People keep asking me: what's different about optimization in RL? Seemingly nothing — the pre-training stack just works (Adam, even SGD 👀 @saagnikkk). Bringing some answers from my last work (sorry for the delay — been cooking 🚀). We introduce ISO: Isospectral Optimization: an RLVR-native optimization stack. Built on one simple observation, spectral inheritance: RLVR can reuse the base model's spectrum and acquire new behavior purely through the singular frames. 🧩 Offline: ISO-Merger — consolidates RL experts into one model with no data, no rollouts, no OPD. Checkpoints only. ⚙️ Online: ISO-Optimizer — a drop-in wrapper on AdamW / Muon that matches AdamW's accuracy with ~2.7× fewer steps on Qwen3-8B-Base. 📄 🌐 🧵👇

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models.

hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)

Just want to add that @tydsh himself is a great example of down-to-earth do-er! Last week he asked me for some model checkpoints, and immediately ran a bunch of experiments to discover some important findings. You can’t be a great LLM researcher without being down-to-earth! 🫡

Moonshot AI, the company behind Kimi, has four core founders. Their backgrounds are unusually strong: - Founder and CEO Yang Zhilin studied computer science at Tsinghua before earning his PhD from Carnegie Mellon. He was the first author of Transformer-XL and XLNet, and previously worked at FAIR and Google Brain. - Co-founder and CTO Zhang Yutao earned his PhD in computer science from Tsinghua. His earlier work covered knowledge graphs and AMiner, and he previously co-founded Recurrent AI with Yang. - Co-founder Wu Yuxin studied at Tsinghua and CMU before joining FAIR. He worked with Kaiming He on Group Normalization and also created Detectron2. - Co-founder Zhou Xinyu studied computer science at Tsinghua and later joined Megvii, where he worked on turning research algorithms into production systems and co-authored ShuffleNet. They all share one root: Tsinghua University. Tsinghua is widely regarded as one of China’s top universities. In the latest U.S. News Best Global Universities ranking, it reached No. 6 worldwide. Its influence on China’s AI industry extends well beyond Moonshot. the company behind the GLM models, also grew out of Tsinghua. Its co-founder and chief scientist, Tang Jie, was once Yang Zhilin’s teacher. There is also a more personal connection. Yang and Zhou formed a rock band together at Tsinghua. Moonshot AI’s Chinese name, 月之暗面, comes from Pink Floyd’s album The Dark Side of the Moon, one of Yang’s favorites. Kimi may look like a young AI company. Behind it is a much older network of classmates, teachers, research labs, and friendships.

Recursive just came out of stealth, and the team has been cooking 🔥 Our first results: an automated AI research system that can improve AI across 3 very different settings across training and GPU kernel optimization.

AI is now doing our AI research. At Recursive we set out to build recursive self-improving superintelligence (RSI) to automate knowledge discovery. The best way to expand humanity’s knowledge is through the scientific method. RSI leads to better ideas, explanations and inventions which lead to better RSI. Automating the scientific method requires closing the loop between ideation, implementation and validation, and being able to run it over extended periods of time. Today, we are excited to share the first outputs of Recursive’s automated open-ended discovery system. To be clear, this system is merely a milestone towards RSI, a v0.1 of what I sometimes call the “Eureka Machine”. It is one program that you can point at any hard problem and get useful inventions out. Though it’s still very early, we've run it on three AI tasks and achieved state-of-the-art results on all three. These results demonstrate that even this early version of the system can solve a variety of autoresearch problems in AI and improve over prior state of the art. Concretely, it did this on the community benchmarks NanoGPT speedrun, NanoChat, and NVIDIA's Sol-ExecBench. AI is code and AI can code. The code and ideas that lead to these results were not invented by our team but by the AI system itself. To do RSI safely, we need to investigate its inventions. That's best done transparently with the community. @Recursive_SI we are open-sourcing the system’s discoveries, demonstrating that it finds creative and benign solutions instead of focusing on obvious optimizations or dangerous ideas. Link below.