Karpathy's argument is bigger than 'AI writes code': software is on its third rewrite, natural language is the new programming interface, and today's LLMs are an operating system in its 1960s timesharing phase. The part most retellings drop is the discipline — partial autonomy, a fast generation-verification loop, and infrastructure built for agents as first-class users.

Software 3.0 is not autocomplete bolted onto 2.0. The programming interface moved to natural language, and that dissolves the wall between user and programmer.

“remarkably we're now programming computers in English”

watch at 4:16

His most load-bearing analogy: LLMs are an operating system in its 1960s timesharing era — compute centralized in the cloud, every user a thin client. Personal computing's equivalent moment has not arrived.

“we're all just sort of thin clients that interact with it over the network”

watch at 11:00

The psychology section is the talk's sharpest tool: LLMs are superhuman in spots, with failure modes no human has and no memory consolidation between sessions. Design for the deficits, not the demo.

“they're kind of like people spirits. They are stochastic simulations of people”

watch at 14:48

His pushback on agent hype is a timeline correction, not a denial: the transition runs ten years, and products that price that in are the ones that survive it.

“this is the decade of agents”

watch at 27:30

The working pattern he endorses: AI generates, the human verifies, and the product's job is to compress that loop — GUIs built for audit speed, autonomy exposed as a slider the user sets.

“usually they are doing the generation and we as humans are doing the verification”

watch at 22:15

Agents are a new class of user. Ship docs in markdown, put llms.txt beside robots.txt, make actions reachable without a mouse — the sites that do this get read by the machines that now answer questions.

“a simple markdown that's telling LLMs what this domain is about”

watch at 34:17