocabulary projection entirely. Crucially it solves the collaboration bandwidth problem via Lossless KV-Cache Inheritance: physically transplanting the entire high-dimensional working memory tensor from one agents neural stack to" [YouTube Link](https://youtube.com/watch?v=NreIscoJe8o) 2025-12-01T13:45Z 82.9K followers, [----] engagements "AI visualizes insight from Accenture's TechVision [----] (SBERT 5) An AI explores a tech vision document. AI induced insights. Augment your human tech vision with AI. #sbert #nlproc #nlptechniques #clustering #semantic #bert #climatechange #3danimation #3dvisualization #topologicalspace #deeplearning #machinelearningwithpython #pytorch #sentence #embedding #complex #umap #insight #computerscience #SentenceTransformer #ai #networkx #plotly #3dvisualization #visualization #3danimation" [YouTube Link](https://youtube.com/watch?v=NxnIpFmcW1I) 2021-04-09T05:15Z 82.6K followers, [---] engagements "New: AI Agent Self-Improvement + Self-Fine-Tune Reinforcement Self-Training (REST) and Fine-Tuning of LLMs meet ReACT-style LLM agent for reasoning and action on external data by Google on the topic of Medicine. AI Agent to self-improve + self-fine-tune. Reward policy optimization and ranking code now evolves to a simple prompt: Prompt Engineering in [----] continues Advanced Local LLM Update Mechanism The described system introduces a mechanism for overnight self-updating of local Large Language Models (LLMs) such as those on Mac Mini or Mac Studio (192GB unified mem) devices. Users can" [YouTube Link](https://youtube.com/watch?v=O5iLfzSFptg) 2023-12-26T13:00Z 80.6K followers, 10.2K engagements "COMPASS: The Cognitive Upgrade for Multi-Agent AI We've all seen it happen: you give an AI agent a complex long-horizon task and after a few steps it starts to drift. It forgets critical constraints gets stuck in repetitive loops and ultimately loses the plot. The problem isn't the agent's raw intelligence; it's a crisis of context. We need context engineering. Today we're diving in my video into COMPASS a groundbreaking new framework that tackles this head-on by giving agents what they've been missing: a dedicated strategic brain to supervise the work and a smart context manager to keep them" [YouTube Link](https://youtube.com/watch?v=OFJNxpAC9EM) 2025-10-21T14:00Z 84.5K followers, [----] engagements "Unified Theory of Agentic Reasoning (Berkeley NVIDIA) New video: Unified Theory of Agentic Reasoning - The Geometric Edition. Q-Learning Gradient policy RL Large Reasoning Models SFT Reasoning manifold Low-dimensional subspaces complex reasoning agentic reasoning agentic reasoning graph GPT-5 DeepSeek V3. All rights w/ authors: GSM-AGENT: UNDERSTANDING AGENTIC REASONING USING CONTROLLABLE ENVIRONMENTS Hanlin Zhu [--] Tianyu Guo [--] Song Mei [--] Stuart Russell [--] Nikhil Ghosh [--] Alberto Bietti [--] Jiantao Jiao [--] from [--] UC Berkeley [--] Flatiron Institute [--] Nvidia REMA: A UNIFIED REASONING MANIFOLD FRAME-" [YouTube Link](https://youtube.com/watch?v=P2_nACjEcq4) 2025-10-01T12:01Z 82.1K followers, [----] engagements "Base LLM can Reason: Activation Switch found This video presents compelling causal evidence that the core machinery of advanced reasoning - the ability to backtrack verify and compute - already exists fully formed but dormant within the base models themselves. If the reasoning skills are already there in the base LLMs then what exactly have we been training (RLVR) all this time and what does it mean when the primary act of "learning" is simply the art of orchestration Ground breaking new study. All rights w/ authors: BASE MODELS KNOW HOW TO REASON THINKING MODELS LEARN WHEN Constantin Venhoff" [YouTube Link](https://youtube.com/watch?v=PESHaTTnC6M) 2025-10-12T14:00Z 80.8K followers, [----] engagements "Data Scientists devastated - Databricks AI networks create themselves Latest dev on Deep Learning (DL) w/ Databricks: TensorFlow Keras Hyperopt and MLflow auto-tune a neural network. Jupyter NB provided by Databricks code segments include TensorFlow Keras Hyperopt MLflow and other common python frameworks. Code execution on Databricks Community Edition for educational and demonstration purposes only. Real time coding. Yes a Python Jupyter Notebook which creates a Neural Network model with TensorFlow (trains w/ TensorBoard online visualization) performs automated hyperparameter tuning with" [YouTube Link](https://youtube.com/watch?v=Q-nahhtS4dc) 2021-12-03T09:00Z 84.1K followers, [--] engagements "Riemann Liquid Spatio-Temporal Graph The End of "Flat Earth" AI Both new research pre-prints deliver a unified wake-up call: Current Deep Learning architectures are too "flat" to model reality. Whether it is the geometric flatness of Euclidean space (RLSTG) or the attentional flatness of the Context Window (Chimpanzee paper) treating the world as a uniform instantly accessible grid is causing SOTA models to fail. We are learning that simply feeding more data into a Transformer does not magically teach it the structural laws of physics or the privacy boundaries of a human mind. The "Bitter" [YouTube Link](https://youtube.com/watch?v=Q1ugd9NYPUA) 2026-01-22T14:15Z 85.2K followers, [----] engagements "AI calculates Your Future Professional Career Choices (MIT) AI Avatars for professional choices: Simulating Your Future Professional Career Choices with AI. Can AI Predict Your Future Professional Career Choices You are human. Just a simple pattern for AI Let's explore. All rights w/ authors: "Simulating Life Paths with Digital Twins: AI-Generated Future Selves Influence Decision-Making and Expand Human Choice" Rachel Poonsiriwong MIT Media Lab Cambridge MA USA rachelpo@mit.edu Chayapatr Archiwaranguprok MIT Media Lab Cambridge MA USA pub@mit.edu Constanze Albrecht MIT Media Lab Cambridge MA" [YouTube Link](https://youtube.com/watch?v=QU15HjhvUu0) 2025-12-09T13:45Z 82.5K followers, [----] engagements "NEW AI Models: Hierarchical Reasoning Models (HRM) Explore a new AI architecture that combines recurrent neural networks (RNN) with Transformers (but not GPT). A new optimization framework for advanced reasoning AI models. @TsinghuaUniversity_official All rights w/ authors: Hierarchical Reasoning Model Guan Wang [--] Jin Li [--] Yuhao Sun [--] Xing Chen [--] Changling Liu [--] Yue Wu [--] Meng Lu [--] Sen Song [--] Yasin Abbasi Yadkori [--] from [--] Sapient Intelligence Singapore [--] Tsinghua University #transformer #airesearch #ainews #aiexplained #science #transformers #singapore #recurrent" [YouTube Link](https://youtube.com/watch?v=QWD55guu0So) 2025-07-02T14:00Z 84K followers, 28.7K engagements "NEW TextGrad by Stanford: Better than DSPy In this TEXTGRAD framework each AI system is transformed into a computation graph where variables are inputs and outputs of complex (not necessarily differentiable) function calls. The feedback to the variables (dubbed textual gradients) are provided in the form of informative and interpretable natural language criticism to the variables; describing how a variable should be changed to improve the system. The gradients are propagated through arbitrary functions such as LLM API calls simulators or external numerical solvers. (Stanford Univ) Stanford" [YouTube Link](https://youtube.com/watch?v=Qks4UEsRwl0) 2024-06-16T12:00Z 80.6K followers, 19.7K engagements "Seed-Prover vs Deep Think (IMO) Google AI Deep Think IMO ( available to Ultra users) won Gold Medal in IMO [----] - Math Olympiad. ByteDance Seed-Prover Achieves Silver Medal Score in IMO [----]. OPENai was officially not participating. New detailed tech report by ByteDance on the training and the methods of their Seed-Prover. All main new insights presented in this video. All rights w/ authors: "Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving" ByteDance Seed AI4Math https://github.com/ByteDance-Seed/Seed-Prover #reasoning #airesearch #programming #lean4 #olympiadmathematics" [YouTube Link](https://youtube.com/watch?v=R628msJolV0) 2025-08-02T14:00Z 84.6K followers, [----] engagements "After Diffusion & FLOW Models: Equilibrium Matching (MIT Oxford Harvard) NEW Equilibrium Matching is a glimpse into the future of image generation AI. All rights w/ authors: Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive Tyler Farghly Peter Potaptchik Samuel Howard George Deligiannidis Jakiw Pidstrigach from Department of Statistics University of Oxford EQUILIBRIUM MATCHING: GENERATIVE MODELING WITH IMPLICIT ENERGY-BASED MODELS Runqian Wang MIT Yilun Du Harvard University @harvard @mit @oxforduniversity #airesearch #imageai #artificialintelligence" [YouTube Link](https://youtube.com/watch?v=RKUDo42GwI0) 2025-10-06T12:45Z 82.5K followers, [----] engagements "Stable Diffusion x10: LCM-LoRA (CODE & Theory) We don't stop at LCM We go further: LCM-LoRA Turbo charge SDXL x10 LCM-LoRA explained with Python code examples: New developments in stable diffusion especially in Latent Consistency Models (LCM) enable a speed boost for calculating the PF-ODE for reverse diffusion. Stable Diffusion now 10x faster for text-to-image generation w/ LCM-LoRA. Python code examples for LCM-LoRA w/ acceleration vector and style vector integration via multiple LoRA Adapters. Quality comparison of SDXL (classical) with LCM-LoRA SDXL adapter (w/ and w/o style vector" [YouTube Link](https://youtube.com/watch?v=S9-nNAz-7fM) 2023-11-22T15:00Z 82.1K followers, [----] engagements "Gemini [--] PRO in [---] seconds 2nd test on the causal reasoning capabilities of the new GEMINI [--] PRO. A different logic test - with GEMINI [--] PRO on HIGH. My 1st live test of GEMINI [--] PRO is available here https://youtu.be/pBKtjSbNWLk #gemini3pro #test #artificialintelligence #aiexplained #googledeepmind #google" [YouTube Link](https://youtube.com/watch?v=SMdaxjkTIJA) 2025-11-19T06:00Z 81.6K followers, [----] engagements "AI creates Video Game: Genie [--] . and PaliGemma [--] CODE: ---------- https://colab.research.google.com/github/merveenoyan/smol-vision/blob/main/Fine_tune_PaliGemma.ipynbs=03#scrollTo=EB0gv8OzHfLV https://github.com/google-gemini/gemma-cookbook 00:00 Genie [--] Prompt-to-Game 02:38 Autoregressive Latent Diffusion 05:11 SIMA Agents in 3d virtual world 06:37 Genie [--] DEMO 08:46 OpenAI o1 PRO model 09:14 PaliGemma [--] Intro 11:47 Pre-training datasets 13:11 Finetune PaliGemma2 w JAX 14:00 PT and FT models on HF 16:22 PaliGemma [--] DEMO 20:22 PaliGemma [--] CODE NB #ai #airesearch #vision #gamedev #coding" [YouTube Link](https://youtube.com/watch?v=SQRu0cqmWqE) 2024-12-06T13:00Z 84.6K followers, [----] engagements "Finally: Grokking Solved - It's Not What You Think Grokking or the sudden generalization by AI models to new knowledge - that occurs after prolonged overfitting in LLMs is a surprising phenomenon that has challenged our understanding of deep learning and AI in general. While a lot of progress has been made in understanding grokking finally we get some answers -we have been waiting for [--] months to be discovered. GROKKING - Finally understood Part II is available here https://youtu.be/H3OofROzlA0 All rights w/ authors: GROKKING AT THE EDGE OF NUMERICAL STABILITY Lucas Prieto Melih Barsbey Pedro" [YouTube Link](https://youtube.com/watch?v=SRfJQews1AU) 2025-01-14T15:00Z 83.9K followers, 22.2K engagements "NEW Qwen3 MAX - Performance TEST Sept [--] #qwen3 Alibaba's Most Powerful AI Model: Qwen [--] MAX released Sept [--] [----]. #qwenai #reasoning" [YouTube Link](https://youtube.com/watch?v=STluOuEBTKU) 2025-09-26T06:15Z 84.5K followers, [----] engagements "CRM Integration: Salesforce - Einstein GPT Current status of GPT AI integration in a complete Commercial Ecosystem focus on Generative AI application and business integration. Example: Einstein GPT Salesforce. Not sponsored. Just interested in current level of AI integration in business processes. All rights with the sources and their authors: https://www.salesforce.com/news/stories/salesforce-gpt-resources/ https://developer.salesforce.com/blogs/2023/08/bring-your-own-ai-models-to-salesforce-with-einstein-studio" [YouTube Link](https://youtube.com/watch?v=SXF6lH5Isew) 2023-08-20T12:00Z 80.6K followers, [----] engagements "AGI Finally Achieved: o4-mini Comparing the performance of OpenAi's o4 models I performed my causal reasoning test (elevator test) where o4 refused to accept the correct results (as given by Gemini [---] PRO) and trying to justify its own first answer which failed in the performance rating by inventing new rules. o4-mini was not hallucinating but strategically lying to me. Almost convincingly. o4-mini tried to hold on to its first answer rejecting correct answers neglecting arguments ignoring facts and strategically constructing new tactical rules to justify its first answer given. Holding on" [YouTube Link](https://youtube.com/watch?v=SgknC9B1dm8) 2025-06-24T14:00Z 82.6K followers, [----] engagements "Latest on AI for Scientific Discovery (SAGA) Automated Scientific Discovery w/ AI We explore AI Scientist as multi-agent self-learning systems at the end of [----]. What have we achieved with AI Science how autonomous are pure AI scientists Can AI scientists really augment human professionals in medicine or pharmacology Today we have a detailed look at Scientific Autonomous Goal-evolving Agent (SAGA). Solving Reward Hacking in Scientific AI. SAGA: "System 2" for Automated Discovery by AI. All rights w/ authors: Accelerating Scientific Discovery with Autonomous Goal-evolving Agents Yuanqi Du1*" [YouTube Link](https://youtube.com/watch?v=T4g5uSaY3Ko) 2025-12-30T14:00Z 85.2K followers, [----] engagements "DEEPSEEK: NEW Paper (MLA MTP FP8T EP) before R2 This new AI research paper by DeepSeek (May [--] 2025) presents an in-depth analysis of the next DeepSeek model architecture and its AI infrastructure highlighting key innovations such as Multi-head Latent Attention (MLA) for enhanced memory efficiency Mixture of Experts (MoE) architectures for optimized computation-communication trade-offs FP8 mixed-precision training to unlock the full potential of hardware capabilities and a Multi-Plane Network Topology to minimize cluster-level network overhead. All rights w/ authors: "Insights into" [YouTube Link](https://youtube.com/watch?v=T8Ty99O4m0w) 2025-05-15T11:45Z 84.6K followers, 18.3K engagements "Understanding 4bit Quantization: QLoRA explained (w/ Colab) QLoRA 4bit Quantization for memory efficient fine-tuning of LLMs explained in detailed. 4-bit quantization QLoRA for beginners theory and code. PEFT - parameter efficient fine-tuning methods. Based on my first videos on the theory of LoRA and other PEFT methods (https://youtu.be/YVU5wAA6Txo) and the detailed code implementation of LoRA in my video (https://youtu.be/A-a-l_sFtYM) now my third video on 4-bit quantization and QLoRA. An additional Colab NB with code to fine-tune FALCON 7B with QLoRA 4-bit quantization and Transformer" [YouTube Link](https://youtube.com/watch?v=TPcXVJ1VSRI) 2023-06-11T12:15Z 81.7K followers, 49.2K engagements "Google Stanford AI Co-Scientist: The SCIENCE YOU want All rights w/ authors: Towards an AI co-scientist Juraj Gottweis1 Wei-Hung Weng2 Alexander Daryin1 Tao Tu3 Anil Palepu2 Petar Sirkovic1 Artiom Myaskovsky1 Felix Weissenberger1 Keran Rong3 Ryutaro Tanno3 Khaled Saab3 Dan Popovici2 Jacob Blum7 Fan Zhang2 Katherine Chou2 Avinatan Hassidim2 Burak Gokturk1 Amin Vahdat1 Pushmeet Kohli3 Yossi Matias2 Andrew Carroll2 Kavita Kulkarni2 Nenad Tomasev3 Vikram Dhillon4 Eeshit Dhaval Vaishnav5 Byron Lee5 Tiago R D Costa6 Jos R Penads6 Gary Peltz7 Yunhan Xu3 Annalisa Pawlosky1 Alan Karthikesalingam2 and" [YouTube Link](https://youtube.com/watch?v=TUo1VeeBgOU) 2025-02-22T15:00Z 80.4K followers, [----] engagements "Understand DSPy: Programming AI Pipelines The origin and evolution of DSPy: Programming AI Pipelines introduces the idea its link to ColBERT v2 retriever models modular pipeline generation descriptive programming the evolution and the use case of DSPy (DSPy == Declarative Self-improving Language Programs pythonically). Q answered: Is DSPy only a Prompt Engineering optimization Q answered: Is DSPy expensive for my AI pipeline optimization Q answered: Can I substitute DSPy with a simple many-shot In-Context Learning prompt #airesearch" [YouTube Link](https://youtube.com/watch?v=U5ZuMZkZBSY) 2024-05-07T12:00Z 80.4K followers, [----] engagements "Hierarchical Reasoning HRM 2.0: NEW Attractor Dynamics in AI Hierarchical Reasoning Models are a powerful alternatives to autoregressive Ai models like ChatGPT. Today we further optimize these HRM for improved reasoning performance and discover a fixed point trap on their manifolds. Plus the explanation for Grokking - which also happens to HRM. All rights w/ authors: "Are Your Reasoning Models Reasoning or Guessing A Mechanistic Analysis of Hierarchical Reasoning Models" Zirui Ren [--] [--] Ziming Liu [--] [--] from [--] Shanghai Qi Zhi Institute Shanghai China [--] Department of Physics Tsinghua University" [YouTube Link](https://youtube.com/watch?v=UETxlAf0BOA) 2026-01-19T14:15Z 85.2K followers, [----] engagements "NEW: Multiple Training DATASETS to fine-tune your SBERT model in [----] (SBERT 33) Python code on how to train multiple training datasets for your specific Sentence Transformer model. Combine Datasets of SNLI with MS MARCO and Reddit for training your SBERT model. Example on COLAB with PyTorch. #datascience #datastructure #dataset #datasets #pytorch #deeplearning #colab #sbert #semantic #search" [YouTube Link](https://youtube.com/watch?v=UZOzh6-js8I) 2022-07-14T13:30Z 82.6K followers, [----] engagements "AI for HealthCare Are you sure OpenAI AI HealthCare for FREE Linguistic HealthCare for $20 a month Does Sam's promises hold w/ OpenAI What about mental Health AI Human conversations - are they safe Medical advice by AI - Is this just a new business model for Ai companies or does it provide real value to human patients What are the risks involved in Medical AI What can go wrong with medical advice by AI Who is legally liable New research study by Duke and Stanford Univ. All rights w/ authors: "MedRedFlag: Investigating how LLMs Redirect Misconceptions in Real-World Health Communication"" [YouTube Link](https://youtube.com/watch?v=UewplNGNqPI) 2026-01-18T14:30Z 85.2K followers, [----] engagements "AI AGENTS Evolve: New TOPOLOGY for Multi-Agents NEW Autonomous agents a novel framework modeling agentic workflows as self-organized graphs with the Semantic-Topological Evolution (STEV) algorithm using textual gradients as discrete-domain surrogates for backpropagation. Multi-agent system multi-agent reasoning multi-agent topological optimization. All rights w/ authors: "HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution" Jinzhou Tang 1* Jusheng Zhang 1* Qinhan Lv 1* Sidi Liu [--] Jing Yang [--] Chengpei Tang1 Keze Wang1 from [--] Sun Yat-sen University" [YouTube Link](https://youtube.com/watch?v=VMsJ4me5Q3o) 2025-09-05T12:01Z 84.5K followers, [----] engagements "NEW Distributed Neural Graph Architecture for AI (Stanford) What of we get rid of the layer architecture in our transformers What if we operate a dynamic distributed graph network with different modules What if we combine transformer blocks with mamba blocks for an adaptive architecture of more complex tasks Can we improve reasoning New insights into AI. @meta @stanford All rights authors: Towards Distributed Neural Architectures Aditya Cowsik [--] Tianyu He [--] Andrey Gromov [--] from [--] FAIR at Meta [--] Stanford University [--] University of Maryland College Park #stanford #metaai #airesearch" [YouTube Link](https://youtube.com/watch?v=Vs5gSudsMS8) 2025-07-01T13:15Z 82.5K followers, [----] engagements "Free RAG (File Search) w/ App dev by Google: TEST Notebook LM - Good News: Google launching the File Search Tool a fully managed RAG system built directly into the Gemini API that abstracts away the retrieval pipeline so you can focus on building. File Search provides a simple integrated and scalable way to ground Gemini