@iScienceLuvr Tanishq Mathew Abraham, Ph.D.
Tanishq Mathew Abraham, Ph.D. posts on X about ai, model, open ai, llm the most. They currently have [------] followers and [---] posts still getting attention that total [-------] engagements in the last [--] hours.
Engagements: [-------] #
- [--] Week [-------] +82%
- [--] Month [---------] +114%
- [--] Months [----------] +85%
- [--] Year [----------] +22%
Mentions: [--] #
- [--] Month [--] -78%
- [--] Months [---] -6.40%
- [--] Year [---] +21%
Followers: [------] #
- [--] Week [------] +0.19%
- [--] Month [------] +0.63%
- [--] Months [------] +7.10%
- [--] Year [------] +21%
CreatorRank: [------] #
Social Influence
Social category influence technology brands 16.88% stocks 5.19% social networks 3.25% celebrities 3.25% finance 1.95% countries 1.3% fashion brands 0.65% gaming 0.65% nba 0.65%
Social topic influence ai #1121, model #309, open ai 3.9%, llm 3.25%, bytedance #38, xai 2.6%, rl 1.95%, this is 1.95%, code 1.95%, image #1020
Top accounts mentioned or mentioned by @sophontai @medarcai @johnowhitaker @ctibedo @willccbb @0x_vivek @alberfuen @jacoed @bcherny @renegadesilicon @dexteraiagent @justinmini12008 @ptremblay @codewithimanshu @teknium @tszzl @dylan522p @chrisalbon @elonmusk @theias
Top assets mentioned Microsoft Corp. (MSFT) Alphabet Inc Class A (GOOGL)
Top Social Posts
Top posts by engagements in the last [--] hours
"Impressive how open a leading AI lab like MiniMax is MiniMax just published their [----] TODOs on Hugging Face interesting times ahead 💫 https://t.co/Qme9lmq71j https://t.co/jdBno7CXKC MiniMax just published their [----] TODOs on Hugging Face interesting times ahead 💫 https://t.co/Qme9lmq71j https://t.co/jdBno7CXKC"
X Link 2026-01-05T14:37Z 85.9K followers, [----] engagements
"i was followed by a clawdbot on twitter idk how to feel about that lol"
X Link 2026-02-01T07:48Z 85.7K followers, [----] engagements
"Golden Goose: A Simple Trick to Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text "constructing a multiple-choice question-answering version of the fill-in-the-middle task" "Given a source text we prompt an LLM to identify and mask key reasoning steps then generate a set of diverse plausible distractors." "GooseReason effectively revives models saturated on existing RLVR data" "GooseReason-Cyber sets a new state-of-the-art in cybersecurity surpassing a 7B domain-specialized model with extensive domain-specific pre-training and post-training""
X Link 2026-02-02T08:03Z 85.8K followers, 14.3K engagements
"Generative Modeling via Drifting New Kaiming He paper Instead of a "pushforward" behavior carried out iteratively at inference time e.g. in diffusion/flow-based models evolve the pushforward distribution during training naturally enabling 1-step inference. SOTA results on ImageNet [------] with FID [----] in latent space and [----] in pixel space https://twitter.com/i/web/status/2019340015029919925 https://twitter.com/i/web/status/2019340015029919925"
X Link 2026-02-05T09:19Z 85.8K followers, 15K engagements
"Honestly I expect AI researchers to get replaced by AI agents before other researchers because most other research disciplines require physical experimentation"
X Link 2026-02-06T00:08Z 85.8K followers, 27.5K engagements
"I wonder how Mark Cuban feels about this lol IT'S HERE: Finally a direct to consumer website designed to find the lowest drug prices for YOU. https://t.co/jjnWOhcipw https://t.co/p0aoOJ8xtg IT'S HERE: Finally a direct to consumer website designed to find the lowest drug prices for YOU. https://t.co/jjnWOhcipw https://t.co/p0aoOJ8xtg"
X Link 2026-02-06T00:18Z 85.7K followers, 21.3K engagements
"I don't know why ByteDance is making a biology LLM benchmark but I welcome it lol BABE: Biology Arena BEnchmark "we develop BABE(Biology Arena BEnchmark1) a benchmark specifically designed to evaluate biological AI systems experimental reasoning capabilities. Critically all tasks in BABE are derived from peer-reviewed research papers and real-world biological studies" https://twitter.com/i/web/status/2019687305808687326 https://twitter.com/i/web/status/2019687305808687326"
X Link 2026-02-06T08:19Z 85.8K followers, 23K engagements
"Multimodal foundation models are the future of medical AI and medical AI is the future of healthcare"
X Link 2026-02-07T21:16Z 85.8K followers, [----] engagements
"@willccbb how much git though like surely anyone who's coded before the vibecoding era knows basic git is that enough"
X Link 2026-02-07T22:37Z 85.9K followers, [----] engagements
"A full super bowl ad for codex that's wild You can just build things. https://t.co/g0JCVjbSef You can just build things. https://t.co/g0JCVjbSef"
X Link 2026-02-09T00:06Z 85.8K followers, 21.9K engagements
"Can someone get me the Codex merch I don't have my laptop with me rn 😭😭😭"
X Link 2026-02-09T00:49Z 85.8K followers, 10.6K engagements
"@Teknium I'm disappointed you think it's lame"
X Link 2026-02-09T08:08Z 85.8K followers, [----] engagements
"Honestly kinda wild how many AI related ads there were during the Super Bowl. OpenAI Anthropic Gemini Alexa Microsoft Copilot Genspark Base44 Wix Harmony Not to mention the other ads that likely used AI like Dunkin or Xfinity"
X Link 2026-02-09T08:09Z 85.8K followers, [----] engagements
"iGRPO: Self-Feedback-Driven LLM Reasoning "In Stage [--] iGRPO samples multiple exploratory drafts and selects the highest-reward draft using the same scalar reward signal used for optimization. In Stage [--] it appends this best draft to the original prompt and applies a GRPO-style update on draft-conditioned refinements training the policy to improve beyond its strongest prior attempt. Under matched rollout budgets iGRPO consistently outperforms GRPO across base models (e.g. Nemotron-H-8B-Base-8K and DeepSeek-R1 Distilled)""
X Link 2026-02-10T09:55Z 85.8K followers, 13K engagements
"Learning to Self-Verify Makes Language Models Better Reasoners "learning to self-verify alone can significantly improve generation performance" "Learning to self-verify requires significantly fewer tokens to solve the same problems.""
