AI-Wiki

Andrej Karpathy

7 min read

Andrej Karpathy

Confidence 0.92 · 7 sources · last confirmed 2026-05-28

AI researcher and educator; co-founder of OpenAI (2015–2017); led Tesla Autopilot computer-vision (2017–2022); founder of Eureka Labs (2024–) for AI-native education. The wiki’s most cross-cutting individual contributor: he is both the upstream-spec author for this entire repo (llm-wiki.md, llm-wiki-v2.md, llm-wiki-v2-plan.md) and a substantive source on agentic-engineering paradigm vocabulary as of 2026 (Sequoia AI Ascent interview).

Why this page exists (cross-page-presence promotion)

Per CLAUDE.md §Lifecycle “Author-entity promotion”, first-source mentions are usually deferred to the dangling list. Karpathy is promoted on first source-page mention via the same cross-page-presence judgment that produced Jack Clark: he is named explicitly in repo-root specs (llm-wiki.md, llm-wiki-v2.md, llm-wiki-v2-plan.md) and in CLAUDE.md as the originator of the LLM-wiki idea this repo implements. That means the wiki has been implicitly citing Karpathy as the upstream-spec author since instantiation — promoting him formalises a citation chain that already exists.

The promotion is a one-off judgment call (the same caveat noted in the Jack Clark log entry) — the audit script (scripts/lint-dangling-authors.mjs) tracks only the strict ≥2-source-frontmatter rule and was not modified.

Career snapshot

  • PhD, Stanford (2011–2016, advisor: Fei-Fei Li). Convolutional neural networks for visual recognition; co-instructor of the canonical Stanford CS231n Convolutional Neural Networks for Visual Recognition course. Authored the influential Hacker’s guide to neural networks and the long-running karpathy.ai writing.
  • OpenAI co-founder and research scientist (Dec 2015 – Jun 2017). One of the original group named in the founding announcement.
  • Tesla, Director of AI / Senior Director of AI (Jun 2017 – Jul 2022). Led the Autopilot computer-vision team; widely credited with the “vision-only / no-LiDAR” path Tesla pursued. Coined and popularised Software 2.0 during this period — the framing that programming neural networks (datasets + architectures + training) is itself a form of programming.
  • OpenAI, returning (Feb 2023 – Feb 2024). Worked on improving GPT-class assistants.
  • Eureka Labs (founded Jul 2024 – present). AI-native education company; the LLM 101n undergraduate-level course is the public-facing artifact.

What he is the wiki’s source for (vocabulary and conceptual artifacts)

Karpathy is unusually prolific at naming things that the field then adopts. The wiki is using or planning to use several of his coinages directly:

  • Software 3.0 — the 1.0 / 2.0 / 3.0 trilogy framing (rules → weights → prompts). 1.0 was the standard reference; 2.0 was Karpathy’s coinage in 2017 (the famous “Software 2.0” Medium post); 3.0 is his 2025–26 extension as foundation-model prompting consolidates as a programming paradigm.
  • Vibe coding — coined by Karpathy in 2024 for intuitive, exploratory code-with-LLM workflows. Now in widespread practitioner use.
  • The wiki’s jagged-frontier explanatory layer — Karpathy contributes the cause (verifiability + labs care + RL training mechanism) for the task-level observation described by Dell’Acqua et al. The “are LLMs jagged?” framing in the wiki sits across both pages.
  • “Animals vs ghosts” — Karpathy’s mindset framing for what kind of intelligence LLMs are: “We’re not building animals; we are summoning ghosts.” He himself flags this as “a little bit of philosophising”; the wiki captures it as a sub-section inside jagged-frontier.
  • Agentic engineering as a discipline — Karpathy positions this as the “preserve quality bar at agent speed” complement to vibe coding’s “raise the floor”. Names it explicitly as an engineering discipline.
  • The LLM-wiki / LLM-knowledge-base pattern itself — the llm-wiki.md conceptual spec at the root of this repo is Karpathy’s articulation of the workflow. As of 29 April 2026 he confirms continued use of the pattern: “I really enjoy whenever I read an article I have my wiki that’s being built up from these articles… anytime I see a different projection onto information, I always feel like I gain insight.” On 4 April 2026 Karpathy posted a GitHub Gist that crystallised the pattern as an “idea file” — a markdown document designed to be copy-pasted into an LLM agent. The gist hit 17 million views, 5,000 stars, and 4,282 forks within days (per Raju 2026). Three explainer articles ingested in this batch ( ex-brain, Raju, Liu) collectively describe the pattern, build a reference implementation, and place it in the RAG / Wiki / Skills hybrid landscape. The wiki now anchors the construct as a dedicated concept page: llm-wiki.

Key claims attributable to Karpathy (in this wiki)

From the Sequoia AI Ascent interview:

  • The December 2025 phase change: agentic coding tipped from “useful but needs corrections” → “I just trust it now” in a single month. “I’ve never felt more behind as a programmer.”
  • Software 3.0 is a new computing paradigm, not a faster way to do existing programming. “It’s not just about programming becoming faster — there are new things available now.”
  • LLMs automate what you can verify, just as classical computers automate what you can specify in code. The mechanism is RL training in verification-rewarded environments.
  • Two factors of jaggedness: verifiable + labs care. Either alone is insufficient; both together explain why models fly in some circuits and struggle in others. “You’re slightly at the mercy of whatever the labs are doing.”
  • Agentic engineering pushes far past 10×. The 10× engineer used to be the upper bound; Karpathy’s claim is that agentic engineering enlarges that bound substantially.
  • Agents are intern entities — remarkable but error-prone in surprising ways. Humans must own aesthetics, judgment, taste, oversight, the spec/plan.
  • You can outsource your thinking, but you can’t outsource your understanding. Humans become the bottleneck for what’s worth building, why, and how to direct agents.

Affiliations

  • Founder, Eureka Labs (2024 – present, dangling — not yet promoted to entity page).
  • Co-founder, OpenAI (2015 – 2017 + 2023 – 2024).
  • Former Director / Senior Director of AI, Tesla (2017 – 2022, dangling — not yet promoted).
  • Stanford PhD (2011 – 2016).

Mentioned in

Open questions

  • Eureka Labs as an entity page — currently a dangling reference. If a second source surfaces (interview, blog post, course materials), promote.
  • Tesla as an entity page — currently dangling but mentioned in the AI Index reports too; could be promoted on cross-page-presence grounds in a future pass.
  • The original “outsource thinking ≠ outsource understanding” tweet — Karpathy quotes it but doesn’t name the author. Worth a citation upgrade if the original surfaces.
  • Karpathy’s hinted “valuable RL environments not in the labs’ mix” — declined on stage. If a follow-up writeup names the domains, that’s a high-value second source.

Relationships

Authored by

Graph View

Backlinks