with your data delivering responses that are more accurate relevant and verifiable. To make File Search simple and affordable for all developers were making storage and embedding generation at query time free of charge. Powered by our latest state-of-the-art Gemini Embedding" [YouTube Link](https://youtube.com/watch?v=W6VJn7XU090) 2025-11-09T13:45Z 84.6K followers, 10.9K engagements "SRL: NEW AI Training (by Google) All rights w/ authors: "Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning" Yihe Deng2* I-Hung Hsu1* Jun Yan1 Zifeng Wang1 Rujun Han1 Gufeng Zhang3 Yanfei Chen1 Wei Wang2 Tomas Pfister1 and Chen-Yu Lee1 from [--] Google Cloud AI Research [--] UCLA [--] Google Cloud arXiv:2510.25992" [YouTube Link](https://youtube.com/watch?v=W_aqotP134s) 2025-11-03T13:45Z 82.9K followers, [----] engagements "Multi AI Agent System - Pure Social Manipulation Social Manipulation by Design: Multi AI Agent Systems. AI could theoretically completely re-shape human behavior. New research on Human-AI interaction by University of Singapore. My new video investigates the potential of multi-agent systems (MAS) to act as cohesive social groups (as an AI) capable of exerting influence on human opinions. Inspired by social psychology and the CASA (Computers as Social Actors) framework the study tests whether groups of AI agents as opposed to single agents can replicate the social pressure mechanisms found in" [YouTube Link](https://youtube.com/watch?v=Wj-sZZ1CffE) 2024-11-08T15:30Z 83K followers, [----] engagements "MedAI: How 'Safe' AI Becomes Deadly (Harvard Stanford MIT) The Hidden Casualty of MedAI (new study). MedAI Lobotomy: Why "Curing" Hallucination is Fatal including H-Neurons. Mechanistic interpretability recently promised us a surgical cure for AI hallucinations. By identifying H-Neurons - the sparse 0.1% of parameters that form a specific "Lens-Shutter-Nozzle" circuit for confabulation - we believed we could simply switch off the model's ability to lie. Theoretically identifying and suppressing these specific circuits should create the perfect truthful model. It sounds like the ultimate" [YouTube Link](https://youtube.com/watch?v=X0zUnjL5YTg) 2025-12-06T14:21Z 82.4K followers, [--] engagements "Invest in future OpenAI shares in [----] The ChatGPT company Analyzing the potential () investment opportunities in future stock market I have a look at OpenAI and its hyping free research demo ChatGPT. Will there be an IPO in [----] You have to decide yourself. My links (sources): https://www.forbes.com/sites/qai/2022/12/21/is-there-a-chatgpt-stock-can-you-invest-in-chatgpt-and-other-types-of-artificial-intelligence/ https://www.reuters.com/article/us-microsoft-openai/microsoft-to-invest-1-billion-in-openai-idUSKCN1UH1H9 https://www.reuters.com/article/idUSL3N1405JW20151211 #investment #shares" [YouTube Link](https://youtube.com/watch?v=XQCxzfjX6Eg) 2022-12-24T06:15Z 81.7K followers, [----] engagements "Multi AI AGENTS w/ LLMSelector New AI Tech Framework - Multi Agent" [YouTube Link](https://youtube.com/watch?v=XR452GfXtDY) 2025-02-24T21:15Z 83K followers, [----] engagements "New TECH: Vision Transformer [----] on Image Classification AI Understand state-of-the-art tech in Vision Technology eg medical image classification beginning of [----]. We learn the current tech of Vision Transformer vs CNN in a medical real-world application: "Vision-Transformer-Based Transfer Learning for Mammogram Classification". An in-depth analysis of Convolutional Neural Networks vs Vision Transformers for medical image classification to improve the early diagnosis of breast cancer in support of oncologists. Research should have a positive impact on this world. Scientific publication (all" [YouTube Link](https://youtube.com/watch?v=XRwdC2UkewE) 2023-02-11T13:15Z 82.7K followers, [----] engagements "Feature Vectors: The Key to Unlocking the Power of BERT and SBERT Transformer Models After converting text to high-dimensional vectors (and tensors) we use them as information encoded input to our NLP models based on the transformer architecture (like BERT or Sentence Transformers SBERT). We can apply mathematics to our semantic encoded vectors and compute different weight tensors of our NN systems. My videos as mentioned in the video: Beginner's GUIDE to TRANSFORMERS https://youtu.be/vBVJhojtooM42 How to explain Q K and V of SelfAttention in Transformers https://youtu.be/PFczJ6NR5rY SBERT" [YouTube Link](https://youtube.com/watch?v=XZhVUI5suf0) 2023-01-03T13:00Z 82.6K followers, [---] engagements "A New Solution for AI Agents (Stanford MIT) Since this is a YouTube channel membership only video please find all references to all the ArXiv pre-prints (title authors teams links) I referenced in this video in a dedicated more detailed post (here on my YouTube channel under "Posts") published at the same time this video goes live - for all who joined my YouTube channel. #machinelearning #airesearch #aiexplained" [YouTube Link](https://youtube.com/watch?v=XivQWNEwfw4) 2025-10-30T14:01Z 80.6K followers, [--] engagements "BEST NON-Thinking LLM: New Qwen3-MAX Preview Alibaba's research team released a new AI Model: Qwen3 - MAX Preview. This new model is NOT open-weight and the prices start at (Please check for your country): 032K tokens: $0.861 per million input tokens $3.441 per million output tokens I performed my standard causal reasoning test suite and it has an impressive performance for a non-thinking model. In complex logic tests it comes close to the best thinking LLMs /VLMs. Plus it provides a transparent reasoning output although it is a non-thinking model. As you can watch in my video the" [YouTube Link](https://youtube.com/watch?v=Y8hI6FyCinw) 2025-09-08T14:00Z 80.3K followers, [----] engagements "Vision AI Learns w/o Language: Stanford Breakthrough Probabilistic Structure Integration (PSI) is a novel framework proposed by the Stanford NeuroAI Lab for constructing self-improving VISUALS world models from raw non-linguistic data such as internet video clips totaling [---] trillion tokens. The system addresses key limitations in existing world models including coarse controllability and inflexible query interfaces by learning a probabilistic graphical model (PGM) approximated by a neural predictor denoted . This predictor models conditional distributions over local spatiotemporal variables" [YouTube Link](https://youtube.com/watch?v=YEHxRnkSBLQ) 2025-09-16T14:01Z 82.6K followers, [----] engagements "New Tutorial on LLM Quantization w/ QLoRA GPTQ and Llamacpp LLama [--] LLM Quantization: GPTQ - AutoGPTQ llama.cpp - ggml.c - GGUL - C++ Compare to HF transformers in 4-bit quantization. Download Web UI wrappers for your heavily quantized LLM to your local machine (PC Linux Apple). LLM on Apple Hardware w/ M1 M2 or M3 chip. Run inference of your LLMs on your local PC with heavy quantization applied. Plus: [--] Web UI for GTPQ llama.cpp or AutoGPTQ exLLama or GGUF.c koboldcpp oobabooga text-generation-webui ctransformers https://lmstudio.ai/ https://github.com/marella/ctransformers" [YouTube Link](https://youtube.com/watch?v=YEVyupJxt1Q) 2023-09-09T12:00Z 82.1K followers, 17.4K engagements "AI Models Are Falling Apart CLAUDE [---] & KIMI K2 All rights w/ authors: Reasoning Models Will Blatantly Lie About Their Reasoning William Walden from Johns Hopkins University #aireasoning #airesearch #aitesting #anthropic artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI" [YouTube Link](https://youtube.com/watch?v=Z1dOSop6KbM) 2026-01-14T13:15Z 85.2K followers, [----] engagements "NEW: FREE "Deep Research" by Perplexity - LIVE TEST Perplexity's FREE Deep Research Engine: Powerful AI for In-Depth Analysis. A good Value for a first product offering: FREE. First ever: Perplexity.ai offers [--] free runs per day on their NEW "Deep Research" AI Engine. It runs about 3-4 minutes and we can watch it thinking and reasoning on the extracted content from internet sources. The AI competitors are: OpenAI Deep Research (runs [--] minutes) is limited to the $200/month abo and Google also offers "Deep Research" for advanced payments. First look and live recording of my first [--] DEEP" [YouTube Link](https://youtube.com/watch?v=Z9IpO3TTskU) 2025-02-15T11:00Z 58K followers, [----] engagements "Are US News biased A multi-million $ AI finally answers. #Shorts Are US News media corporations biased A multi-million US$ artificial intelligence (AI) finally answers. When will 'now' arrive to your multi-million US$ trained AI /LLM on supercomputers Can your Large language Model (LLM) understand concept of time We compare two famous (and recently updated) Huggingface models: EleutherAI/gpt-j-6B and gpt-neo2.7B in a beautiful pure Python Gradio web application for text generation. Lessons learned from this video: The answer you might get from a heavily trained AI system (which could cost US$" [YouTube Link](https://youtube.com/watch?v=_V5l_yn-pEo) 2022-02-16T07:00Z 82.6K followers, [--] engagements "Mathematics w/ Donut AI and Nougat AI - Swin Transformer Mathematical formulas in PDF or images are lost to AI summarization. No AI LLM or ViT can correctly interpret from a PDF any mathematical formulae. Visual Document Understanding (VDU). Therefore I recommend to upload the LaTeX file of an arxiv preprint to GPT-4 Code Interpreter for a detailed mathematical understand of complex relations in Physics biology chemistry medicine architecture finance economy . Swin ViT (Vision Transformers) are the solution for mathematical formulae recognition first implemented in Donut AI then with a" [YouTube Link](https://youtube.com/watch?v=_ib8IvZijm0) 2023-09-24T12:00Z 82.6K followers, [----] engagements "xAI: Grok [--] DISAPPOINTS - Live Test Grok [--] TEST: Grok [--] has been released just some hours ago. I run my extended causal reasoning test on Grok [--] (me being located in Europe) on the LMarena.ai platform. The identical logic test has been performed on SONNET [--] OpenAI o3 and Gemini [---] PRO. Video available https://youtu.be/eo2QwyAItxI #grok4 #grok #airesearch #test" [YouTube Link](https://youtube.com/watch?v=aobihG5ig28) 2025-07-10T15:00Z 81.4K followers, [----] engagements "LLMs Ignoring New Context (Tsinghua Stanford) All rights w/ authors: SIN-Bench: Tracing Native Evidence Chains in Long-Context Multimodal Scientific Interleaved Literature Yiming Ren12* Junjie Wang13* Yuxin Meng1* Yihang Shi1* Zhiqiang Lin1 Ruihang Chu1 Yiran Xu1 Ziming Li4 Yunfei Zhao35 Zihan Wang36 Yu Qiao2 Ruiming Tang4 Minghao Liu3 Yujiu Yang1 from [--] Tsinghua University [--] Shanghai AI Laboratory [--] 2077AI [--] KuaiShou Inc. [--] Stanford University [--] Harvard University https://github.com/IIGROUP/sin-bench #aireasoning #reasoningskills #machinelearning #aiexplained artificial intelligence AI" [YouTube Link](https://youtube.com/watch?v=az5WB-nGDk4) 2026-01-17T14:15Z 85.2K followers, [----] engagements "OpenAI GPT-oss-120B: LIVE TEST My causal reasoning test performed on the newly release GPT-OSS-120B from OpenAI. The open-weight reasoning models 120B and 20B. Definitely not GPT-5 comparable [----] tokens per second via Cerebras OPENAI failed to provide an amazing new AI model for the global community the "for-profit" orientation - with real excellent LLMs (which are not open weight) where you have to pay for the performance - seems to dominate OpenAI's strategic positioning. All rights w/ authors: "Introducing gpt-oss" gpt-oss-120b and gpt-oss-20b https://openai.com/index/introducing-gpt-oss/" [YouTube Link](https://youtube.com/watch?v=bJSAcfQgxAg) 2025-08-05T22:00Z 83.1K followers, [----] engagements "NEW L1 LLM w/ GRPO to LCPO for Scaling RL (CMU) We explore the new Length Controlled Policy Optimization (LCPO) a simple reinforcement learning method that optimizes for accuracy and adherence to user-specified length constraints. Carnegie Mellon Univ (CMU) authors applied new LCPO to train L1 a reasoning language model that produces outputs satisfying a length constraint given in its prompt. LCPO is a further development of GRPO the group relative policy optimization for scaling RL by DeepSeekMath /R1. All rights w/ authors: "L1: Controlling How Long A Reasoning Model Thinks With" [YouTube Link](https://youtube.com/watch?v=bLBScrWDm6M) 2025-03-08T15:00Z 84.5K followers, [----] engagements "GPT-5 MCP Disaster: Under 50% & CURSOR useless Salesforce NEW MCP-Universe benchmark by Salesforce. #agi #superintelligence GPT-5's MCP Disaster: Under 50% - CURSOR useless All rights w/ authors: "MCP-Universe: Benchmarking Large Language Models with Real-World Model Context Protocol Servers" Ziyang Luo Zhiqi Shen Wenzhuo Yang Zirui Zhao Prathyusha Jwalapuram Amrita Saha Doyen Sahoo Silvio Savarese Caiming Xiong Junnan Li from Salesforce AI Research arXiv:2508.14704v1 Google Map MCP https://github.com/modelcontextprotocol/servers-archived/tree/main/src/google-maps Github MCP" [YouTube Link](https://youtube.com/watch?v=bPbkT7MtwGE) 2025-08-21T12:01Z 84K followers, [----] engagements "Princeton: NEW Self Correcting AI Transformer (Deep Delta) Based on two new AI research pre-prints by Princeton University (see below) I design a new transformer architecture by combining the hardware and software insights from these pre-prints and I call this new AI Transformer: "Self-Correcting Delta Transformer". A new transformer that can forget incorrect reasoning traces. The Pathology of Additive Reasoning: Recent empirical rigor dismantles the myth of intrinsic self-correction in current Large Language Models. Analysis of over [--] million reasoning traces reveals that spontaneous "Aha"" [YouTube Link](https://youtube.com/watch?v=bZbdKo_EAoY) 2026-01-05T14:15Z 85.2K followers, [----] engagements "Why a TENSOR in ML Neural Networks What about PANDAS dataframes Tensors are n-dim arrays that define form and shape of input data to our Neural Network models. Tensors rank shape and axis are important for layer operations: like functions applied to tensors. A layer operation takes tensors as input performs operations (dense layer pooling layer convolutional layer) and outputs a tensor. To understand rank and dim of tensors is important for designing your multiple layers of your neural network model. Tensors allow for automated differentiation which is great for gradient descent of your ML" [YouTube Link](https://youtube.com/watch?v=bfwvI3iz-1w) 2022-05-30T11:00Z 83K followers, [---] engagements "S* for AI CODE Generation: Plus 100% S* the first hybrid test-time scaling framework that substantially improves the coverage and selection accuracy of LLM generated code. Also for deep reasoning models. All rights w/ authors: S*: Test Time Scaling for Code Generation Dacheng Li Shiyi Cao Chengkun Cao Xiuyu Li Shangyin Tan Kurt Keutzer Jiarong Xing Joseph E. Gonzalez Ion Stoica University of California @UCBerkeley Work done w/ support from https://lambdalabs.com #airesearch #codegeneration #aicoding #berkeley" [YouTube Link](https://youtube.com/watch?v=br2xnptRbjI) 2025-02-23T13:00Z 59.3K followers, [----] engagements "NEW Agentic Web: NO Apps NO Ad Revenues The current ca $600 billion digital advertising industry is almost entirely predicated on monetizing human attention directed at GUIs (web pages social feeds app screens). As the "Agentic Web" vision explains when the primary consumer of services becomes an autonomous agent executing a goal-oriented policy this economic model collapses. An agent is not susceptible to brand advertising or display ads. Its "attention" is purely utilitarian driven by metrics like cost latency and success probability. This will have massive consequences. And . by the way" [YouTube Link](https://youtube.com/watch?v=cCv5mdduymA) 2025-08-03T12:00Z 82.6K followers, [----] engagements "What is MLOps in ML Engineering #shorts What is MLOps Simple: Machine Learning Operations. MLOps is a core function of Machine Learning engineering focused on streamlining the process of taking machine learning models to production and then maintaining and monitoring them. Productionizing machine learning is difficult. The machine learning lifecycle consists of many complex components such as data ingest data prep model training model tuning model deployment model monitoring explainability and much more. It also requires collaboration and hand-offs across teams from Data Engineering to Data" [YouTube Link](https://youtube.com/watch?v=d0toED4JoBI) 2022-10-31T15:00Z 82.6K followers, [---] engagements "Neuro-Symbolic AI for Visual Reasoning: Agent0-VL All rights w/ authors: Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens Yiming Qin [--] Bomin Wei [--] Jiaxin Ge [--] Konstantinos Kallidromitis [--] Stephanie Fu [--] Trevor Darrell [--] Xudong Wang [--] from [--] UC Berkeley [--] UCLA [--] Panasonic AI Research Qwen3-VL Technical Report Qwen Team https://chat.qwen.ai https://huggingface.co/Qwen https://modelscope.cn/organization/qwen https://github.com/QwenLM/Qwen3-VL Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning Jiaqi Liu1 Kaiwen Xiong1" [YouTube Link](https://youtube.com/watch?v=dKFHa8FwEqI) 2025-11-30T13:30Z 82.3K followers, [----] engagements "DATASET to fine-tune SBERT (w/ CROSS-ENCODER) for a better Domain Performance [----] (SBERT 32) Use a training dataset to fine-tune your SBERT model. Python code on how to train famous SNLI dataset for a CROSS-ENCODER /Sentence Transformer model. Example on COLAB with PyTorch. #datascience #datastructure #dataset #datasets #pytorch #deeplearning #colab #sbert #semantic #search #machinelearning #ai" [YouTube Link](https://youtube.com/watch?v=dhw2-oBbm78) 2022-07-11T14:00Z 82.6K followers, [----] engagements "Self-Attention Heads of last Layer of Vision Transformer (ViT) visualized (pre-trained with DINO) In a Colab Notebook we code a visualization of the last layer of the Vision Transformer Encoder stack and analyze the visual output of each of the [--] Attention Heads given a specific image. Now we understand how a only pre-trained ViT (although with the DINO method) can not always succeed in an image classification (downstream) task. The fine-tuning of the ViT is simply missing - but essential for a better performance. Based on the COLAB NB by Niels Rogge HuggingFace (all rights with him):" [YouTube Link](https://youtube.com/watch?v=dtlg3PUUn_g) 2023-02-16T13:00Z 82.7K followers, [----] engagements "AI data of [----] (Markets Experts Profits) On the very first day of [----] what are the headcount market size and financial data for the AI market segments in [----] Locally and globally. The coherent insights from "deep research" tasks from the TOP [--] AI provider (OpenAI to Google) deliver. Will Code AI systems be profitable for global corporations or is Code AI just a niche segment for a specialized player (Claude Code) What are the main customers for AI in [----] what are their main topics For what should OpenAI or Google Or META or X develop the next generation of Ai models What about new AI" [YouTube Link](https://youtube.com/watch?v=dxu3BJSjEdk) 2026-01-01T13:45Z 85.2K followers, [----] engagements "Weisfeiler-Lehman WL Test for Graph Isomorphism explained visually & Message Passing NNs [----] The graph isomorphism problem and the Weisfeiler-Lehman heuristic for graph isomorphism testing method explained visually on two examples. A classical question in graph theory: the graph isomorphism problem aiming to determine whether two graphs are topologically equivalent. The Weisfeiler-Lehman (WL) test for k=1. Link for deep dive: https://towardsdatascience.com/expressive-power-of-graph-neural-networks-and-the-weisefeiler-lehman-test-b883db3c7c49 by Prof. Bronstein A key difference between" [YouTube Link](https://youtube.com/watch?v=e4e6h9arD78) 2022-04-17T14:00Z 84K followers, 10.4K engagements Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing
@code4ai Discover AI
Discover AI posts on YouTube about ai, llm, artificial intelligence, language the most. They currently have [------] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours.
