Anthropic Engineer Andrej Karpathy: "LLM models writes code - and now it writes it better than you." Now LLMs write 70–90% of production code at top AI labs. Agents need the same thing: a loop that verifies its own output. No baseline = no benchmark = no improvement. In... this 15-min talk, Karpathy breaks down how AI models actually works.show more

Generation is a commodity. Verification is the only alpha left.

is the 70-90 number from anthropic + openai internal or an industry average?

In general from industry

He always know that model more important than scaling

70 to 90 percent of code written by AI now is a huge shift

Solid take. Verification loops are the real unlock for agents generation is now table stakes.

@0xRicker That verify-its-own-output loop is the whole game, and it's the exact step humans quietly drop the second a deadline shows up. Funny that we're now engineering into machines the discipline we keep talking ourselves out of.

the verification loop is the missing layer between demos and production

benchmarks make agents actually improve

andrej's 15 minute talk is like a mini lecture

Writing code was never the bottleneck. It was the joy.

agents need the same verification loop that got LLMs to 70-90% of production code. most people skip this step

写代码的重心都挪到验收上了

That much scale doesnt necessarily mean more accuracy, though it just means the model can fit a lot of noise and still produce plausible looking code.

Writing code is only half the job

karpathy quietly saying LLMs already write more code than the engineers watching is peak 2026 energy

the verification loop is the real unlock - without tests and evals, 90% generated code just means 90% of the risk is automated too