Andrej Karpathy on Software 3.0, Jagged Intelligence, and What You Still Have to Own
Software 3.0 collapses app layers into prompts: Karpathy’s full MenuGen app was rendered obsolete by a single Gemini prompt that did the same thing in one neural-net pass. “That app shouldn’t exist” - and most layer-heavy products face the same problem Andrej @ 5:37.
LLM capability peaks trace verifiability: Opus 4.7 can find zero-day vulnerabilities and refactor 100k-line codebases, but tells you to walk to a car wash 50 meters away - because RL trains hard on verifiable domains and skips everything else Andrej @ 11:29.
Vibe coding raises the floor for everyone; agentic engineering (AI-assisted professional development) extends the ceiling - Karpathy says the 10x engineer framing is an understatement for practitioners who master the second Andrej @ 17:00.
Taste, judgment, and spec design remain the human bottleneck - agents still make structural errors, like trying to match Stripe and Google accounts by email address Andrej @ 19:45.
Andrej Karpathy built an app. Then one prompt made it obsolete.
“That app shouldn’t exist,” he said - and then spent 30 minutes at Sequoia’s AI Ascent 2026 explaining why most of the software being built today has the same problem.
The argument is not subtle. Software 3.0 does not speed up the old paradigm. It eliminates layers. And the implications run from what to build, to how engineers get hired, to what humans are actually for when agents do most of the work. Karpathy, the OpenAI co-founder who built Tesla Autopilot and now runs Eureka Labs, is the rare person who has lived all three software eras. He is not optimistic in a vague way. He is precise about what breaks and what does not.
This post is part of ongoing Software Rewrite Supercycle series of posts that helps investors understand how to evaluate potential investment in software companies who adopt AI the right way.
The Paradigm Break
Karpathy’s software 1.0/2.0/3.0 framework is not new. The MenuGen story makes it land.
