Mark Zuckerberg gets it right: “The defining question of our age isn’t whether superintelligence will exist, but who will have access to it. Will it be centralized and restricted to a few institutions, or will it be a tool that empowers everyone?” Concentration of power is the biggest risk of AI. When a small number of labs (working hand-in-glove with the administrative state) decide who has access to which model capabilities, they inevitably shape what can be said, known, and built. That’s not “safety.” It’s control. As Mark points out, the history of open source shows that broad access and transparency are usually the best path to actual security and resilience. Decentralization creates checks and balances on power. By contrast, centralized alternatives, like bureaucratic approval regimes and mandatory gatekeeping, typically produce regulatory capture and reinforce cartels. Personal superintelligence in everyone’s hands, with competing models and real data sovereignty, is a far better check on a dystopian future than self-appointed guardians who claim to be “aligned” with all of humanity.

Hey sir. We are not asking you to open source Anthropic. Just don’t lobby the government to shut down others who do. Jensen never framed other chips as “dangerous” or decides who can use CUDA based on who’s “safe”.

It continues to boggle the mind how many people, who otherwise seem to have a functioning intellect, appear to lose all cognitive capacity when it comes to thinking about actions that impact the company that pays them.

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.

I still remember the days when it was basically us & a few other people pushing this topic (on the top of my head @allen_ai @ylecun @jeremyphoward @soumithchintala @BlancheMinerva @fchollet @GuillaumeLample,…). We’ve gone a long way for it to be the ceo of the most valuable company in the world’s first tweet!

We made Claude Code and Codex surf docs and counted every 404 they hit before finding the answer. We served the same content 4 ways: raw HTML, plain markdown, markdown + an llms.txt index, and markdown with the llms.txt inlined on every page. We found that agents reach for markdown and llms.txt on their own, and HTML-only finished last by a wide margin (15-30x more 404s than markdown + llms.txt).

Growth of AI data labelling industry in China: If your mental model is distillation is the major route for getting data for China based labs, where do you think SeeDance 2 of ByteDance got their data? They certainly did not distill SORA 2. SeeDance 2 is much superior to Sora 2. ByteDance has large internal data labelling operation curating data for their models. SeeDance is closed source, but the point to note is they curate their own data and are world leader in that space. China has both: AI data labelling companies like Datatang and teams at large companies doing it in-house like at ByteDance, Alibaba, Baidu, Tencent etc. Furthermore, there are large crowd sourcing platforms like Baidu Crowd and Alibaba Cloud PAI-iTAG. In fact, data labelling is one of the fastest growing job roles in China. Quote from Global Times: "According to the Digital China Development Report (2025) released by the National Data Administration in May, seven data annotation bases have been established in the country, with a total workforce of 95,000 data annotation professionals as of the end of 2025" Contrary to perception - data is not so difficult to get/curate. It requires lot of diligence though and may cost a lot. While in the west Mercor, Turing etc. have rebranded themselves as "Intelligence Clouds" they have roots in IT body shopping/providing elite IT human resources. China has multiple such companies. These companies provide the human labor for AI tasks that works on hourly basis. For example: 1. Proginn (程序员客栈 - Chuangyeyuan Kezhan) 2. Yuanwuxian (猿森林 / 猿五线) & Jiedan Platforms They provide highly technical, engineering-grade human resources for the most complex stages of the AI testing and alignment lifecycle Also, companies all over the world are accessible to labs from China (closed, as well as, open source). There are no trade restrictions on buying and selling data, as far as, I know. This is a crowded space and many providers are looking for buyers. Let us review task types and how to get data for them: Verifiable domains: For post training particularly, there are already vast amount of problem statements and solutions in millions of text books. Also, synthetically generating data (e.g. introducing bugs in programs) is an effective strategy. This works for verifiable domains like Code. Math problem statements can also be generated systematically and their solutions can be verified automatically. Knowledge work domains: As for tasks like GDPEval etc. (e.g. spreadsheet making, making presentations etc.) you do need experts curating problem statements and ideal solutions; but it is not a rocket science. Complex problems in science and engineering, you need to hire bunch of professors and phds (e.g. SpaceX hires Olympiad winners globally) to make AI models fail and discover their weaknesses and design tasks to train them to remove the weaknesses. China has more of them that the entire developed world combined. They also have very diverse expertise due to China's vast industrial base. RL environments can be somewhat complex to make: you have build mock web services, mock apps etc. But, with strong models these days such environments can be designed with ease. Collecting computer use data and egocentric data is very very human labor intensive. There too China has a huge advantage, with large number of youth available, it is easy and not that expensive. The reason you don't hear a lot about Chinese data labelling companies is the same why you don't hear a lot about Chinese IT service companies: they have limited themselves to the mainland where they have ample business. E.g. Chinasoft International has 80K employees, but I bet most haven't heard about them. Also, many large Chinese tech companies have their own data labelling units or they directly manage their contractors. One of the reason one should care about this topic is because, safe open source AI is important for a more just & equitable world.

Congratulations to Zhilin Yang, founder and CEO of @Kimi_Moonshot, on the latest Kimi release. What a huge win for the open-source community! It feels like just yesterday Zhilin was graduating from my lab at CMU, jointly co-advised with William Cohen. Not only did he complete his Ph.D. in just four years, but he also made truly fundamental contributions to ML during his time at CMU. What a spectacular career! Congrats again Zhilin, and thank you and the entire Kimi team for everything you're doing for the open-source community.

Big news: Kimi-K3 by @Kimi_Moonshot is now #1# in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18# -> #1#). In Frontend, Kimi-K3 ranked #1# in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2# only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!

I don't think Anthropic realizes how disruptive these changes are to users. I appreciate the extension, but please stop playing games. Either keep it under the subscriptions or put it under the API already.

Obvious in retrospect, but I didn't really anticipate: Fable: Performing Final Review of <Awesome Feature> Also Fable: I appear to have introduced a critical security vulnerability. <This model's safeguards flagged this message.> Opus 4.8: Doesn't look like anything to me.

Is Muon as good as they say? We looked beyond training speed and found a hidden cost: Muon loses the simplicity bias of older optimizers like gradient descent — and this matters for generalization.

Introducing SWE-Together: a multi-turn benchmark built from real user–agent coding sessions. Coding agents are often benchmarked like exam-takers: given the full spec up front, then graded on the final code. But real coding help is a conversation — users clarify goals, add constraints, and correct course along the way. SWE-Together turns real coding work into a reproducible, verifiable benchmark: 109 repo-level tasks curated from 11,260 recorded sessions, replayed with a reactive LLM user simulator that preserves the original user’s intent. We evaluate agents as collaborators, not just patch generators: final pass rate and how many user interventions were needed to get there. In this evaluation snapshot, claude-opus-4.8 currently leads among the 7 agents we tested — achieving the highest pass rate while requiring the fewest user interventions. 📄 Paper: 💻 Code: 🌐 Website:

People are missing out on how big a deal Longcat 2.0 by Meituan (aka "Chinese Doordash") is. Near frontier performance, trained on 50k Chinese domestic accelerators! The first ever to achieve this!