🕶️ OpenAI's co-founder just switched teams

The scarce resource in AI isn't GPUs anymore, it's the people who know how to aim a trillion of them, and the physics-aware models that can turn design cycles into competitive moats. Andrej Karpathy just moved to Anthropic's pre-training team, betting that smarter algorithms will beat bigger clusters.

Mistral bought a Viennese simulation lab to make sovereign AI actually useful in foundries and packaging plants. And somewhere between those two moves, the frontier shifted: from who can afford the most compute to who can make training cheaper, inference faster, and industrial deployments regulator-friendly.

Docs piling up faster than you can write them? Same.

Every team knows the feeling — product ships, docs don't. Changelogs get forgotten. Style violations quietly accumulate. Broken links go unnoticed for months.

Mintlify's new Workflows feature fixes this. Define automation rules, and the agent handles the recurring maintenance work for you — on your schedule, by your rules.

Draft docs when a PR merges. Generate changelogs every Friday. Run a style audit on every push. Flag translation lag before it becomes a problem. Each workflow is version controlled, fully configurable, and fits into your existing review process.

You decide when it runs, what it checks, and whether changes get committed directly or opened as a pull request for review.

The result: documentation that actually keeps up with your product, without someone manually chasing it down.

Try Workflows

Mistral is buying its way into the hardest part of European AI: factories and foundries. The Paris lab has acquired Vienna-based Emmi, a specialist in airflow, heat, and material-stress simulation, in an industrial AI acquisition aimed squarely at aerospace, automotive, and semiconductor design teams.

Emmi sits in “physics‑aware AI”: neural nets trained on CFD/FEA outputs that trade a bit of resolution for orders‑of‑magnitude faster runs. For customers whose design cycles are gated by simulation throughput, that’s not a nice-to-have; it’s a new ceiling on how fast products can ship.

Combined with Mistral’s earlier infra buy, the Emmi deal brings both deployment and physics modelling capabilities inside one European perimeter, a pattern sovereign‑AI advocates have been demanding. The bet is that EU industrials will pay a premium for tightly scoped, regulator-friendly systems even as they comparison-shop generic frontier models from US labs.

Andrej Karpathy, a co‑founder of OpenAI and former Tesla AI lead, has joined Anthropic to work on pre‑training, the large‑scale runs that give models their core skills. He started this week and will build a team that uses Claude to speed up pre‑training research, one of the most compute‑intensive phases of model development. He posts personal update, and TechCrunch outlines team mandate.

Least obvious consequence: if technique and AI‑assisted research, not just raw compute, become the edge, budgets shift toward data quality, evaluation, and training‑orchestration vendors. Anthropic also strengthened model security by adding veteran researcher Chris Rohlf to its frontier red team, which stress‑tests advanced systems; he expands frontier red team. Expect faster capability pushes paired with stricter safety gates across the lab ecosystem, raising the bar for external integrations.

For founders selling AI infrastructure, data tooling, or eval platforms: tune your pitch to lower cost per training token and shorten iteration loops, and be ready with proofs that show efficiency gains, not just throughput. For applied‑AI product teams, plan for quicker model refresh cycles and tighter safety reviews before deployments. Edtech builders should watch whether Karpathy resumes education work later, which could reset expectations for tutor‑style agents and curricula.