Bio
Andrej Karpathy's framework centers on a three-era model of software development. Software 1.0 is traditional code written by humans. Software 2.0 is neural networks where the "code" is learned weights, which he observed eating Software 1.0 at Tesla Autopilot. Software 3.0 is the current era where natural language prompts become programs and LLMs become a new kind of computer.
Operating themes
- Operating thesis: We are in the 1960s of a new computing paradigm where LLMs are programmable in English, software is being rewritten from scratch as Software 3.0, and the barrier to building apps is dropping toward zero so that anyone can create software as easily as they make a TikTok.
- Building With Llms
- Agentic Ai Design
- No Code Product Building
- Ai Product Strategy
Cards
- An LLM should maintain a wiki, not re-derive knowledge per query — LLM should maintain a wiki, not re-derive knowledge per query [Tier A]
- You can outsource thinking, but not understanding — verification is the new human job — You can outsource thinking, but not understanding — verification is the new human job [Tier A]
- AI capability is not evenly distributed — it spikes where labs have data, rewards, and verification loops — AI capability is not evenly distributed — it spikes where labs have data, rewards, and verification loops [Tier A]
Sources captured
- 2026-04 — "LLM Wiki" gist (
raw/essays/karpathy--llm-wiki--2026-04.md) - 2026-04 —
vibe-coding-menugen-karpathy.md(operator essay archive) - 2026-04 —
andrej-karpathy-on-software-30-software-in-the-age-of-ai.md(operator essay archive) - 2026-04 —
andrej-karpathy-software-is-changing-again.md(operator essay archive) - 2026-04-30 — Sequoia AI Ascent 2026 fireside (Software 3.0)