🛠️Karpathy: Software 3.0 Is Prompts, Agents, Memory, Verification
TL;DR
Andrej Karpathy's Sequoia Ascent recap argues LLMs are no longer chatbots or autocomplete but a new programmable layer. He names December 2025 as the inflection when agents stopped being demos and started doing real work, and sketches what Software 3.0 looks like.
Andrej Karpathy's Sequoia Ascent recap argues LLMs are no longer chatbots or autocomplete but a new programmable layer. He names December 2025 as the inflection when agents stopped being demos and started doing real work, and sketches what Software 3.0 looks like.
Key Points
Posted late May 2026 on Karpathy's personal blog (talk delivered April 29)
Software 1.0 = explicit code, 2.0 = trained weights, 3.0 = prompts + agents + tools + memory + verification
December 2025 named as the moment agents became practical for real work
Example: 'menugen' — an app fully built and run by an LLM behind a thin UI
Argues new AI-native companies will look more like ops + verification loops than SaaS
Why It Matters
If Karpathy's frame holds, every team needs a verification layer as serious as their test suite — and the next defensible companies will be the ones that own that loop, not the ones with the best wrapper UI.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,483 builders reading daily.