#product

Andrej Karpathy on why OpenAI’s first computer-use agents failed, why model capability had to come first, and why agent builders should think in decades rather than demos.

Notes from Lenny's Podcast with Andrew Ambrosino on Codex, cheap implementation, taste, role collapse, dogfooding, product planning, and why the future of work may look less like chat and more like a home base for delegated action.