The article is reprinted from Ba Chi AI

Yesterday, Dario Amodei, the founder of Anthropic, published a long article on his personal blog entitled "We Must Control the Pace of AI Frontier Development."

The article is emotionally charged, starting with the story of his father dying from a disease that could have been conquered by AI, and goes on to discuss how recursive self-improvement could allow an uncontrollable AI agent to take over the Internet within six to twelve months.

This appears to be an AI safety manifesto filled with humanitarian concerns. Sam Altman replied beneath it, "I agree with Dario," and Musk chimed in, "Dario is right." Three people who usually refuse to share a photo together, unexpectedly found common ground.

However, if you flip to the latter part of the article, you'll find something amiss. Dario has listed three containment measures specifically targeting China in the section about how to brake the frontier, banning chip sales, cracking down on distillation (the original text says "artificial intelligence technology refinement"), and weight theft.

A proposal to limit technology speed suddenly becomes a manual for a technological Cold War against China.

Just two days before this lengthy article was published, Anthropic had released a long threat report naming seven Chinese AI laboratories: Alibaba, Dark Side of the Moon, DeepSeek, Zhiyu, Xiaomi, SenseTime, and MiniMax. They accused these labs of launching industrial-scale distillation attacks on Claude, accumulating nearly 200 million interactions.

Is it a coincidence that these two events are so close together? Probably not.

When a company valued at nearly a trillion dollars and racing towards the largest IPO in history begins to frame widely used conventional technology as a national security threat, one must question Dario's motivations.

1

The Origin of Distillation

To clarify whether distillation is theft, we must first understand what it actually is.

In March 2015, Geoffrey Hinton, Oriol Vinyals, and Jeff Dean published a paper on arXiv titled "Knowledge Distillation in Neural Networks."

This paper conveys a simple principle: the knowledge gained by a large model (teacher) is not only its parameters but also the probability distribution it generates in response to inputs, which is soft targets. Training a smaller model (student) using these soft targets enables the smaller model to approximate the performance of the larger model using far less computational power.

Hinton even clearly stated in the paper that the transfer set used to train the student model can consist entirely of unlabeled data.

In other words, from day one, distillation has been a legitimate, open, and widely recognized method for model compression and capability transfer in academia. Its original intention was to enable AI to run on mobile phones and embedded devices with limited computing power.

Hinton himself used this method to improve Google's voice recognition system, with significant results.

Eleven years have passed, and distillation is no longer a niche technique; it is the infrastructure of the large model industry.

In October 2024, OpenAI directly launched an official distillation feature in its API, encouraging developers to fine-tune cheaper models using the outputs of o1-preview and GPT-4o.

Google distilled Gemini Ultra into Gemini Flash, focusing on low-latency scenarios. Meta open-sourced Llama, and the entire community has been using it for distillation. DeepSeek distilled a complete set of small models (32B, 14B, 8B, 1.5B) using R1 to allow inference capabilities to run on consumer-grade GPUs.

Even Anthropic itself is internally using distillation to produce a lightweight version of Claude. This is no secret; industry analysis firm Ertas AI made it clear in an article titled "Model Distillation is Not Theft" that every lab with a cutting-edge model has derived smaller models from it.

Distillation plays a role in AI that is akin to leveraging resources; it allows knowledge to flow from expensive large models to more accessible smaller models, from the computational barriers of a few companies to broader application scenarios.

So the question arises: why is the same technology referred to as capability transfer by OpenAI, model compression by Google, open-source ecosystem by Meta, but becomes an industrial-scale distillation attack and a significant national security risk when Chinese companies are involved?

This double standard is more intriguing than the distillation itself.

1

Dario's Narrative Technique

To understand why Dario elevates distillation to the level of national security, we first need to look at how he constructs his narrative.

In the article "We Must Control the Pace of AI Frontier Development," Dario employs a very classic rhetorical structure, establishing a large premise that no one can oppose: AI is developing too quickly and needs safety.

This logic translates to, "I can limit speed, but you cannot overtake me while I am slowing down."

And distillation is one of his core pieces of evidence to prove that China is getting a head start. The other is chips, but Nvidia isn't engaging with him yet, possibly not planning to share his focus.

The report by Anthropic reads with great seriousness. Seven Chinese labs used fraud accounts, residential proxies, disposable emails, and virtual payments to bypass risk controls, extracting a large number of Claude Opus's thinking chains (CoT), not just the final answers.

Alibaba alone had 151 million interactions between May and July, using over 3,500 fraud accounts and peaking at nearly 3 million daily interactions.

These figures look frightening. However, with a bit more thought, one can find inconsistencies.

First, if these interactions were truly fraudulent, why did Anthropic's risk control system fail to intercept them immediately? 151 million interactions, 3,500 accounts, over three months, this is not a secret hacker attack; it's like setting up a stall at your door, collecting money for three months, then claiming the other party is a thief.

Anthropic did not mention how much API fees they collected during these interactions in the report.

Second, is the thinking chain (CoT) actually theft? CoT is the intermediate steps generated during the model's reasoning process, which users can obtain through normal API calls. If Anthropic did not want users to see CoT, they could completely shield it at the API level. They did not do so.

Third, and most critically, violating terms of service and breaking the law are two different matters. It may be written in Anthropic's API service terms that batch generation of training data is prohibited, but this is a civil contract issue, not a criminal offense, and certainly not a national security threat.

If Chinese companies truly violated the terms, Anthropic could ban accounts, initiate lawsuits, and demand compensation. However, they chose the most drama-filled route, releasing a threat report that framed a commercial dispute as a national security event.

This narrative technique is not new. Historically, each time a technological hegemon has faced a rising competitor, they have raised the banner of national security.

Nineteenth-century Britain attempted to ban the export of textile machinery and skilled workers in response to America's industrial rise. In the twentieth century, the United States, facing Japan's semiconductor catch-up, used national security as a justification to force Japan to sign the Plaza Accord and restrict semiconductor exports.

San Hongyang mentioned in "Salt and Iron Discussion": Those who want grain must act in accordance with the times, those who want to govern must follow the conditions. The national security narrative of hegemonic countries has always adapted to the time and situation. When they are in the lead, they talk about free trade; when they are being chased, they speak of national security.

Dario's brilliance lies in his ability to package this old narrative into a new bottle labeled AI. He constructed a complete logical chain, from distillation allowing Chinese companies to narrow the gap at minimal cost, to that narrowing gap threatening the United States' AI leadership, to the idea that AI leadership is a matter of national security, and ultimately concluding that distillation represents a national security threat.

Upon careful reflection, each link in this chain is somewhat misleading. Distillation reduces gaps in engineering implementation, not foundational research. The progress of Chinese large models mainly comes from their own research and investment and algorithmic innovation, with distillation being just one part of it. Whether AI leadership directly equals national security is a topic worth discussion in itself.

However, Dario does not need this chain to be logically flawless; he only needs it to be emotionally persuasive. The phrase "national security" itself is a weapon of emotional appeal.

Interestingly, while Dario talks extensively about democratic countries uniting, one of the largest investors in his company is Google, which has significant business and interests in the Chinese market.

He speaks of cracking down on authoritarian states' distillation but does not mention the allegations of mutual distillation among American companies, like how OpenAI allegedly used Anthropic's outputs to train its own models.

1

The Performance on the Eve of a $2 Trillion IPO

Dario's attack at this time seems quite calculating.

In 2021, when Dario left with a group of former OpenAI researchers to establish Anthropic, the company's valuation was only $5.5 billion. In five years, this figure skyrocketed like a rocket, reaching $965 billion after a $65 billion financing in May 2026, surpassing OpenAI's $852 billion valuation historically.

This is not the end. On June 1, 2026, Anthropic secretly submitted an S-1 prospectus to the SEC, formally starting the IPO process. Morgan Stanley, Goldman Sachs, and JPMorgan are co-underwriting, with the earliest listing possible in October.

Investors have set a target valuation starting at $2 trillion, and some optimistic observers even see $3 trillion. If it is issued at $2 trillion, it will surpass SpaceX's record of $1.77 trillion set in June, becoming the largest IPO in global business history.

What story does Anthropic need to tell to support this valuation?

They need to tell a story that AI is the most important technological revolution of our time, but OpenAI is also telling the same story, and OpenAI's brand recognition and user base are larger.

Anthropic needs a story that only it can tell, and the threat from China is the best material for that story.

When an AI company tells investors that it is not only developing cutting-edge technology but also safeguarding the technological hegemony of the Western world, its valuation logic shifts. Its competitors become not just OpenAI and Google, but the entire Chinese AI industry.

This narrative has a fatal appeal to Wall Street.

It packages a highly uncertain technological wager as an inevitable geopolitical choice. Regardless of how AI technology evolves or which company's model is stronger, democratic nations need a reliable AI supplier, and that demand is certain.

And Anthropic is positioning itself as the only answer to that demand.

Dario can control the pace for his frontier, but he cannot control the wheels of the entire industry. No matter how compelling the $2 trillion valuation story is, it ultimately comes back to a simple question: how good is your model?

If China's open-source models provide comparable capabilities at one-tenth the cost, global developers will vote with their feet, choosing more open and cheaper solutions. All narratives and accusations will ultimately be dissolved by market realities.

After all, capital markets can pay a premium for narratives, but users only pay for products.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。