You don’t need a PhD to be a great AI researcher. Even @OpenAI’s Chief Research Officer doesn’t have a PhD.

The odds of OpenAI shipping GPT-5 just went up

Give someone a fish, and you feed them for a day; teach someone to fish, and you feed them for a lifetime. Elevating from Weak to Strong with Self-Play Fine-Tuning (SPIN) for All LLMs. Empower, Evolve, SPIN! arxiv.org/pdf/2401.01335.pdf

DeepMind just trained a 270M transformer that can play like a grandmaster without searching through moves (MCTS), It was trained on Stockfish’s assessment of ~10M games so it didn’t really learn the game from scratch. Impressive that transformers generalize to a “logic” task

🚀🚀

Llama 4 Maverick was 400B(70B active) and >30T tokens = 429 tokens / active param Llama 4 Behemoth was 2T(288B active) and > 30T tokens = 104 tokens / active param DeepSeek v3 is 671B(37B active) and 14.8T tokens = 400 tokens / active param Kimi K2 is 1T(32B active) and 15.5T tokens = 484 tokens / active param

trying GPT-4.5 has been much more of a "feel the AGI" moment among high-taste testers than i expected!

X is a $44 billion propaganda machine. Yet it attempts to disguise itself as a defender of unfettered free speech, a source of factual information, and a substitute for professional journalism.

Reminder instruct-gpt (OpenAI) was trained on 16 epochs (even after overfitting on just 1). SFT for LLMs is wild

Popping this up: a response to a question about what I consider reasoning & planning, why current Auto-Regressive LLMs can't do it, why that would require AI systems with world models, and why we still have a lot of progress to do towards AI systems that can learn and reason.

This is yet another reminder that if you are like freshman into ML, before you question: "hmm... maybe this part of transformer is useless/can be replaced?" just read these paper. arxiv.org/abs/2109.08668v2 arxiv.org/abs/2001.08361 arxiv.org/abs/2002.05202 arxiv.org/abs/2404.05405 and realize there is practically no 'fundamental' aspect of current transformer that can give you dramatic benefit. You can speed up training time but arch probably aint it.

We achieved gold medal level performance on this year's IMO! Our model thinks and writes proofs in clear, plain‑English - no formal code required. Unlike the narrower systems used in past competitions, our model is built to reason broadly, far beyond contest problems.

DeepSeek has now moved up to the second spot from third since this morning among all of the top iPhone apps!

what yann lecun said: "the more tokens an llm generates, the more likely it is to go off the rails and get everything wrong" what actually happened: "we get extremely high accuracy on arc-agi by generating billions of tokens, the more tokens we throw at it the better it gets"

1/ Introducing “Tensor Product Attention Is All You Need” (TPA) and Tensor ProducT ATTenTion Transformer (T6)! 🚀 Ever wondered if there’s a more memory-efficient way to handle long contexts in LLMs? Homepage: tensorgi.github.io/T6

SB 1047 has been vetoed in California by @GavinNewsom—Thank you! This is a huge relief and a win for the future of tech innovation in the state.

Another triumph for Self-Play! Self-Play Preference Optimization (SPPO) has surpassed (iterative) DPO, IPO, Self-Rewarding LMs, and others on AlpacaEval, MT-Bench, and the Open LLM Leaderboard. Remarkably, Mistral-7B-instruct-v0.2 fine-tuned by SPPO achieves superior performance to GPT-4 0613 without relying on any GPT-4 responses. Explore the roadmap of LLM fine-tuning techniques: Supervised Fine-tuning: SFT --> SPIN Preference Fine-tuning: PPO --> DPO --> SPPO Paper: arxiv.org/pdf/2405.00675

Becoming a better researcher is easy. Just spend your time like so: 40% reading papers 40% writing code 40% coming up with new ideas 40% studying textbooks 40% talking to other researchers

In the old days, the term “RL” by default meant the “true RL” used in AlphaZero. Now, the term “RL” by default means the “fake RL” used in RLHF (nothing is negative about RLHF here. RLHF is a great innovation).

1/8 ⭐General Preference Modeling with Preference Representations for Aligning Language Models⭐ arxiv.org/abs/2410.02197 As Huggingface Daily Papers: huggingface.co/papers/2410.0… We just dropped our latest research on General Preference Modeling (GPM)! 🚀

1/6 We introduce RPG, a principled framework for deriving and analyzing KL-regularized policy gradient methods, unifying GRPO/k3-estimator and REINFORCE++ under this framework and discovering better RL objectives than GRPO: Paper: arxiv.org/abs/2505.17508 Code: github.com/complex-reasoning… Webpage: complex-reasoning.github.io/… @yifanzhang_, @HuiZhuoY, @QuanquanGu

Introducing… the winners of this year’s Sloan Research Fellowship! These extraordinary researchers represent some of the most exciting young minds working today—and we are thrilled to support them. Meet the winners here: sloan.org/fellowships/2022-F… #SloanFellow

GPT-5 failed twice. Scaling laws are coming to an end. Open-source AI will have the Mandate of Heaven.

