BeCuriouzLearn something new every day
Karpathy says he's 'never felt more behind as a programmer.' Here's what changed.
💡 Karpathy sees three eras of software. We're in Software 3.0 where context windows are programs. Human value shifts to taste, judgment, and specs — not writing code.
In December 2025, Andrej Karpathy — one of AI's most respected researchers — had a moment that shook him. Testing the latest models during vacation, he started requesting larger code blocks and stopped correcting the output. The agent's code was more reliable than what he'd write by hand.
He defines a framework with three eras: Software 1.0 (human-written rules), Software 2.0 (learned neural network weights), and now Software 3.0 — where context windows are programs and LLMs are interpreters. Text replaces bash scripts. Agents handle environmental adaptation.
But he identifies a critical problem: jagged intelligence. Frontier models can refactor 100,000-line codebases but recommend walking 50 meters to a car wash, missing the obvious intent. Models excel where labs applied reinforcement learning (code, math) and fail where they didn't.
His advice for what humans should focus on: taste, judgment, and specification clarity. "You can outsource thinking but not comprehension." The enduring human role is architectural decisions, design tradeoffs, and understanding why work matters.