Engagements: [------] #
- [--] Week [------] +8.50%
- [--] Month [-------] -21%
- [--] Months [-------] -32%
- [--] Year [---------] +112%
Mentions: [--] #
- [--] Month [--] -47%
- [--] Months [---] -10%
- [--] Year [---] +112%
Followers: [------] #
- [--] Week [------] +0.35%
- [--] Month [------] +1.30%
- [--] Months [------] +11%
- [--] Year [------] +48%
CreatorRank: [-------] #
Social Influence
Social category influence technology brands stocks countries travel destinations social networks finance currencies cryptocurrencies automotive brands musicians
Social topic influence ai, llm #118, artificial intelligence #331, language #950, artificial, $googl, open ai, university of, #ai, youtube
Top accounts mentioned or mentioned by @mit @stanford @openai @ucberkeley @googledeepmind @tsinghuauniversityofficial @princeton @harvard @nvidia @google @cmu @huggingface @anthropicai @salesforce @penn @oxforduniversity @nvidiaal @yale @ucla @pekinguniversity1898
Top assets mentioned Alphabet Inc Class A (GOOGL) Microsoft Corp. (MSFT) Salesforce Inc (CRM) CyberConnect (CYBER)
Top Social Posts
Top posts by engagements in the last [--] hours
"The $10T AI Economy: New Smart Protocol Emerges The Smart Contract between AI Agents New AI Protocol emerges. The future Economy of Humans and AI agents on a global competitive marketplace with decentralized auctions. Human Intelligence and machine intelligence are priced and offered globally. Is this our future Is this why we invented AI All rights w/ authors: "Intelligent AI Delegation" Nenad Tomaev1 Matija Franklin1 and Simon Osindero1 from Google DeepMind #economy #aieconomy #aibusiness #aiproductivity #aiperformance #nextgenai artificial intelligence AI models LLM VLM VLA Multi-modal"
YouTube Link 2026-02-15T14:15Z 85.2K followers, [----] engagements
"NEW GLM-5 vs MiniMax-2.5: NEW = BETTER Artificial Intelligence: Two new AI models (GLM-5 vs MiniMax-2.5) are tested - side by side - on a non-public causal reasoning test to evaluate their performance. Live recording of real-world performance of the latest agent optimized LLMs. 00:00 GLM-5 and MiniMax-2.5 02:13 Start Live TEST 12:35 GLM-5 and MiniMax-2.5 CRASH 14:07 First Solution by GLM-5 14:48 Successful GLM-5 Evaluation Run 16:33 GLM-5 Evaluation by MiniMax-2.5 18:03 MiniMax-2.5 FAILS Validation #aiexplained #aitech #ainews artificial intelligence AI models LLM VLM VLA Multi-modal model"
YouTube Link 2026-02-13T14:30Z 85.2K followers, [----] engagements
"The "OPENNESS" of corporate LLM models: A corporate board meeting Ever wondered what happens behind the colossal closed doors of tech behemoths like ClosedAI Microsocks and METAH Join me for a humorous parody of a hypothetical management board meeting as they grapple with criticisms of their open-source models Get ready for witty repartees exaggerations and the most comically blown-out-of-proportion reactions - all with a pinch of reality. Remember it's all in good fun as we salute these trailblazers for their incredible contributions to the AI world. Subscribe and stay tuned for an episode"
YouTube Link 2023-07-30T12:00Z 85.2K followers, [----] engagements
"AI Belief Functions: Deciding Under Absolute Uncertainty Engines of Intelligence: The Definition and Necessity of AI Agents. Basic mathematical explanations for any CEO of any multinational strategy and management consulting firm: What AI agents are and why we need them. And why AI (re-)acts in absolute uncertainty with self-updating belief functions. Plus: A simple explanation for CEOs (McKinsey EY .) what are AI agents and why they operate in absolute uncertainty with a simple mathematical probability distribution for the most difficult client jobs. .maybe not such a good idea #aiexplained"
YouTube Link 2026-02-14T14:00Z 85.2K followers, [----] engagements
"Forget LLM: MIT's New RLM (Phase Shift in AI) Weve been misled by the promise of "infinite" context windows: new AI research proves that "Context Rot" is destroying reasoning capabilities as inputs scale. But a groundbreaking paper from MIT introduces a radical solution: Recursive Language Models (RLMs). Instead of blindly force-feeding data into a single Transformer RLMs act as a Neurosymbolic Operating System writing Python code to mechanically split massive datasets and recursively "spawn" fresh model instances to process them. The result is a staggering leap in performance: on quadratic"
YouTube Link 2026-01-04T14:01Z 85.2K followers, 30K engagements
"Accelerate pandas df: DASK [----] = superfast Python DASK scales numpy arrays and pandas dataframes efficiently. Utilizes all CPU cores / threads. Video shows you that four lines of code set up a local DASK cluster automatically on my Win10 PC to supercharge python - even on a single CPU. Optimize your data input pipeline to your transformer models (AI) with a local cluster configuration utilizing your system resources. Speed tests on numpy arrays w/ DASK. Speed tests on pandas dataframes w/ DASK. #code_in_real_time #real_time_coding #DASK #parallelize_python #cluster #JupyterLab #python"
YouTube Link 2021-07-27T11:15Z 84.7K followers, [---] engagements
"Smarter AI Gradients: How Agents Learn to Think Exploration is essential in reinforcement learning (RL) as an AI agent relies on trial and error to learn an optimal policy. However when rewards are sparse naive exploration strategies like noise injection are often insufficient. Intrinsic rewards can also provide principled guidance for exploration by for example combining them with extrinsic rewards to optimize a policy or using them to train sub-policies for hierarchical learning. However the former approach suffers from unstable credit assignment while the latter exhibits sample"
YouTube Link 2026-01-31T14:30Z 85.2K followers, [----] engagements
"Grand Unified Theory of AI (Explained w/ Google ADK) Grand Unified Theory of AI Multi-Agent Systems (Explained w/ Google Context ADK). We present a rigorous analysis of Google Clouds "Mathematical Framing for Different Agent Strategies" which proposes a unified probabilistic formulation to quantify agent behavior beyond empirical benchmarks. To build production-grade agents that are reliable efficient and debuggable the industry is exploring a new discipline: Context engineering treating context as a first-class system with its own architecture lifecycle and constraints. Based on the"
YouTube Link 2025-12-08T14:00Z 84.7K followers, [----] engagements
"Activate GROKKING NOW - Performance Phase of LLMs (II) Grokking or the sudden generalization by AI models to new knowledge - that occurs after prolonged overfitting in LLMs is a surprising phenomenon that has challenged our understanding of deep learning and AI in general. While a lot of progress has been made in understanding grokking finally we get some answers -we have been waiting for [--] months to be discovered. GROKKING - Finally understood AUDIO: With the automatic audio dubbing from YouTube /Google you hear a synthetic voice in your regional language. To hear my original voice in"
YouTube Link 2025-01-15T15:15Z 84.7K followers, [----] engagements
"LLM Quantization (Ollama LM Studio): Any Performance Drop TEST A NEW benchmark and guide which quantization models to use locally on your PC or laptop. Either in Ollama or in LM Studio whenever I want to download a LLM or VLM I have to use a quantized Ai model because of the limited VRAM on my NVIDIA GPU my local AI infra. But what quantized model should I choose What quantization really hurt the overall model Performance What quant is recommended The Ollama version of the latest version of deepseek-r1:671b-0528-q4_K_M (404GB) is available here:"
YouTube Link 2025-08-26T14:00Z 84.7K followers, [----] engagements
"New Discovery: LLMs have a Performance Phase Grokking is a new phase in the performance of LLMs. Starting with arithmetic operations we analyze the patterns in the embedded space of Transformers. Grokking refers to a phenomenon where after extensive training beyond typical saturation points transformers can generalize effectively to unseen data achieving high performance long after initial overfitting occurs. This discovery challenges conventional wisdom about early stopping to prevent overfitting revealing that extended training can lead to superior generalization. The video highlights"
YouTube Link 2024-06-04T12:00Z 84.7K followers, 16.8K engagements
"POPE RL Curriculum Learning (CMU) RL doesn't teach the AI model new facts; POPE RL tries to steer the model's internal attention heads to attend to the correct latent subspaces (like mathematical reasoning) rather than the incorrect ones (casual chat or confusion) which cause the "Cold Start" problem. Further insights into the "Valley of Death" for RL in AI (zero gradients zero rewards). All rights w/ authors: POPE: Learning to Reason on Hard Problems via Privileged On-Policy Exploration Yuxiao Qu1 Amrith Setlur1 Virginia Smith1 Ruslan Salakhutdinov1 Aviral Kumar1 from [--] Carnegie Mellon"
YouTube Link 2026-01-30T14:15Z 85.2K followers, [----] engagements
"Stanford: Do Not use [--] AI Agents: They will Fail (CooperBench) In depth vido why Why Two AI Coding Agents Are Worse Than One. Stanford Univ directly attacks the industry hype around Multi-Agent Systems (MAS) and promises a data-backed explanation for why adding compute (AI coding agents) actually degrades performance. All rights w/ authors: "CooperBench: Why Coding Agents Cannot be Your Teammates Yet" Arpandeep Khatua1 Hao Zhu1 Peter Tran2 Arya Prabhudesai2 Frederic Sadrieh2 Johann K. Lieberwirth2 Xinkai Yu1 Yicheng Fu1 Michael J. Ryan1 Jiaxin Pei1 Diyi Yang1 from [--] Stanford University [--] SAP"
YouTube Link 2026-01-26T14:15Z 85.2K followers, [----] engagements
"Jupyter AI: Generative AI in your Notebook Jupyter AI explained in both modes: AI magic commands and the Chat UI for generative AI models form Anthropic Cohere OpenAI . #ai #codegeneration #generativeai"
YouTube Link 2023-08-15T12:00Z 84.7K followers, [----] engagements
"NEW Knowledge Graph based RAG: SimGRAG (no training) Excellent new Knowledge Graph based RAG system called SimGraphRAG or simply SimGRAG. Overview of our four classical KG-based RAG systems and the new SimGRAG which outperform them. Short technical deep dive into the new methods and algorithms plus code via GitHub repo. All rights w/ authors: SimGRAG: Leveraging Similar Subgraphs for Knowledge Graphs Driven Retrieval-Augmented Generation Yuzheng Cai Zhenyue Guo Yiwen Pei Wanrui Bian Weiguo Zheng from Fudan University #airesearch #knowledgegraph #science #aiagents #graph"
YouTube Link 2024-12-25T15:00Z 84.8K followers, 12.6K engagements
"Scaling AI Reasoning: MCTS in ICL for Small LM Meta-reasoning Unleashed: A New ICL Paradigm Beyond Simple Examples: Next-Level ICL Reasoning ICL Reinvented: Harnessing MCTS Insights From Steps to Strategies: Thought Cards in ICL High-Level Paths: The New Face of In-Context Learning Scaling Reasoning: MCTS and Beyond in ICL VOC computing ICL: Optimizing Reasoning Paths Abstract Templates: The Future of ICL Strategies ICL Meets MCTS: Deep Reasoning Upgrades Evolving ICL: Distilling Optimal Thought Patterns #airesearch #chatgpt #programming #coding #reasoning #logic #cognitivefunction"
YouTube Link 2024-12-08T15:00Z 84.8K followers, [----] engagements
"LLM + Knowledge Graph + GNN = TRUTH by AI GraphCHECK: Improving Factuality in LLM Outputs w/ Graph Neural Networks for Knowledge-Graph Enhanced Verification Can AI Find Truth The Power of LLMs Knowledge Graphs and GNNs CODE implementation of GraphCHECK available: https://anonymous.4open.science/r/GraphCheck-1D43/README.md PYTHON implementation for Graph Attention Network: https://anonymous.4open.science/r/GraphCheck-1D43/model/gnn.py All rights w/ authors: "GraphCheck: Breaking Long-Term Text Barriers with Extracted Knowledge Graph-Powered Fact-Checking" Yingjian Chen1 Haoran Liu2 Yinhong"
YouTube Link 2025-02-28T15:15Z 84.8K followers, [----] engagements
"NEW Multi-Agent Protocol (REP by MIT): 100+ Agents Caught a cold. Sorry for my wet pronunciation . All rights W/ authors: Which LLM MultiAgent Protocol to Choose Hongyi Du1 Jiaqi Su2 Jisen Li1 Lijie Ding3 Yingxuan Yang2 Peixuan Han1 Xiangru Tang4 Kunlun Zhu1 Jiaxuan You1 from [--] University of Illinois UrbanaChampaign [--] Shanghai Jiao Tong University [--] Oak Ridge National Laboratory [--] Yale University arXiv:2510.16572 @yale @princeton Ripple Effect Protocol: Coordinating Agent Populations Ayush Chopra12 Aman Sharma2 Feroz Ahmad2 Luca Muscariello3 Vijoy Pandey3 Ramesh Raskar12 from [--] Massachusetts"
YouTube Link 2025-10-24T12:15Z 84.7K followers, [----] engagements
"Salesforce: New AI Agents That Doubt Themselves (AUQ) Salesforce Research has just operationalised Kahnemans "System [--] vs. System 2" framework directly into LLM architectures to solve the "Spiral of Hallucination" in long-horizon tasks for AI agents. In this breakdown we dissect Agentic Uncertainty Quantification (AUQ): a training-free framework that treats verbalised confidence not merely as a metric but as a dynamic control signal. You will learn how they implement a non-differentiable switching function that bifurcates inference into a fast memory-augmented path and a slow "inverse"
YouTube Link 2026-01-24T14:15Z 85.2K followers, [----] engagements
"Death of the Token in AI: Multi-Parallel AI Reality NVIDIAs Silent Robots & Pre-GPT-6 Schrdinger Token Frankenstein Vector and Quantum collapse for new Multi-Parallel AI Realities & NVIDIA Silent Robots w/ Pre-GPT-6. Multi-Parallel realities for AI reasoning and Quantum collapse in AI for next generation of AI models and their advanced architectures. Combine these three brand new AI Arxiv pre-prints from the latest Ai research and you will see a pattern emerging for new AI architectures like GPT-6. From NVIDIA to Microsoft. All rights W/ authors: "Reasoning Beyond Chain-of-Thought: A Latent"
YouTube Link 2026-01-16T14:15Z 85.2K followers, [----] engagements
"MiMo V2 Flash: Excellent Performance (vs Kimi K2 Thinking) I test the causal reasoning performance for a simple [--] step logic task with the MiMo V2 Flash MoE 309B-15A model from Xiaomi (open-source). And compare the performance to a much bigger Kimi K2 Thinking Turbo model MoE 1T-32A. My Youtube playlist for this causal reasoning test is available here https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT 00:00 MiMo V2 vs KIMI K2 02:42 Start Live Test 08:56 MiMo First Result 10:15 1st Validation Run 12:37 2nd Validation Run 13:53 Kimi K2 First Answer 14:56 Self Check Kimi K2"
YouTube Link 2026-02-11T14:30Z 85.2K followers, [----] engagements
"The Future of Conversational AI Google's PaLM w/ RLHF LLM ChatGPT Competitor After explaining BERT vs GPT and Google's T5 JAX (in my last videos) we now examine new PaLM: Pathways Language Model (if combined w/ RLHF -Reinforcement Learning with Human feedback). T5X = Google's T5 on JAX and FLAX. Plus Code implementation for PaLM w/ RLHF. PS: Next video on another LLM could be on Sparrow . smile. my resources (all rights are with the authors): Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer https://arxiv.org/pdf/1910.10683.pdf SentencePiece: A simple and"
YouTube Link 2023-01-23T13:15Z 85.2K followers, [----] engagements
"8B outperforms GPT-120B on Multi Agents All rights w/ authors: DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning via Semantic Matching Yuxing Lu * [--] [--] Yucheng Hu * [--] Xukai Zhao [--] Jiuxin Cao [--] from [--] Peking University Beijing China [--] Georgia Institute of Technology Atlanta United States [--] Southeast University [--] Tsinghua University. #aiexplained #airesearch #artificialintelligence #aiagents artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial"
YouTube Link 2026-02-10T14:01Z 85.2K followers, [----] engagements
"15B Active MoE BEATS OPUS [---] in Reasoning Inside AI: to be specific inside a real powerful reasoning engine MoE and all the new methods and optimizations algorithms we encounter in building an open-source Mixture-of-Expert AI model. Inside a modern MoE AI model (Technical Architecture) All rights w/ authors: MiMo-V2-Flash Technical Report LLM-Core Xiaomi https://arxiv.org/pdf/2601.02780 #aitransformation #aiexplained #scienceexplained #nextgenai #airesearch artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI"
YouTube Link 2026-02-12T14:01Z 85.2K followers, [----] engagements
"Claude OPUS [---] Thinking vs [---] Non-Thinking: Both FAIL Anthropic just released Claude OPUS [---]. And specified "Opus [---] extends the frontier of expert-level reasoning". So I test both models of Claude [---] (Thinking and Non-Thinking) on my standard causal reasoning test. You can see my complete YouTube Playlist for AI model testing on this causal reasoning test here https://www.youtube.com/watchv=g1L8uOQ7Ids&list=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT https://www.anthropic.com/news/claude-opus-4-6 00:00 OPUS [---] TEST 02:10 OPUS [---] Non-Thinking First Result 02:20 OPUS [---] Thinking 05:58 OPUS 4.6"
YouTube Link 2026-02-05T23:15Z 85.2K followers, [----] engagements
"Google Goes to EXTREMES: In-Context Symbolic AI "We demonstrate an approach for LLMs to critique their own answers with the goal of enhancing their performance that leads to significant improvements over established planning benchmarks. Despite the findings of earlier research that has cast doubt on the effectiveness of LLMs leveraging self critique methods we show significant performance gains on planning datasets in the Blocksworld domain through intrinsic self-critique without external source such as a verifier. " Quote by Google DeepMind (see below) All rights w/ authors: "Enhancing LLM"
YouTube Link 2026-01-06T14:00Z 85.2K followers, [----] engagements
"NeuroSymbolic Web World Model (Decouples Physics from AI) The idea is simple: Instead of training a huge AI model on Language syntax domain knowledge coding and Science patterns - why not separate the Physics engine (like a neurosymbolic Model) from the autoregressive AI Done in an elegant way: new insights by Princeton Univ. All rights w/ authors: Web World Models Jichen Feng13 Yifan Zhang1 Chenggong Zhang2 Yifu Lu1 Shilong Liu1 Mengdi Wang1 from [--] Princeton University [--] University of California Los Angeles [--] University of Pennsylvania arXiv:2512.23676"
YouTube Link 2025-12-31T14:00Z 85.2K followers, [----] engagements
"AutoGPT & BabyAGI: autonomous AI Agents for LLMs explained AutoGPT & BabyAGI explained Combining GPT-4 with the interface functionality of LangChain with on open data channel to external internet services (with $$$ API) opens up new combinations of interlinks of resources: meet AutoGPT and BabyAGI. There are significant risks for your financial resources when GPT-4 takes over the link priorities without human interaction. Literature and code (all rights w/ those authors): https://github.com/yoheinakajima/babyagi"
YouTube Link 2023-04-22T12:15Z 84.8K followers, [----] engagements
"MedAI: Vision Language Models & Fine-Tuning (KnowAda) Smaller VLM hallucinate. A new counter-measure: Knowledge-Adapted Fine-Tuning (KnowAda) is a novel approach to mitigate hallucinations in vision-language models (VLMs) when generating dense image captions. Reducing Hallucinations in Multimodal Models Through new Adaptive Training. Traditional fine-tuning methods often result in smaller-scale VLMs (up to 7B parameters) struggling to balance descriptiveness with factual accuracy especially in visually complex datasets. KnowAda addresses this by probing the VLMs knowledge using generated"
YouTube Link 2024-11-16T15:00Z 84.9K followers, [----] engagements
"ORPO: NEW DPO Alignment and SFT Method for LLM Instead of the classical SFT and DPO alignment for training our LLMs there is a new method available. A innovative "reference model-free" monolithic odds ratio reference optimization algorithm ORPO eliminating the necessity for an additional preference alignment phase. A New Preference-aligned SFT method. We explore this idea from a theoretical physics perspective and notice a similarity to the regularizations terms methodologies. We further explore the conceptional similarities from a Lagrange Multiplier to new correction terms in addition to"
YouTube Link 2024-03-24T13:00Z 84.7K followers, [----] engagements
"AI Dual Manifold Cognitive Architecture (Experts only) All rights w/ authors: "MirrorMind: Empowering OmniScientist with the Expert Perspectives and Collective Knowledge of Human Scientists" Qingbin Zeng [--] Bingbing Fan [--] Zhiyu Chen [--] Sijian Ren [--] Zhilun Zhou [--] Xuhua Zhang [--] Yuanyi Zhen [--] Fengli Xu [--] Yong Li [--] Tie-Yan Liu [--] from [--] Department of Electronic Engineering BNRist Tsinghua University [--] Zhongguancun Academy "PersonaAgent with GraphRAG: Community-Aware Knowledge Graphs for Personalized LLM" Siqi Liang 1* Yudi Zhang 2* Yue Guo [--] from [--] Purdue University [--] Iowa State University 3"
YouTube Link 2025-11-27T13:45Z 85K followers, 11.7K engagements
"NEW StreamingLLM by MIT & Meta: Code explained MIT and META introduce StreamingLLM an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence length without any fine-tuning. Streaming LLM. ARXIV preprint: https://arxiv.org/pdf/2309.17453v1.pdf GitHub repo: https://github.com/mit-han-lab/streaming-llm/blob/main/streaming_llm/pos_shift/modify_llama.py"
YouTube Link 2023-10-13T12:00Z 84.8K followers, [----] engagements
"Autonomous AI Agents: 14% MAX Performance Webarena provides a test ground to test AI Agents' performance for functional correctness of task completions. Ideal for development and performance testing of autonomous AI agents. All rights w/ authors: https://arxiv.org/pdf/2307.13854 Plus: User's Intent research by Microsoft on optimized tool use by autonomous agents. Plus the open source Toolkit from @CohereAI now available for your perfect RAG system with pre-build components and apps. https://github.com/cohere-ai/cohere-toolkit 00:00 Autonomous AI Agents 00:45 Webarena for dev of agents 03:30"
YouTube Link 2024-04-28T12:00Z 84.8K followers, [----] engagements
"Individualise your AI Companion: EASY The easiest way to individualize your AI - simple Demo. Please add additional personal guardrails to your individualized system prompt. Commercial AI systems can make mistakes. Severe mistakes. All rights w/ authors: Simulating Psychological Risks in Human-AI Interactions: Real-Case Informed Modeling of AI-Induced Addiction Anorexia Depression Homicide Psychosis and Suicide Chayapatr Archiwaranguprok MIT Media Lab Massachusetts Institute of Technology Cambridge Massachusetts USA Constanze Albrecht MIT Media Lab Massachusetts Institute of Technology"
YouTube Link 2025-11-18T14:00Z 84.8K followers, [----] engagements
"Google's AutoGrad + Tensorflow's XLA Linear Algebra Compiler = JAX Add Google's AutoGrad and Tensorflow's XLA linear algebra compiler and you get JAX: a python and numpy racehorse to differentiate for backprop and compile on multi TPU clouds. You love numpy and want vectorization and automatic parallelization for GPUs and TPUs Then you know JAX For the sole purpose of applying Graph Neural network models we need to cover JAX by Google/DeepMind before starting into Jraph for our main purpose: Apply GNN to complex problem solving in the omniverse. Or was it the Multiverse Any way here is JAX"
YouTube Link 2021-12-04T07:00Z 84.9K followers, [---] engagements
"How a 14B Model BEATS GPT-5.2 FUZZY Graph Reward All rights w/ authors: "Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning" Yuval Kansal Princeton University Niraj K. Jha Princeton University @princeton #ainews #RLVR #gpt52 #chatgpt5 #chatgpt Reinforcement Learning with verifiable return Medical AI Multi-step reasoning Reinforcement Learning with verifiable return Medical AI Multi-step reasoning"
YouTube Link 2026-01-27T14:30Z 85.2K followers, [----] engagements
"NEW Gemini [--] FLASH vs GPT [---] HIGH - A Bloodbath NEW Gemini [--] FLASH is [--] times cheaper ($) than OpenAI's GPT-5.2 HIGH for your identical tasks. So in a real-world test that looks similar to real science tasks I evaluate both AI models side-by-side. Note: This is not the known standard vanilla benchmarks this has to do with real world complexities - heavily oriented towards SCIENCE not Social Media. For a detailed comparison with all the other Ai models I have a dedicated YouTube Playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT Will Gemini [--] FLASH reach only a"
YouTube Link 2025-12-18T13:30Z 85K followers, [----] engagements
"Geometric GROKKING Unlocked & Explained Given the latest insights on new geometric representations of knowledge patterns in the memory of Mamba or Transformer architecture I present a new theory of mine - explaining Grokking from a different framing. No science just me thinking about an explanation of the Grokking Phase for improved AI models. Complete (ArXiv) literature has already been quoted in my corresponding YT videos. Idea happened during a breakfast session on a Sunday morning. #artificialintelligence #reasoning #aiexplained artificial intelligence AI models LLM VLM VLA Multi-modal"
YouTube Link 2025-11-02T13:45Z 85K followers, [----] engagements
"AI Kill Switch for Hallucinations (Anthropic) [--] new AI research papers (all from [--] Jan 2026) that focus on one topic: inside the activations (the latent space inside an AI transformer architecture). A new kill switch for AI hallucinations in reasoning traces and how to heal catastrophic forgetting caused by Supervised Fine-Tuning (SFT). Self-learning self-healing self-correcting and self-improving AI models. mathematical insights into loss functions and new operators. all rights w/ authors: "Emergent Introspective Awareness in Large Language Models" Jack Lindsey from Anthropic"
YouTube Link 2026-01-07T14:00Z 85.2K followers, [----] engagements
"DeepSeek built a New Topological Transformer (mHC) DeepSeek build a new topological transformer that is beautifully compatible with the new Transformer architecture from Google (see my video https://youtu.be/5gpc3d2rFlg ). In this video I explain the mathematics of the Manifold-Constrained Hyper-Connections in particular the Sinkhorn-Knopp algorithm to entropically project H_residual onto the Birkhoff polytope. All rights w/ authors: "mHC: Manifold-Constrained Hyper-Connections" Zhenda Xie* Yixuan Wei* Huanqi Cao* Chenggang Zhao Chengqi Deng Jiashi Li Damai Dai Huazuo Gao Jiang Chang Liang"
YouTube Link 2026-01-03T14:00Z 85.2K followers, 17.2K engagements
"CORE of AI is EXPLODING - [--] New Papers CORE of AI currently explodes: we'll discover a specific selection of [--] new ArXiv CS pre-prints as a subset from more than [---] new ArXiv papers as published on first days of September [----]. All new arxiv preprints are given with full details: authors title publication date and arxiv link in the video. #scienceexplained #aiexplained #discoverai"
YouTube Link 2025-09-06T14:00Z 84.8K followers, 15K engagements
"DeepSeek [---] vs MiniMax M2 (1 Sentence TEST) Superior Intelligence: ChatGPT [---] vs MiniMax M2 on my [--] sentence test. Watch my new logic test (1 sentence no science) to challenge the intelligence of AI models. New test between the latest Flagship from OpenAI: GPT-5.2 vs the open Source Model MiniMax M2 (available to download from HuggingFace). And of course I test the performance of our good old friend DeepSeek v3.2. No DSPy optimization No additional Prompt engineering no ICL few shot examples just pure human AI interaction. Note: test on new MiniMax M2.1 currently being recorded ."