Mitigating racial bias from LLMs is a lot easier than removing it from humans! Can’t believe this happened at the best AI conference @NeurIPSConf We have ethical reviews for authors, but missed it for invited speakers? 😡

No, these 70B models are not better than GPT-4

See below on what Zuckerberg is looking for in star recruits worth $100m pay packages for Meta’s plans in Artificial Intelligence. But weren’t some people saying calculus is no longer useful in the AI age? 🤔

LoRA makes fine-tuning more accessible, but it's unclear how it compares to full fine-tuning. We find that the performance often matches closely---more often than you might expect. In our latest Connectionism post, we share our experimental results and recommendations for LoRA. thinkingmachines.ai/blog/lor…

New paper by Google provides evidence that transformers (GPT, etc) cannot generalize beyond their training data

What's your preferred optimizer for LLMs?

🚀 Introducing DeepSeek-V3! Biggest leap forward yet: ⚡ 60 tokens/second (3x faster than V2!) 💪 Enhanced capabilities 🛠 API compatibility intact 🌍 Fully open-source models & papers 🐋 1/n

Give someone a fish, and you feed them for a day; teach someone to fish, and you feed them for a lifetime. Elevating from Weak to Strong with Self-Play Fine-Tuning (SPIN) for All LLMs. Empower, Evolve, SPIN! arxiv.org/pdf/2401.01335.pdf

Fantastic Pretraining Optimizers and Where to Find Them "we conduct a systematic study of ten deep learning optimizers across four model scales (0.1B-1.2B parameters) and data-to-model ratios (1–8× the Chinchilla optimum)." "we find that all the fastest optimizers such as Muon and Soap, use matrices as preconditioners" "However, the speedup of matrix-based optimizers is inversely proportional to model scale, decreasing from 1.4× over AdamW for 0.1B parameter models to merely 1.1× for 1.2B parameter models." Observations made in the paper: 1. Hyperparameter transfer between optimizers is non-trivial. 2. The speedup of new optimizers is lower than claimed and diminishes with model size. 3. Early-stage loss curves can mislead significantly. 4. Matrix-based optimizers consistently outperform scalar-based optimizers for small models. 5. Optimal choice of optimizer shifts depends on data-to-model ratios.

(1/n)🚀 Introducing Accelerated Preference Optimization (APO) for Reinforcement Learning from Human Feedback (RLHF)! 🎯 RLHF plays a central role in aligning large language models (LLMs) with human preferences. Direct Preference Optimization (DPO), one of the most popular approaches, simplifies this process by avoiding explicit reward estimation. But what if we could make it even faster and more efficient? The answer: momentum! ⚡ 💡 APO Framework: We introduce Accelerated Preference Optimization (APO), a general framework that unifies several preference optimization methods and incorporates Nesterov's momentum to accelerate the convergence process. By framing preference optimization as a proximal point method, we demonstrate APO's ability to speed up the RLHF process. 📊 Theoretical Insights: APO achieves a faster convergence rate compared to standard methods like iterative DPO and Self-Play Preference Optimization (SPPO), which a significant improvement in the efficiency of training LLMs. 🚀 Empirical Results: On the AlpacaEval 2.0 benchmark, APO outperforms DPO, iterative DPO, and other strong baselines in aligning LLMs with human preferences. 🔗 Paper: arxiv.org/pdf/2410.02197

1/ Very excited to present cryoSTAR, a novel approach for continuous heterogeneity in SPA cryo-EM. In brief, cryoSTAR leverages the prior knowledge from a user-given atomic model to better find dynamics in the final reconstruction. bytedance.github.io/cryostar… biorxiv.org/content/10.1101/…

The question I got asked most frequently during COLM this year was what research questions can be studied in academia that will also be relevant to frontier labs. So I’m making a talk for this. What topics / areas should I cover? RL/eval/pretraining,?

Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost. thinkingmachines.ai/blog/on-…

If you don’t think conference codes of conduct are bad for free speech, read this.