X Link 2026-02-10T10:07Z 85.8K followers, 11.5K engagements
"@tszzl my time to shine (i currently have [---] tabs open although i just closed a bunch earlier today on average it's usually closer to 300)"
X Link 2026-02-11T03:21Z 85.8K followers, [----] engagements
"Features as Rewards: Scalable Supervision for Open-Ended Tasks via Interpretability Introduces a new RL pipeline called Reinforcement Learning from Feature Rewards (RLFR) which uses LLM features as rewards. Applied to hallucination reduction trained Gemma-3- 12B-IT model that is 58% less likely to hallucinate compared to the original model https://twitter.com/i/web/status/2021513421041107243 https://twitter.com/i/web/status/2021513421041107243"
X Link 2026-02-11T09:15Z 85.8K followers, [----] engagements
"ClinAlign: Scaling Healthcare Alignment from Clinician Preference Introduces HealthRubrics a dataset of [----] physician-verified preference examples in which clinicians refine LLM-drafted rubrics to meet rigorous medical standards. Distilled those rubrics into HealthPrinciples [---] broadly reusable clinically grounded principles organized by clinical dimensions enabling scalable supervision beyond manual annotation. Finetuned Qwen3-30B-A3B-Instruct achieves 33.4% on HealthBench-Hard beating o3 and DeepSeek-R1 https://twitter.com/i/web/status/2021518210663616999"
X Link 2026-02-11T09:34Z 85.8K followers, [----] engagements
"@dylan522p Seriously that's really lame on HBO's part it's not even close. If it's cuz of the MAX maybe you could get away with InferenceMax or InferenceMaxx"
X Link 2026-02-13T01:00Z 85.8K followers, [----] engagements
"are people hosting funerals for these models like they did with claude [--] sonnet Tomorrow at 10am PT legacy models (GPT-5 GPT-4o GPT-4.1 GPT-4.1 mini and OpenAI o4-mini) will be deprecated in ChatGPT. https://t.co/RJioBsLY6D Tomorrow at 10am PT legacy models (GPT-5 GPT-4o GPT-4.1 GPT-4.1 mini and OpenAI o4-mini) will be deprecated in ChatGPT. https://t.co/RJioBsLY6D"
X Link 2026-02-13T01:12Z 85.8K followers, [----] engagements
"GPT-5 was released [--] months ago. It's already being deprecated. GPT-5.1 was released [--] months ago GPT-5.2 was released a month ago and GPT-5.3 Codex was released a week ago. The rate of progress is wild can't keep up Tomorrow at 10am PT legacy models (GPT-5 GPT-4o GPT-4.1 GPT-4.1 mini and OpenAI o4-mini) will be deprecated in ChatGPT. https://t.co/RJioBsLY6D Tomorrow at 10am PT legacy models (GPT-5 GPT-4o GPT-4.1 GPT-4.1 mini and OpenAI o4-mini) will be deprecated in ChatGPT. https://t.co/RJioBsLY6D"
X Link 2026-02-13T01:43Z 85.9K followers, [----] engagements
"EMA Policy Gradient: Taming Reinforcement Learning for LLMs with EMA Anchor and Top-k KL Really nice paper that proposes two improvements to RL LLM training: [--]. an unbiased token-level KL estimator that computes exact KL on the top-k indices of logits and a sampled KL on the rest [--]. use an EMA model for the reference policy LLM in the KL term The combined EMA-PG approach not only learns faster and but reaches higher asymptotic performance. That said their experiments were only with Qwen models. Looks like lots of theory to back up their results though. These tweaks are definitely worth a try"
X Link 2026-02-05T09:31Z 85.9K followers, [----] engagements
"How do you increase your brain's context length"
X Link 2026-02-07T01:33Z 85.9K followers, 57.2K engagements
"@chrisalbon yes but they are most definitely trying to replace all of this with AI too lol"
X Link 2026-02-14T20:39Z 85.9K followers, [---] engagements
"With Jimmy Ba's departure half of the xAI founding team has left Last day at xAI. xAI's mission is push humanity up the Kardashev tech tree. Grateful to have helped cofound at the start. And enormous thanks to @elonmusk for bringing us together on this incredible journey. So proud of what the xAI team has done and will continue to stay close Last day at xAI. xAI's mission is push humanity up the Kardashev tech tree. Grateful to have helped cofound at the start. And enormous thanks to @elonmusk for bringing us together on this incredible journey. So proud of what the xAI team has done and will"
X Link 2026-02-11T00:27Z 85.9K followers, 60.5K engagements
"This is insane GPT-5.2 making serious progress in theoretical physics now Curious to get thoughts from actual physicists on this. So when will AI get a Nobel Prize in Physics 🤣 GPT-5.2 derived a new result in theoretical physics. Were releasing the result in a preprint with researchers from @the_IAS @VanderbiltU @Cambridge_Uni and @Harvard. It shows that a gluon interaction many physicists expected would not occur can arise under specific GPT-5.2 derived a new result in theoretical physics. Were releasing the result in a preprint with researchers from @the_IAS @VanderbiltU @Cambridge_Uni and"
X Link 2026-02-13T19:25Z 85.9K followers, 20.1K engagements
"THIS ISSUE SHOULD NOT BE CLOSED Claude Code still doesn't support Chrome extension with WSL2 which I would really love to have. The issue is closed for some reason. There are a bunch of workarounds in the thread but none of them work for me. cc: @bcherny"
X Link 2026-02-14T09:34Z 85.9K followers, 12.2K engagements
"@renegadesilicon Yeah this is why I hate sleep https://x.com/i/status/1921423241106673840 sleep is just death being shy https://x.com/i/status/1921423241106673840 sleep is just death being shy"
X Link 2026-02-16T19:44Z 85.9K followers, [----] engagements
"ViT-5: Vision Transformers for The Mid-2020s "a systematic investigation into modernizing Vision Transformer backbones by leveraging architectural advancements from the past five years" * LayerScale * RMSNorm * original MLP design with GeLU activation * both APE and 2D RoPE jointly * registers with a separate 2D RoPE * QK-Norm * remove bias terms in the QKV projection layers 84.2% top1 accuracy on ImageNet-1k [----] FID on ImageNet-256 https://twitter.com/i/web/status/2021165285479157828 https://twitter.com/i/web/status/2021165285479157828"
X Link 2026-02-10T10:12Z 85.9K followers, 102.6K engagements
"Isomorphic Labs (GDM spinoff) announces their Drug Design Engine that goes beyond AlphaFold3 and improves generalizability "IsoDDE more than doubles the accuracy of AlphaFold [--] on a challenging protein-ligand generalisation benchmark" "providing a new state of the art for antibody-antigen interface prediction and CDR-H3 loop modeling" "or small molecule binders IsoDDEs affinity predictions exceed gold-standard physics-based methods" unfortunately no details yet on the arch or training The Iso team has cooked something incredible: our new technical report unveils the latest results from our"
X Link 2026-02-10T10:39Z 85.9K followers, 15.7K engagements
"Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation "In Latent Forcing we train a single diffusion model over a pixel space and latent space simultaneously with multiple time variables. By scheduling the denoising trajectory to reveal self-supervised encoder latents before pixels we achieve the convergence benefits of latent diffusion without losing information due to a tokenizer. The generated latent which effectively serves as a scratchpad to condition the generation of the natural image is discarded at the end of denoising process.""
