Who he is
Anthropic pre-training team (since 19 May 2026) — building a group that uses Claude to accelerate pre-training research, under Nick Joseph. Founder, Eureka Labs (2024, on ice — “plan to resume my work on it in time”); OpenAI co-founder; ex-Tesla AI director (Autopilot/FSD); designed Stanford CS231n; PhD under Fei-Fei Li. Coined “vibe coding” (Feb 2025).
One-line position: the best way to understand AI is to build it from scratch — and the frontier’s next few years are formative enough that he went back inside to build it.
Discipline & technical bet
Fully technical — the rare educator who ships CUDA. Current bet: AI-accelerated pre-training research — Claude helping improve how Claude is trained. That is a wager that the meta-loop (AI improving AI research) is now the highest-leverage point in the stack — and a move DOWN the stack, from education (presentation layer) to pre-training (deepest infra).
Key claims (Says)
- Understanding AI requires building it — Zero to Hero, nanoGPT, nanochat: comprehension through construction.
- “Decade of agents,” not year of agents (Dwarkesh, Oct 2025): AGI ∼a decade out; current agents lack continual learning and reliability; the industry is overhyping the near term.
- Teacher + AI symbiosis (Eureka Labs thesis): experts design courses, AI scales instruction — now paused, not renounced.
- The stack matters: Python → C → CUDA, not just the API layer.
- The frontier years are formative (May 2026): where the next era gets decided is inside pre-training, now.
Notable predictions — with falsifiable checks
- (Oct 2025) Agents need a decade to do reliable, economically meaningful multi-day work. Check: annual agent-reliability benchmarks; interim read 2027 — if agents reliably hold week-long jobs by 2028, graded Wrong.
- (2024) AI-assisted education scales world-class teaching to everyone. Check: does LLM101n ship as a complete AI-taught course with measured outcomes by end-2027? Signal weakened — founder is at Anthropic.
- (Ongoing) Capabilities keep improving with scale + data + methods. Check: frontier-model deltas per year. His job move IS the bet, doubled.
Revealed behavior (Does)
- Joined Anthropic pre-training, 19 May 2026 — left founder autonomy for frontier R&D. Actions say: the decisive game is training-run research, not applications, not (yet) education.
- nanochat (Oct 2025) + miniseries v1 (Jan 2026): full ChatGPT-clone pipeline, ~$100 training budget — largely hand-written; he publicly noted coding agents fell short for novel systems work. Practices the comprehension he preaches.
- Keeps shipping free educational artifacts even while employed at a frontier lab — the educator identity persists in behavior.
Feels
Wants to be where history is made and fears watching it from the sidelines — “formative years” is longing, not analysis. Wants comprehension to stay possible for individuals as systems scale.
Hears
Fei-Fei Li lineage; OpenAI founding cohort; deep-learning empiricists; now the Anthropic pre-training team (Nick Joseph). Notably NOT downstream of policy/governance circles.
Sees
Near-unique vantage: inside three frontier orgs (OpenAI founding, Tesla deployment at scale, Anthropic pre-training) plus the open-source classroom. When he says what training runs actually look like, that is testimony, not opinion.
Incentive map
Anthropic salary/equity — expect no public criticism of Anthropic strategy; discount his frontier-lab comparisons accordingly. Educational brand monetizes clarity and candor — strong incentive to stay honest about technical limits. Audience-capture risk on X (millions of followers) remains.
Theories aligned with
- Education-through-construction pedagogy
- AI-accelerates-AI research loop (now his day job)
- Adjacent to techno-optimism in a builder register; a measured skeptic on agent timelines — both at once.
What he’s reacting against
- Black-box consumption of AI; API-layer-only understanding.
- Agent hype — “year of agents” talk without continual learning.
- Both “AI is magic” and “AI is just statistics.”
Where he overlaps / splits (with Rich)
- Overlaps: comprehension through building (the harness ethos); measured agent timelines; individual empowerment through understanding; teach the stack.
- Splits: he went to the infra layer; Rich’s Barbell builds at presentation. Karpathy the person is now a data point FOR infra-layer gravity — the strongest builders get pulled down-stack. Watch whether he resurfaces at education (presentation) as promised.
- Vs Ng: Ng teaches use, Karpathy teaches the stack. Vs Altman/Amodei: cares about comprehension, not just capability — though he now works for Amodei’s company.
Track record
- CS231n — launched thousands of careers. OpenAI founding (2015). Tesla Autopilot/FSD (~2017–2022) — deployment at scale, though the program’s timeline promises were chronically optimistic.
- Zero to Hero / nanoGPT / nanochat — the most-cited technical AI education corpus; verifiable (you can run the code).
- Eureka Labs — thesis unproven; now paused. “Vibe coding” — coined Feb 2025, entered the language within months.
Empirical vs normative
Empirical: code walkthroughs, benchmark-grounded claims — unusually checkable. Normative: understanding should be accessible; AI should augment teachers; the frontier should be built by people who comprehend it.
Weak spots / open questions
- Eureka Labs limbo: is “resume in time” real, or the polite word for shelved?
- Employment filter: his frontier commentary now passes through an Anthropic filter — treat lab-comparison claims as partisan.
- Thin on governance/societal implications by choice — deep on how, quiet on whether.
- Does the man who preaches individual comprehension change inside a scaling lab, or change the lab?
Rich’s take
(your synthesis here)
Delta log
WRONG (ours, not his): file carried “Founder, Eureka Labs” as current — he joined Anthropic pre-training 19 May 2026, a month BEFORE our 16 Jun review. The review missed a month-old job change. Learning filed.
DRIFTED: “AI education is the highest-leverage application” — his revealed behavior now says frontier pre-training outranks it (education “in time”). “Teacher+AI symbiosis” — thesis intact, vehicle paused.
HELD: understanding-requires-building (nanochat Oct 2025, miniseries Jan 2026); capabilities-improve-with-scale (he re-upped his career on it); stack-matters pedagogy.
ADDED: decade-of-agents position (Dwarkesh Oct 2025) with falsifiable check; vibe-coding coinage; empathy facets; incentive map re-based on Anthropic employment; layer_focus=infra; decay=quarterly (justification: three affiliation-scale events in 24 months); wake_triggers: job change, major repo, long-form interview.
Sources
- TechCrunch, 19 May 2026 — Karpathy joins Anthropic pre-training (role, Nick Joseph, quotes)
- Dwarkesh Podcast, Oct 2025 — “AGI is still a decade away” / decade of agents
- karpathy/nanochat — repo (Oct 2025) + miniseries v1 discussion (7 Jan 2026)
- Eureka Labs announcement (Jul 2024, X)
- Prior persona v1 (16 Jun 2026) — archived _archives/andrej-karpathy.html.bak-20260824