SecondBrain
Ask the Brain
Index/Entityupdated Sat Jun 13 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Andrej Karpathy

personmlllmkarpathysoftware-3.0auto-research

Andrej Karpathy

ML researcher and educator. Co-founder of OpenAI, ex-head of AI at Tesla (Autopilot), founder of Eureka Labs. Coined Vibe Coding (2025) and Software 3.0. Author of the gist that seeded this Second Brain.

warning Unverified claim (June 2026): per I Turned Claude Fable Into The Ultimate Second Brain (Nate Herk), Karpathy "did just join Anthropic" — stated in passing by Nate Herk, secondhand, not yet corroborated by any other vault source. Treat as rumor until confirmed.

On Fable 5 (June 2026, via the same source)

Two tweets quoted: "You can give Fable a lot more ambitious tasks than what you're used to and the model just 'gets it'" and "Yes, the benchmarks are great… but this is a major version bump deserving step change forward" — with the caveat that "the model still has quirks… the safeguards are configured to be a little too trigger-happy for launch." Pairs with Task Imagination (ambition as the new constraint).

Connection to this wiki

  • Wrote LLM Wiki (Karpathy gist) — the source for the LLM Wiki Pattern this vault implements.
  • Quote that drives this vault's design: "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase."
  • His own LLM-knowledge-base habit (per Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent)): every article he reads is re-projected into a personal wiki — "different projection on information" as a learning tool.

Recurring framings (worth borrowing)

  • Software 3.0 — programming-by-prompting; LLM as interpreter
  • Vibe Coding vs Agentic Engineering — floor vs ceiling
  • Jagged Intelligence — capabilities are spiky, follow verifiability + lab focus
  • Ghosts not animals — LLMs are statistical simulation circuits; don't anthropomorphize
  • "You can outsource your thinking but you can't outsource your understanding" — tweet that "blew his mind"; the human's irreducible role is direction-setting and understanding
  • auto research (~Q2 2026) — give an agent something to improve + one clear metric, then let it loop (try / measure / keep-or-revert / never stop). The framing: the human can sleep. See Build Self-Improving Claude Code Skills (Simon Scrapes) for Simon Scrapes's lift of the pattern to Claude Code skills.

Predictions worth tracking

  • 2026 equivalent of "websites in the 90s" — neural-first computers, diffusion-rendered UIs from raw audio/video; classical CPUs as co-processors. He stresses it's TBD and piecewise.
  • Hiring for agentic engineers — whiteboard puzzles obsolete; replace with "build this big project, then 10 codecs try to break it."
  • Founder advice — find verifiable domains the labs aren't covering; build RL environments; fine-tune to spiky capability.

2026-06-13 — a concrete RSI datapoint (Economist)

Per How AI Got Better at Building Itself (Economist), Karpathy trained Nanochat (a GPT-2-class model) on 8 GPUs in ~3h, then handed optimization to his own agent, which cut training to 1h39m (~18%) autonomously — no human in the loop. Cited as a concrete instance of Recursive Self-Improvement in the wild, and a practical realization of his Auto Research Loop (Karpathy) framing (give an agent a metric, let it loop, the human sleeps).

Sources