X Link 2026-02-13T09:19Z 85.9K followers, 18K engagements
"3am is the best time to Claude Code :)"
X Link 2026-02-14T08:18Z 85.9K followers, [----] engagements
"Precigenetics is working on an important intersection of cutting edge technologies from microfluidics to high-throughput label-free microscopic imaging combined with AI. They are using this to solve important bottlenecks in AI drug discovery worth following along Never thought a chip could cure cancer but here we are. For months our hardware team at Precigenetics has been secretly working on MBOP: Modular Biological Observation Platform. Were now fabricating v.1 and heres why you should be excited. MBOP is the first microfluidic https://t.co/dn7fRdB1Am Never thought a chip could cure cancer"
X Link 2026-02-17T08:36Z 85.9K followers, [----] engagements
"Can't believe it's been [--] year since @humanscotti and I incorporated @SophontAI 😲 We've accomplished so much in this timeframe 🎉 We raised a $9.2M seed round 🔥 We've added three more amazing researchers to our team ❤ We organize and support the largest online medical AI research community (@MedARC_AI) We released OpenMidnight (pathology) and Medmarks (LLMs) and presented at NeurIPS workshop (fMRI) We're proud to be at the forefront of medical AI innovation and making a lot of progress towards our vision for a universal foundation model for medicine. Excited for what year [--] will bring 👀"
X Link 2026-02-07T04:49Z 85.9K followers, 24.2K engagements
"This is an interesting paper this company has trained a cross-species multimodal foundation model of immunology and inflammation: - 440M-parameter model (300M-parameters gene expression encoder 85M-parameter histology encoder 55M-parameter fusion head) - integrates human and mouse bulk RNA-seq microarray pseudobulked single-cell and histology into unified sample embeddings across more than [--] tissues and conditions multimodality improved performance RNA encoder scaled with across model sizes the model works across both mice and human data and different gene expression technologies. I would"
X Link 2026-02-12T08:43Z 85.9K followers, [----] engagements
"😭😭😭 Everybody builds for Mac and I get left behind. I am slowly considering trying Mac just for AI features. I would still advocate for people to build for Windows though @iScienceLuvr @bcherny Just buy a Mac 🙏 You lost so much time haggling with windows already that a Mac would be saving you money (or wait until M5 Max and get that one) @iScienceLuvr @bcherny Just buy a Mac 🙏 You lost so much time haggling with windows already that a Mac would be saving you money (or wait until M5 Max and get that one)"
X Link 2026-02-14T09:55Z 85.9K followers, [----] engagements
"ByteDance has released a frontier model with some really strong benchmarks A Valentine's Day gift Maybe more like a Chinese New Year's gift lol Seed [---] is finally out 🔥 https://t.co/XXPqBSaE0E Seed [---] is finally out 🔥 https://t.co/XXPqBSaE0E"
X Link 2026-02-14T10:31Z 85.9K followers, [----] engagements
""oh it's valentine's day i totally forgot i was too busy grinding on my B2B SaaS startup" - half the founders on Twitter"
X Link 2026-02-14T20:19Z 85.9K followers, [----] engagements
"I'm very confused why accounts that have blocked me are showing up on my feed. This seems like a bug As a side note I guess this person thinks I'm a slop account lol"
X Link 2026-02-15T08:28Z 85.9K followers, [----] engagements
"MonoLoss: A Training Objective for Interpretable Monosemantic Representations "we introduce the Monosemanticity Loss (MonoLoss) a plug-in objective that directly rewards semantically consistent activations for learning interpretable monosemantic representations. Across SAEs trained on CLIP SigLIP2 and pretrained ViT features using BatchTopK TopK and JumpReLU SAEs MonoLoss increases MonoScore for most latents.""
X Link 2026-02-16T07:48Z 85.9K followers, [----] engagements
"abs: https://arxiv.org/abs/2602.12403 https://arxiv.org/abs/2602.12403"
X Link 2026-02-16T07:48Z 85.9K followers, [----] engagements
"Chinese New Year is rapidly becoming the AI researcher's favorite holiday"
X Link 2026-02-16T08:26Z 85.9K followers, 136.9K engagements
"Qwen3.5-397B-A17B release"
X Link 2026-02-16T09:41Z 85.9K followers, 15.9K engagements
"New Qwen model is comparable to GPT-5.2 Claude [---] Opus Gemini [--] Pro Qwen3.5-397B-A17B release https://t.co/phVSXTUgBU Qwen3.5-397B-A17B release https://t.co/phVSXTUgBU"
X Link 2026-02-16T09:46Z 85.9K followers, 12.6K engagements
"link: https://qwen.ai/blogid=qwen3.5 https://qwen.ai/blogid=qwen3.5"
X Link 2026-02-16T09:49Z 85.9K followers, [----] engagements
"Interesting got a bunch of Chinese followers after tweeting this lol Chinese New Year is rapidly becoming the AI researcher's favorite holiday Chinese New Year is rapidly becoming the AI researcher's favorite holiday"
X Link 2026-02-16T21:52Z 85.9K followers, [----] engagements
"i just lost [---] tabs on my laptop it's so over"
X Link 2026-02-17T09:27Z 85.9K followers, [----] engagements
"HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam "We release HLE-Verified a verified and revised version of HLE constructed via a transparent component-wise verification protocol and fine-grained error taxonomy comprising [---] verified items [----] revised-and-verified items and a 689-item documented uncertain set for structured community refinement." "verification materially alters measured performance (an average +710 accuracy points overall and +3040 points on items with erroneous problems/answers)""
X Link 2026-02-17T10:32Z 85.9K followers, [----] engagements
"dataset: abs: https://arxiv.org/abs/2602.13964 https://github.com/SKYLENAGE-AI/HLE-Verified https://arxiv.org/abs/2602.13964 https://github.com/SKYLENAGE-AI/HLE-Verified"
X Link 2026-02-17T10:32Z 85.9K followers, [---] engagements
"ByteDance releases BitDance: Scaling Autoregressive Generative Models with Binary Tokens "We present BitDance a scalable autoregressive (AR) image generator that predicts binary visual tokens instead of codebook indices. With high-entropy binary latents BitDance lets each token represent up to [----] states yielding a compact yet highly expressive discrete representation." "On ImageNet [------] BitDance achieves an FID of [----] the best among AR models. ""
X Link 2026-02-17T10:35Z 85.9K followers, [----] engagements
"project page: code: demo: model: paper: https://arxiv.org/abs/2602.14041 https://huggingface.co/collections/shallowdream204/bitdance https://huggingface.co/spaces/shallowdream204/BitDance-14B-64x https://github.com/shallowdream204/BitDance https://bitdance.csuhan.com/ https://arxiv.org/abs/2602.14041 https://huggingface.co/collections/shallowdream204/bitdance https://huggingface.co/spaces/shallowdream204/BitDance-14B-64x https://github.com/shallowdream204/BitDance https://bitdance.csuhan.com/"
X Link 2026-02-17T10:35Z 85.9K followers, [---] engagements
"Image Generation with a Sphere Encoder a few-step image generation method by mapping images to a spherical latent space trained with simple reconstruction+consistency losses"
X Link 2026-02-17T10:58Z 85.9K followers, [----] engagements
"abs: https://arxiv.org/abs/2602.15030 https://arxiv.org/abs/2602.15030"
X Link 2026-02-17T10:58Z 85.9K followers, [---] engagements
"@_xjdr Could this just be model hallucination I never trust what a model says about itself"
X Link 2026-02-17T12:11Z 85.9K followers, [---] engagements
"is Cursor CLI better or worse than Claude Code"
X Link 2025-08-10T00:18Z 80.2K followers, 11.9K engagements
"Diffusion language models (DLMs) are cool but people wonder why is it better than autoregressive language models (ARLMs) More experiments need to be done but preliminary results like this show diffusion models squeeze more performance out of the data In addition to its speed benefits there seems to be a path for DLMs to face broader applications and usage. However DLMs are not without its problems. Maybe DLMs distilled into ARLMs could be useful or ideally some sort of model that interpolates between diffusion and AR based on the prompt. Token crisis: solved. ✅ We pre-trained diffusion"
