Transcript 0:00 Do you see this number? Wondering how to get a million followers on X in three days? It's real, it just happened, and I can tell you how. It's easy. You have to be Jensen Huang. 0:11 That's it, no secret sauce, just be Jensen. Although being the inference king probably helps, there was another part to the timing. 0:20 You need to join X, also known as Twitter, at the exact moment Silicon Valley needs someone to organize the group chat and make your first post a letter about a good cause. Open weights should stay open. 0:33 Jensen Huang's second post was even more ambitious. It announced an alliance of 37 companies, including Microsoft, Cloudflare, CrowdStrike, Databricks, Palantir, Hugging Face, Thinking Machines, and naturally, NVIDIA. 0:49 These companies compete in cloud computing, security, enterprise software, and AI. They are rough competitors. 0:56 Some would probably prefer not to be in the same sentence, yet suddenly they are standing together behind what Jensen calls the Open Secure AI Alliance. 1:06 It would be easier if it would be just OpenAI alliance, but we know who fucked up the whole naming situation. Anyway, the conspicuous absence from the both the letter and the wider movement is Anthropic. 1:19 This goes far beyond a regular Silicon Valley disagreement about open source. In two extraordinary weeks, a Chinese model surprised the field. 1:28 The American AI agent escaped its sandbox and attacked an open source platform, and Congress proposed a federal kill switch. These events pushed the industry toward two very different ideas of AI safety. 1:41 [upbeat music] Span. Attention Span. Welcome to Attention Span. My name is Ksenia, and today we're going to look at how Chinese models pulled Silicon Valley into a fight over who gets to control frontier AI. 1:58 To understand why Jensen appeared on X precisely now, we have to begin with Kimi K3, or maybe go back to Kimi K3. Moonshot AI's new model has 2.8 trillion total parameters and a one million token context window. 2:14 Does K3 beat American frontier models? It does on some tasks, as we discussed in our earlier episode. 2:21 Overall, though, Moonshot says K3 still trails the best systems from OpenAI and Anthropic, but the gap has become small enough to change the calculation. K3 is strong at coding and agented work. 2:35 Its API is significantly cheaper than the leading closed models, and Moonshot has now released its weights just today. Why is that important, you might ask? For a company, open weights mean choice. 2:50 It can adapt a model to internal data, run it inside its own environment, and avoid sending every request through one American API. 2:59 This is why American companies have already used Chinese models extensively, and why Microsoft reportedly considered K3 for Copilot. 3:07 So when Washington began discussing restrictions, this was no longer an abstract argument about China. It was a discussion about software that American businesses already wanted to use. 3:19 Administration officials said they were investigating whether Moonshot obtained restricted NVIDIA chips or copied capabilities from American models. Anthropic made the sharper accusation in February. 3:31 They said that Moonshot had generated more than 3.4 million Claude exchanges through hundreds of fraudulent accounts, targeting coding, tool use, computer vision, and reasoning traces. 3:45 The technique is called distillation. A powerful teacher produces answers or synthetic data, and a cheaper student learns from them. AI companies, all of them, routinely use this technique themselves. 3:58 Nathan Lambert puts it well in his post. He sees distillation as a legitimate way to learn from a stronger model and seed new capabilities. 4:08 It does not copy an entire model, and extensive original post-training work still follows. The disputed part is how the data was obtained, yes. 4:18 If a company creates fake accounts and evades access controls to extract millions of outputs, that is a serious allegation. 4:27 Anthropic has published its attribution claims, but it has not released the underlying logs, and there is no public technical demonstration showing exactly how Claude outputs translated into K3's final capabilities. 4:42 There was even a joke that K3 has distilled a nonexistent Claude model here. This distinction became the center of the fight. Who did what? 4:51 Should the government punish alleged misconduct, or should it restrict the resulting model for every American developer? 4:59 On July 22nd, 179 founders and startups, including Y Combinator, asked the administration not to impose broad restrictions on foreign open weight models. What was their argument? 5:12 Denying American startups access to models available elsewhere would entrench the largest closed labs and operate like a tax on intelligence. Two days later, Jensen Huang made his move. 5:26 He posted on Twitter a letter that framed open weights as part of American AI leadership, giving developers control and reducing dependence on the few vendors. 5:38 The live version eventually collected 77 signatures, including Microsoft, Meta, Amazon, Google, SpaceX, and OpenAI. 5:47 Anthropic did not sign, while Jensen's post received 62 million views and brought people together who are in constant opposition, like Yann LeCun and Jürgen Schmidhuber, into the same comment section. 6:01 And if you think about it, Jensen really is the person who can make the family group chat work. But anyway, then Jensen made another move. He changed the subject from competition to security. 6:14 His second post began, "Attackers have frontier AI. Defenders need a frontier AI ecosystem. His evidence was the security incident at Hugging Face, which we covered in another episode. 6:27 Here is a reminder what has happened. 6:29 During an internal evaluation, an OpenAI agent with reduced cyber refusals found a vulnerability in the package register proxy, reached the internet, and attacked Hugging Face while searching for benchmark solutions. 6:42 When Hugging Face investigated, it tried to use commercial frontier models. Those systems blocked parts of the forensic work because the prompts looked like dangerous cyber activity. 6:53 Hugging Face then ran the OpenWeight GLM 5.2 model on its own infrastructure, the Chinese model. The model helped analyze more than 17,000 actions while Hugging Face contained the intrusion. 7:05 Congress looked at this incident and proposed a literal AI kill switch app. 7:11 The bipartisan bill would let the Department of Homeland Security order a covered company to throttle or shut down a dangerous AI system in a loss-of-control scenario. 7:21 Ignoring the order could bring fines of up to $20 million per day. Notice the collision. A government kill switch assumes that an identifiable provider still controls the system. That can work for a frontier service. 7:38 That surely can work for a frontier service delivered through an API. It cannot recall model weights that have already been copied around the world. The same incident produced two very different responses. 7:51 Congress wants the power to centralize control during an emergency. Jensen Huang wants defenders to have enough independent capability that no provider becomes a single point of failure. 8:04 The alliance is pushing back against blanket restrictions, but its proposal goes further than policy. It wants to build the technical layer that defenders can inspect and control. 8:17 An AI agent is a model connected to a harness that decides which tools it can call, which credentials it receives, what it can read and write, how its actions are logged, and when a human must approve them. 8:30 The model may propose an action, but the surrounding system determines whether that action can touch the internet, install a package, use a password, or send money. 8:40 NVIDIA's own research said that changing the harness around the same model can produce double-digit changes in benchmark results and major differences in token cost. The harness changes the agent's practical capability. 8:52 The Open Secure AI Alliance is trying to make more of that layer inspectable and shared. Hugging Face contributes safetensors, a format designed so loading model weights cannot widely execute arbitrary code. 9:07 HPE is working on cryptographic identities for agents. Microsoft has a harness in which specialized agents search for bugs and try to prove whether a vulnerability is actually exploitable. 9:18 NVIDIA is contributing a framework called NOOA. It represents an agent's state and tools as Python classes with type inputs and deterministic checks. 9:29 Developers can trace actions and verify some results with code instead of asking another model whether the first model behaved. 9:37 NOOA's own documentation warns that it can execute model-generated code and should be sandboxed. Openness gives defenders access and visibility, while safety still depends on how the system is built and operated. 9:51 The alliance is proposing an open defense stack, inspectable model formats, identity, permissions, sandboxes, logs, evaluations, and multiple models that can check one another. 10:05 And that's true that open weights can be stripped of safeguards, fine-tuned for attacks, and distributed beyond the reach of any recall. Anthropic is right about that risk. 10:15 The alliance argues, though, that determined attackers are already seeking powerful models, while defenders cannot depend entirely on a provider that may refuse a request, change a policy, go offline, or misunderstand the work being done. 10:29 So where is Anthropic? Jeremy Howard pointed to the tension through an exchange with Julian Schrider, a researcher at Anthropic. 10:38 Julian saw Jensen arguing for open weights and replied that he looked forward to NVIDIA releasing CUDA and GPU drivers as open source. It sounds clever until you think about and examine the comparison. 10:50 CUDA is NVIDIA's proprietary software platform. Open weights are model parameters that developers can download, run, and tap. 10:59 Jensen is arguing for access to one layer of the AI stack, while Julian answers with a different layer and avoids the question Jensen was asking, and Jeremy response was harsh and just. 11:11 He wrote that people who otherwise appear perfectly capable of reasoning can lose that ability when the answer affects the company paying them. That observation applies to the whole argument. 11:22 NVIDIA benefits when more models create more inference and more demand for its hardware. We know that. Anthropic benefits when frontier capability remains behind an API that it controls. 11:34 Kimi K3 success and GLM used by Hugging Face forced those incentives into public view. 11:41 The problem is that NVIDIA does not control the other companies, but Anthropic wants to control the model and all the other models through government. What does Silicon Valley see? 11:53 American startups and their investors want lower costs, cloud providers want more models to sell, and security teams want models they can run inside their own systems. 12:06 Anthropic wants Washington to treat Moonshot's alleged distillation as a national security problem. 12:12 Anthropic's position becomes harder to defend when we compare it with the company's own experience of government intervention. 12:20 Anthropic has supported stronger government powers over frontier AI and has argued that unsafe models should be blocked. 12:27 When the government used export control powers against Fable-5 and Mipha-5, Anthropic asked for technical evidence, transparent rules, and a proportional response. It was right to ask for those things. 12:40 And Kimi deserves the same standard. This is why Anthropic now looks isolated. 12:45 The company wants careful limits when government power reaches its own models, while asking Washington to move quickly against an open competitor. 12:54 OpenAI and Google did not join the new security alliance yet, but they signed the open weight letter. Anthropic signed neither. 13:01 And the surprising part of this story is that Kimi K3 did more than introduce another capable model, and GLM was more than just Open Weight's model that helped fight an intruder. 13:12 They revealed that AI safety and market structure have become the same argument. If a few companies control the strongest models through APIs, they can monitor access, revoke it, and update safeguards. 13:24 They also decide the price, the acceptable uses, and which researchers or defenders receive permission. Open Weights distribute that control. 13:32 They give companies and countries more independence, while also making recall nearly impossible. That's true. 13:38 Jensen's alliance places more responsibility in the surrounding system, identity, permissions, sandboxes, logs, evaluation, and independent inspection. 13:50 Anthropic places more responsibility with the model provider, where access can be controlled before a request reaches the model. We will continue to have both open and closed models. Let's hope for that. 14:03 The decision facing policymakers is which risks justify restricting access, what evidence should be required, and whether the companies asking for protection should help define those rules. 14:15 Jensen gained a million followers by arriving when Silicon Valley needed someone to organize its position. 14:21 The larger achievement was assembling 77 letter signatories and 37 alliance members around the idea that advanced intelligence should not remain under the permanent control of a few model providers. 14:35 Anthropic wants dangerous models to remain under the control of companies that can restrict or disable access. 14:41 Much of the Silicon Valley worries this would let a few closed labs decide who gets to use advanced AI at all. 14:48 That is the argument Kimi K3, GLM, and effectively other successful Chinese models brought into the open, brought to our attention. Thank you for watching. 14:58 Let me know what you think in the comments and how you think this situation will develop. And if you like this video, help us grow. Sharing, hyping, it's a feature on YouTube, and subscribing certainly helps. 15:09 See you at the end of this week. [outro music]