Andrej Karpathy joined Anthropic on 19 May 2026. Most people read it as an OpenAI story — he co-founded the lab, left twice, and landed at the rival.
That's not what this is. Eight weeks before he made the call, Karpathy built a 630-line Python script that ran 300 experiments on his own code in two days and found improvements he'd missed after months of hand-optimizing.
He called it "the final boss battle." Anthropic then offered him the mandate to run that exact loop at the most consequential scale in the industry: using Claude to accelerate its own pre-training research.
In this episode of In The Loop, I'm tracing what actually changed Karpathy's mind — from calling AI agents "terrible" in October 2025 to joining the lab pushing recursive self-improvement six months later.
I cover autoresearch, what it found, and why Dario Amodei's quote about the AI-assistance multiplier is the most important thing said in frontier AI right now.
⏭️ Episode highlights
(01:15) – Karpathy calls AI agents "terrible" in Oct 2025
(02:30) – What changed: coding agents that basically work now
(03:30) – 700 experiments, two days, 11% faster training
(05:00) – What pre-training is and why it costs hundreds of millions
(06:15) – AlphaEvolve and Codex: every lab running the same loop
(07:10) – Dario Amodei: the multiplier going from 5% to 40%
(08:20) – The METR study: experienced devs were 19% slower
🔗 Links & resources
Episode transcript with more resources on the Mindset AI blog
If you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.
🤝 We're social
Stay in the loop, even when you're not listening to this podcast.
Mindset AI
