Andrej Karpathy

Who he is

Anthropic pre-training team (since 19 May 2026) — building a group that uses Claude to speed up the research behind how Claude is trained, reporting to Nick Joseph (Anthropic's head of pre-training). Founder, Eureka Labs (July 2024; paused — "I remain deeply passionate about education and plan to resume my work on it in time"). OpenAI co-founder (2015); ex-Tesla AI director (Autopilot/FSD, ~2017–2022); designed Stanford CS231n; PhD under Fei-Fei Li. Coined "vibe coding" (Feb 2025) and, a year later, its grown-up successor "agentic engineering" (Sequoia AI Ascent, 30 Apr 2026).

Role check run 2026-10-03: no public role change since 19 May 2026. No new public repo, blog post or long-form interview found between 24 Aug and 3 Oct 2026 — the first quiet stretch of this length in two years, consistent with a frontier-lab employee under a publication filter. Treat silence as the observable, not as absence of news.

One-line position: the best way to understand AI is to build it from scratch — and the thing most worth building right now is the loop where AI does the building: agents that run the experiments that make the next model better.

Discipline & technical bet

Fully technical — the rare educator who ships CUDA. Current bet: AI-accelerated pre-training research, i.e. Claude helping improve how Claude is trained. This review found the missing middle of that story: he had already prototyped the bet in public before Anthropic hired him for it. autoresearch (7 Mar 2026) is a ~630-line single-GPU training script plus a program.md brief; a coding agent edits train.py, trains for five minutes, keeps the change if validation loss (a score for how well the model predicts text) improves, and loops — about 100 experiments while he sleeps. First public run: 700 experiments in two days, 20 keepers, an 11% speed-up when ported to a bigger model (Fortune, 17 Mar 2026). His framing: "All LLM frontier labs will do this. It's the final boss battle." The Anthropic job is that repo with a frontier budget. It is also 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, and now microgpt (12 Feb 2026): a complete GPT in ~200 lines of dependency-free Python, which he called the end of "a decade-long obsession to simplify LLMs to their bare essentials."
• "Decade of agents," not year of agents (Dwarkesh, 17 Oct 2025): AGI ~a decade out; agents lack continual learning (the ability to keep learning on the job) and reliability. See Delta log — his own 2026 behavior has re-scoped this claim.
• Verifiable vs unverifiable (No Priors, 21 Mar 2026): "All unverifiable domains still belong to humans; all verifiable domains either already belong to machines or will soon belong to them." Code, math and anything with a test suite goes first. Sequoia version: "Traditional software automates what you can specify… LLMs automate what you can verify."
• Vibe coding raises the floor; agentic engineering raises the ceiling (Sequoia, 30 Apr 2026): the professional job is now specs, eval loops (automatic checks that catch regressions), diff review and permission management for fallible agents — "You can outsource your thinking, but you can't outsource your understanding."
• Teacher + AI symbiosis (Eureka Labs thesis): experts design courses, AI scales instruction — paused as a company, and by March 2026 his own practice had flipped: "I am no longer explaining things to people; I am explaining things to agents."
• The stack matters: Python → C → CUDA, not just the API layer.
• The frontier years are formative (19 May 2026): where the next era gets decided is inside pre-training, now. Unfalsifiable as written — rewritten below as a stay/leave check.

Notable predictions — with falsifiable checks

• (Oct 2025) Agents need a decade to do reliable, economically meaningful multi-day work. Check: METR's time-horizon series (how long a task a model can finish half the time). Reading: public frontier models at ~12 human-hours at 50% success, ~1–4 hours at 80%, Feb–Mar 2026; doubling every ~105 days on the post-2024 fit; the test suite tops out near 16 hours (METR Frontier Risk Report, 19 May 2026). Interim read 2027; if agents reliably hold week-long jobs by 2028, graded Wrong. Watch the 80% line, not the 50% — reliability is his actual claim.
• (Oct 2025) Continual learning is the missing piece and is ~a decade out. Check: a frontier model that measurably improves on a user's tasks across weeks without retraining, by 2028. 2026 benchmarks (Continual Learning Bench, AhaBench, Sept 2026) exist precisely because nobody passes them yet — Held so far.
• (Mar 2026) "All LLM frontier labs will do this" — agent swarms autoresearching training recipes. Check: a lab other than Anthropic publicly credits agent-run experiment loops in a model card or technical report by end-2027.
• (2024) AI-assisted education scales world-class teaching to everyone. Check: LLM101n ships as a complete AI-taught course with measured outcomes by end-2027. Signal weakening — no LLM101n commits of note in 2026; founder is at Anthropic.
• (Ongoing) Capabilities keep improving with scale + data + methods. Check: frontier-model deltas per year. METR's 105-day doubling says Held; his job move IS the bet, doubled.
• (May 2026, rewritten) The frontier years are formative → Check: is he still inside a frontier lab's training org on 19 May 2028? Leaving early for education or a startup grades this Drifted.

