Author|桦林舞王
Editor| Jingyu
If someone had 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 probably would have thought they were describing a plot from a sci-fi novel.
But this is something that really happened today.
Andrej Karpathy, a name that hardly needs introduction in the AI community. Main lecturer of Stanford's CS231n course, the most popular science communicator in the field of deep learning, OpenAI co-founder, former head of Tesla's Autopilot team. A single tweet from him can make a technical direction skyrocket in popularity; a video he uploads to YouTube explaining Transformers easily gets millions of views.
And it is this very person who announced today that he is joining Anthropic.
Karpathy's official announcement on X|Image Source: X
Karpathy's role at Anthropic will focus on pre-training research, leading a new team whose core mission is to use Claude to accelerate exploration in pre-training directions.
Pre-training is the foundation of large model capabilities. Whoever makes a breakthrough at this level holds the initiative in the competition for the next few years. Placing Karpathy here makes Anthropic's intent abundantly clear.
But if we only understand this event as "a talented person changing jobs," we are severely underestimating its significance.
Karpathy embodies something extremely rare in the AI community—a dual overlay of technical credibility and mass influence. He is not just a researcher who can write great code and publish papers; he is the kind of person whom other top researchers willingly follow.
There's a saying in the industry: the addition of a prestigious researcher often prompts a group of people to reassess their career choices. Karpathy's arrival could be the signal flare for an impending wave of talent influx at Anthropic.
Even more intriguing is his motivation. In 2015, he was one of the co-founders of OpenAI, witnessing firsthand the company's entire transformation from its non-profit idealistic beginnings. He later went to Tesla, then briefly returned to OpenAI, before leaving again to start his own venture.
This choice of Anthropic carries a certain degree of "statement".
01 Anthropic on a Winning Streak
Viewing Karpathy's joining in isolation misses an important context: Anthropic has recently been in a state of uncommon upward momentum.
Two weeks ago, data from the Ramp AI Index quietly went viral among tech media circles.
The data showed that Anthropic's enterprise adoption rate rose by 3.8 percentage points in April to 34.4%, while OpenAI's fell by 2.9 percentage points during the same period, dropping to 32.3%. This marks the first time in history that Anthropic has surpassed OpenAI in enterprise adoption rate. Though the gap isn't yet wide, the directional significance is strong.
In the same week, Anthropic launched a version of Claude for small businesses, integrating tools like QuickBooks, PayPal, HubSpot, Canva, and DocuSign that small and medium-sized enterprises rely on daily, embedding AI capabilities directly into these users' workflows. This is a clear signal of moving down-market; Anthropic is no longer just targeting large enterprise clients, it's moving into a broader market.
Just the day before that, Anthropic announced a partnership with the Gates Foundation, committing $200 million in funding, Claude usage credits, and technical support over four years for global health, education, and economic development, among other areas. The financial amount of this collaboration isn't the most eye-catching, but its narrative value is high. A company originally focused on "AI safety" is increasingly solidifying its identity around "responsible AI."
At the moment when financing valuations are nearing the trillion-dollar mark and enterprise adoption has just completed an overtake, Karpathy's joining is the highlight at the end of all this.
Fortune magazine's headline put it bluntly: "Anthropic Can't Seem to Stop Winning."
02 Why Not Return to OpenAI?
Where there are winners, there are inevitably those feeling the pressure.
Karpathy is not the first person to leave the OpenAI system for Anthropic.
Anthropic's founding team itself—Dario Amodei, Daniela Amodei, and a group of core researchers at the time—were a collective that left OpenAI to found the company in 2021. To some extent, from its very first day, Anthropic was the product of an internal divergence in philosophy at OpenAI.
In the years since, as OpenAI has moved faster and faster on commercialization and productization—accelerating releases, chasing revenue, increasingly aligning with Microsoft—some researchers who place greater value on "pure research" or "safety first" have begun voting with their feet.
Karpathy's choice of Anthropic comes at a sensitive time. OpenAI has been quite intensive in its external narrative recently, with the GPT series, the o series, Sora, and Operator all advancing simultaneously. The internal pace is so fast that some in the industry privately describe it as "running three marathons at once." Amid such rapid expansion, how to retain those who truly care about research depth and not just valuation is a difficult puzzle to solve.
Of course, OpenAI still possesses immense talent density and resource scale; a single departure won't shake its foundation. But if such movement becomes a trend, what truly warrants attention is the shift in industry expectations it signals.
A tech analyst put it bluntly: "AI development is no longer just a technical race; it's a war of intellectual leadership. The movement of an influential researcher can reshape the entire industry's research priorities."
Karpathy's own influence in the deep learning community perfectly fits this assessment. His Stanford lectures and YouTube videos are introductory material for many researchers now working in top AI labs. Where he goes carries a certain endorsement of "this direction is worth betting on."
03 Pre-training: Fighting for the Future
Returning to the specific focus of Karpathy's joining Anthropic—pre-training.
Over the past two years, the industry's attention has heavily concentrated on relatively "application-layer adjacent" directions like inference, multimodal capabilities, agents, and RAG. Breakthroughs in foundational model capabilities are seen by some as having entered a stage of "fine-tuning and optimization" rather than fundamental leaps.
Anthropic clearly doesn't see it that way. Tasking Karpathy with specifically forming a team to explore "using Claude to accelerate pre-training research" is a bet on a more foundational, longer-cycle, but potentially higher-reward direction.
There's an interesting logic hidden here: using existing large models to assist in the pre-training of the next generation of large models is an "AI helping AI evolve" approach. This path is still very new, without a mature roadmap, but if it can be made to work, it could mean nonlinear improvements in training efficiency and capability boundaries.
Entrusting this to Karpathy is a bold bet by Anthropic on a technical direction.
The talent war in the AI industry has reached an intensity today that can no longer be described as simply poaching a few engineers. It's more like a battle for "narrative control." Whoever attracts the people who can define research directions is sending a signal to the entire industry: we are the future protagonists of this game.
Karpathy's choice is perhaps just such a signal.
