That is roughly the number of people on the planet who actually understand the visceral, gritty reality of pre-training a frontier model at scale without wasting a billion dollars in compute. Most of the industry is just tweaking hyperparameters and hoping for the best, but Andrej Karpathy is one of the few who treats the process like an actual science. Now, he’s taking that expertise to Anthropic.
For a while, Anthropic has played the role of the “safe” lab. They’ve spent a lot of time on Constitutional AI and making sure their models don’t hallucinate too wildly. But safety is essentially a layer of polish applied to a base model. If the base model is mediocre, you’re just putting a very expensive, very polite coat of paint on a rusted engine. To actually win the scaling war, you need to get meaner about how the core is built.
The move to join the pre-training team suggests that Anthropic is pivoting back to the hard science of data curation and token efficiency. According to TechCrunch AI, Karpathy is stepping directly into the machinery that builds the model from scratch. This isn’t about alignment or RLHF; it’s about the raw ingredients. The industry loves to talk about “emergent properties” as if they happen by magic, but anyone who has actually managed a cluster of H100s knows it’s mostly about fighting hardware instability and cleaning messy datasets.
Karpathy has essentially become the unofficial professor of the AI era. His “Zero to Hero” series is the gold standard for anyone who actually wants to understand the math instead of just calling an API. The real question is whether he can maintain that public-facing, educational persona while locked inside the vault of a frontier lab (and likely under a mountain of NDAs).
It is hard to imagine him continuing to release deep-dive tutorials while working on future versions of Claude. The level of secrecy around pre-training recipes has become obsessive. If he’s designing the next big jump in efficiency, he can’t exactly post a Colab notebook about it on X. We might be seeing the end of the “educator” era and the return of the “stealth” era. Or maybe not—he might find a way to teach the basics without giving away the secret sauce. But for now, the priority is the weights, not the students.
Watching a co-founder move to a direct competitor is like watching a star quarterback switch teams mid-season. It’s not just about the loss of talent—though losing someone who understands the architecture from the ground up is a blow—it’s about the signal. When the people who built the foundation start moving to the new house, you have to wonder if the original foundation has started to crack.
OpenAI has shifted toward a product-first company. They are focused on the ecosystem, the apps, and the enterprise deals. Anthropic, meanwhile, has always felt more like a research lab that happened to build a product. By bringing Karpathy in, they are doubling down on the research side. It’s a bet that the next leap in intelligence won’t come from better UI or more plugins, but from a fundamentally better pre-training run.
It’s a massive win for Anthropic.
Within six months, we’ll see a shift in Anthropic’s technical output toward aggressive data curation strategies rather than just alignment research. If they can systematize the “art” of pre-training that Karpathy is known for, the gap between the top labs will shrink almost overnight. Does OpenAI still have the talent to counter this? Probably. But the momentum has shifted.