Revealed behavior (Does)

• Joined Anthropic pre-training, 19 May 2026 — left founder autonomy for frontier R&D. Announcement post: 13.6M views. Actions say: the decisive game is training-run research, not applications, not (yet) education.
• Stopped writing code, Dec 2025. On 26 Dec 2025 (14M views): "I've never felt this much behind as a programmer. The profession is being dramatically refactored." By March: "I don't think I've typed like a line of code probably since December" — the split went from 80% him / 20% agent in November to ~0 / 100 (No Priors, 21 Mar 2026). He describes the state as "psychosis… trying to figure out what's possible." Our Aug-2026 file still carried the Oct-2025 snapshot — "largely hand-written; coding agents fell short" — as current. Wrong by ten months.
• Shipping cadence before the hire (all missed by the Aug review): nanochat miniseries v1 (7 Jan 2026); microgpt (12 Feb); autoresearch (7 Mar; ~97k stars by Oct); nanochat GPT-2 speedrun leaderboard with "autoresearch round 1/2" entries (9 and 14 Mar). 6 Mar post on an unattended run: "ah yes, this is what post-agi feels like :) i didn't touch anything. brb sauna."
• Shipping cadence after the hire: nothing public found 19 May → 3 Oct 2026. nanochat's last leaderboard entry is 14 Mar 2026. The educator identity is now dormant in behavior, not just in company status.
• Side-build: home-automation agents ("claw psychosis," Jan 2026, OpenClaw) — he uses the tools he theorizes about, on his own house.

Feels

Wants to be where history is made and fears watching from the sidelines — "formative years" is longing, not analysis. New since March 2026: a builder's vertigo. "Never felt this much behind," "psychosis," and he admits he is "not on the forefront of the field" and nervous about it — unusual candor from someone with three frontier-lab stints. 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) and, by his own account, the coding agents themselves — his input diet is increasingly what Claude and Codex produce overnight. 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. He is also the one person who ran the agents-improve-training loop in public (autoresearch) and then went inside to run it with a frontier budget — when he says what that loop does and does not find, it is testimony, not opinion.

Incentive map

Anthropic salary/equity — expect no public criticism of Anthropic strategy; discount his frontier-lab comparisons accordingly, and expect the quiet period to continue (the filter is structural, not personal). Educational brand monetizes clarity and candor — still a strong incentive to stay honest about technical limits, now in tension with the lab filter. Audience-capture risk on X (millions of followers; 13–14M-view posts) remains — the "post-agi feels like" register plays to the crowd. What he can't say: anything about Claude training runs; anything comparing Anthropic to OpenAI, where he has friends and history.

Theories aligned with

• Education-through-construction pedagogy
• AI-accelerates-AI research loop (now his day job; prototyped as autoresearch)
• Verifiable-domains-first: automation lands wherever there is a reward signal (tests, proofs, metrics) — a clean, checkable theory of which jobs move first
• Adjacent to techno-optimism in a builder register; a measured skeptic on general agent timelines — both at once.

What he's reacting against

• Black-box consumption of AI; API-layer-only understanding.
• "Year of agents" hype without continual learning — AND, since Dec 2025, programmers who dismiss agents because the Oct-2025 version was "slop."
• Both "AI is magic" and "AI is just statistics."
• Apps as the unit of software — "MenuGen should stop existing as an app" once a model can do the transformation directly (Sequoia, Apr 2026).

Where he overlaps / splits (with Rich)

• Overlaps: comprehension through building (the harness ethos); the agentic-engineering job description — specs, eval loops, permissions — is Rich's harness work in other words; measured timelines on general agents; 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. His "apps should stop existing" line is a direct challenge to the presentation-layer end of the barbell; worth a counter-argument rather than a nod.
• New split: he has stopped writing code; Rich still reads and edits what the agents produce. Karpathy's line "you can't outsource your understanding" is the test of whether his own practice holds.
• 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 / microgpt — the most-cited technical AI education corpus (~120k combined stars on the nano repos by Jul 2026, plus ~97k on autoresearch); verifiable, you can run the code.
• Coinages that stuck: "Software 2.0" (2017), "vibe coding" (Feb 2025), "agentic engineering" (Apr 2026). Three for three on naming the moment.
• Eureka Labs — thesis unproven; paused. "Decade of agents" (Oct 2025) — partly overtaken by his own December, see Delta log.

Empirical vs normative

Empirical: code walkthroughs, benchmark-grounded claims, autoresearch numbers — unusually checkable. Normative: understanding should be accessible; AI should augment teachers; the frontier should be built by people who comprehend it; humans must keep "directional judgment" even when they stop typing.

Weak spots / open questions

• The October/December whiplash: "decade of agents" (17 Oct 2025) to "never felt this much behind" (26 Dec 2025) in ten weeks, with no explicit retraction. Is this one coherent view (general agents far, coding agents here) or audience-driven mood? He has not reconciled them in public.
• Eureka Labs limbo: "resume in time" — the LLM101n repo shows no 2026 progress. Polite word for shelved until proven otherwise.
• Employment filter: frontier commentary now passes through Anthropic — treat lab-comparison claims as partisan, and expect long silences.
• Autoresearch's ceiling: HN critics called it "inefficient hyperparameter optimization." Does the loop find genuinely new methods, or only tune knobs? The Anthropic team is the experiment; results will not be public.
• 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)

2026-10-03 — validator's note, not Rich's voice: two prompts if useful — (1) is "agentic engineering" just the harness with a better name, and does Rich want to adopt or resist the term? (2) his "apps should stop existing" claim vs the presentation end of the Barbell.

Delta log

Sources

• Anthropic hire: TechCrunch, 19 May 2026 · CNBC, 19 May 2026 · Fortune, 19 May 2026 · HN thread — "I've joined Anthropic" (autoresearch skepticism)
• Autoresearch: karpathy/autoresearch (7 Mar 2026) · Fortune — "The Karpathy Loop," 17 Mar 2026 (700 experiments, 11%, "final boss battle") · The New Stack — 50 experiments overnight · X, 6 Mar 2026 — "this is what post-agi feels like"
• Coding flip: X thread, 26 Dec 2025 — "never felt this much behind" (Futurum, 30 Dec 2025) · Fortune, 21 Mar 2026 — No Priors quotes · No Priors (Sarah Guo) transcript summary, 21 Mar 2026
• Agentic engineering: karpathy.bearblog.dev — Sequoia Ascent 2026 summary, 30 Apr 2026 · Sequoia — "From Vibe Coding to Agentic Engineering" w/ Stephanie Zhan
• Repos: microgpt, 12 Feb 2026 · karpathy/nanochat (leaderboard to 14 Mar 2026) · miniseries v1, 7 Jan 2026 · karpathy/LLM101n
• Decade of agents: Dwarkesh Podcast, 17 Oct 2025 · Zvi review, 21 Oct 2025 · 80,000 Hours — "What happened with AGI timelines in 2026?" (recorded 3 Jul 2026)
• Agent reliability checks: METR Frontier Risk Report, Feb–Mar 2026 (19 May 2026) · Continual Learning Bench (Jun 2026) · AhaBench (Sep 2026)
• Eureka Labs: announcement, 16 Jul 2024 · pause confirmed in the 19 May 2026 announcement
• Prior versions: v2 (24 Aug 2026) — archived _archives/andrej-karpathy.html.bak-20261003 · v1 (16 Jun 2026) — _archives/andrej-karpathy.html.bak-20260824

migrated v2→v2.5 2026-10-03 · backup: _archives/andrej-karpathy.html.bak-20261003 · converts-from: andrej-karpathy.html (v2, 2026-08-24) · Singularity 🧠 / People