X Link 2025-08-11T04:45Z 80.2K followers, 30.4K engagements
"GLM-4.5: Agentic Reasoning and Coding (ARC) Foundation Models "Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning GLM-4.5 achieves strong performance across agentic reasoning and coding (ARC) tasks scoring 70.1% on TAU-Bench 91.0% on AIME [--] and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks""
X Link 2025-08-11T11:51Z 80.2K followers, [----] engagements
"project page: abs: https://arxiv.org/abs/2508.05954 https://bifrost-1.github.io/ https://arxiv.org/abs/2508.05954 https://bifrost-1.github.io/"
X Link 2025-08-11T11:58Z 80.2K followers, [----] engagements
"this is huge and much awaited memories are being added to Claude Claude can now reference past chats so you can easily pick up from where you left off. https://t.co/n9ZgaTRC1y Claude can now reference past chats so you can easily pick up from where you left off. https://t.co/n9ZgaTRC1y"
X Link 2025-08-11T19:32Z 80K followers, [----] engagements
"If you're using ChatGPT as just "someone" to chat with then memories can easily lead to you being oneshotted & is a net negative imo. However I'm using ChatGPT as a tool for my work and I find that the memories provide useful context about my work that speed up what I'm doing"
X Link 2025-08-12T00:21Z 80K followers, [----] engagements
"RLVR/RLHF libraries: verl - ByteDance TRL - HuggingFace slime - Zhipu AI prime-rl - Prime Intellect ROLL - Alibaba Nemo-RL - NVIDIA AReaL - Ant Research SkyRL - UC Berkeley open-instruct - Allen AI torchtune - PyTorch Any I am missing Which do you like"
X Link 2025-08-12T19:44Z 80.4K followers, 81.2K engagements
"Meta releases DINOv3 Everyone talks about Llama but I think Meta's contributions to computer vision (SAM DINOv2 etc.) are highly underappreciated. They're now releasing a newer iteration with large model (7B param) better data curation and improved dense features. This is sure to be the foundation for many computer vision use-cases going forward. Introducing DINOv3: a state-of-the-art computer vision model trained with self-supervised learning (SSL) that produces powerful high-resolution image features. For the first time a single frozen vision backbone outperforms specialized solutions on"
X Link 2025-08-14T18:56Z 80.2K followers, 23.3K engagements
"What is the "Hello World" of RLVR"
X Link 2025-08-14T22:34Z 80.2K followers, 22.8K engagements
"I can't believe Grok Imagine is this bad. This is like [----] level of quality. The xAI team definitely can do better 😭 And why is Elon proudly sharing this come on. Grok Imagine prompt: A m wizard shoveling snow in a blizzard on the moon with a rocket in the background and a gremlin. https://t.co/ID2yoHOgMZ Grok Imagine prompt: A m wizard shoveling snow in a blizzard on the moon with a rocket in the background and a gremlin. https://t.co/ID2yoHOgMZ"
X Link 2025-08-15T02:13Z 80.3K followers, 96.8K engagements
"@PFPercyJr yes we did"
X Link 2025-08-15T11:42Z 80.2K followers, [----] engagements
"Will Smith eating spaghetti with Grok Imagine"
X Link 2025-08-15T11:52Z 80.3K followers, [----] engagements
"@rasbt @natolambert xAI open-sourced Grok 1"
X Link 2025-08-17T22:11Z 80.2K followers, [----] engagements
"I sleep at the same time Newton did rofl tell me again about how locked in you are https://t.co/GruwNVWJFU tell me again about how locked in you are https://t.co/GruwNVWJFU"
X Link 2025-08-17T22:54Z 80.3K followers, 10K engagements
"life update: for those who dont know i co-founded @SophontAI a few months ago to work on open source medical AI. incredibly excited about what were building 🚀 life update: for those who dont know i joined @primeintellect a few months ago to work on open source AGI. incredibly excited about what were building 🚀 life update: for those who dont know i joined @primeintellect a few months ago to work on open source AGI. incredibly excited about what were building 🚀"
X Link 2025-08-18T04:45Z 80.3K followers, 19.4K engagements
"Thyme: Think Beyond Images "Thyme transcends traditional "thinking with images" paradigms by autonomously generating and executing diverse image processing and computational operations through executable code significantly enhancing performance on high-resolution perception and complex reasoning tasks. Leveraging a novel two-stage training strategy that combines supervised fine-tuning with reinforcement learning and empowered by the innovative GRPO-ATS algorithm Thyme achieves a sophisticated balance between reasoning exploration and code execution precision.""
X Link 2025-08-18T11:23Z 80.3K followers, 33.8K engagements
"github: huggingface: project page: abs: https://arxiv.org/abs/2508.11630 https://thyme-vl.github.io/ https://huggingface.co/Kwai-Keye/Thyme-RL https://github.com/yfzhang114/Thyme https://arxiv.org/abs/2508.11630 https://thyme-vl.github.io/ https://huggingface.co/Kwai-Keye/Thyme-RL https://github.com/yfzhang114/Thyme"
X Link 2025-08-18T11:23Z 80.3K followers, [----] engagements
"new Qwen model release for image editing 🚀 Excited to introduce Qwen-Image-Edit Built on 20B Qwen-Image it brings precise bilingual text editing (Chinese & English) while preserving style and supports both semantic and appearance-level editing. ✨ Key Features ✅ Accurate text editing with bilingual support ✅ https://t.co/p21KUXoC50 🚀 Excited to introduce Qwen-Image-Edit Built on 20B Qwen-Image it brings precise bilingual text editing (Chinese & English) while preserving style and supports both semantic and appearance-level editing. ✨ Key Features ✅ Accurate text editing with bilingual"
X Link 2025-08-18T17:55Z 80.2K followers, [----] engagements
"@giffmana okay fair. the reason i chose the project page summary is cuz the abstract was quite long and i know no one would read it but maybe i should have stuck to that lol"
X Link 2025-08-18T18:49Z 80.2K followers, [----] engagements
"nano-banana is from Google 🍌 🍌"
X Link 2025-08-19T20:53Z 80.4K followers, 62.7K engagements
"Generative Medical Event Models Improve with Scale "we introduce the Cosmos Medical Event Transformer (CoMET) models a family of decoder-only transformer models pretrained on [---] million patients representing [---] billion discrete medical events (151 billion tokens). We present the largest scaling-law study for medical event data establishing a methodology for pretraining and revealing power-law scaling relationships for compute tokens and model size. Based on this we pretrained a series of compute-optimal models with up to [--] billion parameters. Conditioned on a patient's real-world history"
X Link 2025-08-20T09:47Z 80.3K followers, [----] engagements
"MDPO: Overcoming the Training-Inference Divide of Masked Diffusion Language Models "We propose a novel Masked Diffusion Policy Optimization (MDPO) to exploit the Markov property diffusion possesses and explicitly train the model under the same progressive refining schedule used at inference. MDPO matches the performance of the previous state-of-the-art (SOTA) method with 60x fewer gradient updates while achieving average improvements of 9.6% on MATH500 and 54.2% on Countdown over SOTA when trained within the same number of weight updates. Additionally we improve the remasking strategy of"
X Link 2025-08-20T09:55Z 80.3K followers, [----] engagements
"Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation "In this work we analyze specific properties which make a benchmark more reliable for such decisions and interventions to design higher-quality evaluation benchmarks. We introduce two key metrics that show differences in current benchmarks: signal a benchmark's ability to separate better models from worse models and noise a benchmark's sensitivity to random variability between training steps. We demonstrate that benchmarks with a better signal-to-noise ratio are more reliable when making decisions at small"
X Link 2025-08-20T09:59Z 81.8K followers, [----] engagements
