Why Anthropic CEO is hostile to DeepSeek and Chinese AI

Source: Silicon-based Standpoint

A scientist wrote an article a few months ago, declaring his and his company's good intentions to solve many aspects of human physical and mental health, mental illness, poverty, peace, the meaning of work and life through powerful artificial intelligence. A few months later, the same scientist suddenly published an article, strongly calling for no American chips to be exported to China, in order to restrict the development of China's artificial intelligence and maintain the "unipolar world" of artificial intelligence (I was shocked that he would use this term nakedly). This matter makes people feel divided and a little hypocritical no matter how you look at it.

This person is Dario Amodei, the founder and CEO of the artificial intelligence company Anthropic, an Italian-American, a doctor of neurophysics, a senior AI scientist, one of the souls of the former OpenAI research team, an early employee of the Deep Learning Laboratory of the Chinese company Baidu, an idealist who claims to build the most powerful and secure AI, the founder of OpenAI's most important competitor, and now, the most ardent advocate of the United States' all-round and no-dead-angle embargo on Chinese AI, without a doubt.

Although Anthropic and its Claude series of models are not well-known among the Chinese public, after all, it is the provider of the world's most popular large language model for AI application developers, and it also has a considerable number of fans among Chinese AI researchers and developers. But overnight, many Chinese AI practitioners publicly stated that Anthropic and Amodei himself had lost their most basic respect.

This is the effect of a "manifesto".

In this article entitled "About DeepSeek and Export Control", Dario Amodei on the one hand calmly pointed out that DeepSeek's achievements in shaking the advantages of American AI were exaggerated. While expressing his affirmation of the innovation of the DeepSeek-V3 model, he resolutely refused to acknowledge the breakthrough of the DeepSeek reasoning model - R1, which caused a greater sensation (his thoughts on this issue are the focus of this article later). What he is even more reluctant to admit is the achievements of the DeepSeek model in terms of computing power cost and algorithm efficiency. He used a rumor that DeepSeek, which he himself admitted was "unverified", had 50,000 smuggled Nvidia A100, H100 and H800 graphics cards to prove that the DeepSeek-V3 model could not have been trained at a low cost of $6 million. Obviously, Amodei could not accept that DeepSeek replaced the increasingly widely recognized path of computing power stacking with algorithm efficiency innovation, so he did not hesitate to use the premise that DeepSeek, which he himself knew was unverified, smuggled a large number of high-end graphics cards to achieve this argument. But he also stated that the US export control of computing power to China has not failed. He may have just forgotten that his previous argument was based on the assumption that DeepSeek smuggled.

Source: https://darioamodei.com/on-deepseek-and-export-controls

Let’s restore the logical line of his argument: DeepSeek’s influence has been exaggerated—— V3 is indeed an innovation, but it is impossible to spend so little money——I heard that they smuggled chips——So they did spend more training costs—— DeepSeek is not original, it is based on our research and of course the cost is lower—— The R1 reasoning model is absolutely not innovative, it just reproduces the results of o1 (pretending not to see that OpenAI has acknowledged that DeepSeek’s achievements in reasoning are independent discoveries)—— Export control has not failed, that's right (I forgot that the premise of my previous argument was that DeepSeek obtained smuggled graphics cards) - we want to build a unipolar world of AI, and China must not make models comparable to ours (I forgot that I said at the beginning that DeepSeek is not to be feared) - so let alone H100 and H800, even the lowest-end H20 cannot be exported to China, so China can't win.

You see, a scientist who always talks about logic and reasoning, trying to use a 10,000-word article to prove a conclusion that cannot be justified, but must be logical in form, he will appear so clumsy and hypocritical.

This is not the first time that Dario Amodei has called for stricter control over China's computing power. You can't expect an American AI scientist to have an innate goodwill towards China. However, in the context of DeepSeek's widespread attention, affirmation and a certain degree of panic in Silicon Valley, he specifically called for further control over China's computing power exports, and strongly denied DeepSeek's innovation in computing power efficiency optimization and model reasoning methods. This is a phenomenon that is very worthy of attention and analysis. No one expects his goodwill towards China, but his malice and resentment towards China and DeepSeek, an AI company born in China, is so heavy that it is worth pondering.

Why does Dario Amodei "look down on" DeepSeek-R1?

Although it is strongly speculated that the training cost of DeepSeek-V3 is more than $6 million, fortunately Dario Amodei did admit that V3 is a real innovation, but he insisted on emphasizing that this is not a breakthrough, but "an expected point on the continuous cost reduction curve." He believes that "the difference is that the first company to demonstrate the expected cost reduction is a Chinese company, which has never happened before and has geopolitical significance." This kind of compliment without really wanting to be sincere is really tiring to watch. I would rather see Amodei say directly: "American companies are all innovating in reducing model costs, it's just that DeepSeek happened to be the first to do it", but straightforwardness is not a quality he possesses.

When it came to DeepSeek-R1, Amodei became straightforward. He absolutely refused to admit that R1 was a breakthrough achievement. He left no room for this issue, regardless of the fact that even OpenAI, which trained the reinforcement learning models o1 and o3, admitted that R1 made an original breakthrough in the reinforcement learning method. He also pretended to turn a blind eye to the research results that pointed out that DeepSeek's reinforcement learning got rid of the intervention of human feedback and was the "AlphaGo moment" of the large language model. He insisted that R1 was just a reinforcement learning based on V3, and all its actions were to reproduce o1. Every American AI company is trying to make inferences in this area. This is a technological trend and has nothing to do with open source. It's just that DeepSeek happened to be made first.

We don't have to be indignant because of Amodei's stubbornness. After all, as a recognized researcher with outstanding achievements in the field of AI, Amodei's views on some key issues can greatly influence the views of the AI ​​industry, venture capital, Wall Street and even Washington, DC on the DeepSeek phenomenon. This is why he had to jump out. He was not speaking up for OpenAI (the grudge between him and OpenAI is deep), but at this time, he had to come out to find a step-by-step preparation for the next move of Anthropic, which he founded.

A very significant fact is that Anthropic has not yet officially released an inference model to the public. Although Dario Amodei publicly stated in an interview that he disdained independent reasoning models - at that time, his main target was of course OpenAI.

Amodei's point of view is: reasoning is not that difficult, the base model is more important. Just like he secretly praised DeepSeek-V3 for its innovation but it was still weaker than his Claude 3.5 Sonnet model in programming and other aspects, he publicly acknowledged the breakthrough made by o1, but did not think that reinforcement learning was the best way to achieve enhanced model reasoning capabilities. He said that in some specific scenarios and practices, Claude 3.5 Sonnet, as a pre-trained model, showed reasoning capabilities that were not weaker than o1. Therefore, he does not think that reasoning models and ordinary models should be separated, and the pre-trained base model is still more important, and reasoning capabilities can be included.

Therefore, it is very likely that Anthropic plans to achieve a leap in model reasoning capabilities in a different way from OpenAI and DeepSeek, which will most likely be reflected in Claude's next-generation flagship base model, and will still adopt a method based on human feedback-based reinforcement learning (RLHF), supplemented by other reinforcement learning methods (Amodei himself said) - this path is significantly different from the breakthroughs of OpenAI o1's Chain of Thought (CoT) and DeepSeek R1 in AI autonomous reinforcement learning.

Anthropic, which was completely born out of OpenAI and regards OpenAI as its most direct (almost only) competitor, is, in a sense, the most fundamentalist believer in OpenAI's series of large language model concepts in the pre-GPT-4 era. Amodei has repeatedly stood up to deny that as training data is exhausted, pre-training has "hit the wall" and the scale effect has decreased, and has repeatedly emphasized the importance of the classic "Scaling Law" (that is, the continuous expansion of the model scale can lead to performance enhancement). AI researchers and developers are really looking forward to Anthropic breaking the bottleneck of Scaling Law and pre-trained models and launching a new generation of flagship pre-trained models with stronger reasoning capabilities.

But so far, Anthropic has not launched this thing. With its excellent model training and a history of never releasing futures, there is reason to believe that Anthropic is nervously preparing this pre-trained model with stronger reasoning capabilities to prove that OpenAI's o1 is not the best path to achieve improved reasoning capabilities. But with the launch of DeepSeek-V3, they suddenly have more things to prove.

First, DeepSeek-V3 further proves after R1 that the path of independent reasoning models of reinforcement learning is reliable and may even be the best; secondly, DeepSeek-V3 verifies that reinforcement learning can enable AI to think deeply without human feedback (Dario Amodei is one of the main inventors of reinforcement learning based on human feedback); thirdly, DeepSeek-V3 proves that the training cost to achieve all this can be significantly reduced.

This means that once Anthropic launches a new pre-trained model with stronger reasoning capabilities, it will have to answer more complex questions than in the past: Why not use reinforcement learning as the main training mode? What are the advantages of reinforcement learning based on human feedback over autonomous reinforcement learning represented by R1? And what is your training cost? Is there a cheaper and more efficient way? Can the API price be reduced? (Claude API is the most expensive in the world, and Deep Seek is almost the cheapest)

And these thorny problems and troubles are all caused by DeepSeek.

Therefore, before launching its own new model with stronger reasoning capabilities, Anthropic's "soul figure" Dario Amodei can only take the initiative to jump out and try his best to reduce and dispel people's preconceived good impression of DeepSeek-R1: it is absolutely impossible to admit that it is an innovation and breakthrough, and it is also difficult to accept that its cost has really been reduced.

Dario Amodei (Source: Wikipedia)

This is a question of two routes, a bit of a "life-and-death" flavor. And these two routes, to some extent, are also different representations of the two paths of classic Silicon Valley model training and Chinese model training in the "post-pre-training era" of large language models: the former relies on the advantages of computing power resources to improve model performance through the extensive and violent aesthetics of computing power stacking; the latter focuses on algorithm efficiency, reduces training costs through architecture and engineering innovation, and improves model performance at the same time.

Anthropic is even more representative of computing power scale, model scale and violent aesthetics than OpenAI, which also led to Dario Amodei's newly published article, which not only secretly released the malice against DeepSeek, but also undisguisedly projected this malice to the entire Chinese AI field.

Why is Dario Amodei so obsessed with computing power export control?

This is not the first time that Dario Amodei has publicly called for strengthening China's computing power export control. He has previously expressed in an interview that China's computing power export control is necessary and needs to be strengthened. American friends should not regret this, and Chinese friends do not need to be angry about it. He has always been like this.

But taking advantage of the "DeepSeek effect", Amodei took the opportunity to write a few thousand words, saying that DeepSeek is behind the trend that China's artificial intelligence may keep pace with the United States, and called for further strengthening of China's computing power control, which is very interesting. Believe me, when an American scientist or entrepreneur publicly expresses an attitude that is too close or hostile to China, their personal demands come first.

Let's first re-examine what Anthropic is.

There is no doubt that it is the best artificial intelligence company in the United States and the world at present - sometimes even the best, and Dario Amodei is its technical soul. Compared with the self-contradiction and coyness when belittling DeepSeek and talking about the export control of computing power, when he talked about the vision and limitations of artificial intelligence and explained specific artificial intelligence terms and theories, he presented a convincing rationality, restraint, clarity and precision, which is much more convincing than his former colleague, Sam Altman, the CEO of OpenAI who really doesn't understand technology.

Of course, as the main competitor of OpenAI, the label that Anthropic impressed the outside world the most is "safety", which is also the most criticized part of OpenAI. Of course, it has done a lot for security, such as the principle of "Constitutional AI" that ubiquitously embeds reinforcement learning based on human feedback (RLHF) into the entire model training process. "Safety" is Anthropic's selling point, and sometimes it becomes its burden.

In 2024, Anthropic took away 15% of OpenAI's market share in the enterprise market, of course, because the Sonnet 3.5 model is indeed powerful, and on the other hand, it is thanks to the amulet of "safety". But think about it carefully, the main focus on "safety", in addition to targeting corporate users, who should be the main buyer?

The answer is obvious: the government. To be precise, the US government.

But in participating in projects of the federal government and related departments, Anthropic, as a latecomer, is obviously not as popular as OpenAI. The first major AI project in the Trump 2.0 era, "Stargate", was decided by the White House, and the main participants were OpenAI and SoftBank, and Anthropic had no share.

Although Dario Amodei immediately ridiculed the Trump administration's "Stargate" as "a mess" at the Davos Forum, it is clear that no AI company wants to participate in the US government-led project more than Anthropic. To this end, he has also done a series of contradictory things:

On the one hand, on January 6, before Trump was about to officially take office, Dario Amodei published a signed article in the Wall Street Journal titled "Trump Can Ensure America's Leading AI", which was a clear test of the waters for active cooperation.

On the other hand, the controversial "Frontier Artificial Intelligence Model Safety and Security Innovation Act" launched at the end of the last Democratic administration's term, which is committed to strengthening supervision and requiring artificial intelligence companies to actively share model research results with the government, was almost unanimously opposed by the progressive and conservative camps in Silicon Valley, and was finally refused to be signed by California Governor Newsom. And our Dario Amodei is almost the only founder of an artificial intelligence company in Silicon Valley who supports this bill.

In the past, I naively thought that Anthropic had the shadow of early Google, because this company put transparency, explainability and ethics at the bottom of technology and products, and had the brilliance of idealism. However, early Google built this principle into the core value of the founders and the team, and did not advocate relying on supervision and administrative will to achieve all this. The two founders of Google never tried to discipline themselves as compradors of the White House. But our Dario Amodei is not like that.

Unfortunately, the Trump cabinet, which is full of new supporters in Silicon Valley, has very different ideas from the Biden cabinet on the development and regulation of artificial intelligence. At least for now, this group does not buy Dario Amodei's account. After Amodei published the bizarre article calling for stricter control of computing power in China, Marc Andreessen, the founder of the venture capital firm Andreessen Horowitz, which supports Trump, came out to slap him in the face: "Closed source, opaque, nitpicking, seeking political manipulation versus open source and free, is not the way the United States needs to win."

In a sense, Dario Amodei, who is determined to obtain large orders from the federal government and hopes to participate in national artificial intelligence "big projects", unconditionally supports AI regulation during the Biden cabinet, and touts Trump as the savior who ensures the leading position of American AI after Trump's election, is actually trapped in ecological isolation. He is not in the core circle of American AI policy making, but he very much wants to get in, which requires him to show a more radical and resolute attitude to get this ticket.

At this time, DeepSeek appeared, which made him somewhat passive in the path of reinforcement learning, but also gave him a good opportunity to express his radical stance to curb the development of China's artificial intelligence. However, Anthropic's model training path depends on the scale expansion of computing power, which makes him unwilling to believe that algorithm efficiency and engineering optimization can really reduce computing power costs, and believe that strangling the neck of computing power can cut off the future of China's AI. And this proposition is the easiest for the White House to understand and most likely to accept. Therefore, it is not difficult to understand why Amodei is so obsessed with calling for stricter computing power export controls.

I still can't help but sigh: the core figures of the new generation of artificial intelligence companies in the United States - whether it is Sam Altman of OpenAI or Dario Amodei of Anthropic, or even Zuckerberg of Meta and Alexandr Wang of Scale.ai, they and their careers accept the discipline of American "nationalism" so naturally and quickly. However, the "discipline" accepted by most Chinese AI entrepreneurs, the latest representative of which is DeepSeek and its founder Liang Wenfeng, is cosmopolitanism and globalization. This is really an interesting phenomenon.