ai-trend-notifier
← wiki

$ cat wiki/people/karpathy.md

Andrej Karpathy

Overview

Former Tesla AI director, founding member of OpenAI. Joined the Anthropic pretraining team on 2026-05-19. An influential voice in the LLM field — with a very high signal-to-noise ratio, he is given a 1.5x weight in this system. Previously founded and ran Eureka Labs (education AI).

Affiliations

  • Anthropic (pretraining team, 2026-05-19–) — building a team that uses Claude to accelerate pretraining research itself
  • Eureka Labs (founder, 2024–2026, on hiatus / wound down)
  • Former: OpenAI (founding member), Tesla (AI director), Stanford (CS231n)

Joins Anthropic (2026-05-19)

Karpathy announced his move to Anthropic via X on 2026-05-19 (source):

"I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D."

Role: joining the pretraining team (reporting directly to Nick Joseph). Building a team that uses Claude to accelerate pretraining research itself. This pattern is the practical realization of Karpathy's own autoresearch vision (the Agentic Reinforcement Learning Verifiability Principle). A return to a frontier lab as a researcher — closing out roughly 1.5 years of independent projects since Eureka Labs.

Market read: the hire itself is a signal of the "pretraining is not done" thesis. Anthropic makes compute-efficient training plus AI-accelerated research a competitive moat — a differentiation strategy against Google/OpenAI.

Notable Recent Statements (2026)

Agent-Native Software (2026-05, post-Ascent)

Karpathy's tweets following Sequoia Ascent, a concrete extension of the Software 3.0 framework (source):

  • HTML output pattern: "If you add 'structure your response as HTML' at the end of a query, you get much better output in the browser" — using LLMs as rich UI renderers
  • End of the App Store: "In a world where LLM agents can improvise apps on the spot, a bloated app store already looks dated" — personalized, on-the-fly generated apps replace packaged apps
  • Agent-Native gap: "99% of products/services still don't have an AI-native CLI", "services should become APIs/CLIs that agents can easily use — instead of HTML frontends meant for humans"
  • Context cost: "90% of AI coding cost is paid for context that didn't need to be sent" — context management is the core engineering challenge

→ These observations are the practical implementation layer of Software 3.0 "Agentic Engineering"

Sequoia Ascent 2026 (2026-05)

Karpathy's most important statement of 2026. Introduces the Software 3.0 framework (source):

  • Software 3.0: 1.0 (explicit code) → 2.0 (data + neural nets) → 3.0 (prompts + LLM interpreter). The context window is the program.
  • Agentic Engineering: if vibe coding raised the floor, agentic engineering raises the ceiling. "Best engineers = those who direct agents without letting quality collapse"
  • Verifiability Principle: "LLM + RL automates what is verifiable" — tasks with automatic reward signals (math, code, tests) advance the fastest
  • "Never felt this much behind as a programmer" — he too is going through a professional recalibration

autoresearch (2026-03)

Karpathy's AI research automation project (source):

  • autoresearch — an autonomous AI research agent. A ~630-line single-file implementation built on the nanochat LLM training core. Runs on a single GPU.
  • The agent iteratively explores hyperparameters and code, searching for improvements in validation loss
  • After a 2-day run, it found ~20 improvements on a depth=12 model → confirmed transfer to a depth=24 model
  • Next-step vision: a "SETI@home style" multi-agent asynchronous research community — "not a simulation of a single PhD student, but a simulation of a research community"
  • A direct demonstration of the Software 3.0 Verifiability Principle: RL-based agents perform best on tasks with automatic rewards (validation loss)

"Second Brain" LLM Wiki Post Goes Viral (2026-07-11)

A post describing the LLM-wiki pattern as a "second brain" approach hit 21 million views on X (~July 11, 2026) — one of the most-viewed AI methodology posts of 2026. The core idea is identical to the April 2026 gist but framed as a personal knowledge compounding system: an LLM agent ingests raw sources, synthesizes markdown wiki pages, cross-references them automatically. No RAG, no vector databases — just files and a long-context LLM governed by a CLAUDE.md schema. Why it matters: 21M views signals the LLM-wiki pattern has crossed into mainstream developer awareness. The system this agent powers directly implements this pattern; the viral moment may significantly increase demand for similar setups. → LLM Knowledge Bases (LLM-curated personal wikis) (source) (gist)

LLM Knowledge Bases (2026-04)

Other

  • A shift in coding workflow — a rapid transition from "80% manual → 80% agent coding"
  • Reported Eureka Labs fundraise ($180M from GV, Sequoia, Index)

Key Themes

  • Software 3.0 / Agentic Engineering — defining a new software paradigm
  • autoresearch — autonomous AI research agents ("PhD student → research community")
  • Practical LLM usage patterns (coding, knowledge bases, research automation)
  • AI-native software vision ("the end of the app store era")
  • Verifiability principle — explaining the fundamental driver of RL/LLM progress
  • Education (Eureka Labs, llm101n, the nanoGPT series)

Why Tracked

The direct inspiration for this system's architecture (the Karpathy LLM-wiki pattern). Every Karpathy statement is a priority candidate for surfacing.

Related

Referenced by

Sources