TrustLLM: Trustworthiness in Large Language Models paper page: huggingface.co/papers/2401.0… Large language models (LLMs), exemplified by ChatGPT, have gained considerable attention for their excellent natural language processing capabilities. Nonetheless, these LLMs present many challenges, particularly in the realm of trustworthiness. Therefore, ensuring the trustworthiness of LLMs emerges as an important topic. This paper introduces TrustLLM, a comprehensive study of trustworthiness in LLMs, including principles for different dimensions of trustworthiness, established benchmark, evaluation, and analysis of trustworthiness for mainstream LLMs, and discussion of open challenges and future directions. Specifically, we first propose a set of principles for trustworthy LLMs that span eight different dimensions. Based on these principles, we further establish a benchmark across six dimensions including truthfulness, safety, fairness, robustness, privacy, and machine ethics. We then present a study evaluating 16 mainstream LLMs in TrustLLM, consisting of over 30 datasets. Our findings firstly show that in general trustworthiness and utility (i.e., functional effectiveness) are positively related. Secondly, our observations reveal that proprietary LLMs generally outperform most open-source counterparts in terms of trustworthiness, raising concerns about the potential risks of widely accessible open-source LLMs. However, a few open-source LLMs come very close to proprietary ones. Thirdly, it is important to note that some LLMs may be overly calibrated towards exhibiting trustworthiness, to the extent that they compromise their utility by mistakenly treating benign prompts as harmful and consequently not responding. Finally, we emphasize the importance of ensuring transparency not only in the models themselves but also in the technologies that underpin trustworthiness. Knowing the specific trustworthy technologies that have been employed is crucial for analyzing their effectiveness.

This is an unwise statement that can only make people confused about what LLMs can or cannot do. Let me tell you something: Math is NOT about solving this kind of ad hoc optimization problems. Yeah, by scraping available data and then clustering it, LLMs can sometimes solve some very minor math problems. It's an achievement, and I applaud you for that. But let's be honest: this is NOT the REAL Math. Not by 10,000 miles. REAL Math is about concepts and ideas - things like "schemes" introduced by the great Alexander Grothendieck, who revolutionized algebraic geometry; the Atiyah-Singer Index Theorem; or the Langlands Program, tying together Number Theory, Analysis, Geometry, and Quantum Physics. That's the REAL Math. Can LLMs do that? Of course not. So, please, STOP confusing people - especially, given the atrocious state of our math education. LLMs give us great tools, which I appreciate very much. Useful stuff! Go ahead and use them AS TOOLS (just as we use calculators to crunch numbers or cameras to render portraits and landscapes), an enhancement of human abilities, and STOP pretending that LLMs are somehow capable of replicating everything that human beings can do. In this one area, mathematics, LLMs are no match to human mathematicians. Period. Not to mention many other areas. Calling on my friend @ericweinstein and @GaryMarcus, who has been one of the few sane expert voices on these matters lately. 🙏 h/t @hellheff

We've just open-sourced the code and data for Self-play Fine-Tuning (SPIN)! Time to SPIN every model out there! 🚀🚀🚀 Code: github.com/uclaml/SPIN Data: huggingface.co/collections/U… Models: huggingface.co/collections/U… Project Page: uclaml.github.io/SPIN/ Many thanks to @Yihe__Deng, @HuizhuoY, and @Kaixuan_Ji_19 for their tremendous efforts in preparing these.

🔥Excited to share our recent research on query-efficient RLHF! Introducing Active Direct Preference Optimization (ADPO), a new approach that improves DPO performance on Open-LLM-Benchmark with just half the queries. Discover how ADPO eliminates the significant demand for querying preference labels.🚀[1/4] Paper: arxiv.org/pdf/2402.09401.pdf A joint work with @JiafanHe , and @QuanquanGu 👏

Deep Think in 2.5 Pro has landed. 🤯 It’s a new enhanced reasoning mode using our research in parallel thinking techniques - meaning it explores multiple hypotheses before responding. This enables it to handle incredibly complex math and coding problems more effectively.

Personal news: I have graduated from UCLA and officially become Dr. Xu. I will join Duke University @DukeU as a tenure track assistant professor in the Department of Biostatistics and Bioinformatics in August 2022. In the meantime, I will be a CAST Postdoctoral Fellow at Caltech.

🎯 Just released a new preprint that proves LR transfer under μP. -> The Problem: When training large neural networks, one of the trickiest questions is: what learning rate should I use? [1/n]🧵 Link: arxiv.org/abs/2511.01734

1/ Introducing “Tensor Product Attention Is All You Need” (TPA) and Tensor ProducT ATTenTion Transformer (T6)! 🚀 Ever wondered if there’s a more memory-efficient way to handle long contexts in LLMs? Homepage: tensorgi.github.io/T6

Paper - "LLMs Will Always Hallucinate, and We Need to Live With This" Podcast format generated with Google's new illuminate tool (illuminate is trained to produce short podcast from research papers)

My prediction for the next bio/ML trend at NeurIPS. DPO and RLHF for protein design. Protein language models in particular. 😉

🚀 Day 1 of #OpenSourceWeek: FlashMLA Honored to share FlashMLA - our efficient MLA decoding kernel for Hopper GPUs, optimized for variable-length sequences and now in production. ✅ BF16 support ✅ Paged KV cache (block size 64) ⚡ 3000 GB/s memory-bound & 580 TFLOPS compute-bound on H800 🔗 Explore on GitHub: github.com/deepseek-ai/Flash…

Doing research in a ChatGPT Group Chat with friends*, alternating between proofs and simulations, so so much fun ... *I refrain from using the term "vibes research", because obviously every step eventually needs to be very carefully checked!

Terence Tao makes $700k a year, what’s your excuse?

Anyone who has trained a Transformer has viscerally felt its O(T^2) cost. It is not tractable to train Transformers end-to-end on long contexts. Here's a writeup of the research direction I believe is most likely to solve this: linear transformers. manifestai.com/blogposts/fas… 1/7

The hottest thing a man can be is exceptionally good at math

Announcing Stable Diffusion 3, our most capable text-to-image model, utilizing a diffusion transformer architecture for greatly improved performance in multi-subject prompts, image quality, and spelling abilities. Today, we are opening the waitlist for early preview. This phase is crucial for gathering insights to improve its performance and safety ahead of open release. You can sign up to join the waitlist and learn more here: bit.ly/3OR2qQF #stablediffusion3 Prompt: Epic anime artwork of a wizard atop a mountain at night casting a cosmic spell into the dark sky that says "Stable Diffusion 3" made out of colorful energy

Introducing Gemma: a family of lightweight, state-of-the-art open models for developers and researchers to build with AI. 🌐 We’re also releasing tools to support innovation and collaboration - as well as to guide responsible use. Get started now. → dpmd.ai/3UJu1Y1

🚀 Day 0: Warming up for #OpenSourceWeek! We're a tiny team @deepseek_ai exploring AGI. Starting next week, we'll be open-sourcing 5 repos, sharing our small but sincere progress with full transparency. These humble building blocks in our online service have been documented, deployed and battle-tested in production. As part of the open-source community, we believe that every line shared becomes collective momentum that accelerates the journey. Daily unlocks are coming soon. No ivory towers - just pure garage-energy and community-driven innovation.

Another triumph for Self-Play! Self-Play Preference Optimization (SPPO) has surpassed (iterative) DPO, IPO, Self-Rewarding LMs, and others on AlpacaEval, MT-Bench, and the Open LLM Leaderboard. Remarkably, Mistral-7B-instruct-v0.2 fine-tuned by SPPO achieves superior performance to GPT-4 0613 without relying on any GPT-4 responses. Explore the roadmap of LLM fine-tuning techniques: Supervised Fine-tuning: SFT --> SPIN Preference Fine-tuning: PPO --> DPO --> SPPO Paper: arxiv.org/pdf/2405.00675

I'm teaching a new "Intro to Modern AI" course at CMU this Spring: modernaicourse.org. It's an early-undergrad course on how to build a chatbot from scratch (well, from PyTorch). The course name has bothered some people – "AI" usually means something much broader in academic contexts – but I think the time has come where the first thing that many students interested in AI should see is how the AI they are familiar with actually works (because it's really simple!) The more people who understand it the better. I'll be trying to put as much material as I can that we develop online (assignments + autograding, hopefully lecture videos), though as a first-time course there are also likely to be some bumps along the way. Hopefully it becomes a good resource over time, though. Feedback welcome.

Imagine where Microsoft would be now if they hadn't destroyed the amazing WizardLM team, causing them to leave and go to Tencent… …where they're now releasing models far beyond anything Microsoft has created.

🔗 Thoughts on Research Impact in AI. Grad students often ask: how do I do research that makes a difference in the current, crowded AI space? This is a blogpost that summarizes my perspective in six guidelines for making research impact via open-source artifacts. Link below.

Qwen3-30B-A3B is de facto on par with Qwen3-32B dense and the greatest vindication of finegrained MoEs the world has seen in the open.

People studying misinformation lean left for two reasons: 1. scientists lean left, regardless of specialty, because they care about facts. 2. misinformation today primarily comes from the Right ("they're eating the dawwwgs!") which makes it worth studying and fighting against for people leaning left.

🚨BREAKING; APPLE HELD INTERNAL TALKS ABOUT BUYING PERPLEXITY