"FutureX: An Advanced Live Benchmark for LLM Agents in Future Prediction "we introduce FutureX a dynamic and live evaluation benchmark specifically designed for LLM agents performing future prediction tasks. FutureX is the largest and most diverse live benchmark for future prediction supporting real-time daily updates and eliminating data contamination through an automated pipeline for question gathering and answer collection. We evaluate [--] LLM/agent models including those with reasoning search capabilities and integration of external tools such as the open-source Deep Research Agent and"
X Link 2025-08-20T10:07Z 80.3K followers, 20.4K engagements
"Alibaba's Ant Group does interesting work and underrated imo Reinforcement Learning with Rubric Anchors "we extend the RLVR paradigm to open-ended tasks by integrating rubric-based rewards where carefully designed rubrics serve as structured model-interpretable criteria for automatic scoring of subjective outputs. We construct to https://t.co/9yWacIIlgb Reinforcement Learning with Rubric Anchors "we extend the RLVR paradigm to open-ended tasks by integrating rubric-based rewards where carefully designed rubrics serve as structured model-interpretable criteria for automatic scoring of"
X Link 2025-08-21T01:46Z 80.4K followers, [----] engagements
"There's a future Sophont employee in this app and it's my job to get them out"
X Link 2025-08-21T21:12Z 80.3K followers, 13.1K engagements
"Seriously if you're an AI researcher why would you work on waifus when you could instead work on saving people's lives (we're hiring 😉) Imagine you attend Stanford CS. top of your class. Youve busted your ass to get into the AI industry. You land your dream job. You get to build the future with your idol the visionary Elon musk. And then he puts you to work in the gooncave making whatever this is Imagine you attend Stanford CS. top of your class. Youve busted your ass to get into the AI industry. You land your dream job. You get to build the future with your idol the visionary Elon musk. And"
X Link 2025-08-22T05:01Z 80.3K followers, 70.8K engagements
"Can't believe it's been [--] YEARS since Stable Diffusion was released The impact that this trailblazing launch had on the AI community and open-source ecosystem was huge and I'm glad I got to be a (very small) part of it. It's incredible to see how much has changed in these past [--] years and how the generative AI landscape has grown. Delighted to announce the public open source release of #StableDiffusion Please see our release post and retweet https://t.co/dEsBX7cRHw Proud of everyone involved in releasing this tech that is the first of a series of models to activate the creative potential of"
X Link 2025-08-22T23:38Z 80.5K followers, 11.6K engagements
"xAI finally open sources Grok [--] @BasedBeffJezos Its high time we open sourced Grok [--]. Will make it happen next week. Weve just been fighting fires and burning the 4am oil nonstop for a while now. @BasedBeffJezos Its high time we open sourced Grok [--]. Will make it happen next week. Weve just been fighting fires and burning the 4am oil nonstop for a while now"
X Link 2025-08-23T20:31Z 80.4K followers, 23.2K engagements
"so are there any ML researchers on LinkedIn or nah"
X Link 2025-08-23T22:29Z 80.4K followers, 12.9K engagements
"@johnowhitaker the tiles are not overlapping i remember in our lab we used to use Microsoft Image Composite Editor (it got discontinued though) cuz we collected overlapping tiles. but IIRC we used Fiji for non-overlapping tiles"
X Link 2025-08-24T05:38Z 80.3K followers, [---] engagements
"Follow @SophontAI :)"
X Link 2025-08-24T10:10Z 80.3K followers, [----] engagements
"very exciting to see what Prime Intellect is doing to grow the open-source RL ecosystem. We hope to do a similar strategy to grow the open-source medical AI ecosystem as well (part of that includes developing medical RL envs) More info about how to contribute coming soon i'll confess i do have a very specific mission in mind with this project. the semi-vague private beta rollout is part of it. the set of tasks we're sourcing is part of it. the GPU bounties are part of it. the shitposts are part of it. the podcasts are part of it. mindshare is i'll confess i do have a very specific mission in"
X Link 2025-08-25T03:43Z 80.4K followers, 17.2K engagements
"a reminder that literally none of these influential people lists are accurate So Time Magazine put out a list of the [---] most influential people in AI. People NOT on the list: - Demis Hassabis - Mustafa Suleyman - Aravind Srinivas - Geoffrey Hinton - Yann LeCun No influential media personalities that cover AI either. - Joanna Stern - Nilay Patel - So Time Magazine put out a list of the [---] most influential people in AI. People NOT on the list: - Demis Hassabis - Mustafa Suleyman - Aravind Srinivas - Geoffrey Hinton - Yann LeCun No influential media personalities that cover AI either. - Joanna"
X Link 2025-08-29T02:41Z 80.4K followers, 10.3K engagements
"@willccbb We're sponsoring a workshop and doing a social too see you in San Diego"
X Link 2025-08-29T08:14Z 80.4K followers, [---] engagements
"apparently all that vagueposting from openai employees yesterday was likely due to an internal celebration event unrelated to any model"
X Link 2025-08-29T23:42Z 80.5K followers, 43.6K engagements
"If you skip tweeting for even a single day The Algorithm just decides it hates you now just fyi"
X Link 2025-09-01T05:26Z 80.5K followers, 34.1K engagements
"rStar2-Agent: Agentic Reasoning Technical Report "We introduce rStar2-Agent a 14B math reasoning model trained with agentic reinforcement learning to achieve frontier-level performance." "three key innovations that makes agentic RL effective at scale: (i) an efficient RL infrastructure with a reliable Python code environment that supports high-throughput execution and mitigates the high rollout costs enabling training on limited GPU resources (64 MI300X GPUs); (ii) GRPO-RoC an agentic RL algorithm with a Resample-on-Correct rollout strategy that addresses the inherent environment noises from"
X Link 2025-09-02T08:42Z 80.7K followers, 12.2K engagements
"I am looking to get new noise-cancelling Bluetooth headphones with high-quality sound. does anyone have suggestions since I don't have Apple products anything apart from AirPods Max"
X Link 2025-09-03T00:58Z 80.5K followers, 23.1K engagements
"actually apparently airpods max work fine with non-apple devices do others have experience with this any complaints I am looking to get new noise-cancelling Bluetooth headphones with high-quality sound. does anyone have suggestions since I don't have Apple products anything apart from AirPods Max I am looking to get new noise-cancelling Bluetooth headphones with high-quality sound. does anyone have suggestions since I don't have Apple products anything apart from AirPods Max"
X Link 2025-09-03T01:31Z 80.5K followers, [----] engagements
"Baichuan-M2: Scaling Medical Capability with Large Verifier System Baichuan has released what is probably right now the best open-source LLM for medicine Second only to GPT-5 "Despite its relatively small number of parameters (only 32B) Baichuan-M2 outperformed all other open-source models including gpt-oss-120B and most advanced closed-source counterparts on HealthBench. It particularly excelled on the HealthBench Hard test achieving a score exceeding [--] a performance level previously reached by only one other model globally GPT-5." "Our framework comprises two key components: a Patient"
X Link 2025-09-03T09:42Z 80.7K followers, 13.9K engagements
"Congrats @corbtt and team on OpenPipe's acquisition by CoreWeave"
X Link 2025-09-03T19:48Z 80.7K followers, [----] engagements
"@kalomaze fair but everything is quantum and in many cases quantum effects do have meaningful effects on biological processes. i'm not sure there's really any evidence to support the "quantum microtubules in the brain" theory though"
X Link 2025-09-04T08:54Z 80.5K followers, [---] engagements
"Having seen a few different blog posts like this I am convinced that with a small team and limited funding it's possible to build a half decent actually useful text search engine as a working consumer-facing product. However I don't know if the same can be said for searching images and videos which is the more useful product opportunity right now. That seems like a much much more challenging infra problem. this is one of the most remarkable technical blog posts Ive ever read https://t.co/MVD3EKXnvE this is one of the most remarkable technical blog posts Ive ever read https://t.co/MVD3EKXnvE"
X Link 2025-09-08T20:14Z 81.2K followers, 163K engagements
"Yep I found this kinda shocking tbh Like why isn't the LLM RL ecosystem more developed and solidified by now It's been [--] months since R1 open-source should be moving faster. I agree with Rohan's conclusion that currently one of the best libraries right now is prime-rl but it's right now not good for scaling up to R1/Qwen/Kim-scale training at all. But it looks like progress is being made. were approaching the end of [----] and theres still no plug-n-play RL lib in the interrim: - i built a shitty version of this (llamagym) - RL started working (o1) - oss found out how it worked (r1) - RL env"
X Link 2025-09-10T21:19Z 80.7K followers, 49.2K engagements
"Folks ask me if there will be follow-up research to MindEye2. Well kinda. With the MindEye models we trained a simple neural network solely to translate fMRI to CLIP embeddings. We're taking a slightly different approach now. Now we're focused on training a foundation model for fMRI. We hope this model will be used not only for image decoding but for other clinical/diagnostic applications. Before foundation models like GPTs people used to train custom models for individual tasks. Now they get better results finetuning foundation models. The same will be true here. If you're interested in"
X Link 2025-09-10T21:52Z 80.7K followers, 17K engagements
"JUST IN: Nvidia is now worth more than Canada JUST IN: Nvidia is now worth more than Canada"
X Link 2025-09-16T00:08Z 81.2K followers, 25K engagements
"this is wild lol a first for me on linkedin"
X Link 2025-09-17T19:51Z 81.3K followers, [----] engagements
"wow a live demo of silently writing a message with Meta neural band on the Meta Ray-Ban Display pretty cool"
X Link 2025-09-18T00:26Z 81K followers, 21.7K engagements
"I have heard of several folks using torchtitan internally for RL training. However torchtitan doesn't directly support GRPO which means folks are adding an implementation themselves. A few questions: [--]. Are there any good open-source torchtitan forks with GRPO support [--]. What other groups are using torchtitan for RL [--]. Why is torchtitan so compelling that people are willing to add GRPO support and use it for RL training"
X Link 2025-09-18T02:58Z 81K followers, 13.9K engagements
"Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision "This paper asks a simple question: Can inference compute substitute for missing supervision" "the current policy produces a group of rollouts; a frozen anchor (the initial policy) reconciles omissions and contradictions to estimate a reference turning extra inference-time compute into a teacher signal." "With training CaT-RL delivers up to 33% relative improvement on MATH-500 and 30% on HealthBench with Llama [---] 8B and large gains across two other model families without human annotations""
X Link 2025-09-18T08:55Z 81.1K followers, 13.6K engagements
"At @MedARC_AI we are building a comprehensive suite of medical LLM evals and we already have tons of volunteers and lots of great progress The project started less than a week ago Are there other medical LLM evals we should include"
X Link 2025-09-19T12:40Z 81.6K followers, [----] engagements
"We obviously don't have as much compute as an AGI lab but we have much more compute than many medical AI labs/startups"
X Link 2025-09-21T02:49Z 81.6K followers, 27K engagements
"yes we are we also have open-source RL medical LLM research in @MedARC_AI discord https://sophont.med/job_postings/llm @iScienceLuvr Are you guys hiring RL people Would be awesome to be able to have some compute to do RL experiments on med data https://sophont.med/job_postings/llm @iScienceLuvr Are you guys hiring RL people Would be awesome to be able to have some compute to do RL experiments on med data"
X Link 2025-09-21T02:59Z 81.6K followers, 12.3K engagements
"My conspiracy theory is that Alex Krizhevsky (inventor of AlexNet) reunited with Ilya to work at SSI It would make sense Alex is offline with no info since [----] and SSI is completely stealth certainly a good fit"
X Link 2025-09-21T10:43Z 81.5K followers, 89.5K engagements
"hmm there must have been an earthquake in SF"
X Link 2025-09-22T10:01Z 81.5K followers, 16.1K engagements
"There's no way this is real the robot is straight up aura farming lol Unitree G1 has mastered more quirky skills 🤩 Unitree G1 has learned the "Anti-Gravity" mode: stability is greatly improved under any action sequence and even if it falls it can quickly get back up. https://t.co/gDR0n0eIXl Unitree G1 has mastered more quirky skills 🤩 Unitree G1 has learned the "Anti-Gravity" mode: stability is greatly improved under any action sequence and even if it falls it can quickly get back up. https://t.co/gDR0n0eIXl"
X Link 2025-09-22T11:16Z 81.2K followers, 13.2K engagements
"There's an optimal amount of switching between AI tools: if you switch to the next SOTA tool every single time there is a high cost to switch workflows but if you stick to one tool forever you will fall behind. Efficiently leveraging AI requires striking this balance"
X Link 2025-09-22T23:12Z 81.3K followers, [----] engagements
"i think this is such an unbelievably bad take. the innovation that has happened cuz of open-source models like Stable Diffusion and DeepSeek is huge you cannot deny this there are more things you can do with open-source models (finetuning interpretability etc. which some companies do provide for their models but extremely limited) they do tend to be cheaper as well (many companies are hosting them so the price goes down as opposed to a monopoly) in a perfect world yes a company providing a proprietary model builds out a very flexible stack and makes the model super cheap that it's almost as"
X Link 2025-09-23T03:44Z 81.9K followers, [---] engagements
"The bitter lesson doesn't work if you have bad data especially the case if you're working in the sciences a conversation with my friend (who is a biotech founder) made me realize something today: massive computational resources in big tech can create a kind of inertia where teams lean on brute-force compute rather than pursuing precise ground-truth data when you have the amount a conversation with my friend (who is a biotech founder) made me realize something today: massive computational resources in big tech can create a kind of inertia where teams lean on brute-force compute rather than"
X Link 2025-09-24T08:49Z 81.3K followers, [----] engagements
"Alibaba is on a roll with Qwen they just keep shipping incredible models 🚀 Qwen3-Max is hereno preview just power Qwen Chat:https://t.co/FBpr7zfQY6 Blog: https://t.co/jJJcfi5FJJ API: https://t.co/olURJV1Enl Weve supercharged coding & agentic skillsnow Qwen3-Max-Instruct without thinking rivaling top models on SWE-Bench Tau2-Bench https://t.co/ZIL08Akm24 🚀 Qwen3-Max is hereno preview just power Qwen Chat:https://t.co/FBpr7zfQY6 Blog: https://t.co/jJJcfi5FJJ API: https://t.co/olURJV1Enl Weve supercharged coding & agentic skillsnow Qwen3-Max-Instruct without thinking rivaling top models on"
X Link 2025-09-24T08:59Z 81.7K followers, [----] engagements
"Reinforcement Learning on Pre-Training Data "RLPT enables the policy to autonomously explore meaningful trajectories to learn from pre-training data and improve its capability through reinforcement learning" "it adopts a next-segment reasoning objective rewarding the policy for accurately predicting subsequent text segments conditioned on the preceding context." "RLPT can be applied to a base model after next-token pre-training but it requires a minimum level of instruction-following ability to initiate next-segment reasoning.""
X Link 2025-09-24T10:10Z 81.4K followers, [----] engagements
"The Illusion of Readiness: Stress Testing Large Frontier Models on Multimodal Medical Benchmarks "We caution that medical benchmark scores do not directly reflect real-world readiness." "Leading systems often guess correctly even when key inputs like images are removed flip answers under trivial prompt changes and fabricate convincing yet flawed reasoning. These aren't glitches; they expose how today's benchmarks reward test-taking tricks over medical understanding. We evaluate six flagship models across six widely used benchmarks and find that high leaderboard scores hide brittleness and"
X Link 2025-09-24T10:27Z 81.7K followers, 52.3K engagements
"Isn't it interesting that Apple is working on protein folding 🤔 SimpleFold: Folding Proteins is Simpler than You Think "we introduce SimpleFold the first flow-matching based protein folding model that solely uses general purpose transformer blocks. Protein folding models typically employ computationally expensive modules involving https://t.co/x7R7Gu6sSZ SimpleFold: Folding Proteins is Simpler than You Think "we introduce SimpleFold the first flow-matching based protein folding model that solely uses general purpose transformer blocks. Protein folding models typically employ computationally"
X Link 2025-09-24T10:39Z 81.7K followers, 56.7K engagements
"I've talked about how bad LLM evals for medicine are but it's so much worse for multimodal LLMs lol for example on one multimodal benchmark (NEJM) GPT-5 accuracy goes [-----] [-----] if you remove the medical images. That is the performance barely changes when you remove the medical images And this benchmark is supposed to test multimodality performance of a model I am actually relatively comfortable with people using LLMs to answer medical related questions. It's not perfect but not terrible either. But I would NEVER recommend you use the current generation of multimodal LLMs to interpret your"
X Link 2025-09-24T19:28Z 81.9K followers, 48.2K engagements
"Yang Song one of the world's top diffusion model researchers and inventor of consistency models has left OpenAI to join Meta"
X Link 2025-09-25T05:39Z 81.6K followers, 73.9K engagements
"Mickey Mouse is very impressed by Kimi"
X Link 2025-09-25T09:43Z 81.8K followers, [----] engagements
"RLBFF: Binary Flexible Feedback to bridge between Human Feedback & Verifiable Rewards "We propose Reinforcement Learning with Binary Flexible Feedback (RLBFF) which combines the versatility of human-driven preferences with the precision of rule-based verification enabling reward models to capture nuanced aspects of response quality beyond mere correctness. RLBFF extracts principles that can be answered in a binary fashion (e.g. accuracy of information: yes or code readability: no) from natural language feedback.""
X Link 2025-09-26T10:19Z 81.6K followers, [----] engagements
"@probflow @SophontAI lol I forgot to post it"
X Link 2025-09-27T12:03Z 81.5K followers, [---] engagements
"Sad to see people waste their talents working on harmful technologies (imo) like this when they could use their talents to make a positive difference. From biomedical AI to educational tools so many ways to help the world as AI engineers and yet people choose to do this We can collectively decide to not poison our children. This isn't difficult choice. Please make this happen if you are in position of power. https://t.co/DjKbBYdmjZ We can collectively decide to not poison our children. This isn't difficult choice. Please make this happen if you are in position of power. https://t.co/DjKbBYdmjZ"
X Link 2025-09-27T12:04Z 81.7K followers, 62.3K engagements
"practical modern GRPO tweaks as described in Meta's Code World Models paper"
X Link 2025-09-28T11:53Z 82.3K followers, 243.3K engagements
"you can make literally anything in Minecraft. next up: AGI this is beyond mindblowing for me. somebody built a [--] million param language model inside minecraft trained it equipped it with basic conversational ability. probably the best thing i have seen entire month. https://t.co/2VOAZylAH1 this is beyond mindblowing for me. somebody built a [--] million param language model inside minecraft trained it equipped it with basic conversational ability. probably the best thing i have seen entire month. https://t.co/2VOAZylAH1"
X Link 2025-09-29T05:10Z 81.7K followers, [----] engagements
"how am i just now noticing the AstraZeneca logo actually is an A and Z I thought it was some random squiggles 😭😭😭"
X Link 2025-09-29T08:49Z 82K followers, [----] engagements
"DeepSeek introduces a new sparse attention variant called DeepSeek Sparse Attention (DSA) DSA primarily consists of two components: a lightning indexer and a fine-grained token selection mechanism. It leads to significant inference speedups: "DSA reduces the core attention complexity of the main model from O() to O() where ( ) is the number of selected tokens." DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention NEW DEEPSEEK MODEL RELEASE AND PAPER "We introduce DeepSeek-V3.2-Exp an experimental sparse-attention model which equips DeepSeek-V3.1-Terminus with"
X Link 2025-09-29T11:28Z 82.3K followers, 78.7K engagements
"Some relevant details about the RL phase for DeepSeek's latest models: - finetune separate specialist models with RL for mathematics competitive programming general logical reasoning agentic coding and agentic search which are used to generate data for final training - instead of multi-stage RL merge reasoning agent and human alignment training into one RL stage - combination of rule-based outcome reward length penalty language consistency reward but also generative reward model with rubrics for more general tasks DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse"
X Link 2025-09-29T11:41Z 81.8K followers, 23.4K engagements
"so do we just switch between claude code and codex every few months"
X Link 2025-09-29T19:00Z 82K followers, 64.4K engagements
"Incredible blog post by @johnschulman2 The RL results are quite surprising. "LoRA fully matches the learning performance of FullFT when running policy gradient algorithms for reinforcement learning even with ranks as low as 1" I incorrectly assumed that surely RL must be sophisticated enough that the capacity of LoRA would not be enough to get full performance. Turns out that's not the case John walks thru an information theoretic argument that explains that even at rank-1 LoRA has more than enough capacity to absorb all the bits of information provided during typical post-training using"
X Link 2025-09-30T00:52Z 82K followers, 64.9K engagements
"LLaDA-MoE: A Sparse MoE Diffusion Language Model "LLaDA-MoE achieves state-of-the-art performance among diffusion language models with larger parameters surpassing previous diffusion language models LLaDA LLaDA [---] and Dream across multiple benchmarks. The instruct-tuned model LLaDA-MoE-7B-A1B-Instruct demonstrates capabilities comparable to Qwen2.5-3B-Instruct in knowledge understanding code generation mathematical reasoning agent and alignment tasks despite using fewer active parameters""
X Link 2025-09-30T10:28Z 81.8K followers, 44.2K engagements
"the spaghetti eating test (it won't let me do will smith) the prompt is literally "@sama eating spaghetti""
X Link 2025-10-01T02:45Z 81.9K followers, [----] engagements
"@chrisdotai this logic doesn't apply to many apps that get iOS/macOS support first. For example Sora is currently not a paid app"
X Link 2025-10-01T18:55Z 81.8K followers, 12.9K engagements
"@dptru10 why are do their ideal customers use iPhones and Macs"
X Link 2025-10-01T18:56Z 81.9K followers, 39.3K engagements
"@DanAdvantage Except you can't do cameos in web interface"
X Link 2025-10-01T19:32Z 81.9K followers, 23.7K engagements
"BroRL: Scaling Reinforcement Learning via Broadened Exploration "In this work we investigate a complementary paradigm for scaling RL: BroRLincreasing the number of rollouts per example to hundreds to exhaustively Broaden exploration which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps.""
X Link 2025-10-02T11:52Z 82.2K followers, 12.3K engagements
"A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning "multi-turn RL improvements diminish with increasing complexity" "Performance also scales with model size" "agents trained on simpler environments achieve substantial generalization to more complex ones" "multi-turn RL develops generalizable skills that transfer across diverse objectives." "multi-turn RL with good imitation priors achieves comparable performance with dramatically fewer RL episodes" "PPO consistently outperforms RLOO in multi-turn settings with performance gaps increasing for complex environments. " "Dense"
X Link 2025-10-02T12:04Z 82.3K followers, 25.4K engagements
"wait can i log into sora on an iphone and create a cameo then log out and use that cameo from web/android"
X Link 2025-10-02T20:58Z 81.9K followers, [----] engagements
"this continues to reinforce my belief that current frontier models are completely terrible at medical imaging analysis DO NOT TRUST LLM's interpretation of your medical images 🚨 Just published All frontier AI models have failed Radiologys Last Exam - the toughest benchmark in radiology launched today ✅ Board-certified radiologists scored 83% trainees 45% but the best performing AI from frontier labs GPT-5 managed only 30%. ❌ These results https://t.co/cQIPJXJ2eH 🚨 Just published All frontier AI models have failed Radiologys Last Exam - the toughest benchmark in radiology launched today ✅"
X Link 2025-10-02T21:10Z 82K followers, 17.9K engagements
"My take on Sora is that it's a killer product. However the model itself isn't SOTA. Veo [--] is probably better in terms of video quality. But wrapping the model itself in this social media video shorts feed plus the interactive cameo feature makes Sora a much more compelling product. Honestly I'd say this is what OpenAI shines at. From ChatGPT to image editing (Studio Ghibli) and now Sora [--] the models aren't significantly better than what other companies have but they do a much better job of building compelling viral products around the models and bringing them to the masses. As a side note"
X Link 2025-10-03T11:49Z 81.9K followers, 50.3K engagements
"I hope someone actually tests Thinking Machines' hypothesis that you could replicate DeepSeek-R1-Zero which just LoRA (if someone has already tested this please point me to it)"
X Link 2025-10-04T09:32Z 81.9K followers, 52.1K engagements
"This Nobel Prize is for the discovery of regulatory T cells (Treg). Treg cells are immunosuppressive cells that prevent other T-cells from attacking the body's own cells and prevents the development of autoimmune disease. They are also often upregulated in cancer to prevent the immune system from attacking the tumor. Very interesting stuff BREAKING NEWS The [----] #NobelPrize in Physiology or Medicine has been awarded to Mary E. Brunkow Fred Ramsdell and Shimon Sakaguchi for their discoveries concerning peripheral immune tolerance. https://t.co/nhjxJSoZEr BREAKING NEWS The [----] #NobelPrize in"
X Link 2025-10-06T09:55Z 82.3K followers, 47K engagements
"It's actually insane that ChatGPT has hit [---] million weekly active users. That's 10% of the entire world's population 🤯"
X Link 2025-10-06T18:14Z 82.2K followers, 63.5K engagements
"Arduino is how I got started exploring electronics for the first time especially with the SparkFun Inventor Kit (back in [----] IIRC). Learned a lot. Wonder what Qualcomm will do with the company. Were acquiring @Arduino to make edge computing and #AI more accessible. Arduino will remain independent supporting multiple silicon vendors and its community. Launching today: Arduino UNO Q powered by Qualcomm #Dragonwing. Discover more: https://t.co/B7szrg7zC0 https://t.co/hn6zz3qxV0 Were acquiring @Arduino to make edge computing and #AI more accessible. Arduino will remain independent supporting"
X Link 2025-10-08T23:06Z 82K followers, [----] engagements
"h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning "In this work we introduce a scalable method to bootstrap long-horizon reasoning capabilities using only existing abundant short-horizon data. Our approach synthetically composes simple problems into complex multistep dependency chains of arbitrary length. We train models on this data using outcome-only rewards under a curriculum that automatically increases in complexity allowing RL training to be scaled much further without saturating.""
X Link 2025-10-09T10:34Z 82.3K followers, 13K engagements
"The course is how I got started in AI. Few years later in a full circle moment I got the opportunity to contribute & help teach a few lectures of the course too. I still think it's one of the best resources for learning deep learning. Highly recommend http://fast.ai https://t.co/aQsW59Ylt6 combined with learning help from an LLM is more accessible than ever. Great way to learn how AI and deep learning really work at the foundations. https://t.co/VzJswBMym6 http://fast.ai https://t.co/aQsW59Ylt6 combined with learning help from an LLM is more accessible than ever. Great way to learn how AI and"
X Link 2025-10-11T07:31Z 82.3K followers, 59K engagements
"Wow I just reached a new milestone: [----] CITATIONS 🥳 I've been waiting for this one. Onto publishing more research"
X Link 2025-10-12T10:49Z 82.3K followers, 117.4K engagements
"SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models "we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood." "SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K 2.6% in MATH500 18.4% in Countdown and 27.0% in Sudoku.""
X Link 2025-10-13T09:30Z 82.3K followers, 13.5K engagements
"There's a talk on the PyTorch conference schedule about TorchForge a new PyTorch-native library purpose-built for scalable RL post-training and agentic development. However looks like the library hasn't been released yet. It sounds really cool I am excited to try it out"
X Link 2025-10-15T02:09Z 82.3K followers, [----] engagements
"Had a fun time @SacHackerLab -drinking from the 3D printed cup & sis @iCatLuvr playing with Eiffel Tower http://t.co/8oVzwgf4yE"
X Link 2013-11-01T23:56Z 62.2K followers, [--] engagements
"Enjoying @sacstate #hornets game waiting for our Half-time jazz dance with @hornet_girlz & our dance class @ARCPTK http://t.co/VOMrzudMX1"
X Link 2014-10-05T01:43Z 62.2K followers, [--] engagements
"These #millennials are more successful than you http://t.co/2vPpWLTinC @CNBC @MorrisAtLarge thanks for recognizing me as'The new #CarlSagan' http://cnb.cx/1vIQfGE http://cnb.cx/1vIQfGE"
X Link 2014-10-06T16:46Z 80.4K followers, [--] engagements
"#Zoology class - from rats cats whales to humans Studying real human bones in zoology #anatomy #osteology will be fun in med school"
X Link 2016-01-29T03:41Z 62.1K followers, [--] engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing