This is not just an ordinary job change in Silicon Valley, but a transfer of knowledge power.

Written by: Hulin Dance King

Edited by: Jingyu

If someone told me a few years ago that one of the co-founders of OpenAI would go to Anthropic to help a competitor with pre-training research, I would probably think this person was talking about a science fiction plot.

But this event has really happened today.

Andrej Karpathy, this name needs almost no introduction in the AI circle. The speaker of the Stanford CS231n course, the most popular science communicator in the field of deep learning, a co-founder of OpenAI, and the former head of the Tesla autonomous driving team. A tweet from him can make a certain technical direction's popularity soar, and a video he uploads on YouTube about Transformers can easily exceed a million views.

This is the person who announced today that he is joining Anthropic.

Karpathy's official announcement on X | Image source: X

Karpathy will join Anthropic to focus on pre-training research and lead a new team whose core task is to use Claude to accelerate the exploration in the pre-training direction.

Pre-training is the foundation of large model capabilities. Whoever makes breakthroughs at this level will have an advantage in the competition in the coming years. Anthropic's placing Karpathy here is unmistakably clear.

But if this event is only understood as "an impressive person changing jobs," then it is greatly underestimated.

Karpathy carries a rare quality in the AI circle—an overlap of technical credibility and public influence. He is not just a researcher who can write good code and publish good papers; he is the kind of person who can motivate other top researchers to willingly follow him.

There is a saying in the industry, the joining of a reputable researcher often prompts a group of people to reevaluate their career choices. Karpathy's arrival may signal a wave of talent influx for Anthropic.

What is even more intriguing is his starting point. In 2015, he was one of the co-founders of OpenAI and witnessed the transformation of the company from a nonprofit ideal. Later, he went to Tesla, then briefly returned to OpenAI, and subsequently left to start his own venture.

Choosing to join Anthropic this time carries a certain "statement" meaning.

01 Anthropic has been winning

Looking at Karpathy's joining in isolation misses an important context; Anthropic has been experiencing a rare upward trend recently.

Two weeks ago, data from Ramp AI Index quietly circulated in the tech media.

The data showed that Anthropic's adoption rate among enterprises increased by 3.8 percentage points in April, reaching 34.4%, while OpenAI's dropped by 2.9 percentage points, falling to 32.3%. This is the first time in Anthropic's history that it has surpassed OpenAI in enterprise adoption rates, and although the gap is not drastic, it is highly directional.

In the same week, Anthropic launched a version of Claude for small businesses, integrating a group of tools that small and medium-sized enterprises rely on daily, such as QuickBooks, PayPal, HubSpot, Canva, and DocuSign, embedding AI capabilities directly into these users' workflows. This is a clear signal towards lower-tier markets; Anthropic is no longer just focusing on large enterprise clients, but is moving towards a broader market.

The day before that, Anthropic announced a partnership with the Gates Foundation, committing to invest $200 million over four years, along with Claude usage quotas and technical support, covering global health, education, and economic development sectors. The amount of this partnership is not the most striking, but the narrative value it brings is quite high; a company originally focused on "AI safety" is increasingly embodying the concept of "responsible AI."

At a point where valuation approaches a trillion dollars and enterprise adoption rates have just reversed, Karpathy's joining is the highlight of it all.

Fortune magazine's commentary headline is direct: "Anthropic seems unable to stop winning."

02 Why not return to OpenAI?

If someone wins, someone else must be under pressure.

Karpathy is not the first person to leave the OpenAI system to join Anthropic.

The founding team of Anthropic itself, Dario Amodei, Daniela Amodei, and a group of core researchers at that time, collectively left OpenAI in 2021 to establish this company. In a way, Anthropic has been a product of a route divergence within OpenAI since its birth.

In the following years, as OpenAI accelerated in commercialization and productization, hastening releases, chasing revenue, and gradually aligning itself with Microsoft, some researchers who valued "pure research" or "safety first" began to vote with their feet.

Karpathy's choice to join Anthropic comes at a sensitive time. OpenAI has been quite intensive in its external narrative recently, with multiple lines of GPT series, o series, Sora, and Operator advancing simultaneously; internally, the pace has been described by some in the industry as "like running three marathons at once." Amid such rapid expansion, how to retain those who genuinely care about the depth of research and are not just seeking valuation is a difficult problem to solve.

Of course, OpenAI still has a very strong talent density and resource scale; one departure will not shake its foundation. But if this kind of turnover becomes a trend, what is truly noteworthy is the changes in industry expectations it conveys.

A tech analyst put it very directly: "AI development is no longer just a technical competition, but a war for knowledge leadership. The movement of an influential researcher can reshape the research priorities of the entire industry."

Karpathy's influence in the deep learning community perfectly matches this assessment. His Stanford lectures and YouTube videos are introductory materials for many researchers currently working in top AI labs. Wherever he goes, there is a sort of endorsement that "this direction is worth betting on."

03 Pre-training, hitting the future

Returning to the specific focus of Karpathy's joining Anthropic, it is pre-training.

In the past two years, the industry’s attention has been largely focused on reasoning, multimodal, agents, and RAG, which are relatively "close to the application layer" directions. Breakthroughs in foundational model capabilities have been viewed by some as having entered the "fine-tuning and optimization" phase, rather than a fundamental leap.

Anthropic clearly does not see it this way. Assigning Karpathy to specifically form a team to explore "using Claude to accelerate pre-training research" is a bet on a more foundational, longer-term, but potentially more rewarding direction.

There is an interesting logic hidden here; using existing large models to assist in pre-training the next generation of large models is an idea of "AI helping AI evolve." This path is still quite new and lacks a mature roadmap, but if it can be successful, it means that training efficiency and capability boundaries could see nonlinear improvements.

Entrusting this task to Karpathy represents a bold technical bet by Anthropic.

The talent war in the AI industry has reached a point where it can no longer be described by simply poaching a few engineers. It is more like a struggle for "narrative power"; whoever attracts those who can define research directions sends a signal to the entire industry that we are the protagonists of this game in the future.

Karpathy's choice may just be such a signal.