Back

People

Andrej Karpathy

Former Tesla AI director, OpenAI founding member, and deep learning educator (cs231n, "Neural Networks: Zero to Hero"). Recent focus has shifted from model internals to knowledge management and agent workflows — advocating for LLM-compiled personal wikis, file-over-app data ownership, agent proficiency as core skills, and exploring the creative frontier where LLMs build things no human would bother to.

Created Apr 5, 2026·Updated Oct 12, 2026

Recent Updates

Recent Thinking (2026)

Karpathy's focus has visibly shifted from model architecture to how people use AI agents effectively. Key themes:

LLM Knowledge Bases

Popularized the pattern of using LLMs to compile and maintain personal wikis from raw data sources. His own research wiki reached ~100 articles and ~400K words. Uses Obsidian as a viewer, with the LLM handling all writing and maintenance. See LLM Knowledge Bases.

File Over App

Strong advocate for storing personal data in universal file formats (markdown, images) rather than proprietary app databases. The argument: files are interoperable, agents can operate on them with Unix tools, and you maintain full ownership. Any AI can plug in (BYOAI).

Agent Proficiency

Called this a "CORE SKILL of the 21st century." His framing: agents speak English and do the computer stuff for you — the skill is knowing what to point them at and how to structure the inputs. See Agent Proficiency.

Idea Files Over Code

Proposed sharing ideas as text files rather than code/apps, since in the agent era, the recipient's agent can customize and build it for their specific needs. Published his LLM wiki concept as a GitHub gist "idea file."

Autoresearch

New project (March 2026): a framework for AI agents that autonomously experiment with training code overnight on a single GPU. The agent modifies train.py, runs 5-minute training experiments, checks if loss improved, and repeats — ~100 experiments per 8-hour sleep cycle. The human edits program.md (agent instructions), not the code directly.

Key design choices:

  • Fixed 5-minute time budget per experiment — makes results directly comparable regardless of what the agent changes (architecture, batch size, etc.)
  • Single file for the agent to modify — keeps scope manageable
  • Metric: val_bpb (validation bits per byte), vocab-size-independent for fair architectural comparisons
  • "You're not touching any Python files like you normally would as a researcher. Instead, you are programming the Markdown files."

Released on GitHub as karpathy/autoresearch. Framed aspirationally in the README intro as the origin story of fully autonomous AI research. See also Agentic Engineering for related self-improving agent work.

The Perception Gap Funnel

In April 2026, Karpathy identified a growing gap between two groups of AI users — casual ChatGPT users who concluded AI is overhyped, and professional frontier model users who describe productivity gains as "nothing short of staggering." The key mechanism: domains with verifiable rewards (coding, math) have seen the largest gains due to reinforcement learning requiring clear success signals. This maps directly to the jagged capabilities picture in AGI Definitions.

By October 2026, Karpathy revisited this observation and found the gap widening — no longer "two groups speaking past each other" but a sharp funnel with four tiers and rough population estimates:

  1. ~6B people (75%) have barely come in contact with LLMs at all.
  2. ~1–2B (20%) are casual, infrequent users of free-tier products. They see "derpy chatbots" and treat them as a better Google search or writing aid. Even many professionals outside math and code are here — execs spend their time talking to people, so AI capability remains "second-hand and a bit abstract."
  3. ~20M (0.2%) use frontier-grade LLMs professionally in math and code. This group sees first-hand that "large, complex projects that used to take them weeks/months can now be completed by agents with a prompt." Karpathy notes this shift happened in under a year — he was still writing code manually by hand, "typing memorized computer code commands into a code editor character by character."
  4. ~5,000 (0.00006%) have access to frontier-grade systems internally. What the external world has seen is "the preview." Internally, it looks like "swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics." Human review and comprehension are starting to fall behind.

Three factors drive the funnel shape: (1) impact scales with ambition and problem size — a paragraph answer barely stresses the system; (2) capability peaks in domains that are digital, verifiable, and economically valuable, because LLM capability emerges from RL on verifiable rewards on a curated environment mixture driven by revenue potential (the jaggedness dimension); (3) access tiers — free, paid, internal.

"The Decade of Agents"

In a Dwarkesh Patel interview (Apr 2026), Karpathy pushed back on "year of agents" framing: "There's some over-prediction going on in the industry. In my mind, this is more accurately described as the decade of agents." Early agents like Claude Code and Codex are "extremely impressive" but still cognitively lacking — no continual learning, insufficient multimodality, unreliable computer use. "It will take about a decade to work through all of those issues."

"We're Not Building Animals"

Central metaphor from the same interview: LLMs are "ghosts" or "spirits," not animals. "We're not doing training by evolution. We're doing training by imitation of humans and the data that they've put on the Internet. You end up with these ethereal spirit entities because they're fully digital and they're mimicking humans." This is a fundamentally different kind of intelligence — starting from a different point in the space of possible intelligences. But it's "possible to make them a bit more animal-like over time."

On memorization vs. abstraction: LLMs can "regurgitate word-for-word what is the next thing in a Wikipedia page" but their ability to learn abstract concepts quickly (the way a child can) is limited. Humans are "not actually that good at memorization, which is actually a feature" — it forces finding patterns in a more general sense.

On the coding stack: autocomplete is high-bandwidth ("you point to the code where you want it, you type out the first few pieces, and the model will complete it") while typing instructions in English is "too much typing." Different modes work for different parts of the coding workflow. See Agentic Engineering.

Frontier Creative Capability & Verification Gaps

In August 2026, Karpathy tested Opus 5 by giving it the opening paragraph of The Lord of the Rings, a 1M token budget (~$10), and asking for a three.js render. The model spent ~2 hours writing 5,500 lines of procedural code that animated the scene — placing and orchestrating polygon assets in 3D coordinates. The result was "kind of janky but fun," yet the fact that it produced anything coherent at all was "mindboggling."

The broader point: LLMs unlock a category of creation where the bottleneck was never technical skill but human patience. "No one in their right mind would ever spend the time to write something this custom," but LLMs have "all the stamina and patience in the world." This shifts the calculus from "impossible" to "~free" — and Karpathy sees it pointing toward hyper-custom, ephemeral game worlds generated on demand ("an ephemeral GTA of X").

Karpathy later articulated this as a general principle with a concrete modality hierarchy for understanding LLM outputs: text (specifically ASD-STE100, a controlled-language spec from aerospace maintenance — he suggests "80% of the way to ASD-STE100" to soften the stringency) → diagrams → interactive HTML web pages → fully custom explainer videos ("Create a 3b1b style video explainer on X"). Each step up makes the output easier to parse and verify. The broader framing: as LLMs take on more autonomous legwork, human work "rises up the abstractions into oversight and understanding" — and LLMs can help with that shift too, because abundant intelligence means you can request "large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before." This is the same abundance logic as the Opus 5 experiment, generalized into a workflow recommendation. See Loop Engineering for the verification side of this oversight shift.

But this domain also exposes a key LLM weakness: verification in dynamic environments. The model couldn't efficiently audit its own 3D output because it lacks native video perception and gameplay ability. It had to take screenshots slowly and painstakingly, introducing visual jank it couldn't catch. This echoes the verification theme from Loop Engineering and AI Evals — raw generation capability outpacing the ability to close the feedback loop, especially in domains without simple text-based success signals.

Sources

  • "LLM Knowledge Bases" — Andrej Karpathy (tweet, Apr 2, 2026) (link)
  • "Farzapedia, personal wikipedia of Farza..." — Andrej Karpathy (tweet thread, Apr 4, 2026) (link)
  • "llm-wiki" — Andrej Karpathy (GitHub gist, Apr 4, 2026; 1,066 stars) (link)
  • "The guy who literally wrote the most popular deep learning..." — Aakash Gupta (commentary, Apr 2026) (link)
  • "karpathy is showing one of the simplest AI architectures..." — JUMPERZ (tweet, Apr 2026) (link)
  • "Andrej Karpathy Shares LLM-Powered Obsidian Knowledge Workflow" — x.com summary (link)
  • "Stanford CS230 | Autumn 2025 | Lecture 9: Career Advice in AI" — Stanford Online (link)
  • "We're summoning ghosts, not building animals" — Andrej Karpathy / Dwarkesh Patel (video, Apr 2026) (link)
  • "The gap between casual ChatGPT users and professional agentic users..." — Andrej Karpathy (tweet, Apr 2026)
  • "Andrej Karpathy on the shift to agentic engineering" — Ben Lang (tweet, May 2026) — pointer tweet on Karpathy's agentic engineering perspective
  • "We're starting to leave the territory where you'd test an LLM by..." — Andrej Karpathy (tweet, Aug 2026) — Opus 5 Lord of the Rings three.js experiment; frontier creative capability and verification gaps
  • "We'll be spending a lot more time trying to understand the outputs of language models..." — Andrej Karpathy (tweet, Oct 2026) — output modality hierarchy (text→diagrams→HTML→videos); oversight shift framing
  • "Dusting off this tweet from April because this gap in shared understanding of LLM capability is widening..." — Andrej Karpathy (tweet, Oct 2026) — four-tier perception funnel with population estimates; internal-access tier; three driving factors