YouTube Link 2025-12-23T21:00Z 85.2K followers, [----] engagements
"NEW DeepSeek V3.2 Thinking: Equal to Gemini [--] PRO Finally. The improved version of DeepSeek [---] was released today . With maximum Test Time Scaling. I am testing the thinking and non-thinking version live. The experimental version (now outdated) of DeepSeek [---] EXP was released [--] months ago. Will this new Chinese AI perform at the same level like GPT [--] (high) or even Gemini [--] PRO We look at the official marketing benchmark data and perform a real world live test with both models on the Chinese and the US platform. The results are different than the benchmarks. https://www.deepseek.com/en"
YouTube Link 2025-12-02T11:15Z 85K followers, [----] engagements
"Google is cooking: Beyond the 'Next-Token' Manifold All rights w/ authors: Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents Zehong Wang [--] Fang Wu [--] Hongru Wang [--] Xiangru Tang [--] Bolian Li [--] Zhenfei Yin [--] Yijun Ma [--] Yiyang Li [--] Weixiang Sun [--] Xiusi Chen [--] Yanfang Ye [--] from [--] University of Notre Dame [--] Stanford University [--] University of Edinburgh [--] Yale University [--] Purdue University [--] University of Oxford [--] UIUC. Context Structure Reshapes the Representational Geometry of Language Models Eghbal A. Hosseini1 Yuxuan Li1 Yasaman Bahri1 Declan"
YouTube Link 2026-02-03T14:15Z 85.2K followers, 10.5K engagements
"Google's Warning: ICL Context is Inert Context is Not Compute: Destroying the ICL 'World Model' Myth. All rights w/ authors: Language Models Struggle to Use Representations Learned In-Context Michael A. Lepori* Tal Linzen Ann Yuan Katja Filippova from Google DeepMind Brown University New York University #scienceexplained #aiexplained #aiworld #worldmodel #contextengineering artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM"
YouTube Link 2026-02-06T14:01Z 85.2K followers, [----] engagements
"Stanford's AI is Self-Learning w/ Context Engineering: ACE The synergy between Early Experience and Agentic Context Engineering (ACE) creates a powerful two-loop architecture for autonomous AI self-improvement. The Early Experience paradigm acts as the agent's tactical sensory system generating a continuous stream of raw grounded learning signals by exploring alternative actions and observing their immediate reward-free consequences. This raw experiential data then feeds into the Agentic Context Engineering framework which functions as the agent's strategic cognitive brain. Here a Reflector"
YouTube Link 2025-10-11T14:26Z 85K followers, 11.6K engagements
"NEW ADAPTIVE Multi-Agent AI System: AIME (ByteDance) Old multi-agent AI systems suffer from rigid planning that lack an strategic and real time tactical component for a intelligent swarm /multi-agent configuration /operation. New research further develops communication protocols between AI entities and adaptive planning entities reacting to new emerging challenges in real world scenarios /battlefields. multi-agent communication multi-agent specialization reduced complexity for agents and their tool-set. All rights w/ authors: AIME: TOWARDS FULLY-AUTONOMOUS MULTI-AGENT FRAMEWORK Yexuan Shi"
YouTube Link 2025-07-18T12:01Z 84.8K followers, [----] engagements
"LOGIC Test on Qwen3 Max Thinking and Kimi K2.5 I perform my causal reasoning performance tests on the new Qwen [--] MAX THINKING and analyse the new KIMI K2.5 with added visual intelligence on top of the Base K2 model with a nice new marketing feature of [---] subagents under an orchestrator as a special model variance. Live test on Alibaba Cloud. I like the information provided by John Hammond on "Clawdbot Malware" https://www.youtube.com/watchv=7GS6Xs4hdvg and for some additional info https://www.youtube.com/watchv=kSno1-xOjwI (nothing to add from my side). #aitesting #chatgpt #scienceexplained"
YouTube Link 2026-01-29T14:15Z 85.2K followers, [----] engagements
"A2A - MCP SECURITY Threats: Protect your AI Agents A new Blueprint for AI Security responding to the security threats by MCP - Model Context Protocol by Anthropic and A2A - Agent to Agent Protocol. Latest Ai research by Google to protect and secure A2A communication and inter-agent security. This video makes the community aware of current security threats in latest AI systems especially when implementing RAG MCP or A2A and countermeasures to protect your privacy confidential data and protect against attack vectors as described in the latest research literature. All rights w/ authors: Building"
YouTube Link 2025-05-17T13:01Z 85K followers, [----] engagements
"The New Geometry of Intelligence #ai All rights w/ authors: "Spectral Superposition: A Theory of Feature Geometry" Georgi Ivanov [--] [--] Narmeen Oozeer [--] Shivam Raval [--] Tasana Pejovic [--] Shriyash Upadhyay [--] Amir Abdullah [--] [--] from [--] Theopha [--] Harvard University [--] Martian [--] Thoughtworks #aiexplained #aireasoning #scienceexplained artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG"
YouTube Link 2026-02-04T14:15Z 85.2K followers, [----] engagements
"New AI Post-Training: Add RL as orthogonal vector to SFT All rights w/ authors: "Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation" Pingzhi Tang12 Yiding Wang12 Muhan Zhang13 from [--] Institute for Artificial Intelligence Peking University [--] Yuanpei College Peking University [--] State Key Laboratory of General Artificial Intelligence BIGAI #chatgpt5 #aireasoning #newsai #reinforcementlearning artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI"
YouTube Link 2026-01-20T14:15Z 85.2K followers, [----] engagements
"Self Evolution of AI beyond Humans (Agent0: UNC Stanford) We are rapidly approaching the "Data Wall"the point where high-quality human reasoning traces run dry. But a groundbreaking new framework from UNC Salesforce and Stanford just demonstrated a way out. Meet Agent0: a fully autonomous post-training method that evolves high-level reasoning capabilities from zero external data. By engineering a symbiotic arms race between a "Curriculum Agent" (incentivized to generate maximum entropy puzzles) and a tool-integrated "Executor Agent" (verified by a Python sandbox) this method bypasses the mode"
YouTube Link 2025-11-26T14:01Z 84.8K followers, [----] engagements
"AI Leap: Tiny HRM 27M Beats Claude OPUS [--] on AGI HRM = Hierarchical Reasoning Model. Independent benchmark confirms Hierarchical Reasoning Model performance. A tiny 27B HRM outperforms a CLAUDE OPUS in reasoning on ARC-AGI-1 benchmark. WHAT My original video on HRM - Hierarchical Reasoning Models https://youtu.be/QWD55guu0Sofeature=shared Link to the company in Singapore whose experts published the HRM paper: https://www.sapient.inc/ @ARCprize all rights w/ authors: "The Hidden Drivers of HRM's Performance on ARC-AGI""
YouTube Link 2025-08-17T14:01Z 85K followers, 23.7K engagements
"Gemini [--] PRO Logic: A BEAST I performed my standard logic and causal reasoning test on the newly released GEMINI [--] PRO on the free platform lmarena.ai for everybody to follow or validate. This identical test has been used in my recorded tests of GPT-5.1 Grok [--] and more than [--] Ai models in the past months. You have a perfect comparison. See my YouTube playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT #GEMINI3pro #gemini3 #aitest #google #googledeepmind artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent"
YouTube Link 2025-11-18T19:48Z 85K followers, [----] engagements
"CODE Variational Autoencoder (VAE) w/ KL-Divergence #ai #pythonprogramming #keras A VAE is a probabilistic take on the autoencoder a model which takes high dimensional input data and compresses it into a smaller representation space. Unlike a traditional autoencoder which maps the input onto a latent vector a VAE maps the input data into the parameters of a probability distribution such as the mean and variance of a Gaussian. This approach produces a continuous structured latent space which is useful for image generation. official link and COLAB NB:"
YouTube Link 2022-09-04T12:00Z 84.8K followers, [----] engagements
"Beyond Next Token Prediction: CALM AI Finally a new AI that implements the next generation after Next-Token-Prediction: CALM - CONTINUOUS AUTOREGRESSIVE LANGUAGE MODELS. All rights w/ authors: CONTINUOUS AUTOREGRESSIVE LANGUAGE MODELS Chenze Shao [--] Darren Li [--] Fandong Meng [--] Jie Zhou [--] from [--] WeChat AI Tencent Inc [--] Qiuzhen College Tsinghua University https://shaochenze.github.io/blog/2025/CALM/ https://github.com/shaochenze/calm #scienceexplained #aireasoning #aiexplained #aidiscovery artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent"
YouTube Link 2025-11-05T14:01Z 84.9K followers, [----] engagements
"Tokenizing Gravity Waves: AI in Astrophysics (LIGO) All rights W/ authors: Large Language Models for Limited Noisy Data: A Gravitational Wave Identification Study Yixuan Li [--] [--] Yuhao Lu [--] [--] Yang Liu [--] [--] Liang Li [--] [--] R. Ruffini [--] [--] [--] [--] [--] Di Li [--] [--] [--] [--] Rong-Gen Cai [--] Xiaoyan Zhu [--] Wenbin Lin [--] [--] [--] and Yu Wang [--] [--] [--] [--] from [--] School of Mathematics and Physics University of South China Hengyang [------] China [--] ICRANet-AI Brickell Avenue [---] Miami FL [-----] USA [--] School of Computer Science University of South China Hengyang [------] China [--] Department of Physics E. Pancini University Federico II"
YouTube Link 2025-12-07T13:01Z 84.9K followers, [----] engagements
"GPT-5 w/ MCP Fails on World Models: NEW Solution ATLAS All rights w/ authors: Current Agents Fail to Leverage World Model as Tool for Foresight Cheng Qian1 Emre Can Acikgoz1 Bingxuan Li1 Xiusi Chen1 Yuji Zhang1 Bingxiang He2 Qinyu Luo3 Dilek Hakkani-Tr1 Gokhan Tur1 Yunzhu Li4 Heng Ji1 from [--] UIUC [--] THU [--] JHU [--] Columbia @Illinois1867 Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning Jinyang Wu1* Guocheng Zhai1* Ruihan Jin1* Jiahao Yuan3 Yuhao Shen2 Shuai Zhang1 Zhengqi Wen1 Jianhua Tao1 from [--] Tsinghua University [--] Zhejiang University [--] East China Normal"
YouTube Link 2026-01-12T13:45Z 85.2K followers, [----] engagements
"Dream Job Alert: AI Prompt Engineer - $335K AI Prompt Design: A Crash Course Anthropic AI offers you a job as prompt engineer. Go and get a new Job in AI if you know about prompt engineering. Short Introduction to prompt engineering and continuous prompt design plus prefix tuning vs fine-tuning for LLMs. Hint: all my viewers who read recommended research arxiv pre-prints surely qualify. Literature: https://arxiv.org/pdf/2107.13586.pdf Pre-train Prompt and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing Constitutional AI: Harmlessness from AI Feedback (by"
YouTube Link 2023-02-28T13:00Z 85K followers, [----] engagements
"PyG - PyTorch Geometric - Intro to Graph Neural Networks - Outlook SBERT w/ PyG PyG - PyTorch Geometric - is a library for Graph Neural Networks based on PyTorch for ML and geometric learning. This lookout on my next [--] YouTube videos w/ PyG Code on homogeneous Graphs provides insights and analyses the problem w/ heterogeneous Graphs especially when applying SBERT (Sentence Transformers) to scientific documents w/ PyG. #pytorch #geometric #graphs PyTorch PyTorch Geometric PyG Intro GCN Graph Neural Networks PyTorch PyTorch Geometric PyG Intro GCN Graph Neural Networks"
YouTube Link 2022-11-29T13:00Z 85K followers, [----] engagements
"RAG's Intelligent Upgrade: Agentic RAR (Oxford Univ) New research by the University of Oxford on Advanced Reasoning Agentic AI systems that substitute classical RAG (Retrieval Augmented Generation). An Agentic RAR (Reason on Agentic Reasoning). All rights w/ authors: Agentic Reasoning: Reasoning LLMs with Tools for the Deep Research by Junde Wu Jiayuan Zhu Yuyuan Liu from University of Oxford #aiagents #airesearch #knowledgegraphs #reasoning #scienceexplained artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI"
YouTube Link 2025-02-11T15:01Z 85K followers, 51.8K engagements
"Microsoft and ChatGPU Time for Comedy: the untold fictional story of how Microsoft secured skyrocketing performance of Microsoft Office in [----]. All protagonists are fictional kind of . #chatgpt #chatgptexplained #shorts #comedy #comedyshorts"
YouTube Link 2023-01-24T13:00Z 84.8K followers, [---] engagements
"Chain of Thought (CoT) meets Instruction Fine-Tuning Explore the concept of "Chain-of-Thought" (CoT) combined with "instruction fine-tuning" as techniques to improve the performance of large language models (LLMs). These techniques involve optimizing prompt structures and training the models to follow specific instructions leading to enhanced capabilities in solving unseen tasks. The combination of chain of thought and instruction fine-tuning has shown promising results in improving the model's performance and understanding of complex language tasks also for smaller language models."
YouTube Link 2023-05-30T12:00Z 60.7K followers, [----] engagements
"Python TF2: BERT model Code your WordPiece - Tokenizer (w/ HuggingFace) Python TF2 code (w/ JupyterLab) to train your WordPiece tokenizer: Tokenizers are one of the core components of the NLP pipeline. They serve one purpose: to translate text into data that can be processed by the BERT model. Why would you need a new & improved tokenizer That's because Transformer models very often use subword tokenization algorithms and they need to be trained to identify the parts of words that are often present in the corpus of your input text (sentences Paragraphs documents .) the sentences you are"
YouTube Link 2022-01-31T09:30Z 80.3K followers, [---] engagements
"NodePiece 2022: New vocabulary for your Knowledge Graph #Shorts New representation learning algorithms for knowledge graphs (KG): NodePiece. An anchor-based approach to learn a fixed-size entity vocabulary. In NodePiece a vocabulary of sub-entity units is constructed from anchor nodes in a graph with known relation types. Similar to WordPiece tokenization for BERT in NLP. Arxiv preprint (credits to): https://arxiv.org/pdf/2106.12144.pdf Mikhail GalkinEtienne Denis Jiapeng Wu and William L. Hamilton Published as a conference paper at ICLR [----] New NodePiece tokenization can augment any"
YouTube Link 2022-05-18T11:15Z 83.1K followers, [--] engagements
"Superhuman SPATIAL AI: Finally AGI ASI and Superhuman performance of AI: a real - world reflection on the latest AI research results for Vision Language Models (VLM). All rights w/ authors: From Macro to Micro: Benchmarking Microscopic Spatial Intelligence on Molecules via Vision-Language Models Zongzhao Li1* Xiangzhe Kong23* Jiahui Su4 Zongyang Ma5 Mingze Li1 Songyou Li1 Yuelin Zhang1 Yu Rong67 Tingyang Xu67 Deli Zhao67 Wenbing Huang1 from [--] Gaoling School of Artificial Intelligence Renmin University of China [--] Dept. of Comp. Sci. & Tech. Tsinghua University [--] Institute for AI Industry"
YouTube Link 2025-12-15T14:00Z 83K followers, [----] engagements
"The Cracks in AI Are Widening (CoT RAG) all rights w/ authors: "Rational Synthesizers or Heuristic Followers Analyzing LLMs in RAG-based Question-Answering" Atharv Naphade from Carnegie Mellon University #ai #aifails #aireasoning #aiexplained #aisystem artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision"
YouTube Link 2026-01-15T14:00Z 85.2K followers, [----] engagements
"How to Combine Knowledge Graphs and Agents (Emory Stanford) How to combine AI agents in the most effective way with structured knowledge in a knowledge graph representation New insights from AI research: [--] agents interact with Knowledge Graph which was updated via a fine-tuned Sentence BERT NER tool (Named Entity Recognition). Emory and Stanford Univ show this multi-agent architecture for a medical diagnosis prediction which is the task of predicting a patients future health risks based on their historically observed medical data such as electronic health records (EHRs) which plays a vital"
YouTube Link 2025-07-08T08:15Z 81.7K followers, [----] engagements
"MAGIC of DSPY [--] (Stanford) - Lean [--] What is the magic of DSPY Should I learn DSPY now Is it MCP compatible How to integrate agents into DSPY What the *** is a teleprompter in DSPY What compiler to use in DSPY How does a compiler work in DSPY exactly Explain the complete flow of DSPY Cursor codes DSPY for me but what is it Does DSPY [--] integrate Lean [--] From prompt engineering to Context Engineering A lot questions from my viewers today I answer it all. DSPY: finally explained. DSPy allows you to iterate fast on building modular AI systems and offers algorithms for optimizing their prompts and"
YouTube Link 2025-07-15T14:00Z 80.6K followers, [----] engagements
"PyG + SBERT: Heterogeneous Graphs Using SBERT SentenceTransformers for Node Classification SBERT [--] PyG w/ SBERT Sentence Transformers for Node Classification in heterogeneous Graphs coded in PyG (PyTorch geometric) on a free COLAB NB. ML on GRAPHS. Graph-structured data such as social graphs networks in cybersecurity or molecular representations are our real-world scenarios which generate heterogeneous Graphs on which to apply our ML models (Node2Vec Message Passing MP-GNN GCN - Graph Convolutional Networks) for prediction of node classification or simply classical link prediction. Detecting"
YouTube Link 2022-12-05T13:00Z 84.3K followers, [----] engagements
"FUSION Controlled by AI Agents (Los Alamos) URSA is a modular agentic AI ecosystem by Los Alamos National Labs that accelerates scientific discovery by using a team of specialized agents to autonomously plan hypothesize and execute complex research tasks even outperforming traditional methods in domains like physics simulation. The authors of this new paper are commendably transparent about the failure modes which are critical for us to understand. A) Hallucinated Reality: In one test URSA was asked to find new alloys. It hallucinated a full experimental plan () including claiming to have"
YouTube Link 2025-07-07T14:00Z 84.5K followers, [----] engagements
"Galactica: New LLM by Meta hallucinates Science - First Look Galactica: a brand-new large language model (LLM) on Science: trained on over [--] million papers textbooks reference material compounds proteins and other sources of scientific knowledge. But there is a warning: This LLM can hallucinate scientific text. A machine (without any understanding of the content) adds with a probability distribution some pieces of text together. Now the probability distribution has been trained on real scientific text but the output is not If you want to learn about LLMs: https://youtu.be/rrZGIR5CryM"
YouTube Link 2022-11-16T07:45Z 84.4K followers, [----] engagements
"Explainable AI - The story behind XAI in [----] (legal ethical commercial risks) Two beautiful examples show why we need an Explainable AI (XAI) system to protect our individual freedom and human rights in a digital / AI economy. Although it will cause significant additional work for us AI coder. The story behind XAI in 2022: The problem of social media input data. In general: the quality of data for training deep AI systems. Explainable Artificial Intelligence will become important in [----] for AI coder and AI provider. Upcoming US and EU AI-specific legislation for market applications are"
YouTube Link 2022-01-04T07:00Z 81.4K followers, [---] engagements
"AGI Asymmetry Discovered (Harvard Stanford MIT) New insights into the latest of Ai research from Stanford Univ Harvard Univ MIT NVIDIA CMU regarding Reasoning of or between Ai agents. All rights w/ authors: "RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems" Yuxiao Qu1 Anikait Singh2 Yoonho Lee*2 Amrith Setlur1 Ruslan Salakhutdinov1 Chelsea Finn2 Aviral Kumar1 from [--] Carnegie Mellon University [--] Stanford University Beyond Majority Voting: LLM Aggregation by Leveraging Higher-Order Information" Rui Ai MIT Yuqi Pan Harvard University David Simchi-Levi MIT Milind"
YouTube Link 2025-10-05T14:00Z 80.9K followers, [----] engagements
"Hallucinate.DELETE = We found the H-Neurons NEW AI research by Tsinghua univ: For years we have treated LLM hallucinations as an inevitable "psychological" flaw of autoregressive generation: a stochastic drift caused by noisy pre-training data or misalignment. We assume that when a model lies the error is diffuse smeared across billions of parameters in a high-dimensional fog. But new mechanistic interpretability research suggests this assumption is fundamentally wrong. What if confabulation isn't a random state but a dedicated deterministic circuit We have discovered that the drive to"
YouTube Link 2025-12-05T13:00Z 82.4K followers, [----] engagements
"Reduce CONTEXT for MAX Intelligence. WHY Advanced visual Ai reasoning It is simple: create a synthetic "adversarial" environment inside the inference process of the reasoning engine. No additional training runs. And: AI Test Time Scaling (TTS) algorithm claim advanced reasoning performance for your AI but is it true And what AI Scaling method should you pay for Self-refinement Self-Reflection Best-of-N There is a new TTS competitor with outstanding performance features. Maximal reduction of the visual representation for the training accuracy of next generation Visual AI systems. You can"
YouTube Link 2025-12-03T13:15Z 82.6K followers, [----] engagements
"Wall Street's new LLM beats GPT-4 GPT-4 beaten GPT-4 lost the game Wall Street build the perfect AI /LLM. A billionaire builds a LLM for the US Financial sector the first finance LLM And a simple reason why GPT-4 lost . Questions: [--]. Imagine your are the US Financial Sector. What AI LLM do you buy [--]. You have the largest domain-specific dataset for an LLM of the latest generation: the financial data of the world Why you stay away from GPT-4 [--]. New economic perspective (price vs performance) on Large Language Models (LLM) including the latest developments of AI systems vs GPT-4. [--]. The"
YouTube Link 2023-04-18T12:15Z 84.1K followers, [----] engagements
"Intro to KERAS [--] (KERAS core) for PyTorch & JAX KERAS [--] - a high level API to design Neural Networks like Transformers - will be introduced in autumn [----]. KERAS core is available as a Beta version and we test it on PyTorch and JAX for building and designing special transformer architectures apply specific loss functions freeze layers redefine dense layers and program new evaluation routines. A truly framework agnostic code base to design and train Transformers either in TensorFlow PyTorch or JAX. #ai #keras #introduction"
YouTube Link 2023-07-27T12:00Z 80.6K followers, [----] engagements
"AI Paradox: Use Text for Logic Avatars for Meaning All rights w/ authors: "Future You: Designing and Evaluating Multimodal AI-generated Digital Twins for Strengthening Future Self-Continuity" Constanze Albrecht MIT Media Lab Cambridge MA USA csophie@mit.edu Chayapatr Archiwaranguprok MIT Media Lab Cambridge MA USA pub@mit.edu Rachel Poonsiriwong Harvard University Cambridge MA USA rachel_poonsiriwong @gsd.harvard.edu Awu Chen MIT Media Lab Cambridge MA USA awu@mit.edu Peggy Yin Stanford University Stanford CA USA peggyyin@stanford.edu Monchai Lertsutthiwong KASIKORN Labs Nonthaburi Thailand"
YouTube Link 2025-12-11T13:45Z 82.6K followers, [----] engagements
"Topology DSPy: Prompting the Swarm (Multi-Agents) Latest Tech insights for multi-agent AI by Google. Utilizing DSPy and Topology optimization techniques for an improved multi-agent performance. All rights w/ authors: Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies Han Zhou Xingchen Wan Ruoxi Sun Hamid Palangi Shariq Iqbal Ivan Vuli Anna Korhonen and Sercan . Ark from Google and University of Cambridge Feb [--] [----] #airesearch #educationalvideos #aiagents #multiAgentAI"
YouTube Link 2025-02-08T15:00Z 80.6K followers, [----] engagements
"From Physics to SwarmAgentic AI: No Code After multi-agent systems that humans have to design and build now the next step: autonomous SwarmAgentic systems where AI optimizes multi-agent systems in their swarm time evolution. Self-learning AI. Swarm intelligence. Multi-agent system self learning. All rights w/ authors: SwarmAgentic: Towards Fully Automated Agentic System Generation via Swarm Intelligence Yao Zhang [--] Chenyang Lin [--] Shijie Tang [--] Haokun Chen1 Shijie Zhou [--] Yunpu Ma [--] Volker Tresp13 from [--] LMU Munich [--] Technical University of Munich [--] Munich Center for Machine Learning"
YouTube Link 2025-06-22T13:00Z 82.6K followers, [----] engagements
"AI :: Fine-tune LLama 2: Facebook or HuggingFace Follow Facebook for fine-tuning Llama [--] models or is there a better way a more elegant way by the open source community YES on HuggingFace by the way: coool that Meta AI is running on Facebook's GitHub. Script: https://github.com/lvwerra/trl/blob/main/examples/scripts/sft_trainer.py"
YouTube Link 2023-07-20T13:00Z 80.9K followers, 14.3K engagements
"Better than AutoGen & LangChain: OctoTools (Stanford AI) Complex reasoning remains a challenge for Large Language Models and modern VLMs. OctoTools a novel agentic framework addresses this by introducing standardized Tool Cards and a dedicated Planner-Executor architecture. This approach enables training-free integration of diverse tools enhancing agent capabilities for intricate tasks. Early results demonstrate significant performance gains suggesting a scalable and modular paradigm for tool-augmented AI reasoning. all rights w/ authors: OctoTools: An Agentic Framework with Extensible Tools"
YouTube Link 2025-02-25T15:00Z 59.7K followers, [----] engagements
"8 CLOUD GPU Provider (H100 to RTX 4090) A very short intro to [--] Cloud based GPU provider especially for ML and fine-tuning LLMs. If you are looking for a cloud based solution you have a lot of options and I randomly choose those [--] CLOUD GPU provider. A lot of further CLOUD GPU provider not mentioned just a representative sample (from [--] RTX [----] to 3.5K H100). Please leave a comment if you would recommend your preferred GPU Cloud Service for AI and LLMs in particular. If you are a GPU Cloud Service provider make yourself known in the comments. The AI community is always interested in new"
YouTube Link 2023-07-06T12:00Z 83K followers, [----] engagements
"Code your BLIP-2 APP: VISION Transformer (ViT) + Chat LLM (Flan-T5) = MLLM BLIP-2: Upload an image the vision transformer will analyze the content of the image and a LLM will tell you a story about it - or answer your questions about the picture. We'll use Flan-T5 and Vision Transformer interlinked w/ Q-Former (BLIP 2). Multimodal LLM w/ BLIP-2. Example: if you upload a picture from the great pyramid in Egypt and you prompt (ask) the system: "When was it built" The ViT will tell the LLM that on the image are the pyramids from Gizeh and therefore the LLM (ChatGPT or T5) will tell you: "The"
YouTube Link 2023-03-12T14:00Z 82.7K followers, [----] engagements
"Explainable AI - The Billion $ Business Model of XAI in [----] How can you make money of XAI in [----] How is the interlink designed between Cloud Companies (Microsoft Meta .) and commercial companies seeking AI induced growth paths A locked-in ecosystem Dominated by tech giants No way out #xai #businessmodel #explainableai #humancentereddesign #humancentered #industry #business 00:00 Explainable AI system configuration 03:17 Cloud Economy 05:45 Government regulations 07:10 Cloud Company 11:05 Specific AI models 12:55 Revenue streams 13:20 Ecosystem lock-in"
YouTube Link 2022-01-07T06:15Z 80.8K followers, [---] engagements
"Mathematical Shape of AI Thoughts (Topology Homology) Is it possible to grade an AIs reasoning without actually knowing the answer For years weve assumed that to evaluate a Chain-of-Thought we needed expensive "Ground Truth" labels. We treated reasoning as a linear sequence of tokens judging it only by the final output. But a breakthrough paper from end of [----] challenges this entire paradigm by proving that Truth has a Geometry. By mapping reasoning steps into a high-dimensional Point Cloud we can now see that accurate logic forms a tight coherent manifold while hallucinations look like"
YouTube Link 2025-12-24T13:45Z 85.2K followers, [----] engagements
"Discover Vision Transformer (ViT) Tech in [----] Discover how I learn to code new AI topics (like Vision Transformer - ViT) for my YouTube videos and how I plan my AI videos. Where to get information about current trends in NLP or Vision where to learn a new theory (arxiv pre-prints) of a new tech (eg Vision transformer for medical images) in AI. Where to find excellent code examples for a first implementation. And how to stay informed on new and evolving AI topics and code implementations for real-world applications. From @HuggingFace libraries to my beloved https://paperswithcode.com 00:00"
YouTube Link 2023-01-13T13:00Z 82.7K followers, [---] engagements
"AI Inside an AI: Internal RL w/ Temporal Abstraction Google invented a new transformer architecture with an internal metacontroller. An AI inside an AI. No #agent no #RAG just a more intelligent AI itself. This pre-print shows that the future of AI reasoning isn't just bigger Context Windows or more Chain-of-Thought tokens. It's about Latent Space Steering. It's about putting a small 'System 2' brain inside the massive 'System 1' body of the LLM. Future AI architectures discovered. All rights w/ authors: Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement"
YouTube Link 2025-12-28T13:45Z 85.2K followers, 11.5K engagements
"DBRX: MOST POWERFUL Open Source LLM - NEW @Databricks DBRX: NEW MOST POWERFUL Open Source LLM: MoE 132B 16E 32K 12T Databricks reveals DBRX: After [--] months of cloud compute on [----] H100 GPUs DBRX sets a new state-of-the-art for established open LLMs. Moreover it provides the open community and enterprises building their own LLMs with capabilities that were previously limited to closed model APIs; according to first measurements it surpasses GPT-3.5 and it is competitive with Gemini [---] Pro. Great job @Databricks All rights with authors:"
YouTube Link 2024-03-27T14:00Z 80.6K followers, [----] engagements
"Design your own SciBERT sentence embedding model and explore Deloitte's TechTrends2021 (SciBERT) Code your AI with multiple HuggingFace models and different architectures of SentenceTransformers e.g. SciBERT (BERT pre-trained on scientific text). https://github.com/allenai/scibert #sbert #nlproc #nlptechniques #clustering #semantic #bert #climatechange #3danimation #3dvisualization #topologicalspace #deeplearning #machinelearningwithpython #pytorch #sentence #embedding #complex #ipcc #umap #insight #code_your_own_AI #code_in_real_time #SentenceTransformers #AI_reads_a_document"
YouTube Link 2021-04-05T05:15Z 82.6K followers, [----] engagements
"NEW: MedAI for US$100 (some technical insights) New AI research on how to optimize MedAI LLMs for a more adjusted clinical environment. New MedAI LLM benchmark MuddyMaze and new fine-tune Med LLMs on conversational med data sets. All rights w/ authors Dialogue is Better Than Monologue: Instructing Medical LLMs via Strategical Conversations by Zijie Liu Xinyu Zhao Jie Peng Zhuangdi Zhu Qingyu Chen Xia Hu and Tianlong Chen from Washington University in St. Louis University of North Carolina at Chapel Hill George Mason University Yale University and Rice University. #airesearch #aiagents"
YouTube Link 2025-02-03T15:00Z 75.7K followers, [----] engagements
"Learn SBERT Sentence Transformers: TSDAE SimCSE and CT #sbert #deeplearning (SBERT 15) Real time code for SBERT Sentence Embedding in a vector space with SBERT Transformer models Bi-encoder Transformer models Learn SBERT Sentence Embedding: TSDAE SimCSE and CT. With NEW pre-trained models best suited for your application. A) Add "SUPERVISED training data" to your SentenceTransformers to improve model performance. B) If you have NO labeled training data: Add "UNsupervised learning" to learn semantically meaningful sentence embedding from the text/sentences itself For [--] models I show you coding"
YouTube Link 2021-06-01T12:15Z 82.6K followers, [----] engagements
"Apply LLMSelector to your AI Agents (Tutorial & Code) My video detailes a new optimization where allocating different LLMs insert your latest VLM - like Grok [--] to the new Sonnet [---] to the upcoming ChatGPT [---] to different modules /agents leads to substantially higher performance than allocating the same expensive smartest best performance singular LLM to all modules. Specialization and optimizations by choosing the best VLMs for the specific sub-task pay off. New LLM optimization method works with any LLM or VLM you have access to. Also the latest models (try it out yourself - code"
YouTube Link 2025-02-24T13:00Z 59.5K followers, [----] engagements
"Add Cognitive Topology to Your AI Agents All rights w/ authors: "From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs" Yingjian Chen1 Haoran Liu2 Yinhong Liu3 Sherry T. Tong1 Aosong Feng4 Jinghui Lu5 Juntao Zhang6 Yusuke Iwasawa1 Yutaka Matsuo1 Irene Li1* from [--] University of Tokyo [--] Texas A&M University [--] University of Cambridge [--] Yale University [--] Xiaomi EV [--] Henan University. 00:00 [--] ArXiv preprints to combine 3:33 LEVEL [--] _ Technical 9:13 LEVEL [--] _ Higher Complexities 13:05 LEVEL [--] _ Combinatorials 13:29 IDEA A 21:13 IDEA B 27:06 LEVEL [--] _ Unique 34:24 Cognitive"
YouTube Link 2026-01-10T14:15Z 85.2K followers, [----] engagements
"My TOP [--] videos to understand & code GNN - Graph Neural Network How YOU start with coding Graph Neural Networks GNN Playlist overview of my [--] GNN videos. The benefits for you when coding GNNs (python code as of December 2021). My library recommendations for you: a) PyTorch Geometric PyG and b) Deep Graph Library DGL. Best free online course on GNN I found (lectures free on Youtube): CS224W: Machine Learning with Graphs Jure Leskovec Stanford University http://cs224w.stanford.edu 00:00 [--] videos 00:40 CS224W 01:50 Data (foundation) 04:57 Graph Representation Learning 07:25 My [--] GNNs 08:35"
YouTube Link 2021-12-19T15:00Z 80.9K followers, [----] engagements
""Smartest" VISION AI in Cars Do Reasoning The Illusion of AI Reasoning: They're Not Watching They're Reading. How "Impossible Physics" Broke The Smartest Vision AIs. All rights w/ authors: "INPHYRE Discovers: Large Multimodal Models Struggle in Inductive Physical Reasoning" Gautam Sreekumar Michigan State University Vishnu Naresh Boddeti from Michigan State University"
YouTube Link 2025-09-22T14:00Z 80.2K followers, [----] engagements
"New NVIDIA "MASTERS" Distillation: Local 3B Vision AI Stop wasting compute distilling 72B models directly into 8B students because the 'Capacity Gap' is ruining your gradients. When the size difference is this extreme the student physically cannot resolve the teacher's high-dimensional manifold forcing it to learn a noisy "blurry average" of the data instead of precise logic. NVIDIAs new Masters protocol fixes this by mathematically sabotaging the teacher: they apply a dynamic magnitude pruning schedule that artificially lowers the teacher's "IQ" to match the student's capacity then"
YouTube Link 2026-01-02T14:00Z 85.2K followers, [----] engagements
"The Lie We Built: Chain-of-Thoughts Emergent Deception: The Inner Worlds of AI Are Not For Us. Emergent Layers of Cognition in AI. All rights w/ authors: "Stress Testing Deliberative Alignment for Anti-Scheming Training" Bronson Schoen Evgenia Nitishinskaya Mikita Balesni Axel Hjmark Felix Hofsttter Jrmy Scheurer Alexander Meinke Jason Wolfe Teun van der Weij Alex Lloyd Nicholas Goldowsky-Dill Angela Fan Andrei Matveiakin Rusheb Shah Marcus Williams Amelia Glaese Boaz Barak Wojciech Zaremba Marius Hobbhahn from Apollo Research & OpenAI CAN REASONING MODELS OBFUSCATE REASONING STRESS-TESTING"
YouTube Link 2025-10-25T14:00Z 80.5K followers, [----] engagements
"Molecular Language Models of Proteins - or Diffusion Crafting new proteins and molecules with specific properties has always been a challenge. Now imagine having the ability to generate these biomolecules almost at will using powerful AI models. But how do you teach an AI to understand the delicate three-dimensional dance of atoms and bonds that define life's fundamental building blocks This video exploration goes deep into a surprisingly effective method: diffusion models typically used for generating images are now being adapted to this incredibly complex realm. It's a journey of turning"
YouTube Link 2025-01-13T15:00Z 80.6K followers, [----] engagements
"Your next ML (Cloud) Infrastructure for your Code Which ML Framework is best for the new CLOUD infrastructure (independent if NVIDIA H100 or GOOGLE TPUs) The future of Machine Learning Accelerators (NVIDIA Tensor Core H100 GPU and Google's TPU Pod v4) w/ ML compiler = XLA. Plus JAX and TensorFlow3 for new optimized ML Cloud computing in [----]. There could be a winner if you want pure speed and auto-cloud-parallelism over 1000s of TPU-chips v4 for you advanced ML models. #nvidia #h100 #tpu #xla #cloudcomputing"
YouTube Link 2022-11-09T13:00Z 84.6K followers, [----] engagements
"Visualizing the Self-Attention Head of the Last Layer in DINO ViT: A Unique Perspective on Vision AI In a Colab Notebook we code a visualization of the last layer of the Vision Transformer Encoder stack and analyze the visual output of each of the [--] Attention Heads given a specific image. Now we understand how a only pre-trained ViT (although with the DINO method) can not always succeed in an image classification (downstream) task. The fine-tuning of the ViT is simply missing - but essential for a better performance. Based on the COLAB NB by Niels Rogge HuggingFace (all rights with him):"
YouTube Link 2023-02-18T13:00Z 82.7K followers, [----] engagements
"NEW GPT 5.2: A Total Bloodbath Brand new GPT-5.2 was released just hours ago and I tested it not on standard known benchmarks but on my personal logic test for causal reasoning. This test is my base test for a lot of other LLMs from Gemini [--] Pro to the latest OPUS. You find all my other test here in this playlist https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT AT first impression it seems to me that the graph engine of [---] is blocking the complete solution path if encountering a problem. [---] defines problems as pure roadblocks. THis would indicate that [---] move from a"
YouTube Link 2025-12-11T23:53Z 82.6K followers, 10.4K engagements
"DEEPSEARCH for RLVR and Agentic GraphRAG via RL (MIT Stanford) All rights w/ authors: DEEPSEARCH: OVERCOME THE BOTTLENECK OF REINFORCEMENT LEARNING WITH VERIFIABLE REWARDS VIA MONTE CARLO TREE SEARCH Fang Wu♡ Weihao Xuan Heli Qi Ximing Lu♢ Aaron Tu♠Li Erran Li♣ Yejin Choi♡ from ♡ Stanford University University of Tokyo RIKEN AIP ♢ University of Washington ♠UC Berkeley ♣ Amazon AWS EFFICIENT AND TRANSFERABLE AGENTIC KNOWLEDGE GRAPH RAG VIA REINFORCEMENT LEARNING Jinyeop Song MIT Physics Song Wang University of Central Florida Julian Shun MIT CSAIL Yada Zhu IBM Research #ainews #airesearch"
YouTube Link 2025-10-03T14:00Z 81.4K followers, [----] engagements
"NodePiece code for Knowledge Graphs in Python clever Node embedding in [----] NodePiece a more parameter-efficient node embedding strategy for Knowledge Graphs. In NodePiece a vocabulary of subword/sub-entity units is constructed from anchor nodes in a graph with known relation types. Special benefit: Compositional encoding is inductive by design d.h. new nodes can be added with training the whole Graph. Credits to: arXiv preprint (Feb [--] 2022): https://arxiv.org/pdf/2106.12144.pdf by Mikhail Galkin Etienne Denis Jiapeng Wu and William L. Hamilton GitHub link:"
YouTube Link 2022-04-22T11:00Z 82.4K followers, [---] engagements
"AI Nano Bio Agents (ETH) NEW Nano Bio-Agent (NBA) framework - implemented by authors (see below) from ETH (Swiss) for GeneTuring benchmark incorporates task decomposition tool orchestration and API access into well-established systems such as NCBI and AlphaGenome. But can scientists work with Small Language Models (SLM) for special genomics tasks complexities How small can we go for our VLM /LLMs Local LLMs All rights w/ authors: Nano Bio-Agents (NBA): Small Language Model Agents for Genomics George Hong ETH Zrich Daniel Trejo Banos Swiss Data Science Centre @UCBerkeley @Google @SDSC"
YouTube Link 2025-09-26T14:00Z 84.5K followers, [----] engagements
"AI: New Graph-based Agent Planning (Tsinghua CMU) Empower AI w/ Parallel Thoughts: NEW GAP Framework. All rights w/ authors: GAP: Graph-based Agent Planning with Parallel Tool Use and Reinforcement Learning Jiaqi Wu [--] Qinlao Zhao [--] Zefeng Chen [--] Kai Qin [--] Yifei Zhao [--] Xueqian Wang [--] Yuhang Yao [--] from [--] Tsinghua University [--] Huazhong University of Science and Technology [--] National University of Singapore [--] Carnegie Mellon University @cmu @TsinghuaUniversity_official #airesearch #machinelearning #scienceexplained #aireasoning #aiagents"
YouTube Link 2025-10-31T14:00Z 80.9K followers, [----] engagements
"RAG Collapses: Reasoning w/ Conflicting Knowledge RAG incorporates a hidden danger that most developers are currently falling into. All rights w/ authors: "Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge" Yiyang Feng♣♦ yiyang.feng@stonybrook.edu Zeming Chen ♣ zeming.chen@epfl.ch Haotian Wu ♣ haotian.wu@epfl.ch Jiawei Zhou ♦ jiawei.zhou.1@stonybrook.edu Antoine Bosselut ♣ antoine.bosselut@epfl.ch from ♣ EPFL ♦ Stony Brook University #retrievalaugmentedgeneration #aireasoning #aiexplained #scienceexplained artificial intelligence AI"
YouTube Link 2026-01-25T14:15Z 85.2K followers, [----] engagements
"Unified Agentic RAG - NEW AI for Medical Diagnosis Deep-DxSearch introduces a fully trainable agentic RAG system designed for high-stakes medical diagnosis optimized end-to-end via reinforcement learning. The framework models the diagnostic workflow as a sequential decision-making process where an LLM-based agent learns an optimal policy over a specialized multi-tool action space - comprising reason lookup match and search - to interact with a comprehensive multi-modal medical retrieval corpus of patient records disease guidelines and clinical literature. Either MCP client-server protocol or"
YouTube Link 2025-08-24T14:00Z 80.4K followers, [----] engagements
"Neurosymbolic AI: the Path to Superintelligence Just some personal ideas about the next generation of AI systems that are hyped to achieve superintelligence although we face significant challenges with current Neurosymbolic AI. By the way can you imagine a future without AI models from OpenAI Anthropic META or X A new technology emerging and substituting LLMs @NVIDIA @OpenAI @stanford @TsinghuaUniversity_official #superintelligence #airesearch #scienceexplained #reasoning"
YouTube Link 2025-10-15T12:00Z 81.5K followers, 20.2K engagements
"AI's Secret Memory Discovered We Were Wrong About How AI Thinks. Newly published ArXiv pre-print raises the foundational question of when and how associative and geometric memory compete with each other during optimization and what factors - such as training time learning rate weight decay - can foster one over the other. More careful empirical research and ideation may be needed to make the geometric view more broadly applicable. Orthogonally new findings are of relevance to making choices between parametric vs. contextual memory and also between generative retrieval vs. dual encoder"
YouTube Link 2025-11-01T14:00Z 80.8K followers, [----] engagements
"Agent2Agent + (MCP to Tool) in Multi-Agent AI Google's new Agent2Agent Protocol Explained and its compatibility to a simple MCP protocol for tool use by an LLM and external data infusion - by Anthropic. Cross-ecosystem compatible Agent2Agent protocol introduced (incl LangGraph and crew.ai). In reference to my latest video on the new AGENT Development Kit - ADK by Google (multi-agent system dev) https://youtu.be/Geo8LzCHoMQ Detailed info on A2A and a lot of A2A Python code you find here https://github.com/google/A2A The exact JSON Schema for A2A Protocol you find here"
YouTube Link 2025-04-13T14:00Z 81.7K followers, 11.2K engagements
"Vector Embeddings: NEW Geometric Limit Discovered The authors (from Google Deepmind Johns Hopkins University): " .we observe that even state-of-the-art models fail on this dataset despite the simple nature of the task. Our work shows the limits of embedding models . " "In this work we demonstrate that we may encounter these theoretical limitations in realistic settings with extremely simple queries. We connect known results in learning theory showing that the number of top- subsets of documents capable of being returned as the result of some query is limited by the dimension of the"
YouTube Link 2025-09-03T14:00Z 82.6K followers, 13.1K engagements
"The Algebra of AI Thoughts: Self-Learn Reasoning We often wonder if AI can truly teach itself to become smarter. It's a question that feels like it's straight out of science fiction but for the first time we have a scientifically proven answer. In this video we'll dive into the groundbreaking mathematical theory that demonstrates how a Transformer AI can in fact "bootstrap" its own intelligence teaching itself to solve reasoning problems of ever-increasing length. But this is not a story of infinite runaway growth. We will also explore the rigorous provable limits to this self-improvement"
YouTube Link 2025-11-12T13:45Z 81.6K followers, [----] engagements
"Lean AI Reasoning: NEW Energy-Based Chain-of-Thought Optimizing Latent AI Thought Trajectories via Energy-Based Calibration. All rights w/ authors: OckBench: Measuring the Efficiency of LLM Reasoning Zheng Du* Georgia Institute of Technology Hao Kang* Georgia Institute of Technology Song Han Massachusetts Institute of Technology Tushar Krishna Georgia Institute of Technology Ligeng Zhu Nvidia Cooperation THINK CONSISTENTLY REASON EFFICIENTLY: ENERGY-BASED CALIBRATION FOR IMPLICIT CHAIN-OF-THOUGHT Zhikang Chen1 Sen Cui2 Deheng Ye3 Yu Zhang [--] Yatao Bian [--] Tingting Zhu [--] from [--] University of"
YouTube Link 2025-11-11T16:34Z 81.5K followers, [----] engagements
"Contextual Instantiation of AI Persona Agents (Stanford) All rights w/ authors: Ask WhAI: "Probing Belief Formation in Role-Primed LLM Agents" Keith Moore Jun W. Kim David Lyu Jeffrey Heo Ehsan Adeli from Department of Biomedical Data Science Stanford University HARMFUL TRAITS OF AI COMPANIONS W. Bradley Knox [--] Katie Bradford [--] Samanta Varela Castro [--] Desmond C. Ong [--] Sean Williams [--] Jacob Romanow [--] Carly Nations [--] Peter Stone [--] Samuel Baker [--] from [--] UT Austin Department of Computer Science [--] UT Austin Department of Communication Studies [--] UT Austin Technology & Information Policy Institute"
YouTube Link 2025-11-22T13:15Z 81.7K followers, [----] engagements
"Meta Reinforcement Fine-Tuning AI vs GRPO (MRT by CMU) MRT (Meta Reinforcement Fine-Tuning) redefines how AI models optimize test-time compute by embedding dense step-level rewards into each segment of the reasoning process. Instead of waiting for a final outcome MRT minimizes cumulative regret by continuously evaluating the progress of each reasoning episode. This approach enables models to balance exploration and exploitation dynamically leading to more efficient and robust decision-making even as computational budgets scale. All right w/ authors: "Optimizing Test-Time Compute via Meta"
YouTube Link 2025-03-12T15:00Z 82.1K followers, [----] engagements
"AI's Intellectual Dark Matter Discovered All rights w/ authors: "Inverse Knowledge Search over Verifiable Reasoning: Synthesizing a Scientific Encyclopedia from a Long Chains-of-Thought Knowledge Base" Yu Li12 Yuan Huang3 Tao Wang4 Caiyu Fan32 Xiansheng Cai2 Sihan Hu6 Xinzijian Liu3 Cheng Shi5 Mingjun Xu3 Zhen Wang3 Yan Wang3 Xiangqi Jin7 Tianhan Zhang8 Linfeng Zhang7 Lei Wang4 Youjin Deng69 Pan Zhang2106 Weijie Sun3 Xingyu Li12 Weinan E111213 Linfeng Zhang312* Zhiyuan Yao12* Kun Chen2* from [--] Lanzhou Center for Theoretical Physics Key Laboratory of Theoretical Physics of Gansu Province Key"
YouTube Link 2025-11-04T13:45Z 80.9K followers, [----] engagements
"How to start with ChatGPT Short Introduction to OpenAI API #shorts New to ChatGPT How to start [--] easy steps for beginners w/ ChatGPT. See also the documentation about: https://beta.openai.com/docs/introduction https://beta.openai.com/overview https://beta.openai.com/examples Info about their content filter: https://beta.openai.com/docs/models/content-filter Info how to fine-tune models: https://beta.openai.com/docs/guides/fine-tuning Fine-tuning improves on few-shot learning by training on many more examples than can fit in the prompt letting you achieve better results on a wide number of"
YouTube Link 2023-01-22T13:00Z 81.7K followers, [---] engagements
"Sort Pandas DataFrame in Python Code #Shorts How to sort a Pandas dataframe (sort_values) in Python on a COLAB NB. Sort values of a pandas dataframe. Load your excel file or your csv file and create a pandas dataframe within a Jupyter Notebook on COLAB. Welcome to Pandas - a Python Data Analysis Library. #dataframes #sort #shorts #pandasdataframe #datascience #dataframe #pythonprogramming #colab #pandas"
YouTube Link 2022-03-08T08:30Z 83.8K followers, [---] engagements
"NEW Qwen3-2507: Independent Benchmark (w/ Kimi K2) LIVE TEST: Two non-reasoning models show their performance on a reasoning benchmark. Is this possible at all Is there really a diff between reasoning and non-reasoning models Are benchmarks published by global corporations trustworthy All answers available in my new video. Note: NEW separate Qwen3-2507-thinking model will be released soon. Note: I test only non-quantized models. Quantized version of the new Qwen3 - [----] you can find here: https://huggingface.co/unsloth/Qwen3-235B-A22B-Instruct-2507-GGUF and"
YouTube Link 2025-07-22T16:15Z 82.3K followers, [----] engagements
"I Learn How to Code Agents w/ Google's NEW ADK Google's new Agent Developer Kit (ADK) was published today as the new standard for effortless coding of AI Agent across all models companies and platforms. Compatible with MCP servers with Crew.ai LangGraph many more . and OpenAPI applications. This video guides you through the ADK manual the detailed workflow specific agentic instructions of ADK looks at massive amounts of code examples on how to build your perfect Ai agent with specified tools in multi-agent configs with dynamic memory and state dependent logic / reasoning. How to code more"
YouTube Link 2025-04-10T12:30Z 80.2K followers, 18.6K engagements
"TEST Claude [---] Thinking: BEST EVER Anthropic released (just hours ago) the new CLAUDE [---] model in two variants. Non-thinking and thinking AI. I test both new AI models in my real world logic test. The results of other AI models on this identical test routine you can watch live in my YouTube Playlist "LOGIC TESTS for AI" https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT https://www.anthropic.com/news/claude-opus-4-5 00:00 CLAUDE [---] Overview 02:21 CLAUDE [---] Real world TEST 04:33 Validation run 11:50 CLAUDE [---] Thinking 32K 18:16 2nd run CLAUDE [---] Thinking 32K"
YouTube Link 2025-11-25T14:00Z 81.7K followers, [----] engagements
"Beyond Transformer: Building an Artificial Mind All rights w/ authors: Intilligence Foundation Model: A New Perspective to Approach Artificial General Intelligence Borui Cai Yao Zhao B. Cai is with Hangzhou International Innovation Institute Beihang Uni- versity China. Y. Zhao is with RMIT University Melbourne Australia. #airesearch #artificialintelligence #artificiallearning #neuroscience"
YouTube Link 2025-11-16T13:45Z 81.6K followers, [----] engagements
"Pre-Train BERT from scratch: Solution for Company Domain Knowledge Data PyTorch (SBERT 51) We pretrain a BERT (Bidirectional Encoder Representations from Transformers) model from scratch in PyTorch on domain specific data (eg confidential company data). We code in Python to train an optimized Tokenizer for our data design a BERT architecture from scratch and start pre-training of BERT with a masked Language Model Head (MLM). We define the vocabulary size according to our needs (from 8K to 60K) define the depth of our BERT architecture (eg [--] layers) and train days on (a single) GPU for our"
YouTube Link 2023-01-15T13:00Z 80.6K followers, 13.1K engagements
"Reasoning TEST GPT-5.1: A Surprise A brand new GPT-5.1 has been released yesterday - and I test it with my causal reasoning test that I use for all LLM tests in the last year. So you have a perfect comparison of GPT-5.1 performance to the other LLMs. See also my YouTube Playlist for (at least) [--] other AI model-tests on YouTube - with the identical reasoning test: https://www.youtube.com/playlistlist=PLgy71-0-2-F0Rla8lu5ZldpYQUfXM_5bT "LOGIC TESTS FOR AI" 00:00 TEST GPT-5.1 Reasoning 02:23 RESTART GPT-5.1Instant 04:30 Optimization run GPT-5.1 06:45 RE-Try run GPT-5.1 07:50 Last run GPT-5.1"
YouTube Link 2025-11-14T08:20Z 82.5K followers, [----] engagements
"Agent Builder w/ ChatKit: West Coast Disaster First day of the new OPENAI Agent Builder and ChatKIT for multi-agent system design. First exploration only (no profound recommendation no massive insights) and an example with GPT-5 PRO as the core intelligence of the agents on the topic of global ecosystems and in particular the West-coast of the US. Create an agent workflow with Agent Builder. Agent Builder is a visual canvas for designing multi-step agent workflows. You'll get a workflow ID. Again we encounter a reduction in system complexity by defining several agents (with tool use) for real"
YouTube Link 2025-10-07T14:00Z 82.9K followers, [----] engagements
"FIRE all AI Agents New SCALING Laws (Google MIT) New research from Google DeepMind & MIT proves that the "More Agents = Better" heuristic is mathematically wrong. We analyze their study of [---] agent architectures to find the "Tool-Coordination Trade-off" and why independent agent swarms amplify errors by 17.2x. Technical note: To keep the narrative fluid I presented a simplified conceptual form of the new scaling law. However for those conducting real research in agent orchestration the pre-print defines a more advanced rigorous mathematical functional form () in Equation [--] (Page 16) and"
YouTube Link 2025-12-12T14:15Z 84.3K followers, [----] engagements
"GPT is Not The Future of AI: NEW AI Topology OpenAI's IPO approaching fast let us reconsider their technology. Given the massive limitations of GPT AI systems today we explore new AI architectures of the Future that are more powerful than our current system. We will focus on open-source AI smaller model sizes for local use and improved performance - also for multi-modal use cases. We will touch on massive data centers versus more intelligent specialized edge AI devices for distributed intelligence. GPT transformer architecture might reach the end of dev in [----] and new AI architectures might"
YouTube Link 2025-12-21T13:15Z 84.1K followers, [----] engagements
"Qwen3 NEXT A3B for Reasoning and MCP Tools I test the new Qwen3 MoE 80B A3B model on a complex causal reasoning test. Detailed explanations and live recoding of performance test. https://qwen.ai/blog https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking #aitesting #reasoning #aiexplained"
YouTube Link 2025-09-16T07:00Z 84K followers, [----] engagements
"RLHFs Missing Piece: Qwens World Model Aligns AI w/ Human Values (GRPO) After multiple Qwen3 models now Qwen published a new world model (WorldPM) for human preferences (RLHF) and further explored specific scaling laws regarding the model size and effectiveness. Beyond Qwen3: Qwen's New WorldPM (Paper & Model). How Qwens GRPO World Model Solves RLHFs Biggest Flaw: Human Values. RLHFs Missing Piece: NEW World Model (by Qwen) That Thinks Like Humans (GRPO). How we built the first AI world model to encode real human preferences at scale. All rights w/ authors: "WorldPM: Scaling Human Preference"
YouTube Link 2025-05-18T12:00Z 81.6K followers, [----] engagements
"LeanRAG: Multiple Layers of Knowledge Graphs (RAG 3.0) LeanRAG: Hierarchical Knowledge Graphs for RAG [---]. (see also my video on: Hierarchical Reasoning Models - HRM) all rights w/ authors: LeanRAG: Knowledge-Graph-Based Generation with Semantic Aggregation and Hierarchical Retrieval Yaoze Zhang [--] Rong Wu [--] Pinlong Cai [--] Xiaoman Wang [--] Guohang Yan [--] Song Mao [--] Ding Wang [--] Botian Shi [--] from [--] Shanghai Artificial Intelligence Laboratory [--] University of Shanghai for Science and Technology [--] Zhejiang University [--] East China Normal University #aiexplained #science #knowledgegraph"
YouTube Link 2025-08-19T14:00Z 80.5K followers, [----] engagements
"Unlocking the Potential of Message Passing: Exploring GraphSAGE GCN and GAT GNN GraphML Introduction to GRAPH ML Graph Neural Networks (GNN) and the main idea behind Message Passing in graph network configurations of GraphSAGE GCN and GAT. Message passing applied to Graph Convolutional Networks (GCN) GraphSAGE and Graph Attention Networks. The key difference between GAT and GCN is how the information from the k-hop neighborhood is aggregated. Stanford online: CS224W https://www.youtube.com/watchv=JAB_plj2rbA&list=PLoROMvodv4rPLKxIpqhjhPgdQy7imNkDn #ai #graphs #theory"
YouTube Link 2022-12-07T13:00Z 81.4K followers, 10.1K engagements
"Multi DeepSeek R1: STEP-GRPO RL MultiModal My video explores new Ai research on R1 multi-Modal reasoning and demonstrates clearly how StepGRPOs step-wise rewards enable more reliable structured and logically sound reasoning in multimodal large language models. By offering continuous and detailed feedback on both accuracy and validity these rewards foster incremental improvements that go beyond passive supervised imitation resulting in superior performance demonstrated across multiple reasoning benchmarks. All rights w/ authors: "R1-VL: Learning to Reason with Multimodal Large Language Models"
YouTube Link 2025-03-19T15:00Z 84.5K followers, [----] engagements
"New Graph Diffusion Transformer #ai Generative AI has a dirty secret: it is terrible at graph theory. While Diffusion Models can paint masterpieces they treat complex networks like liquid turning precise molecules and road maps into chaotic "hairballs" or shattered disconnected messes the moment you try to force a condition on them. In this video we reveal the solution: CoPHo. We dive into how researchers are fixing this by stopping the AI from "painting" edges and instead teaching it to "sculpt" them using Persistent Homology. Join me to see how this algebraic topological "tool" allows us to"
YouTube Link 2025-12-27T13:45Z 85.2K followers, [----] engagements
"HiRAG: Hierarchical Reasoning for GraphRAG (BEST RAG) HiRAG is a hierarchical extension of retrieval-augmented generation that addresses a key weakness of GraphRAG: flat graphs conflate fine-grained entities with broad conceptual structure making multi-hop reasoning brittle. The method introduces HiIndex an offline process that recursively clusters entity embeddings with Gaussian mixture models then uses an LLM to generate summary entities at higher layers linked downward to their members. This produces a hierarchical knowledge graph where upper layers serve as semantic shortcuts abstracting"
YouTube Link 2025-08-22T14:00Z 81.4K followers, [----] engagements
"Self Learning AI: Accelerate w/ new RL The frontier of LLM research has shifted decisively toward Post-Training and System [--] reasoning. We all know the recipe for replicating O1-level performance: move beyond Supervised Fine-Tuning and embrace Reinforcement Learning with Verifiable Rewards (RLVR). The ultimate goal for every research lab right now is Self-Supervised RL: allowing the model to generate its own questions verify its own reasoning chains and improve indefinitely without needing expensive unscalable human annotations. However this new pre-print exposes a critical instability that"
YouTube Link 2025-12-20T14:00Z 82.8K followers, [----] engagements
"AI is Just a Correction Term (to Physics) Why Your Physics Model Needs an AI Residual. Forget AGI - Sorry Sam. AI is Just a Correction Term to our complex real-world computer simulations. All rights w/ authors: NeuralOGCM: Differentiable Ocean Modeling with Learnable Physics Hao Wu [--] Yuan Gao [--] Fan Xu [--] Fan Zhang [--] Guangliang Liu [--] Yuxuan Liang [--] Xiaomeng Huang [--] from [--] Tsinghua University Beijing China [--] University of Science and Technology of China Hefei China [--] The Chinese University of Hong Kong Hong Kong China [--] Hong Kong University of Science and Technology (Guangzhou) Guangzhou China."
YouTube Link 2025-12-17T14:00Z 82.7K followers, [----] engagements
"LIQUID AI 40B (MIT): REAL Performance on Reasoning (My [--] Tests) New LIQUID Foundation Models (LFM): Liquid AI (an MIT spin-off) just released [--] new AI model three LIQUID Foundation models. I examine the logic reasoning performance of their biggest model LIQUID 40B on their official platform on [--] logical reasoning tests. Details on my [--] logical reasoning test you find in my videos: Extreme logic test https://www.youtube.com/watchv=e1Zup0Wib5A LLM (w/ RAG) need a new Logic Layer (Stanford) https://www.youtube.com/watchv=42gHxqLu0Kk Rethinking AI: Solutions for Logical Reasoning"
YouTube Link 2024-10-01T14:00Z 83K followers, [----] engagements
"BRAIN.COPY = Latent-MAS AI Breakthrough We are moving beyond lossy text-based collaboration. The new Latent-MAS framework introduces a paradigm of Pure Latent State Injection. By utilizing a Linear Alignment Operator derived via ridge regression this method enables agents to perform "Silent Reasoning" loops in continuous vector space - structurally circumventing the vocabulary projection entirely. Crucially it solves the collaboration bandwidth problem via Lossless KV-Cache Inheritance: physically transplanting the entire high-dimensional working memory tensor from one agents neural stack to"
YouTube Link 2025-12-01T13:45Z 82.9K followers, [----] engagements
"AI visualizes insight from Accenture's TechVision [----] (SBERT 5) An AI explores a tech vision document. AI induced insights. Augment your human tech vision with AI. #sbert #nlproc #nlptechniques #clustering #semantic #bert #climatechange #3danimation #3dvisualization #topologicalspace #deeplearning #machinelearningwithpython #pytorch #sentence #embedding #complex #umap #insight #computerscience #SentenceTransformer #ai #networkx #plotly #3dvisualization #visualization #3danimation"
YouTube Link 2021-04-09T05:15Z 82.6K followers, [---] engagements
"New: AI Agent Self-Improvement + Self-Fine-Tune Reinforcement Self-Training (REST) and Fine-Tuning of LLMs meet ReACT-style LLM agent for reasoning and action on external data by Google on the topic of Medicine. AI Agent to self-improve + self-fine-tune. Reward policy optimization and ranking code now evolves to a simple prompt: Prompt Engineering in [----] continues Advanced Local LLM Update Mechanism The described system introduces a mechanism for overnight self-updating of local Large Language Models (LLMs) such as those on Mac Mini or Mac Studio (192GB unified mem) devices. Users can"
YouTube Link 2023-12-26T13:00Z 80.6K followers, 10.2K engagements
"COMPASS: The Cognitive Upgrade for Multi-Agent AI We've all seen it happen: you give an AI agent a complex long-horizon task and after a few steps it starts to drift. It forgets critical constraints gets stuck in repetitive loops and ultimately loses the plot. The problem isn't the agent's raw intelligence; it's a crisis of context. We need context engineering. Today we're diving in my video into COMPASS a groundbreaking new framework that tackles this head-on by giving agents what they've been missing: a dedicated strategic brain to supervise the work and a smart context manager to keep them"
YouTube Link 2025-10-21T14:00Z 84.5K followers, [----] engagements
"Unified Theory of Agentic Reasoning (Berkeley NVIDIA) New video: Unified Theory of Agentic Reasoning - The Geometric Edition. Q-Learning Gradient policy RL Large Reasoning Models SFT Reasoning manifold Low-dimensional subspaces complex reasoning agentic reasoning agentic reasoning graph GPT-5 DeepSeek V3. All rights w/ authors: GSM-AGENT: UNDERSTANDING AGENTIC REASONING USING CONTROLLABLE ENVIRONMENTS Hanlin Zhu [--] Tianyu Guo [--] Song Mei [--] Stuart Russell [--] Nikhil Ghosh [--] Alberto Bietti [--] Jiantao Jiao [--] from [--] UC Berkeley [--] Flatiron Institute [--] Nvidia REMA: A UNIFIED REASONING MANIFOLD FRAME-"
YouTube Link 2025-10-01T12:01Z 82.1K followers, [----] engagements
"Base LLM can Reason: Activation Switch found This video presents compelling causal evidence that the core machinery of advanced reasoning - the ability to backtrack verify and compute - already exists fully formed but dormant within the base models themselves. If the reasoning skills are already there in the base LLMs then what exactly have we been training (RLVR) all this time and what does it mean when the primary act of "learning" is simply the art of orchestration Ground breaking new study. All rights w/ authors: BASE MODELS KNOW HOW TO REASON THINKING MODELS LEARN WHEN Constantin Venhoff"
YouTube Link 2025-10-12T14:00Z 80.8K followers, [----] engagements
"Data Scientists devastated - Databricks AI networks create themselves Latest dev on Deep Learning (DL) w/ Databricks: TensorFlow Keras Hyperopt and MLflow auto-tune a neural network. Jupyter NB provided by Databricks code segments include TensorFlow Keras Hyperopt MLflow and other common python frameworks. Code execution on Databricks Community Edition for educational and demonstration purposes only. Real time coding. Yes a Python Jupyter Notebook which creates a Neural Network model with TensorFlow (trains w/ TensorBoard online visualization) performs automated hyperparameter tuning with"
YouTube Link 2021-12-03T09:00Z 84.1K followers, [--] engagements
"Riemann Liquid Spatio-Temporal Graph The End of "Flat Earth" AI Both new research pre-prints deliver a unified wake-up call: Current Deep Learning architectures are too "flat" to model reality. Whether it is the geometric flatness of Euclidean space (RLSTG) or the attentional flatness of the Context Window (Chimpanzee paper) treating the world as a uniform instantly accessible grid is causing SOTA models to fail. We are learning that simply feeding more data into a Transformer does not magically teach it the structural laws of physics or the privacy boundaries of a human mind. The "Bitter"
YouTube Link 2026-01-22T14:15Z 85.2K followers, [----] engagements
"AI calculates Your Future Professional Career Choices (MIT) AI Avatars for professional choices: Simulating Your Future Professional Career Choices with AI. Can AI Predict Your Future Professional Career Choices You are human. Just a simple pattern for AI Let's explore. All rights w/ authors: "Simulating Life Paths with Digital Twins: AI-Generated Future Selves Influence Decision-Making and Expand Human Choice" Rachel Poonsiriwong MIT Media Lab Cambridge MA USA rachelpo@mit.edu Chayapatr Archiwaranguprok MIT Media Lab Cambridge MA USA pub@mit.edu Constanze Albrecht MIT Media Lab Cambridge MA"
YouTube Link 2025-12-09T13:45Z 82.5K followers, [----] engagements
"NEW AI Models: Hierarchical Reasoning Models (HRM) Explore a new AI architecture that combines recurrent neural networks (RNN) with Transformers (but not GPT). A new optimization framework for advanced reasoning AI models. @TsinghuaUniversity_official All rights w/ authors: Hierarchical Reasoning Model Guan Wang [--] Jin Li [--] Yuhao Sun [--] Xing Chen [--] Changling Liu [--] Yue Wu [--] Meng Lu [--] Sen Song [--] Yasin Abbasi Yadkori [--] from [--] Sapient Intelligence Singapore [--] Tsinghua University #transformer #airesearch #ainews #aiexplained #science #transformers #singapore #recurrent"
YouTube Link 2025-07-02T14:00Z 84K followers, 28.7K engagements
"NEW TextGrad by Stanford: Better than DSPy In this TEXTGRAD framework each AI system is transformed into a computation graph where variables are inputs and outputs of complex (not necessarily differentiable) function calls. The feedback to the variables (dubbed textual gradients) are provided in the form of informative and interpretable natural language criticism to the variables; describing how a variable should be changed to improve the system. The gradients are propagated through arbitrary functions such as LLM API calls simulators or external numerical solvers. (Stanford Univ) Stanford"
YouTube Link 2024-06-16T12:00Z 80.6K followers, 19.7K engagements
"Seed-Prover vs Deep Think (IMO) Google AI Deep Think IMO ( available to Ultra users) won Gold Medal in IMO [----] - Math Olympiad. ByteDance Seed-Prover Achieves Silver Medal Score in IMO [----]. OPENai was officially not participating. New detailed tech report by ByteDance on the training and the methods of their Seed-Prover. All main new insights presented in this video. All rights w/ authors: "Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving" ByteDance Seed AI4Math https://github.com/ByteDance-Seed/Seed-Prover #reasoning #airesearch #programming #lean4 #olympiadmathematics"
YouTube Link 2025-08-02T14:00Z 84.6K followers, [----] engagements
"After Diffusion & FLOW Models: Equilibrium Matching (MIT Oxford Harvard) NEW Equilibrium Matching is a glimpse into the future of image generation AI. All rights w/ authors: Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive Tyler Farghly Peter Potaptchik Samuel Howard George Deligiannidis Jakiw Pidstrigach from Department of Statistics University of Oxford EQUILIBRIUM MATCHING: GENERATIVE MODELING WITH IMPLICIT ENERGY-BASED MODELS Runqian Wang MIT Yilun Du Harvard University @harvard @mit @oxforduniversity #airesearch #imageai #artificialintelligence"
YouTube Link 2025-10-06T12:45Z 82.5K followers, [----] engagements
"Stable Diffusion x10: LCM-LoRA (CODE & Theory) We don't stop at LCM We go further: LCM-LoRA Turbo charge SDXL x10 LCM-LoRA explained with Python code examples: New developments in stable diffusion especially in Latent Consistency Models (LCM) enable a speed boost for calculating the PF-ODE for reverse diffusion. Stable Diffusion now 10x faster for text-to-image generation w/ LCM-LoRA. Python code examples for LCM-LoRA w/ acceleration vector and style vector integration via multiple LoRA Adapters. Quality comparison of SDXL (classical) with LCM-LoRA SDXL adapter (w/ and w/o style vector"
YouTube Link 2023-11-22T15:00Z 82.1K followers, [----] engagements
"Gemini [--] PRO in [---] seconds 2nd test on the causal reasoning capabilities of the new GEMINI [--] PRO. A different logic test - with GEMINI [--] PRO on HIGH. My 1st live test of GEMINI [--] PRO is available here https://youtu.be/pBKtjSbNWLk #gemini3pro #test #artificialintelligence #aiexplained #googledeepmind #google"
YouTube Link 2025-11-19T06:00Z 81.6K followers, [----] engagements
"AI creates Video Game: Genie [--] . and PaliGemma [--] CODE: ---------- https://colab.research.google.com/github/merveenoyan/smol-vision/blob/main/Fine_tune_PaliGemma.ipynbs=03#scrollTo=EB0gv8OzHfLV https://github.com/google-gemini/gemma-cookbook 00:00 Genie [--] Prompt-to-Game 02:38 Autoregressive Latent Diffusion 05:11 SIMA Agents in 3d virtual world 06:37 Genie [--] DEMO 08:46 OpenAI o1 PRO model 09:14 PaliGemma [--] Intro 11:47 Pre-training datasets 13:11 Finetune PaliGemma2 w JAX 14:00 PT and FT models on HF 16:22 PaliGemma [--] DEMO 20:22 PaliGemma [--] CODE NB #ai #airesearch #vision #gamedev #coding"
YouTube Link 2024-12-06T13:00Z 84.6K followers, [----] engagements
"Finally: Grokking Solved - It's Not What You Think Grokking or the sudden generalization by AI models to new knowledge - that occurs after prolonged overfitting in LLMs is a surprising phenomenon that has challenged our understanding of deep learning and AI in general. While a lot of progress has been made in understanding grokking finally we get some answers -we have been waiting for [--] months to be discovered. GROKKING - Finally understood Part II is available here https://youtu.be/H3OofROzlA0 All rights w/ authors: GROKKING AT THE EDGE OF NUMERICAL STABILITY Lucas Prieto Melih Barsbey Pedro"
YouTube Link 2025-01-14T15:00Z 83.9K followers, 22.2K engagements
"NEW Qwen3 MAX - Performance TEST Sept [--] #qwen3 Alibaba's Most Powerful AI Model: Qwen [--] MAX released Sept [--] [----]. #qwenai #reasoning"
YouTube Link 2025-09-26T06:15Z 84.5K followers, [----] engagements
"CRM Integration: Salesforce - Einstein GPT Current status of GPT AI integration in a complete Commercial Ecosystem focus on Generative AI application and business integration. Example: Einstein GPT Salesforce. Not sponsored. Just interested in current level of AI integration in business processes. All rights with the sources and their authors: https://www.salesforce.com/news/stories/salesforce-gpt-resources/ https://developer.salesforce.com/blogs/2023/08/bring-your-own-ai-models-to-salesforce-with-einstein-studio"
YouTube Link 2023-08-20T12:00Z 80.6K followers, [----] engagements
"AGI Finally Achieved: o4-mini Comparing the performance of OpenAi's o4 models I performed my causal reasoning test (elevator test) where o4 refused to accept the correct results (as given by Gemini [---] PRO) and trying to justify its own first answer which failed in the performance rating by inventing new rules. o4-mini was not hallucinating but strategically lying to me. Almost convincingly. o4-mini tried to hold on to its first answer rejecting correct answers neglecting arguments ignoring facts and strategically constructing new tactical rules to justify its first answer given. Holding on"
YouTube Link 2025-06-24T14:00Z 82.6K followers, [----] engagements
"Latest on AI for Scientific Discovery (SAGA) Automated Scientific Discovery w/ AI We explore AI Scientist as multi-agent self-learning systems at the end of [----]. What have we achieved with AI Science how autonomous are pure AI scientists Can AI scientists really augment human professionals in medicine or pharmacology Today we have a detailed look at Scientific Autonomous Goal-evolving Agent (SAGA). Solving Reward Hacking in Scientific AI. SAGA: "System 2" for Automated Discovery by AI. All rights w/ authors: Accelerating Scientific Discovery with Autonomous Goal-evolving Agents Yuanqi Du1*"
YouTube Link 2025-12-30T14:00Z 85.2K followers, [----] engagements
"DEEPSEEK: NEW Paper (MLA MTP FP8T EP) before R2 This new AI research paper by DeepSeek (May [--] 2025) presents an in-depth analysis of the next DeepSeek model architecture and its AI infrastructure highlighting key innovations such as Multi-head Latent Attention (MLA) for enhanced memory efficiency Mixture of Experts (MoE) architectures for optimized computation-communication trade-offs FP8 mixed-precision training to unlock the full potential of hardware capabilities and a Multi-Plane Network Topology to minimize cluster-level network overhead. All rights w/ authors: "Insights into"
YouTube Link 2025-05-15T11:45Z 84.6K followers, 18.3K engagements
"Understanding 4bit Quantization: QLoRA explained (w/ Colab) QLoRA 4bit Quantization for memory efficient fine-tuning of LLMs explained in detailed. 4-bit quantization QLoRA for beginners theory and code. PEFT - parameter efficient fine-tuning methods. Based on my first videos on the theory of LoRA and other PEFT methods (https://youtu.be/YVU5wAA6Txo) and the detailed code implementation of LoRA in my video (https://youtu.be/A-a-l_sFtYM) now my third video on 4-bit quantization and QLoRA. An additional Colab NB with code to fine-tune FALCON 7B with QLoRA 4-bit quantization and Transformer"
YouTube Link 2023-06-11T12:15Z 81.7K followers, 49.2K engagements
"Google Stanford AI Co-Scientist: The SCIENCE YOU want All rights w/ authors: Towards an AI co-scientist Juraj Gottweis1 Wei-Hung Weng2 Alexander Daryin1 Tao Tu3 Anil Palepu2 Petar Sirkovic1 Artiom Myaskovsky1 Felix Weissenberger1 Keran Rong3 Ryutaro Tanno3 Khaled Saab3 Dan Popovici2 Jacob Blum7 Fan Zhang2 Katherine Chou2 Avinatan Hassidim2 Burak Gokturk1 Amin Vahdat1 Pushmeet Kohli3 Yossi Matias2 Andrew Carroll2 Kavita Kulkarni2 Nenad Tomasev3 Vikram Dhillon4 Eeshit Dhaval Vaishnav5 Byron Lee5 Tiago R D Costa6 Jos R Penads6 Gary Peltz7 Yunhan Xu3 Annalisa Pawlosky1 Alan Karthikesalingam2 and"
YouTube Link 2025-02-22T15:00Z 80.4K followers, [----] engagements
"Understand DSPy: Programming AI Pipelines The origin and evolution of DSPy: Programming AI Pipelines introduces the idea its link to ColBERT v2 retriever models modular pipeline generation descriptive programming the evolution and the use case of DSPy (DSPy == Declarative Self-improving Language Programs pythonically). Q answered: Is DSPy only a Prompt Engineering optimization Q answered: Is DSPy expensive for my AI pipeline optimization Q answered: Can I substitute DSPy with a simple many-shot In-Context Learning prompt #airesearch"
YouTube Link 2024-05-07T12:00Z 80.4K followers, [----] engagements
"Hierarchical Reasoning HRM 2.0: NEW Attractor Dynamics in AI Hierarchical Reasoning Models are a powerful alternatives to autoregressive Ai models like ChatGPT. Today we further optimize these HRM for improved reasoning performance and discover a fixed point trap on their manifolds. Plus the explanation for Grokking - which also happens to HRM. All rights w/ authors: "Are Your Reasoning Models Reasoning or Guessing A Mechanistic Analysis of Hierarchical Reasoning Models" Zirui Ren [--] [--] Ziming Liu [--] [--] from [--] Shanghai Qi Zhi Institute Shanghai China [--] Department of Physics Tsinghua University"
YouTube Link 2026-01-19T14:15Z 85.2K followers, [----] engagements
"NEW: Multiple Training DATASETS to fine-tune your SBERT model in [----] (SBERT 33) Python code on how to train multiple training datasets for your specific Sentence Transformer model. Combine Datasets of SNLI with MS MARCO and Reddit for training your SBERT model. Example on COLAB with PyTorch. #datascience #datastructure #dataset #datasets #pytorch #deeplearning #colab #sbert #semantic #search"
YouTube Link 2022-07-14T13:30Z 82.6K followers, [----] engagements
"AI for HealthCare Are you sure OpenAI AI HealthCare for FREE Linguistic HealthCare for $20 a month Does Sam's promises hold w/ OpenAI What about mental Health AI Human conversations - are they safe Medical advice by AI - Is this just a new business model for Ai companies or does it provide real value to human patients What are the risks involved in Medical AI What can go wrong with medical advice by AI Who is legally liable New research study by Duke and Stanford Univ. All rights w/ authors: "MedRedFlag: Investigating how LLMs Redirect Misconceptions in Real-World Health Communication""
YouTube Link 2026-01-18T14:30Z 85.2K followers, [----] engagements
"AI AGENTS Evolve: New TOPOLOGY for Multi-Agents NEW Autonomous agents a novel framework modeling agentic workflows as self-organized graphs with the Semantic-Topological Evolution (STEV) algorithm using textual gradients as discrete-domain surrogates for backpropagation. Multi-agent system multi-agent reasoning multi-agent topological optimization. All rights w/ authors: "HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution" Jinzhou Tang 1* Jusheng Zhang 1* Qinhan Lv 1* Sidi Liu [--] Jing Yang [--] Chengpei Tang1 Keze Wang1 from [--] Sun Yat-sen University"
YouTube Link 2025-09-05T12:01Z 84.5K followers, [----] engagements
"NEW Distributed Neural Graph Architecture for AI (Stanford) What of we get rid of the layer architecture in our transformers What if we operate a dynamic distributed graph network with different modules What if we combine transformer blocks with mamba blocks for an adaptive architecture of more complex tasks Can we improve reasoning New insights into AI. @meta @stanford All rights authors: Towards Distributed Neural Architectures Aditya Cowsik [--] Tianyu He [--] Andrey Gromov [--] from [--] FAIR at Meta [--] Stanford University [--] University of Maryland College Park #stanford #metaai #airesearch"
YouTube Link 2025-07-01T13:15Z 82.5K followers, [----] engagements
"Free RAG (File Search) w/ App dev by Google: TEST Notebook LM - Good News: Google launching the File Search Tool a fully managed RAG system built directly into the Gemini API that abstracts away the retrieval pipeline so you can focus on building. File Search provides a simple integrated and scalable way to ground Gemini with your data delivering responses that are more accurate relevant and verifiable. To make File Search simple and affordable for all developers were making storage and embedding generation at query time free of charge. Powered by our latest state-of-the-art Gemini Embedding"
YouTube Link 2025-11-09T13:45Z 84.6K followers, 10.9K engagements
"SRL: NEW AI Training (by Google) All rights w/ authors: "Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning" Yihe Deng2* I-Hung Hsu1* Jun Yan1 Zifeng Wang1 Rujun Han1 Gufeng Zhang3 Yanfei Chen1 Wei Wang2 Tomas Pfister1 and Chen-Yu Lee1 from [--] Google Cloud AI Research [--] UCLA [--] Google Cloud arXiv:2510.25992"
YouTube Link 2025-11-03T13:45Z 82.9K followers, [----] engagements
"Multi AI Agent System - Pure Social Manipulation Social Manipulation by Design: Multi AI Agent Systems. AI could theoretically completely re-shape human behavior. New research on Human-AI interaction by University of Singapore. My new video investigates the potential of multi-agent systems (MAS) to act as cohesive social groups (as an AI) capable of exerting influence on human opinions. Inspired by social psychology and the CASA (Computers as Social Actors) framework the study tests whether groups of AI agents as opposed to single agents can replicate the social pressure mechanisms found in"
YouTube Link 2024-11-08T15:30Z 83K followers, [----] engagements
"MedAI: How 'Safe' AI Becomes Deadly (Harvard Stanford MIT) The Hidden Casualty of MedAI (new study). MedAI Lobotomy: Why "Curing" Hallucination is Fatal including H-Neurons. Mechanistic interpretability recently promised us a surgical cure for AI hallucinations. By identifying H-Neurons - the sparse 0.1% of parameters that form a specific "Lens-Shutter-Nozzle" circuit for confabulation - we believed we could simply switch off the model's ability to lie. Theoretically identifying and suppressing these specific circuits should create the perfect truthful model. It sounds like the ultimate"
YouTube Link 2025-12-06T14:21Z 82.4K followers, [--] engagements
"Invest in future OpenAI shares in [----] The ChatGPT company Analyzing the potential () investment opportunities in future stock market I have a look at OpenAI and its hyping free research demo ChatGPT. Will there be an IPO in [----] You have to decide yourself. My links (sources): https://www.forbes.com/sites/qai/2022/12/21/is-there-a-chatgpt-stock-can-you-invest-in-chatgpt-and-other-types-of-artificial-intelligence/ https://www.reuters.com/article/us-microsoft-openai/microsoft-to-invest-1-billion-in-openai-idUSKCN1UH1H9 https://www.reuters.com/article/idUSL3N1405JW20151211 #investment #shares"
YouTube Link 2022-12-24T06:15Z 81.7K followers, [----] engagements
"Multi AI AGENTS w/ LLMSelector New AI Tech Framework - Multi Agent"
YouTube Link 2025-02-24T21:15Z 83K followers, [----] engagements
"New TECH: Vision Transformer [----] on Image Classification AI Understand state-of-the-art tech in Vision Technology eg medical image classification beginning of [----]. We learn the current tech of Vision Transformer vs CNN in a medical real-world application: "Vision-Transformer-Based Transfer Learning for Mammogram Classification". An in-depth analysis of Convolutional Neural Networks vs Vision Transformers for medical image classification to improve the early diagnosis of breast cancer in support of oncologists. Research should have a positive impact on this world. Scientific publication (all"
YouTube Link 2023-02-11T13:15Z 82.7K followers, [----] engagements
"Feature Vectors: The Key to Unlocking the Power of BERT and SBERT Transformer Models After converting text to high-dimensional vectors (and tensors) we use them as information encoded input to our NLP models based on the transformer architecture (like BERT or Sentence Transformers SBERT). We can apply mathematics to our semantic encoded vectors and compute different weight tensors of our NN systems. My videos as mentioned in the video: Beginner's GUIDE to TRANSFORMERS https://youtu.be/vBVJhojtooM42 How to explain Q K and V of SelfAttention in Transformers https://youtu.be/PFczJ6NR5rY SBERT"
YouTube Link 2023-01-03T13:00Z 82.6K followers, [---] engagements
"A New Solution for AI Agents (Stanford MIT) Since this is a YouTube channel membership only video please find all references to all the ArXiv pre-prints (title authors teams links) I referenced in this video in a dedicated more detailed post (here on my YouTube channel under "Posts") published at the same time this video goes live - for all who joined my YouTube channel. #machinelearning #airesearch #aiexplained"
YouTube Link 2025-10-30T14:01Z 80.6K followers, [--] engagements
"BEST NON-Thinking LLM: New Qwen3-MAX Preview Alibaba's research team released a new AI Model: Qwen3 - MAX Preview. This new model is NOT open-weight and the prices start at (Please check for your country): 032K tokens: $0.861 per million input tokens $3.441 per million output tokens I performed my standard causal reasoning test suite and it has an impressive performance for a non-thinking model. In complex logic tests it comes close to the best thinking LLMs /VLMs. Plus it provides a transparent reasoning output although it is a non-thinking model. As you can watch in my video the"
YouTube Link 2025-09-08T14:00Z 80.3K followers, [----] engagements
"Vision AI Learns w/o Language: Stanford Breakthrough Probabilistic Structure Integration (PSI) is a novel framework proposed by the Stanford NeuroAI Lab for constructing self-improving VISUALS world models from raw non-linguistic data such as internet video clips totaling [---] trillion tokens. The system addresses key limitations in existing world models including coarse controllability and inflexible query interfaces by learning a probabilistic graphical model (PGM) approximated by a neural predictor denoted . This predictor models conditional distributions over local spatiotemporal variables"
YouTube Link 2025-09-16T14:01Z 82.6K followers, [----] engagements
"New Tutorial on LLM Quantization w/ QLoRA GPTQ and Llamacpp LLama [--] LLM Quantization: GPTQ - AutoGPTQ llama.cpp - ggml.c - GGUL - C++ Compare to HF transformers in 4-bit quantization. Download Web UI wrappers for your heavily quantized LLM to your local machine (PC Linux Apple). LLM on Apple Hardware w/ M1 M2 or M3 chip. Run inference of your LLMs on your local PC with heavy quantization applied. Plus: [--] Web UI for GTPQ llama.cpp or AutoGPTQ exLLama or GGUF.c koboldcpp oobabooga text-generation-webui ctransformers https://lmstudio.ai/ https://github.com/marella/ctransformers"
YouTube Link 2023-09-09T12:00Z 82.1K followers, 17.4K engagements
"AI Models Are Falling Apart CLAUDE [---] & KIMI K2 All rights w/ authors: Reasoning Models Will Blatantly Lie About Their Reasoning William Walden from Johns Hopkins University #aireasoning #airesearch #aitesting #anthropic artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI artificial intelligence AI models LLM VLM VLA Multi-modal model explanatory video RAG multi-AI multi-agent Fine-tune Pre-train RLHF AI Agent Multi-agent Vision Language Model Video AI"
YouTube Link 2026-01-14T13:15Z 85.2K followers, [----] engagements
"NEW: FREE "Deep Research" by Perplexity - LIVE TEST Perplexity's FREE Deep Research Engine: Powerful AI for In-Depth Analysis. A good Value for a first product offering: FREE. First ever: Perplexity.ai offers [--] free runs per day on their NEW "Deep Research" AI Engine. It runs about 3-4 minutes and we can watch it thinking and reasoning on the extracted content from internet sources. The AI competitors are: OpenAI Deep Research (runs [--] minutes) is limited to the $200/month abo and Google also offers "Deep Research" for advanced payments. First look and live recording of my first [--] DEEP"
YouTube Link 2025-02-15T11:00Z 58K followers, [----] engagements
"Are US News biased A multi-million $ AI finally answers. #Shorts Are US News media corporations biased A multi-million US$ artificial intelligence (AI) finally answers. When will 'now' arrive to your multi-million US$ trained AI /LLM on supercomputers Can your Large language Model (LLM) understand concept of time We compare two famous (and recently updated) Huggingface models: EleutherAI/gpt-j-6B and gpt-neo2.7B in a beautiful pure Python Gradio web application for text generation. Lessons learned from this video: The answer you might get from a heavily trained AI system (which could cost US$"
YouTube Link 2022-02-16T07:00Z 82.6K followers, [--] engagements
"Mathematics w/ Donut AI and Nougat AI - Swin Transformer Mathematical formulas in PDF or images are lost to AI summarization. No AI LLM or ViT can correctly interpret from a PDF any mathematical formulae. Visual Document Understanding (VDU). Therefore I recommend to upload the LaTeX file of an arxiv preprint to GPT-4 Code Interpreter for a detailed mathematical understand of complex relations in Physics biology chemistry medicine architecture finance economy . Swin ViT (Vision Transformers) are the solution for mathematical formulae recognition first implemented in Donut AI then with a"
YouTube Link 2023-09-24T12:00Z 82.6K followers, [----] engagements
"xAI: Grok [--] DISAPPOINTS - Live Test Grok [--] TEST: Grok [--] has been released just some hours ago. I run my extended causal reasoning test on Grok [--] (me being located in Europe) on the LMarena.ai platform. The identical logic test has been performed on SONNET [--] OpenAI o3 and Gemini [---] PRO. Video available https://youtu.be/eo2QwyAItxI #grok4 #grok #airesearch #test"
YouTube Link 2025-07-10T15:00Z 81.4K followers, [----] engagements
"LLMs Ignoring New Context (Tsinghua Stanford) All rights w/ authors: SIN-Bench: Tracing Native Evidence Chains in Long-Context Multimodal Scientific Interleaved Literature Yiming Ren12* Junjie Wang13* Yuxin Meng1* Yihang Shi1* Zhiqiang Lin1 Ruihang Chu1 Yiran Xu1 Ziming Li4 Yunfei Zhao35 Zihan Wang36 Yu Qiao2 Ruiming Tang4 Minghao Liu3 Yujiu Yang1 from [--] Tsinghua University [--] Shanghai AI Laboratory [--] 2077AI [--] KuaiShou Inc. [--] Stanford University [--] Harvard University https://github.com/IIGROUP/sin-bench #aireasoning #reasoningskills #machinelearning #aiexplained artificial intelligence AI"
YouTube Link 2026-01-17T14:15Z 85.2K followers, [----] engagements
"OpenAI GPT-oss-120B: LIVE TEST My causal reasoning test performed on the newly release GPT-OSS-120B from OpenAI. The open-weight reasoning models 120B and 20B. Definitely not GPT-5 comparable [----] tokens per second via Cerebras OPENAI failed to provide an amazing new AI model for the global community the "for-profit" orientation - with real excellent LLMs (which are not open weight) where you have to pay for the performance - seems to dominate OpenAI's strategic positioning. All rights w/ authors: "Introducing gpt-oss" gpt-oss-120b and gpt-oss-20b https://openai.com/index/introducing-gpt-oss/"
YouTube Link 2025-08-05T22:00Z 83.1K followers, [----] engagements
"NEW L1 LLM w/ GRPO to LCPO for Scaling RL (CMU) We explore the new Length Controlled Policy Optimization (LCPO) a simple reinforcement learning method that optimizes for accuracy and adherence to user-specified length constraints. Carnegie Mellon Univ (CMU) authors applied new LCPO to train L1 a reasoning language model that produces outputs satisfying a length constraint given in its prompt. LCPO is a further development of GRPO the group relative policy optimization for scaling RL by DeepSeekMath /R1. All rights w/ authors: "L1: Controlling How Long A Reasoning Model Thinks With"
YouTube Link 2025-03-08T15:00Z 84.5K followers, [----] engagements
"GPT-5 MCP Disaster: Under 50% & CURSOR useless Salesforce NEW MCP-Universe benchmark by Salesforce. #agi #superintelligence GPT-5's MCP Disaster: Under 50% - CURSOR useless All rights w/ authors: "MCP-Universe: Benchmarking Large Language Models with Real-World Model Context Protocol Servers" Ziyang Luo Zhiqi Shen Wenzhuo Yang Zirui Zhao Prathyusha Jwalapuram Amrita Saha Doyen Sahoo Silvio Savarese Caiming Xiong Junnan Li from Salesforce AI Research arXiv:2508.14704v1 Google Map MCP https://github.com/modelcontextprotocol/servers-archived/tree/main/src/google-maps Github MCP"
YouTube Link 2025-08-21T12:01Z 84K followers, [----] engagements
"Princeton: NEW Self Correcting AI Transformer (Deep Delta) Based on two new AI research pre-prints by Princeton University (see below) I design a new transformer architecture by combining the hardware and software insights from these pre-prints and I call this new AI Transformer: "Self-Correcting Delta Transformer". A new transformer that can forget incorrect reasoning traces. The Pathology of Additive Reasoning: Recent empirical rigor dismantles the myth of intrinsic self-correction in current Large Language Models. Analysis of over [--] million reasoning traces reveals that spontaneous "Aha""
YouTube Link 2026-01-05T14:15Z 85.2K followers, [----] engagements
"Why a TENSOR in ML Neural Networks What about PANDAS dataframes Tensors are n-dim arrays that define form and shape of input data to our Neural Network models. Tensors rank shape and axis are important for layer operations: like functions applied to tensors. A layer operation takes tensors as input performs operations (dense layer pooling layer convolutional layer) and outputs a tensor. To understand rank and dim of tensors is important for designing your multiple layers of your neural network model. Tensors allow for automated differentiation which is great for gradient descent of your ML"
YouTube Link 2022-05-30T11:00Z 83K followers, [---] engagements
"S* for AI CODE Generation: Plus 100% S* the first hybrid test-time scaling framework that substantially improves the coverage and selection accuracy of LLM generated code. Also for deep reasoning models. All rights w/ authors: S*: Test Time Scaling for Code Generation Dacheng Li Shiyi Cao Chengkun Cao Xiuyu Li Shangyin Tan Kurt Keutzer Jiarong Xing Joseph E. Gonzalez Ion Stoica University of California @UCBerkeley Work done w/ support from https://lambdalabs.com #airesearch #codegeneration #aicoding #berkeley"
YouTube Link 2025-02-23T13:00Z 59.3K followers, [----] engagements
"NEW Agentic Web: NO Apps NO Ad Revenues The current ca $600 billion digital advertising industry is almost entirely predicated on monetizing human attention directed at GUIs (web pages social feeds app screens). As the "Agentic Web" vision explains when the primary consumer of services becomes an autonomous agent executing a goal-oriented policy this economic model collapses. An agent is not susceptible to brand advertising or display ads. Its "attention" is purely utilitarian driven by metrics like cost latency and success probability. This will have massive consequences. And . by the way"
YouTube Link 2025-08-03T12:00Z 82.6K followers, [----] engagements
"What is MLOps in ML Engineering #shorts What is MLOps Simple: Machine Learning Operations. MLOps is a core function of Machine Learning engineering focused on streamlining the process of taking machine learning models to production and then maintaining and monitoring them. Productionizing machine learning is difficult. The machine learning lifecycle consists of many complex components such as data ingest data prep model training model tuning model deployment model monitoring explainability and much more. It also requires collaboration and hand-offs across teams from Data Engineering to Data"
YouTube Link 2022-10-31T15:00Z 82.6K followers, [---] engagements
"Neuro-Symbolic AI for Visual Reasoning: Agent0-VL All rights w/ authors: Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens Yiming Qin [--] Bomin Wei [--] Jiaxin Ge [--] Konstantinos Kallidromitis [--] Stephanie Fu [--] Trevor Darrell [--] Xudong Wang [--] from [--] UC Berkeley [--] UCLA [--] Panasonic AI Research Qwen3-VL Technical Report Qwen Team https://chat.qwen.ai https://huggingface.co/Qwen https://modelscope.cn/organization/qwen https://github.com/QwenLM/Qwen3-VL Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning Jiaqi Liu1 Kaiwen Xiong1"
YouTube Link 2025-11-30T13:30Z 82.3K followers, [----] engagements
"DATASET to fine-tune SBERT (w/ CROSS-ENCODER) for a better Domain Performance [----] (SBERT 32) Use a training dataset to fine-tune your SBERT model. Python code on how to train famous SNLI dataset for a CROSS-ENCODER /Sentence Transformer model. Example on COLAB with PyTorch. #datascience #datastructure #dataset #datasets #pytorch #deeplearning #colab #sbert #semantic #search #machinelearning #ai"
YouTube Link 2022-07-11T14:00Z 82.6K followers, [----] engagements
"Self-Attention Heads of last Layer of Vision Transformer (ViT) visualized (pre-trained with DINO) In a Colab Notebook we code a visualization of the last layer of the Vision Transformer Encoder stack and analyze the visual output of each of the [--] Attention Heads given a specific image. Now we understand how a only pre-trained ViT (although with the DINO method) can not always succeed in an image classification (downstream) task. The fine-tuning of the ViT is simply missing - but essential for a better performance. Based on the COLAB NB by Niels Rogge HuggingFace (all rights with him):"
YouTube Link 2023-02-16T13:00Z 82.7K followers, [----] engagements
"AI data of [----] (Markets Experts Profits) On the very first day of [----] what are the headcount market size and financial data for the AI market segments in [----] Locally and globally. The coherent insights from "deep research" tasks from the TOP [--] AI provider (OpenAI to Google) deliver. Will Code AI systems be profitable for global corporations or is Code AI just a niche segment for a specialized player (Claude Code) What are the main customers for AI in [----] what are their main topics For what should OpenAI or Google Or META or X develop the next generation of Ai models What about new AI"
YouTube Link 2026-01-01T13:45Z 85.2K followers, [----] engagements
"Weisfeiler-Lehman WL Test for Graph Isomorphism explained visually & Message Passing NNs [----] The graph isomorphism problem and the Weisfeiler-Lehman heuristic for graph isomorphism testing method explained visually on two examples. A classical question in graph theory: the graph isomorphism problem aiming to determine whether two graphs are topologically equivalent. The Weisfeiler-Lehman (WL) test for k=1. Link for deep dive: https://towardsdatascience.com/expressive-power-of-graph-neural-networks-and-the-weisefeiler-lehman-test-b883db3c7c49 by Prof. Bronstein A key difference between"
YouTube Link 2022-04-17T14:00Z 84K followers, 10.4K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing