# ![@neural_avb Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::1754194661084983296.png) @neural_avb AVB

AVB posts on X about context engineering, deep, rl, youtube the most. They currently have [-----] followers and [---] posts still getting attention that total [-------] engagements in the last [--] hours.

### Engagements: [-------] [#](/creator/twitter::1754194661084983296/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1754194661084983296/c:line/m:interactions.svg)

- [--] Week [-------] +264%
- [--] Month [-------] +1,686%
- [--] Months [---------] +138%
- [--] Year [---------] +655,605%

### Mentions: [--] [#](/creator/twitter::1754194661084983296/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1754194661084983296/c:line/m:posts_active.svg)

- [--] Months [---] -45%
- [--] Year [---] +17,700%

### Followers: [-----] [#](/creator/twitter::1754194661084983296/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1754194661084983296/c:line/m:followers.svg)

- [--] Week [-----] +14%
- [--] Month [-----] +23%
- [--] Months [-----] +84%

### CreatorRank: [------] [#](/creator/twitter::1754194661084983296/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1754194661084983296/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  14.01% [social networks](/list/social-networks)  8.92% [finance](/list/finance)  1.27% [countries](/list/countries)  0.64% [celebrities](/list/celebrities)  0.64% [musicians](/list/musicians)  0.64% [stocks](/list/stocks)  0.64% [products](/list/products)  0.64%

**Social topic influence**
[context engineering](/topic/context-engineering) 7.64%, [deep](/topic/deep) #2598, [rl](/topic/rl) #623, [youtube](/topic/youtube) 6.37%, [if you](/topic/if-you) 5.73%, [agentic](/topic/agentic) 5.73%, [ai](/topic/ai) #1063, [how to](/topic/how-to) 5.1%, [llm](/topic/llm) #582, [code](/topic/code) 5.1%

**Top accounts mentioned or mentioned by**
[@tdatascience](/creator/undefined) [@lateinteraction](/creator/undefined) [@maiyangai](/creator/undefined) [@hesamation](/creator/undefined) [@atulitgaur](/creator/undefined) [@neuralavb](/creator/undefined) [@viditchess](/creator/undefined) [@thdxr](/creator/undefined) [@applywiseai](/creator/undefined) [@farhanhelmycode](/creator/undefined) [@dwarkeshsp](/creator/undefined) [@darioamodei](/creator/undefined) [@sama](/creator/undefined) [@awareoffreedom](/creator/undefined) [@willccbb](/creator/undefined) [@lightshifthq](/creator/undefined) [@sundarpichai](/creator/undefined) [@grok](/creator/undefined) [@andthatto](/creator/undefined) [@crosstensor](/creator/undefined)
### Top Social Posts
Top posts by engagements in the last [--] hours

"Some of yall dont consider this as one of the coolest ML projects of all time so I cant take you seriously"  
[X Link](https://x.com/neural_avb/status/2015377464093262299)  2026-01-25T10:53Z [----] followers, 23K engagements


"About to record the Vision Language Model video Theory intuition implementation code breakdown - this video is going to be the culmination of one whole month of underground work. 🤞🏼🤞🏼"  
[X Link](https://x.com/neural_avb/status/2015846558957437291)  2026-01-26T17:57Z [----] followers, [----] engagements


"Studying this new VLM paper by Tencent: Youtu-VL - Paper argues VLM trainings suffer from an information bottleneck because they treat images only as a background context for generating text - The core contribution of Youtu-VL is the transition from a discriminative vision-as-input paradigm to a generative vision-as-target paradigm. - They expand the LLM's vocabulary to include discrete "visual tokens" (using a learned codebook) - The training objective is updated to include a visual reconstruction loss (L_vision + lamb * L_text) - L_text : usual next token prediction loss - L_image: The"  
[X Link](https://x.com/neural_avb/status/2016872797860163915)  2026-01-29T13:55Z [----] followers, [---] engagements


"@Hesamation The term AGI means different things to different people so every AGI claim is debatable at best and sensationalist at worst. Until we all agree on what the definition is and a process to measure it"  
[X Link](https://x.com/neural_avb/status/2016915022300270994)  2026-01-29T16:43Z [----] followers, [----] engagements


"I sleep like a baby every night moment before disaster https://t.co/rolJkPDVwh moment before disaster https://t.co/rolJkPDVwh"  
[X Link](https://x.com/neural_avb/status/2017202961974464623)  2026-01-30T11:47Z [----] followers, [---] engagements


"After months of prep and testing my audience's patience my VLM video is finally here Goes deep into building VIT-based image encoder pipelines training Q-Former models finetuning LLMs on vision tasks and more. One of my best vids I am super proud of it. Give it a watch"  
[X Link](https://x.com/neural_avb/status/2017482009128661492)  2026-01-31T06:16Z [----] followers, [----] engagements


"VLM training repo is now public 🚨 Contains the code to: - prepare img-text datasets using img2dataset - create VIT + Q-Former image backbone - training Q-Former from a distillbert - train huggingface text LMs on vision tasks - multimodal inference Repo link below 👇🏼"  
[X Link](https://x.com/neural_avb/status/2017502903674777627)  2026-01-31T07:39Z [----] followers, [----] engagements


"This evening I'm just chilling with some classic continual learning literature from [----]. It's crazy to revisit some of these papers in [----]. https://twitter.com/i/web/status/2017620163064172564 https://twitter.com/i/web/status/2017620163064172564"  
[X Link](https://x.com/neural_avb/status/2017620163064172564)  2026-01-31T15:25Z [----] followers, [---] engagements


"Do yall wanna watch a finetuning video about @UnslothAI If yes please leave a like/comment. Genuinely trying to gauge interest levels here. 🙏🏼"  
[X Link](https://x.com/neural_avb/status/2017856143436214504)  2026-02-01T07:02Z [----] followers, [----] engagements


"Why are NDTV editors such assholes"  
[X Link](https://x.com/neural_avb/status/2017866142661153070)  2026-02-01T07:42Z [----] followers, [---] engagements


"Bro how can [--] screenshots carry THIS much knowledge Do yourself a favor and read the full github post Dr Karpathy made last night. Things they did to train nanochat. 🙏🏼"  
[X Link](https://x.com/neural_avb/status/2017908628108513535)  2026-02-01T10:31Z [----] followers, 71.7K engagements


"Our landing page got a huge upgrade. Post launch features pricing updates social proof and copywriting tweaks. Yesterday we saw the largest traffic since launch date so big thanks to all the new people trying out DMs always open for feedback http://paperbreakdown.com http://paperbreakdown.com"  
[X Link](https://x.com/neural_avb/status/2017986792222519460)  2026-02-01T15:41Z [----] followers, [---] engagements


"Deepseek recently published DeepSeek-OCR [--]. There is a cool genius-level intuition behind this paper. "What if you train the image encoder to REORDER the image tokens before processing" - Most VLMs extract patches from an image and present to the LM in a fixed ordering - i.e. left to right top to bottom. Like a printer-like raster scan. - But - human vision does not scan images like a printer. We use "foveal fixations" - locally sharp focus points that move based on semantic importance. - That's what DeepSeek-OCR [--] tries to capture. - OCR-2 replaces the CLIP module with a Qwen-2 based LLM"  
[X Link](https://x.com/neural_avb/status/2018595653027549690)  2026-02-03T08:01Z [----] followers, 24.1K engagements


"This must be a copy pasta tweet from last year coz brother no one is using DeepSeek for anything atm USA has ChatGpt China has Deepseek What does India have USA has ChatGpt China has Deepseek What does India have"  
[X Link](https://x.com/neural_avb/status/2019110600702591067)  2026-02-04T18:07Z [----] followers, [---] engagements


"@ThePrimeagen Reminds me of this Black Mirror episode"  
[X Link](https://x.com/neural_avb/status/2019149127930855888)  2026-02-04T20:40Z [----] followers, [---] engagements


"@TDataScience Article link: https://towardsdatascience.com/how-to-build-your-own-custom-llm-memory-layer-from-scratch/ https://towardsdatascience.com/how-to-build-your-own-custom-llm-memory-layer-from-scratch/"  
[X Link](https://x.com/neural_avb/status/2019291697990324708)  2026-02-05T06:07Z [----] followers, [----] engagements


"Yes. Or you can go to find your paper click Study with AI no manual upload. one click chat. ask any question AI has context/tools to read actual PDF AI highlights relevant sections directly in PDF Free Dont believe Check for yourself. http://paperbreakdown.com How to read a research paper: First of all you do it old school You sit your fucking ass down You turn off your adhd which for most people is just a low attention span in disguise You read the paper word by word and start from the abstract You keep a notebook with http://paperbreakdown.com How to read a research paper: First of all you"  
[X Link](https://x.com/neural_avb/status/2019320771185950924)  2026-02-05T08:02Z [----] followers, [----] engagements


"@atulit_gaur I see this as a product placement opportunity"  
[X Link](https://x.com/neural_avb/status/2019328953442545880)  2026-02-05T08:35Z [----] followers, [----] engagements


"Whats the culture feeling right now"  
[X Link](https://x.com/neural_avb/status/2019336587805618588)  2026-02-05T09:05Z [----] followers, [---] engagements


"@real_deep_ml Loving this arc of gamifying things"  
[X Link](https://x.com/neural_avb/status/2019414436537405951)  2026-02-05T14:14Z [----] followers, [--] engagements


"When tf am I gonna break out of this 4k jail bro"  
[X Link](https://x.com/neural_avb/status/2019420047698280479)  2026-02-05T14:37Z [----] followers, [---] engagements


"If youre into hardcore deep learning and getting shit done at a low gritty technical level Please go give my man a follow right now👇🏾 @neural_avb Stuck on 2.6k for long Hope we both do get outta there @neural_avb Stuck on 2.6k for long Hope we both do get outta there"  
[X Link](https://x.com/neural_avb/status/2019424204631011362)  2026-02-05T14:53Z [----] followers, [----] engagements


"Lets go TDS has bumped this article to their Deep Dives section Thank you to all who are reading it 🙏🏼 https://towardsdatascience.com/how-to-build-your-own-custom-llm-memory-layer-from-scratch/ New article out on @TDataScience implementing custom LLM memory systems from scratch using DSPy. Go give it a read Code is open source. https://t.co/bAxQeCFKOi https://towardsdatascience.com/how-to-build-your-own-custom-llm-memory-layer-from-scratch/ New article out on @TDataScience implementing custom LLM memory systems from scratch using DSPy. Go give it a read Code is open source."  
[X Link](https://x.com/neural_avb/status/2019446060486173044)  2026-02-05T16:20Z [----] followers, [---] engagements


"Gemini [--] Flash is the most under-appreciated model on the market. Not all LLM app is about "chat". There are a lot of other heavy NLP use-cases - summarization tagging structured object query gen etc. Only green model in Artificial Index intelligence vs speed"  
[X Link](https://x.com/neural_avb/status/2019476368174772304)  2026-02-05T18:21Z [----] followers, [---] engagements


"This is like when Kendrick dropped Meet the Grahams [--] mins after Drake dropped Family Matters. GPT-5.3-Codex is here *Best coding performance (57% SWE-Bench Pro 76% TerminalBench [---] 64% OSWorld). *Mid-task steerability and live updates during tasks. *Faster Less than half the tokens of 5.2-Codex for same tasks and 25% faster per token *Good computer use. GPT-5.3-Codex is here *Best coding performance (57% SWE-Bench Pro 76% TerminalBench [---] 64% OSWorld). *Mid-task steerability and live updates during tasks. *Faster Less than half the tokens of 5.2-Codex for same tasks and 25% faster per"  
[X Link](https://x.com/neural_avb/status/2019480035942605023)  2026-02-05T18:35Z [----] followers, [---] engagements


"Introducing Agentic Progress to Paper Breakdown 👏🏼 Auto detects your progress towards core study goals of a research paper Ensures you never miss key concepts from a paper even if you try Straight up trains you to ask better questions Free. Check the demo below"  
[X Link](https://x.com/neural_avb/status/2019638307538497643)  2026-02-06T05:04Z [----] followers, [----] engagements


"Woah anthropic is giving away 50$ credit to all Claude Code users trying Opus [---] If you have a subscription go to this page https://claude.ai/settings/usage https://claude.ai/settings/usage"  
[X Link](https://x.com/neural_avb/status/2019649868256670182)  2026-02-06T05:50Z [----] followers, [---] engagements


"@atulit_gaur Bro what about part where your supervisor says the paper's not good enough"  
[X Link](https://x.com/neural_avb/status/2019653635031109692)  2026-02-06T06:05Z [----] followers, [----] engagements


"Alphaxiv is great. Ill never say anything bad about them. Go use it. Whats extra in Paper Breakdown: - agentic paper search - better customization around recommendations - automatic paper goals - quizzes - diagram generation - auto-highlight relevant sections into the pdf - large tier models. Multi paper chat soon. Check landing page its all there. http://paperbreakdown.com http://paperbreakdown.com"  
[X Link](https://x.com/neural_avb/status/2019666208778514714)  2026-02-06T06:55Z [----] followers, [---] engagements


"Last year when I had 1/4th subscribers I made this lovely Mixture of Experts LM video. Covers dense MOEs sparse MOEs top-K routing noisy routing expert capacity switch transformers load balancing losses Deepseek V3 etc. Full literature review + code examples"  
[X Link](https://x.com/neural_avb/status/2019682778753953851)  2026-02-06T08:01Z [----] followers, [----] engagements


"@viditchess This just broke my heart man"  
[X Link](https://x.com/neural_avb/status/2019767004056666451)  2026-02-06T13:35Z [----] followers, [----] engagements


"A new AGI benchmark is here. LLMs vs Balatro TIL there is a repo to make LLMs play Balatro: And benchmark: https://github.com/coder/balatrobenchtab=readme-ov-file https://github.com/coder/balatrollm https://github.com/coder/balatrobenchtab=readme-ov-file https://github.com/coder/balatrollm"  
[X Link](https://x.com/neural_avb/status/2019793408697471018)  2026-02-06T15:20Z [----] followers, [---] engagements


"calling it now [----] world cup winning squad 👏🏼👏🏼👏🏼👏🏼"  
[X Link](https://x.com/neural_avb/status/2019800618421166527)  2026-02-06T15:49Z [----] followers, [---] engagements


"@leerob I dont get it - what does a sustaining engineering model even mean"  
[X Link](https://x.com/neural_avb/status/2019842049919578217)  2026-02-06T18:34Z [----] followers, [----] engagements


"@GoogleOSS I used to use it heavily for multiple months. I got a TON of shit done with Gemini CLI. Only since last month I have been using a opencode (I still use Gemini [--] pro as the driver). The only reason I switched was lack of built-in worktrees to "rewind"/undo edits"  
[X Link](https://x.com/neural_avb/status/2019843370240340229)  2026-02-06T18:39Z [----] followers, [---] engagements


"So many tribalistic takes recently. Nagging suspicion that certain big accounts are being paid to be shills for OpenAI/Anthropic. I'll urge yall to try both [---] & [---] over the next [--] months. Pick a fav based on your experience"  
[X Link](https://x.com/neural_avb/status/2019979807824703985)  2026-02-07T03:41Z [----] followers, [---] engagements


"Finally studying the "Generative modeling via Drifting". The latest attempt to retire diffusion models for image gen. There is a bit of prior knowledge you need to understand what's happening here. Diffusion/Flow generation works by iterative refinment. Drifting Model is trying to eliminate inference time steps. Drifting Model tries to train the generator network so that it learns to map the noise directly to the final image in a single jump. The training becomes more complex and the reward is generation is super fast. Image-gen remains one of the most terminology heavy areas in DL papers. So"  
[X Link](https://x.com/neural_avb/status/2020039770102251703)  2026-02-07T07:39Z [----] followers, [---] engagements


"Took the Sarvam AI vs Human voice test. Definitely great voices but your boy can smell AI from a mile away"  
[X Link](https://x.com/neural_avb/status/2020075521451975087)  2026-02-07T10:01Z [----] followers, [----] engagements


"@Maverick_Quant This is a really old project that I haven't opened in years. I dont even know if it'll open in current versions of Unity. I do see the repo on my git. I'll check if I can run it this week make corrections and make it open source. 🤞🏼"  
[X Link](https://x.com/neural_avb/status/2020259381670301890)  2026-02-07T22:12Z [----] followers, [---] engagements


"Life update: got a kind comment that made my day 🙏🏼"  
[X Link](https://x.com/neural_avb/status/2020380874324484433)  2026-02-08T06:15Z [----] followers, [----] engagements


"Man I didnt know this would blow up like that. Its kind of poetic justice coz this is a special project. Ill open source the project soon still need to test if it works on latest Unity. Little story time: I started learning Unity in [----] on a whim. Had a background in RL so decided to marry them. Working in this project blew my mind so I decided to start a YouTube channel in [----]. My first [--] videos were all devlogs explaining this project. I had zero video making skills a bad microphone didnt have this X account and those videos barely got [---] views. I felt deflated was supposed to quit. My"  
[X Link](https://x.com/neural_avb/status/2020440465451987027)  2026-02-08T10:11Z [----] followers, [----] engagements


"@thdxr Whats this ai dot com What did I miss"  
[X Link](https://x.com/neural_avb/status/2020880590795591743)  2026-02-09T15:20Z [----] followers, [----] engagements


"@navneet_rabdiya In most cases its fine if synthetic data is a bit messy given its pros. As long as its mostly correct and in the right format 15-20% is fine imo. Generally theres a DPO later on that irons out model responses based on human curated preference data"  
[X Link](https://x.com/neural_avb/status/2020990289381556293)  2026-02-09T22:36Z [----] followers, [--] engagements


"transformers.js v4 is in preview In an era where every website has slow ass AI features imagine how cool yours will be if you run small offline browser models that can automate shit in 100ms. HF and ONNX team cooked this. Theyre reporting 4x speed up on BERT"  
[X Link](https://x.com/neural_avb/status/2021179509517508737)  2026-02-10T11:08Z [----] followers, [----] engagements


"Reading the latest "lets replace RL with better context engineering" paper. This one holds a lot of water. - They actually replaced every major GRPO module with a text-based equivalent. - This is NOT a prompt optimization paper like GEPA. It is more about long-term memory. - ReAct loops to collect experiences/rollouts evaluate trajectories (semantic advantage) and create a library of distilled experiential knowledge. - At query time these distilled experiences are concated and injected directly into the LLM's context. Basically the LLM maintains an ever-evolving list "best-practices" and"  
[X Link](https://x.com/neural_avb/status/2018229742391157016)  2026-02-02T07:47Z [----] followers, [----] engagements


"I love this compartmentalization for Context Engineering methods"  
[X Link](https://x.com/neural_avb/status/2019816947794522591)  2026-02-06T16:54Z [----] followers, [----] engagements


"@ApplyWiseAi pyodide executes the python code the LLM writes. deno orchestrates it and isolates the pyodide environment for safety"  
[X Link](https://x.com/neural_avb/status/2021121265978003797)  2026-02-10T07:17Z [----] followers, [---] engagements


"@farhanhelmycode You can use Docker but itll be slower. You can do a lot with RestrictedPython but its not really a sandbox but thats where id start if I wanted to do something quick. Im choosing deno mainly coz DSPy implementation uses it. There must be alternatives im not aware of"  
[X Link](https://x.com/neural_avb/status/2021496089598361812)  2026-02-11T08:06Z [----] followers, [---] engagements


"Paper Breakdown will reach [---] users in a few hours [----] papers will be studied by end of this week I just can't believe my eyes. 🙏🏼"  
[X Link](https://x.com/neural_avb/status/2021622574904692778)  2026-02-11T16:29Z [----] followers, [----] engagements


"He said [---] lines but the first [--] minutes are imports disclaimers and a cli args parser. My goat didn't sacrifice basic UX to reduce code length. 🙏🏼 New art project. Train and inference GPT in [---] lines of pure dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further. https://t.co/HmiRrQugnP New art project. Train and inference GPT in [---] lines of pure dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify"  
[X Link](https://x.com/neural_avb/status/2021696642500419766)  2026-02-11T21:23Z [----] followers, [----] engagements


"@M_Raj03 @JayShah I hope the Indian public is kinder when our teams lose esp the womens team. The amount of scum written in online circles when they were having a tough patch was shameful. Im so glad that they shut everyone up. I havent felt this emotional about a cricket match in ages"  
[X Link](https://x.com/anyuser/status/1985261945692606662)  2025-11-03T08:25Z [----] followers, 15.3K engagements


"I recently trained a Vision Language Model (VLM) combining a pretrained VIT and a text-only SLM. Wrote a Q-Former adapter from scratch trained on 50k images (4 hrs [--] cents) made a 3b1b-inspired video covering everything Repo: Video in comments. https://github.com/avbiswas/vlm https://github.com/avbiswas/vlm"  
[X Link](https://x.com/neural_avb/status/2020777087045648757)  2026-02-09T08:29Z [----] followers, 19.6K engagements


"My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context). Asked it to find BLEU scores from Attention paper & explain MHA from these articles RLM controlled by the new Minimax [---] Minor prompt changes were needed from the RLM paper. Spends first [--] iterations understanding data format works through errors until it locates the Attention article from the mess. Like a human would use a Jupyter Notebook. Launches subagent on only AIAYN article This subagent launches [--] more subagents to"  
[X Link](https://x.com/neural_avb/status/2022248417003778183)  2026-02-13T09:56Z [----] followers, 36.1K engagements


"@lateinteraction I have a newfound appreciation for RLMs since I began trying to implement it. 🫡"  
[X Link](https://x.com/neural_avb/status/2022263843599323424)  2026-02-13T10:57Z [----] followers, [----] engagements


"@dwarkesh_sp @DarioAmodei This used to how Lex Fridman podcasts used to feel like before he "left" Deep Learning"  
[X Link](https://x.com/neural_avb/status/2022372197004906891)  2026-02-13T18:08Z [----] followers, [----] engagements


"Dev work for the RLM tutorial is almost done entering youtuber mode tonight New: subagents always hit KV caches - drastic cost savings track global token usage (see vid) hierarchical logs - these trajectories are super educational really helps to tune prompts"  
[X Link](https://x.com/neural_avb/status/2022622636904452388)  2026-02-14T10:43Z [----] followers, [----] engagements


"Gemini-2.0-Flash you will be sooo missed. I just generated some instruction-tuning data on 5000+ passages. A dspy module that inputs text - extracts a list of Q&A. Took [--] min costed [--] cents. Insane intelligence for price/latency. Perfect for simple synthetic data gen"  
[X Link](https://x.com/neural_avb/status/2020976891990270304)  2026-02-09T21:43Z [----] followers, [----] engagements


"Article has been updated to include more info about the new prime lab training flow with hosted training 😎"  
[X Link](https://x.com/neural_avb/status/2021695054776070482)  2026-02-11T21:17Z [----] followers, [---] engagements


"The subagent architecture is the same but the difference is how these subagents consume the context. In Claude code they use tool calls and they output answers by literally generating the text. In RLMs they interact using a python repl (think interactive Jupyter kernel) and they can return answers not just by generating it token by token but also storing it inside a python variable through symbolic functions (slicing finding) etc. https://twitter.com/i/web/status/2022281603930890617 https://twitter.com/i/web/status/2022281603930890617"  
[X Link](https://x.com/neural_avb/status/2022281603930890617)  2026-02-13T12:08Z [----] followers, [---] engagements


"@sama I have decided to try Claude Code for [--] month and then Codex for [--] month"  
[X Link](https://x.com/neural_avb/status/2021990297497112991)  2026-02-12T16:50Z [----] followers, [---] engagements


"@rasbt Hey man how do you keep track of the architecture specifications of so many models"  
[X Link](https://x.com/neural_avb/status/2021995275410452908)  2026-02-12T17:10Z [----] followers, [---] engagements


"New article out on @TDataScience implementing custom LLM memory systems from scratch using DSPy. Go give it a read Code is open source"  
[X Link](https://x.com/neural_avb/status/2019291572341514308)  2026-02-05T06:06Z [----] followers, 25.6K engagements


"Hands down the craziest RL project of my life. Competitive self playing agents learning to shoot and evade (jump/dash/duck) bullets. Trained from scratch in [--] hours on my mac. This was in [----] before vibe coding"  
[X Link](https://x.com/neural_avb/status/2020051428899975271)  2026-02-07T08:26Z [----] followers, 358.6K engagements


"I am working on an RLM tutorial. We will implement it from first principles using deno & pyodide (inspired by dspy.RLM). Lot of uber cool things like subagents communicating via global variables file systems etc. I dont think yall get how cool this architecture is"  
[X Link](https://x.com/neural_avb/status/2021105667571130494)  2026-02-10T06:15Z [----] followers, 36.3K engagements


"http://x.com/i/article/2021479918731919364 http://x.com/i/article/2021479918731919364"  
[X Link](https://x.com/neural_avb/status/2021594051686879472)  2026-02-11T14:35Z [----] followers, 34.1K engagements


"Don't just bookmark this. Spend an hour actually reading and understanding what they're cooking here. https://t.co/Bm4U8qEp82 https://t.co/Bm4U8qEp82"  
[X Link](https://x.com/neural_avb/status/2022214457792098473)  2026-02-13T07:41Z [----] followers, 55.5K engagements


"@AwareOfFreedom Yeah will soon - maybe early next week You can use dspy.RLM if you wanna play with RLMs yourself. My version will be optimized to be an educational guide on how to implement them from scratch + a companion piece for the upcoming YouTube video"  
[X Link](https://x.com/neural_avb/status/2022259977386389837)  2026-02-13T10:42Z [----] followers, [---] engagements


"JUST THINK RECURSIVE TREE TRAVERSAL BUT WITH LLMS HAVING A JUPYTER NOTEBOOK. RECEIVES CONTEXT AS A PYTHON VAR. WRITES CODE TO SPLIT CONTEXT INTO SUBTASKS AND ASSIGNS TO SUBAGENTS. SYMBOLICALLY CONSTRUCT ANSWER FROM THE SUBAGENT OUTS. RETURNS TO PARENT AS A VARIABLE OR A STRING. @lateinteraction I have a newfound appreciation for RLMs since I began trying to implement it. 🫡 @lateinteraction I have a newfound appreciation for RLMs since I began trying to implement it. 🫡"  
[X Link](https://x.com/neural_avb/status/2022274552517288420)  2026-02-13T11:40Z [----] followers, [----] engagements


"@MiniMax_AI Awesome man I got to try Minimax [---] on an RLM task. It just flies Insane intelligence+speed for that price. https://x.com/neural_avb/status/2022248417003778183s=20 My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context). Asked it to find BLEU scores from Attention paper & explain MHA from these articles RLM https://t.co/wXKfyZVpzU https://x.com/neural_avb/status/2022248417003778183s=20 My RLM finally went recursive Looking at these logs is way too addictive please send"  
[X Link](https://x.com/neural_avb/status/2022325025794400658)  2026-02-13T15:00Z [----] followers, [----] engagements


"@MaiYangAI @dwarkesh_sp Lol vibe coding founder 😂"  
[X Link](https://x.com/neural_avb/status/2022681660563804349)  2026-02-14T14:37Z [----] followers, [---] engagements


"Dude I think I just got unsloth quantized finetuning working locally on a macbook This is gonna significantly speed up my upcoming tutorial. I can prototype quickly on my Mac and run actual experiments on a rented cuda instance. Thanks to the UnslothAI team for pointing me in the right direction - they have a PR open that should make things "official" for Mac users soon. iiuc the hangup has been coz both triton and bitsandbytes are heavily cuda/nvidia-dependent. This pr doesn't make that problem go away instead it detects mps and switches to using MLX instead. So the usual FastModel becomes"  
[X Link](https://x.com/neural_avb/status/2018342740376142138)  2026-02-02T15:16Z [----] followers, [----] engagements


"The branch: The PR: https://github.com/unslothai/unsloth-zoo/pulls https://github.com/JINO-ROHIT/unsloth/tree/mlx-support https://github.com/unslothai/unsloth-zoo/pulls https://github.com/JINO-ROHIT/unsloth/tree/mlx-support"  
[X Link](https://x.com/neural_avb/status/2018342942516375592)  2026-02-02T15:17Z [----] followers, [---] engagements


"icydk I'm making a comprehensive guide to LLM finetuning with unsloth/trl + special attention to dataset preparation. We will be taking a small lousy pretrained base model and going through all the training stages one by one"  
[X Link](https://x.com/neural_avb/status/2020153222556459336)  2026-02-07T15:10Z [----] followers, [----] engagements


"Diagram gen is one of Paper Breakdown's best chat features Asked how to architect the RLM paper with deno & pyodide Agent explores the paper Explains the algorithm Generates technical diagram in seconds. Diagram gen is rule-based. Its GUARANTEED to be clean in [--] shot"  
[X Link](https://x.com/neural_avb/status/2021286258216186272)  2026-02-10T18:12Z [----] followers, [----] engagements


"Very bullish on Prime Intellect. Literally EVERY deep learning tutorial I have done since July is trained on their GPU instances. I once randomly commented on one of Wills tweets asking if he would hypothetically help me get credits so I can make more content on YT. He responded promptly we had a chat and he was just super chill and kind gave me way more credits than I had asked for. I am training models now and making educational content on YT without worrying about paying cloud bills. Theyre just a cool company with a serious mission statement. .@johannes_hage and @willccbb sat down with"  
[X Link](https://x.com/neural_avb/status/2021343607605850409)  2026-02-10T22:00Z [----] followers, 23.8K engagements


"Reading a Prime Intellect blogpost feels like you're reading a rebel manifesto. I love it ❤🔥"  
[X Link](https://x.com/neural_avb/status/2021472061210886449)  2026-02-11T06:31Z [----] followers, [----] engagements


"Writing this guide made me pause and appreciate how thoughtfully the entire Prime Intellect stack is designed. Yeah they aren't kidding about "frontier AI lab for all". https://t.co/YtrLOVW5nJ https://t.co/YtrLOVW5nJ"  
[X Link](https://x.com/neural_avb/status/2021594946281611774)  2026-02-11T14:39Z [----] followers, [---] engagements


"@LightShiftHQ Its a small coding model for debugging. Hopefully results will be better when I finish. What do you mean by framework"  
[X Link](https://x.com/neural_avb/status/2021985695993409759)  2026-02-12T16:32Z [----] followers, [---] engagements


"@thdxr Wait I dont get it. I've not used Codex much what is the difference Thought they're both just great coding models"  
[X Link](https://x.com/neural_avb/status/2021998775250932046)  2026-02-12T17:24Z [----] followers, [---] engagements


"In the last [--] hour: - MiniMax launched their open source Opus killer - Gemini Deep Think broke Arc AGI [--] - OpenAI released Codex-Spark What"  
[X Link](https://x.com/neural_avb/status/2022021286034325904)  2026-02-12T18:53Z [----] followers, [----] engagements


"@ki_ler @sundarpichai @grok https://arcprize.org/blog/arc-agi-2-technical-report https://arcprize.org/blog/arc-agi-2-technical-report"  
[X Link](https://x.com/neural_avb/status/2022040086364434685)  2026-02-12T20:08Z [----] followers, [--] engagements


"@andthatto @lateinteraction Ooh nice Someone should start running RLM benchmarks"  
[X Link](https://x.com/neural_avb/status/2022264371108618668)  2026-02-13T10:59Z [----] followers, [---] engagements


"@crosstensor RLM paper appendix section has some prompts"  
[X Link](https://x.com/neural_avb/status/2022334172363432323)  2026-02-13T15:36Z [----] followers, [--] engagements


"@sundarpichai I remember when ARC-AGI-2 was supposed to be the impossible roadblock"  
[X Link](https://x.com/neural_avb/status/2022003771526852644)  2026-02-12T17:44Z [----] followers, [----] engagements


"Ngl Dwarkesh is just having a legendary run The @DarioAmodei interview. 0:00:00 - What exactly are we scaling 0:12:36 - Is diffusion cope 0:29:42 - Is continual learning necessary 0:46:20 - If AGI is imminent why not buy more compute 0:58:49 - How will AI labs actually make profit 1:31:19 - Will regulations destroy https://t.co/qsFoNMAy2t The @DarioAmodei interview. 0:00:00 - What exactly are we scaling 0:12:36 - Is diffusion cope 0:29:42 - Is continual learning necessary 0:46:20 - If AGI is imminent why not buy more compute 0:58:49 - How will AI labs actually make profit 1:31:19 - Will"  
[X Link](https://x.com/neural_avb/status/2022386224661696933)  2026-02-13T19:03Z [----] followers, 172.8K engagements


"My next guest prediction model tells me he is bringing @sama next. Ngl Dwarkesh is just having a legendary run https://t.co/MivK7Txiht Ngl Dwarkesh is just having a legendary run https://t.co/MivK7Txiht"  
[X Link](https://x.com/neural_avb/status/2022578815294435690)  2026-02-14T07:49Z [----] followers, [----] engagements


"Yesterday's update has more info: https://x.com/neural_avb/status/2022248417003778183s=20 My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context). Asked it to find BLEU scores from Attention paper & explain MHA from these articles RLM https://t.co/wXKfyZVpzU https://x.com/neural_avb/status/2022248417003778183s=20 My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context). Asked"  
[X Link](https://x.com/neural_avb/status/2022622896502510069)  2026-02-14T10:44Z [----] followers, [---] engagements


"@kalomaze Been playing with it since release to drive an RLM tutorial Im working on. Its actually so good. great speed and really cheap for its value. My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context). Asked it to find BLEU scores from Attention paper & explain MHA from these articles RLM https://t.co/wXKfyZVpzU My RLM finally went recursive Looking at these logs is way too addictive please send help. Notes: Sent it [--] long wikipedia articles about deep learning (2M context)."  
[X Link](https://x.com/neural_avb/status/2022248991300448761)  2026-02-13T09:58Z [----] followers, [----] engagements


"noooooooooooooooooooooooooooooooooooo please dont make this rando my most liked tweet ever Ngl Dwarkesh is just having a legendary run https://t.co/MivK7Txiht Ngl Dwarkesh is just having a legendary run https://t.co/MivK7Txiht"  
[X Link](https://x.com/neural_avb/status/2022750377121321079)  2026-02-14T19:10Z [----] followers, [---] engagements


"@lateinteraction I am curious what you think is the #1 property/detail about RLMs that people dont immediately grasp Like something so fundamental that once you tell them they go "aha I see""  
[X Link](https://x.com/neural_avb/status/2022755018722676883)  2026-02-14T19:29Z [----] followers, [---] engagements


"I have started making slides and illustrations for the Context Engineering tutorial My goal is to upload by Thursday. 🤞🏽 Btw I already have a DSPy tutorial and a pretty comprehensive theoretical overview of RAGs. They are optional pre-requisites"  
[X Link](https://x.com/anyuser/status/1947330319389962574)  2025-07-21T16:18Z [----] followers, [----] engagements


"The best dspy tutorial is actually shipped with dspy. It is the inspect_history function. Tells you everything you need to know about how your signatures are converted into system/user prompts. And how different modules do multi-step/single-step LM calls"  
[X Link](https://x.com/anyuser/status/1947698525535670387)  2025-07-22T16:41Z [----] followers, [----] engagements


"Banger Youtuber"  
[X Link](https://x.com/neural_avb/status/1948052847293088175)  2025-07-23T16:09Z [----] followers, [----] engagements


"Off twitter these days coz I am cooking the most potent Context Engineering hands-on tutorial I can cook. Editing at 50%"  
[X Link](https://x.com/anyuser/status/1948217364824182896)  2025-07-24T03:03Z [----] followers, [----] engagements


"You spend [--] days underground and now theres a new transformer paper from Deepmind on adaptive tokenization"  
[X Link](https://x.com/neural_avb/status/1948420061930103063)  2025-07-24T16:28Z [----] followers, 11.3K engagements


"Full [--] hour [--] minute hands-on course on Context Engineering with DSPy is out now on my YouTube channel I am super proud of the final outcome I hope yall enjoy it. Spread the word if you do"  
[X Link](https://x.com/anyuser/status/1948641466105856390)  2025-07-25T07:08Z [----] followers, 18.1K engagements


"Promise fulfilled Took me [--] days of craziness but my Context Engineering with DSPy tutorial is out now on YouTube. My next goal: getting my sleep schedule back on track. @Hesamation Mental note/Social Promise: I need to make a proper video on Context Engineering. @Hesamation Mental note/Social Promise: I need to make a proper video on Context Engineering"  
[X Link](https://x.com/anyuser/status/1948754725186285735)  2025-07-25T14:38Z [----] followers, 12K engagements


"This is an awesome article. The best part is their note to build around the KV cache. If your system prompt remains consistent your tools remain constant and you always append to conversation json you will hit the KV cache often. Cutting down cost and latency"  
[X Link](https://x.com/anyuser/status/1948774323679830382)  2025-07-25T15:56Z [----] followers, 37.6K engagements


"Dear algorithm please help me find more folks interested in Computer Vision"  
[X Link](https://x.com/neural_avb/status/1949338655748669887)  2025-07-27T05:18Z [----] followers, [----] engagements


"I made a silly task manager terminal app with Sqlite3 and dspy.ReAct. I don't know what I'll use it for but it's cool"  
[X Link](https://x.com/anyuser/status/1949670145103700195)  2025-07-28T03:16Z [----] followers, [----] engagements


"The context engineering with DSPy repo is now public Contains all the code examples I used for my YouTube tutorial. Link below 👇🏽"  
[X Link](https://x.com/anyuser/status/1949736874483781633)  2025-07-28T07:41Z [----] followers, 21K engagements


"You post this and go to sleep. You wake up there's [--] stars. Only on X. The context engineering with DSPy repo is now public Contains all the code examples I used for my YouTube tutorial. Link below 👇🏽 https://t.co/boOGyFHMFR The context engineering with DSPy repo is now public Contains all the code examples I used for my YouTube tutorial. Link below 👇🏽 https://t.co/boOGyFHMFR"  
[X Link](https://x.com/neural_avb/status/1949879639431655692)  2025-07-28T17:08Z [----] followers, [----] engagements


"Whatever happened to Graph Neural Networks"  
[X Link](https://x.com/neural_avb/status/1950581907931959604)  2025-07-30T15:39Z [----] followers, 12.3K engagements


"This video shows how to write transformer architectures explaining each line of code. (Link below)"  
[X Link](https://x.com/anyuser/status/1951307563158454484)  2025-08-01T15:42Z [----] followers, 11.4K engagements


"Anyone looking to get started with PyTorch - there is this really cool article by Sebastian Raschka"  
[X Link](https://x.com/anyuser/status/1951308569124483359)  2025-08-01T15:46Z [----] followers, 59K engagements


"https://sebastianraschka.com/teaching/pytorch-1h/ https://sebastianraschka.com/teaching/pytorch-1h/"  
[X Link](https://x.com/neural_avb/status/1951308615412842888)  2025-08-01T15:46Z [----] followers, [----] engagements


"Anthropic continues doing pretty awesome LLM interpretability/observability papers"  
[X Link](https://x.com/anyuser/status/1951496993487913311)  2025-08-02T04:15Z [----] followers, [----] engagements


"I will be making the most densely packed [--] minute PyTorch tutorial I can design. I already have a lot of advanced tutorials (playlist below) but this one will be about the entire tour - basics to advanced greatest hits. Ill use this article as my guiding light. Anyone looking to get started with PyTorch - there is this really cool article by Sebastian Raschka. https://t.co/MSV0SMZnAg Anyone looking to get started with PyTorch - there is this really cool article by Sebastian Raschka. https://t.co/MSV0SMZnAg"  
[X Link](https://x.com/neural_avb/status/1951542647018934436)  2025-08-02T07:16Z [----] followers, [----] engagements


"8 standard components of every good RAG system I have built: [--]. When preprocessing I generate relevant metadata for each chunk. In the DB I dont just store the text but additional things like what question does this chunk answer and header text. [--]. Always use query rewriting. User queries are flawed and almost always fails in vanilla retrieval. [--]. Embedding + Bm25 is almost always better. Bm-25 is keyword based cosine similarity is semantic. I always go hybrid. [--]. I also use HYDE a lot. In Hyde you generate a hypothetical answer and search this new answer in your DB instead of the query. 5."  
[X Link](https://x.com/neural_avb/status/1951665531133612527)  2025-08-02T15:24Z [----] followers, [----] engagements


"I am super excited about the upcoming PyTorch tutorial. It is going to be about 10-15 minutes - and it's going to cover A LOT of ground"  
[X Link](https://x.com/neural_avb/status/1952258289292480614)  2025-08-04T06:40Z [----] followers, 21K engagements


"I made a video back in [----] that goes over the history of Multimodal Deep Learning till that point. I feel these are essential reads to understand what is up with modern VLMs. A list of research papers/topics I learned (video link in comment). Contrastive Learning - LSTM-CNNs - CLIP - ImageBind Masked Visual LLMs - VisualBERT - VilBERT Unified Architectures - BLIP - VL-T5 Generative LMs - Frozen - Flamingo - PaLM-E"  
[X Link](https://x.com/neural_avb/status/1952782699539664988)  2025-08-05T17:24Z [----] followers, [----] engagements


"YAY I opened my chrome app and guess whose Context Engineering/DSPY article showed up 🎉"  
[X Link](https://x.com/anyuser/status/1953533819715272832)  2025-08-07T19:08Z [----] followers, [----] engagements


"My context engineering + dspy course repo crossed [---] stars Links to the repo blog post and youtube video below"  
[X Link](https://x.com/anyuser/status/1953837134474817870)  2025-08-08T15:14Z [----] followers, [----] engagements


"I started planning my PyTorch video thinking Ill make it just [--] minutes long. It wont be. ☠"  
[X Link](https://x.com/anyuser/status/1955334389769916799)  2025-08-12T18:23Z [----] followers, 59.2K engagements


"I have 90GB more disk space because I ran: rm /.cache/huggingface/* 😇"  
[X Link](https://x.com/anyuser/status/1955518261724057991)  2025-08-13T06:34Z [----] followers, [----] engagements


"My last two blogs have pulled some truly banger page views. The Context Engineering with DSPy one has been up for [--] days The GRPO one for about 30"  
[X Link](https://x.com/anyuser/status/1955683160731332802)  2025-08-13T17:29Z [----] followers, [----] engagements


"PyTorch tutorial is almost done [--] minutes long. [--] chapters. I'll release sometime in the next [--] hours Super excited for this I started planning my PyTorch video thinking Ill make it just [--] minutes long. It wont be. ☠ https://t.co/U46E4NPAUt I started planning my PyTorch video thinking Ill make it just [--] minutes long. It wont be. ☠ https://t.co/U46E4NPAUt"  
[X Link](https://x.com/anyuser/status/1956356796056002610)  2025-08-15T14:06Z [----] followers, 55.5K engagements


"Massive PyTorch tutorial video is now published on my channel Goes from tensor programming basics automatic differentiation to the implementation ideas behind some of the most useful networks"  
[X Link](https://x.com/anyuser/status/1956727478040612911)  2025-08-16T14:39Z [----] followers, 40.2K engagements


"We got to 3K 🙏🏼🎉"  
[X Link](https://x.com/anyuser/status/1956890523458760894)  2025-08-17T01:27Z [----] followers, [----] engagements


"A list of all-timer must-read papers that published between [----] and [----]. Probably the golden era of novelty-based research in Deep Learning. CV: 2014: VGG Inception GANs 2015: ResNets UNet Batch Norm DenseNet YOLO 2017: Mask R-CNN MobileNet NLP: 2014: Word2Vec Seq2Seq GRU 2015: Bahnadau Attention 2017: Attention is All You Need DRL: 2013/4: Atari DQN 2015: Dueling DQNs PER DDPG TRPO 2016: A3C/A2C 2017: HER PPO Also WaveNet Adam GCN GraphSage and so so many GAN variants too"  
[X Link](https://x.com/anyuser/status/1956953378749136957)  2025-08-17T05:36Z [----] followers, [----] engagements


"My last [--] videos have been so overwhelming. Both in terms of the effort on my part but also the positive feedback and kind comments I've received. 🙏 Feels like I'm on a run right now. Next stop: RL"  
[X Link](https://x.com/anyuser/status/1957109722072387790)  2025-08-17T15:58Z [----] followers, [----] engagements


"FYI the DSPy boyz have been saying this for [--] years guys"  
[X Link](https://x.com/anyuser/status/1957114742096498806)  2025-08-17T16:18Z [----] followers, [----] engagements


"In case you missed it I made a [--] minute PyTorch video this week. (Link below)"  
[X Link](https://x.com/anyuser/status/1957814595147489783)  2025-08-19T14:39Z [----] followers, [----] engagements


"ALWAYS SHOOT YOUR SHOT Massive thanks to @willccbb and @PrimeIntellect My upcoming YouTube videos will be using their GPUs for training DL models. My laptop doesnt need to be a room heater anymore. ICE ICE BABY For the past [--] years of doing crazy DL videos on Youtube I've either had to train models locally (which is slow) or pay for GPUs out of pocket. This collaboration literally lets me freely explore and run larger experiments without needing to monitor my billing page every day. Once again big thanks to Will and team. More updates on the exact projects is coming soon"  
[X Link](https://x.com/neural_avb/status/1958181279775502651)  2025-08-20T14:56Z [----] followers, [----] engagements


"Next video: the taxonomy of Deep RL. We will learn the major classes of RL algos the big algorithms from each class and understand how to pick the best tool for your env. I am brainstorming hard to find the simplest words to describe the beauty in all this. (This image is from a GitHub repo btw)"  
[X Link](https://x.com/neural_avb/status/1959867378088460567)  2025-08-25T06:36Z [----] followers, [----] engagements


"My biggest flex as a human is: I can give myself a haircut. When I went to the US as a grad student I thought paying 15$ for a haircut was too steep. So I taught myself to cut my own hair. This requires art bravery and cleaning skills"  
[X Link](https://x.com/anyuser/status/1960249660527783981)  2025-08-26T07:55Z [----] followers, [----] engagements


"My finetuning tutorial is about to hit 100K views Maybe tomorrow HOLYYYY this is a first for me"  
[X Link](https://x.com/anyuser/status/1960727864774680618)  2025-08-27T15:35Z [----] followers, 23.8K engagements


"Dear algorithm only show this to people who are into Machine Learning and also knows what this picture is"  
[X Link](https://x.com/anyuser/status/1960895206548517032)  2025-08-28T02:40Z [----] followers, [----] engagements


"YOOOO LETS GOOOO My LLM fine tuning video hit 100K BAY BAY First time seeing 6-digit views If you are one of them thanks a lot (Video link in comments)"  
[X Link](https://x.com/anyuser/status/1960956059977122117)  2025-08-28T06:42Z [----] followers, [----] engagements


"The Reinforcement Learning video is likely dropping next week. My notes are ready. I am pretty happy with it. I've distilled RL into a set of [--] very specific & simple questions. I will be going deep with each of them individually. This is the RL video I always wanted to make Next video: the taxonomy of Deep RL. We will learn the major classes of RL algos the big algorithms from each class and understand how to pick the best tool for your env. I am brainstorming hard to find the simplest words to describe the beauty in all this. (This image is https://t.co/a6fFjUq7RZ Next video: the taxonomy"  
[X Link](https://x.com/anyuser/status/1962503970020073503)  2025-09-01T13:13Z [----] followers, [----] engagements


"Saw a post on Reddit where someone jailbroke puch ai. This shit is so hard to fix man esp if you want to have super low latency during responses"  
[X Link](https://x.com/anyuser/status/1963856940385071131)  2025-09-05T06:49Z [----] followers, [----] engagements


"Off Twitter coz I'm busy grinding out some generational RL content"  
[X Link](https://x.com/anyuser/status/1965035656142557245)  2025-09-08T12:53Z [----] followers, [----] engagements


"35 minute Deep RL video is likely releasing tomorrow There is a section where I show off my Elden Ring boss fighting skills (to explain exploration vs exploitation). In other words it's a must watch"  
[X Link](https://x.com/neural_avb/status/1966507534669504947)  2025-09-12T14:21Z [----] followers, [----] engagements


"New 30+ minute RL video just went live on my YouTube This is a video I've always wanted to make. It's how my thesis professor explained RL to me when I was in uni. We distill RL down to [--] simple questions. Each one explains an essential concept - model-based vs model-free value vs policy exploration strats TD vs MC etc. Really proud of this one"  
[X Link](https://x.com/neural_avb/status/1967205310445256802)  2025-09-14T12:34Z [----] followers, [----] engagements


"If you truly want to learn RL ditch the readymade gym environments. Make a custom environment on your own. Youll understand how to structure rewards observations random initialization states etc. also how to debug and render. This is the most practical skill you can get in RL. You can literally implement any env you want - a game a robotic control task a text thingy whatever you feel interested in. For the training algorithm start with SB3 modules as youre building the env. This will teach you the important hyperparameters and save you the headache of debugging your env and your training at"  
[X Link](https://x.com/neural_avb/status/1967469439403872342)  2025-09-15T06:04Z [----] followers, 23.9K engagements


"Yesterday we hit [--] videos Special thanks to my girlfriend for pushing me to start this journey [--] years back. I am glad I listened and stayed with it More to come :)"  
[X Link](https://x.com/neural_avb/status/1967505002622288017)  2025-09-15T08:25Z [----] followers, [----] engagements


"The four videos I want to work on next: - Multi agent navigation RL (video below) - Deep dive into prompt optimizers - Training Multimodal LMs/VLMs - Ways to deploy DL models in prod"  
[X Link](https://x.com/anyuser/status/1967844737081348386)  2025-09-16T06:55Z [----] followers, [----] engagements


"Planning a new DSPy video around integrating long term memory into agents. Basically implement the core features of Mem0 from scratch. When you think about it a good memory system has sooo many moving parts. Its the ultimate Context Engineering problem. Indexing filtering keyword/embedding search tool usage query pre/post processing etc with the added knowledge of time"  
[X Link](https://x.com/anyuser/status/1976180369046348225)  2025-10-09T06:58Z [----] followers, 28.2K engagements


"The agentic memory project is coming along nicely Using qdrant to host a vector db locally doing some tool calling some bm-25 and building the core blocks of mem0 with DSPy. I'm having a lot of fun with this one. Tutorial soon"  
[X Link](https://x.com/anyuser/status/1978343658593148997)  2025-10-15T06:14Z [----] followers, 15.5K engagements


"Shot a couple hours of footage for the upcoming from-scratch agentic memory tutorial. The final cut will probably run [--] minutes. Covers: [--]. The mem0 api (to understand what we will eventually be building from scratch) [--]. DSPy basics (signatures and modules) [--]. Extracting memory/factoids from conversations [--]. Indexing embeddings vector db [--]. Hybrid searches [--]. Tool calling for memory upkeep [--]. Making end to end chatbots that integrates with session-scoped user-scoped and long-term memory [--]. Extras and more for Patreon members Video and repo coming next week"  
[X Link](https://x.com/anyuser/status/1979081656821580253)  2025-10-17T07:06Z [----] followers, [----] engagements


"In the end we all have a favourite algorithm"  
[X Link](https://x.com/anyuser/status/1979448784011641225)  2025-10-18T07:25Z [----] followers, 86.1K engagements


"@VBkramnik This is not the time. It actually never was"  
[X Link](https://x.com/neural_avb/status/1980337493913358825)  2025-10-20T18:17Z [----] followers, 99.8K engagements


"50 minute tutorial on building agentic memory systems is dropping soon on my YT. I am 90% there. We will use DSPy to create the core features of mem0 from scratch. Very excited This project just made me happy from the inside"  
[X Link](https://x.com/anyuser/status/1980628343113372044)  2025-10-21T13:32Z [----] followers, 10.6K engagements


"New video on agentic memory systems is out currently out on my channel. This one discusses the challenges of long term memory as a context engineering problem explains the Mem0 api and the proceeds to code the core features of Mem0 from scratch. We use DSPy to extract memories from conversations vector databases contextual searching and filtering tool calls for maintaining a fresh and up-to-date memory state. Also shows how to generate evaluation datasets and evaluate memory systems. This was a super fun project I will be sharing the GitHub repo in a couple of days after finalizing the code"  
[X Link](https://x.com/anyuser/status/1981589315617714303)  2025-10-24T05:11Z [----] followers, 14.7K engagements


"The repo will be public within the next [--] days. Repo got delayed coz I had to travel during the weekend. I had a lot of fun exploring the customizations you can do with query generation attribute tagging tool calls etc. If video does well Ill do another on graph memory. What does it actually take to give an LLM memory @neural_avb explored that question by recreating the architecture described in the Mem0 paper using DSPy showing how extraction indexing retrieval and updates come together inside an agentic memory system. The video distills https://t.co/GZgWPdOORq What does it actually take to"  
[X Link](https://x.com/neural_avb/status/1982323649890934858)  2025-10-26T05:49Z [----] followers, [----] engagements


"Thanks for all the love for the new Long-term Memory video The code is now open-source and you can star fork and play with it. (repo in comments) It is a minimalist feature-complete implementation of the Mem0 memory system using DSPy and Qdrant. It already does user-scoped cross-session long-term memory generate attribute tags for quick filtering agentic query generation for search and a bunch of other cool stuff. Code is deliberately minimal - customize to your will. Note that it's meant to be primarily for educational value and understanding the mechanics of memory retrieval. It's"  
[X Link](https://x.com/neural_avb/status/1982825401928778150)  2025-10-27T15:03Z [----] followers, 21.7K engagements


"icymi I wrote the core components of Mem0 in DSPY and open sourced it. Thanks for all the love for the new Long-term Memory video The code is now open-source and you can star fork and play with it. (repo in comments) It is a minimalist feature-complete implementation of the Mem0 memory system using DSPy and Qdrant. It already does user-scoped https://t.co/ZowUNmG8UE Thanks for all the love for the new Long-term Memory video The code is now open-source and you can star fork and play with it. (repo in comments) It is a minimalist feature-complete implementation of the Mem0 memory system using"  
[X Link](https://x.com/neural_avb/status/1983077409776886078)  2025-10-28T07:44Z [----] followers, [----] engagements


"Last month I made a [--] min video on the fundamentals of Reinforcement Learning. Advanced topics presented in an intuitive visual beginner friendly way. (link below) I'll soon be writing an article about it - bit of a handbook of RL concepts. - Agents environments policies. - Model based vs Model free - Value based vs Policy based - Bellman Equation MDPs - TD Learning vs MC Sampling - Exploration techniques like Curiosity - Policy Gradients - Actor Critics - Trust regions and stability - more. Some of the most important concepts that enable you to understand RL research papers"  
[X Link](https://x.com/neural_avb/status/1983421409457050038)  2025-10-29T06:31Z [----] followers, [----] engagements


"Im very proud of my last [--] videos. They were all a grind but I have given my best every time. This my happy place Ill do it till my wheels fall off"  
[X Link](https://x.com/neural_avb/status/1984533968449085903)  2025-11-01T08:12Z [----] followers, 22.1K engagements


"@jatinsapru I hope the Indian public is kinder when our teams lose esp the womens team. The amount of scum written in online circles when they were having a tough patch was shameful. Im so glad that they shut everyone up. I havent felt this emotional about a cricket match in ages"  
[X Link](https://x.com/neural_avb/status/1985258990574735466)  2025-11-03T08:13Z [----] followers, [----] engagements


"My article on RL fundamentals is now live Thanks @TDataScience for featuring it in the Editor's pick section It goes over everything you need to know to understand research papers and implement RL algorithms. All the academic pre-reqs explained in a simple intuitive way"  
[X Link](https://x.com/anyuser/status/1986475747968758046)  2025-11-06T16:48Z [----] followers, 11.6K engagements


"My channel got approved for the new YT courses system. I've collected some my best videos about Transformers into one short free course. Go give these videos a watch I've been told they're pretty good. The videos cover the intuition theory math DL you'll need. In the coding sections we will implement LMs VITs Sparse MOEs and Speech Transformers from scratch as well"  
[X Link](https://x.com/anyuser/status/1988476181008749006)  2025-11-12T05:17Z [----] followers, [----] engagements


"@bad_piggyy Its for a YouTube tutorial Im making not a production ready model. Ive been training on the cheapest GPU I can find on prime intellect"  
[X Link](https://x.com/neural_avb/status/2009316627138093318)  2026-01-08T17:29Z [----] followers, [--] engagements


"@TheSonOfVinci Huh mustve made a mistake https://github.com/unslothai/unsloth/pull/3856 https://github.com/unslothai/unsloth/pull/3856"  
[X Link](https://x.com/neural_avb/status/2018546252162756864)  2026-02-03T04:45Z [----] followers, [--] engagements


"In prep for the upcoming finetuning/unsloth project I've been digging up some hidden gem survey papers. The amount of papers that came out in 2023-2024 about SFT RLHF function calling is so overwhelming. Back before we went fully into RL/reasoning and the big aha"  
[X Link](https://x.com/neural_avb/status/2018752131445280892)  2026-02-03T18:23Z [----] followers, [----] engagements


"Yesterday I asked people on YT if they wanted an Unsloth finetuning video. Everyone unanimously said yes but. After reading the comments I have understood the main thing people wanna learn is not the finetuning part itself. It is actually dataset preparation. So the upcoming video is going to also cover how to prepare quality data before you go finetune. Continued pretraining instruction tuning tool calling multilingual tokenization QA etc. Part literature review part hands-on tutorial part best-practices guide. We will use Unsloth to train all the models - so you'll simultaneously learn the"  
[X Link](https://x.com/neural_avb/status/2018989753543373182)  2026-02-04T10:07Z [----] followers, [----] engagements


"@goyal__pramod I think Ill make a video soon about scraping thousands of PyTorch repos extracting the top (say 50) functions and explaining them one by one"  
[X Link](https://x.com/neural_avb/status/1942057724574900402)  2025-07-07T03:06Z [----] followers, [----] engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing