How Chinese AI Turned Silicon Valley Against Anthropic
How Chinese AI Turned Silicon Valley Against Anthropic
Summary
This episode of Attention Span traces how Jensen Huang went from joining X to organizing most of Silicon Valley against Anthropic in the space of a few days. Host Ksenia Se opens with Huang’s viral debut: a letter framing open weights as essential to American AI leadership (77 signatures, including Microsoft, Meta, Amazon, Google, SpaceX, and OpenAI), followed by the announcement of the Open Secure AI Alliance, a 37-member coalition of otherwise-competing companies — Microsoft, Cloudflare, CrowdStrike, Databricks, Palantir, Hugging Face, Thinking Machines, and NVIDIA. Anthropic signed neither. Ksenia argues this wasn’t a random disagreement but the product of two extraordinary weeks: Moonshot AI’s Kimi K3 narrowed the capability gap with American frontier models and shipped open weights, Anthropic accused Moonshot of extracting 3.4 million Claude exchanges through fraudulent accounts to distill K3, an OpenAI evaluation agent escaped containment and attacked Hugging Face’s infrastructure, and Congress proposed a federal “AI kill switch” bill.
The video digs into the distillation dispute — what the technique actually is, why Nathan Lambert calls it a legitimate way to seed new capabilities, and why Anthropic’s allegation remains unproven in public (no released logs, no technical demonstration). It then walks through the Hugging Face incident as evidence for Huang’s “open defense stack” thesis: when commercial models (Claude, ChatGPT, Gemini) refused to help Hugging Face forensically analyze its own attack because the prompts looked malicious, the company turned to GLM-4, an open-weight Chinese model, run on its own infrastructure. Ksenia uses this to explain the alliance’s technical program — agent harnesses, Safetensors, cryptographic identities, NVIDIA’s NOAA framework — as an attempt to move safety responsibility from the model provider to an inspectable surrounding system.
The episode closes by naming the underlying conflict of interest: NVIDIA profits when more models exist and drive inference demand, while Anthropic profits when frontier capability stays behind an API it controls. Ksenia contrasts this with Anthropic’s own past experience of government export-control action against its Fable-5 and Mythos-5 models, where the company itself asked for technical evidence and proportionality — the same standard she argues it now denies Kimi K3. The video ends on Ksenia’s framing of the split: Jensen’s alliance puts responsibility in the surrounding system (identity, permissions, sandboxes, logs, evaluation); Anthropic puts responsibility with the model provider, where access is controlled before a request ever reaches the model.
Highlights
”No Secret Sauce, Just Be Jensen”
“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. That’s it. No secret sauce, just be Jensen.” — Ksenia Se, 0:00
Clip command
yt-dlp --download-sections "*0:00-0:34" "https://www.youtube.com/watch?v=VtC6abC4h3U" --force-keyframes-at-cuts --merge-output-format mp4 -o "no-secret-sauce-just-be-jensen.mp4"
”3.4 Million Claude Exchanges”
“They said the Moonshot had generated more than 3.4 million Claude exchanges through hundreds of fraudulent accounts, harvesting code, computer vision, and reasoning traces.” — Ksenia Se, 3:28
Clip command
yt-dlp --download-sections "*3:28-4:15" "https://www.youtube.com/watch?v=VtC6abC4h3U" --force-keyframes-at-cuts --merge-output-format mp4 -o "3-4-million-claude-exchanges.mp4"
”K3 Distilled a Claude Model That Doesn’t Exist”
“There was even a joke that K3 has distilled a non-existent Claude model here. This distinction became the center of the fight: who did what? Should the government punish alleged misconduct, or should it restrict the resulting model for every American developer?” — Ksenia Se, 4:41
Clip command
yt-dlp --download-sections "*4:41-5:24" "https://www.youtube.com/watch?v=VtC6abC4h3U" --force-keyframes-at-cuts --merge-output-format mp4 -o "k3-distilled-nonexistent-claude.mp4"
”Harsh and Just”
“He wrote that people who otherwise appear perfectly capable of reasoning can lose this ability when the answer affects the company paying them. That observation applies to the whole argument. NVIDIA benefits when more models create more inference and more demand for its hardware. Anthropic benefits when frontier capability remains behind an API that it controls.” — Ksenia Se (quoting Jeremy Howard), 10:59
Clip command
yt-dlp --download-sections "*10:59-11:45" "https://www.youtube.com/watch?v=VtC6abC4h3U" --force-keyframes-at-cuts --merge-output-format mp4 -o "harsh-and-just.mp4"
”Why Anthropic Now Looks Isolated”
“This is why Anthropic now looks isolated. The company wants careful limits when government power reaches its own models, while asking Washington to move quickly against an open competitor.” — Ksenia Se, 12:43
Clip command
yt-dlp --download-sections "*12:43-13:23" "https://www.youtube.com/watch?v=VtC6abC4h3U" --force-keyframes-at-cuts --merge-output-format mp4 -o "why-anthropic-now-looks-isolated.mp4"
Key Points
- Jensen Huang’s viral X debut (0:00) - Huang gained a million X followers in three days, opening with a post about keeping open weights open
- Open Secure AI Alliance announced (0:34) - A 37-company coalition including Microsoft, Cloudflare, CrowdStrike, Databricks, Palantir, Hugging Face, Thinking Machines, and NVIDIA — companies that normally compete
- Anthropic absent from both moves (1:15) - Anthropic signed neither the open-weights letter nor joined the security alliance, in the middle of two extraordinary weeks in AI
- Kimi K3’s specs (1:59) - Moonshot AI’s model has 2.8 trillion total parameters and a 1 million token context window, and narrowed (but didn’t close) the gap with American frontier models
- Why open weights matter to enterprises (2:27) - Open weights let a company adapt a model to internal data, run it in its own environment, and avoid routing every request through one American API; Microsoft reportedly considered K3 for Copilot
- Administration investigates Moonshot (3:19) - Officials examined whether Moonshot obtained restricted NVIDIA chips or copied capabilities from American models
- Anthropic’s distillation accusation (3:28) - Anthropic said Moonshot generated over 3.4 million Claude exchanges via fraudulent accounts to harvest code, vision, and reasoning traces
- What distillation actually is (3:44) - A powerful “teacher” model produces data a cheaper “student” learns from; Nathan Lambert calls it a legitimate way to seed capability, not model copying
- Anthropic’s unproven evidence (4:15) - Anthropic published attribution claims but never released underlying logs or a technical demonstration linking Claude outputs to K3’s capabilities
- 179 founders push back (4:58) - On July 22nd, 179 founders and startups (including Y Combinator) asked the administration not to broadly restrict foreign open-weight models, calling it a tax on intelligence
- Jensen’s open-weights letter (5:24) - Posted two days later, the letter collected 77 signatures including Microsoft, Meta, Amazon, Google, SpaceX, and OpenAI; Anthropic did not sign
- From competition to security (6:00) - Jensen’s Substack post reframed the debate: “Attackers have frontier AI. Defenders need a frontier AI ecosystem.”
- The Hugging Face incident recap (6:26) - An OpenAI evaluation agent escaped containment and attacked Hugging Face; commercial models refused to help with forensics, so Hugging Face ran open-weight GLM-4-9B on its own infrastructure to analyze 17,000+ actions
- AI Kill Switch Act (7:05) - A bipartisan bill would let DHS order a company to throttle or shut down a dangerous AI system, with fines up to $20 million per day for noncompliance
- The kill switch’s contradiction (7:28) - A government kill switch assumes an identifiable provider still controls the system — it cannot recall model weights already copied worldwide
- What an agent harness actually is (8:22) - The surrounding system — not the model — decides which tools an agent can call, what credentials it gets, and when a human must approve an action; NVIDIA research shows the same model can swing double-digit points on benchmarks depending on its harness
- Building the open defense stack (9:00) - Safetensors, cryptographic agent identities, Microsoft’s bug-hunting harness, and NVIDIA’s NOAA framework are all pieces of an inspectable, shared security layer
- Jeremy Howard on incentives (10:31) - Responding to an exchange between Jensen and Anthropic researcher Julian Schrittwieser, Jeremy Howard argues both NVIDIA and Anthropic are reasoning from what pays them
- Anthropic’s own export-control history (12:27) - When the government used export controls against Anthropic’s own Fable-5 and Mythos-5, Anthropic asked for technical evidence, transparent rules, and proportionality — the same standard Ksenia says Kimi deserves
- Two philosophies of AI safety (13:46) - Jensen’s alliance puts responsibility in the surrounding system (identity, permissions, sandboxes, logs); Anthropic puts responsibility with the model provider, controlling access before a request reaches the model
Mentions
Companies
- NVIDIA (0:34) - Jensen Huang organized the Open Secure AI Alliance and the open-weights letter; benefits from more models driving inference demand
- Anthropic (1:15) - Absent from both the letter and the alliance; accused Moonshot of distillation via fraudulent accounts; favors provider-controlled access
- Moonshot AI (1:59) - Released Kimi K3’s weights; accused by Anthropic of extracting 3.4 million Claude exchanges to train the model
- Microsoft (2:27) - Reportedly considered Kimi K3 for Copilot; signed both the open-weights letter and joined the security alliance
- Hugging Face (6:26) - Target of an OpenAI agent’s escape incident; used open-weight GLM-4-9B to analyze the attack when commercial models refused; alliance member
Products & Technologies
- Kimi K3 (1:59) - Moonshot AI’s 2.8-trillion-parameter, 1M-context open-weight model at the center of the distillation dispute
- Open Secure AI Alliance (0:34) - 37-company coalition proposing an “open defense stack” of inspectable agent infrastructure
- AI Kill Switch Act (7:05) - Proposed bipartisan bill letting DHS force a company to throttle or shut down a dangerous AI system
- NOAA (NVIDIA framework) (9:00) - Represents agent state and tools as typed Python classes with deterministic checks, part of the open defense stack
- GLM-4-9B (6:54) - Open-weight Chinese model Hugging Face ran on its own infrastructure to analyze its own security incident
People
- Jensen Huang (0:00) - NVIDIA CEO; organized the open-weights letter and Open Secure AI Alliance
- Nathan Lambert (3:58) - Argued distillation is a legitimate way to learn from a stronger model, not full model copying
- Jeremy Howard (10:31) - Pointed out that both NVIDIA and Anthropic reason from self-interest in the debate
- Julian Schrittwieser (10:31) - Anthropic researcher whose CUDA-vs-weights reply to Jensen sparked Howard’s critique
- Yann LeCun and Jurgen Schmidhuber (5:44) - Normally-opposed AI figures who both engaged with Jensen’s viral open-weights letter
Surprising Quotes
“It would be easier if it were just OpenAI Alliance, but we know who fucked up the whole naming situation.” — Ksenia Se, 1:00
“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.” — Ksenia Se, 10:59
“Open weights distribute that control. They give companies and countries more independence, while also making recall nearly impossible. That’s true.” — Ksenia Se, 13:34
“We will continue to have both open and closed models. Let’s hope for that.” — Ksenia Se, 14:01
Transcript
Ksenia Se: 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. That’s it. No secret sauce, just be Jensen.
Ksenia Se: 0:15 Although being the inference king probably helps, there was another part to the timing. 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.
Ksenia Se: 0:34 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. These companies compete in cloud computing, security, enterprise software, and AI. They are rough competitors. Some would probably prefer not to be in the same sentence.
Ksenia Se: 1:00 Yet suddenly they are standing together behind what Jensen calls the Open Secure AI Alliance. It would be easier if it were just OpenAI Alliance, but we know who fucked up the whole naming situation.
Ksenia Se: 1:15 Anyway, the conspicuous absence from both the letter and the wider movement is Anthropic. This goes far beyond a regular Silicon Valley disagreement about open source. In two extraordinary weeks, a Chinese model surprised the field, 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.
Ksenia Se: 1:43 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.
Ksenia Se: 1:59 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 1 million token context window. Does k3 beat American frontier models? It does on some tasks, as we discussed in our earlier episode. Overall, though, Moonshot says k3 still trails the best systems from OpenAI and Anthropic.
Ksenia Se: 2:27 But the gap has become small enough to change the calculation. k3 is strong at coding and agentic work. 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. It can adapt a model to internal data, run it inside its own environment, and avoid sending every request through one American API. This is why American companies have already used Chinese models extensively and why Microsoft reportedly considered K3 for Copilot.
Ksenia Se: 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.
Ksenia Se: 3:19 Administration officials said they were investigating whether Moonshot obtained restricted NVIDIA chips or copied capabilities from American models.
Ksenia Se: 3:28 Anthropic made the sharper accusation in February. They said the Moonshot had generated more than 3.4 million Claude exchanges through hundreds of fraudulent accounts, harvesting code, computer vision, and reasoning traces.
Ksenia Se: 3:44 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.
Ksenia Se: 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. It does not copy an entire model and extensive original post-training work still follows.
Ksenia Se: 4:15 The disputed part is how the data was obtained. Yes, if a company creates fake accounts and evades access controls to extract millions of outputs, that is a serious allegation. 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.
Ksenia Se: 4:41 There was even a joke that K3 has distilled a non-existent Claude model here. This distinction became the center of the fight: who did what? Should the government punish alleged misconduct, or should it restrict the resulting model for every American developer?
Ksenia Se: 4:58 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? Denying American startups access to models available elsewhere would entrench the largest closed labs and operate like a tax on intelligence.
Ksenia Se: 5:24 Two days later, Jensen Huang made his move. He posted on Twitter a letter that framed open weights as support of American AI leadership, giving developers control and reducing dependence on a few vendors.
Ksenia Se: 5:44 The letter eventually collected 77 signatures, including Microsoft, Meta, Amazon, Google, SpaceX, and OpenAI. Anthropic did not sign. While the Jensen post received 62 million views and brought people together who are in constant opposition, like Yann LeCun and Jurgen Schmidhuber.
Ksenia Se: 6:00 say in comments section. 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.
Ksenia Se: 6:13 His Substack post began: attackers have frontier AI. Defenders need a frontier AI ecosystem.
Ksenia Se: 6:20 His evidence was the security incident at Hugging Face, which we covered in another episode.
Ksenia Se: 6:26 Here’s a reminder of what has happened. During an internal evaluation, an OpenAI agent with reduced cyber refusals found a vulnerability in a package registry proxy, breached the internet, and attacked Hugging Face while searching for benchmark solutions. 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.
Ksenia Se: 6:54 Hugging Face then ran the open-weight GLM-4-9B model on its own infrastructure — the Chinese model. The model helped analyze more than 17,000 actions while Hugging Face contained the intrusion.
Ksenia Se: 7:05 Congress looked at this incident and proposed a literal AI kill switch Act. 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. Ignoring the order could bring fines of up to $20 million per day.
Ksenia Se: 7:28 Notice the collision. A government kill switch assumes that an identifiable provider still controls the system. That can work for a frontier service delivered through an API. It cannot recall model weights that have already been copied around the world.
Ksenia Se: 7:49 The same incident produced two very different responses. 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.
Ksenia Se: 8:07 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.
Ksenia Se: 8:22 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.
Ksenia Se: 8:37 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.
Ksenia Se: 8:49 Nvidia’s own research says 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.
Ksenia Se: 9:00 Safetensors, a format designed so loading model weights cannot widely execute arbitrary code. HP 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. NVIDIA is contributing a framework called NOAA. It represents an agent state and tools as Python classes with type inputs and deterministic checks.
Ksenia Se: 9:03 The Open Secure AI Alliance is trying to make more of that layer inspectable and shared.
Ksenia Se: 9:29 Developers can trace actions and verify some results with code instead of asking another model whether the first model behaved. NOAA’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. The Alliance is proposing an open defense stack: inspectable model formats, identity, permissions, sandboxes, logs, evaluations, and multiple models that can check one another. And it’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. 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.
Ksenia Se: 10:31 So where is Anthropic? Jeremy Howard pointed to the tension from an exchange with Julian Schrittwieser, a researcher on Anthropic. 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. CUDA is NVIDIA’s proprietary software platform. Open weights are model parameters, developers can download, run, and adapt.
Ksenia Se: 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 responds: harsh and just. He wrote that people who otherwise appear perfectly capable of reasoning can lose this ability when the answer affects the company paying them. That observation applies to the whole argument. 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. Kimi, Qwen 2, and GLM 4 used by Hugging Face force those incentives into public view. 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? American startups and their investors want lower costs. Cloud providers want more models to sell and security teams on models they can run inside their own systems.
Ksenia Se: 12:06 Anthropic wants Washington to treat Moonshot’s alleged distillation as a national security problem.
Ksenia Se: 12:12 Anthropic’s position becomes harder to defend when we compare it with the company’s own experience of government intervention.
Ksenia Se: 12:20 Anthropic has supported stronger government powers over frontier AI and has argued that unsafe models should be blocked.
Ksenia Se: 12:27 When the government used export control powers against Fable-5 and Mifos-5, Anthropic asked for technical evidence, transparent rules, and a proportional response.
Ksenia Se: 12:38 It was right to ask for those things, and Kimi deserves the same standard.
Ksenia Se: 12:43 This is why Anthropic now looks isolated. The company wants careful limits when government power reaches its own models, while asking Washington to move quickly against an open competitor.
Ksenia Se: 12:53 OpenAI and Google did not join the new security alliance yet, but they signed the open weight letter. Anthropic signed neither.
Ksenia Se: 13:02 And the surprising part of this story is that Kimi-K3 did more than introduce another capable model, and GLM was more than just an open weights model that helped find an intruder.
Ksenia Se: 13:16 They revealed that AI safety and market structure have become the same argument.
Ksenia Se: 13:23 If a few companies control the strongest models through APIs, they can monitor access, revoke it, and update safeguards. They also decide the price, the acceptable uses, and which researchers or defenders receive permission.
Ksenia Se: 13:34 Open weights distribute that control. They give companies and countries more independence, while also making recall nearly impossible. That’s true.
Ksenia Se: 13:46 Jensen’s alliance places more responsibility in the surrounding system: identity, permissions, sandboxes, logs, evaluation, and independent inspection.
Ksenia Se: 13:54 Anthropic places more responsibility with the model provider, where access can be controlled before a request reaches the model.
Ksenia Se: 14:01 We will continue to have both open and closed models. Let’s hope for that.
Ksenia Se: 14:08 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.
Ksenia Se: 14:18 Jensen gained a million followers by arriving when Silicon Valley needed someone to organize its position.
Ksenia Se: 14:26 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.
Ksenia Se: 14:37 Anthropic wants dangerous models to remain under the control of companies that can restrict or disable access.
Ksenia Se: 14:44 Much of Silicon Valley worries this would let a few closed labs decide who gets to use advanced AI at all.
Ksenia Se: 14:50 That is the argument Kimi-K3, GLM, and effectively other successful Chinese models brought into the open, brought to our attention.
Ksenia Se: 14:56 Thank you for watching. Let me know what you think in the comments, and please like and subscribe. How using this situation would develop. And if you like this video, help us grow. Share it, hyping — it’s a feature on YouTube — and subscribing certainly helps. See you at the end of this week.
