Andrej Karpathy
AI researcher, educator, and former OpenAI co-founder and Tesla AI director.
Last updated: 2026-04-13
Overview
Andrej Karpathy is one of the most influential AI educators and researchers of his generation. He was a founding member of OpenAI, led Tesla’s Autopilot AI team, and is known for making deep learning accessible through his lectures (Stanford CS231n, Zero to Hero series on YouTube).
He is active on X (@karpathy) and regularly shares ideas about AI, software, and knowledge systems. Since leaving OpenAI, he has focused on independent research — particularly auto research (autonomous agent loops for model improvement) and educational projects like nanoGPT and micro GPT.
Key Points
- Co-founder of OpenAI; former Director of AI at Tesla
- Creator of the “Zero to Hero” deep learning series; author of micro GPT (200-line complete LLM implementation)
- Proposed the llm-wiki-pattern for LLM-maintained personal knowledge bases
- Strong advocate for plain-text, markdown-first tooling
On Coding Agents and the December Inflection
Karpathy marks December 2024 as the moment his workflow flipped from 80/20 (writing code himself) to nearly 0% manual coding. “I haven’t typed a line of code since December.” The new paradigm: macro actions — delegating whole functionalities to an agent rather than writing functions line by line.
Key framing: token throughput is the new GPU utilization. Just as he felt nervous when his GPUs weren’t fully utilized as a PhD student, he now feels nervous when subscription tokens go unused. “You are the bottleneck” — meaning improvement is always available, which is both empowering and addictive.
On Claws and Home Automation
A “claw” (his term, ~= persistent autonomous agent loop) goes beyond interactive sessions: it has its own sandbox, runs on your behalf even when you’re not looking, and has more sophisticated memory. Karpathy built Dobby the elf claw to manage his home: it found his Sonos system by scanning the LAN, reverse-engineered the API, and now controls lights, HVAC, shades, pool/spa, and security cameras through WhatsApp. “I used to use six apps. Now Dobby handles everything in natural language.” Takeaway: apps that wrap hardware shouldn’t exist — just APIs and agents.
On Auto Research
Karpathy’s central research interest is removing humans from the improvement loop. His project auto-research: define objective + metric + boundaries → let an agent loop indefinitely. Demonstrated on nanoGPT: after years of manual tuning, an overnight run found weight decay on value embeddings was missing and atom betas were under-tuned (these interact — fixing one shifts the optimum for the other).
The program.MD is the markdown file describing how auto research should work — what to try, what’s off-limits. “A research organization is a set of markdown files.” You can tune the program.MD itself, creating a meta-layer of optimization.
Distributed extension: untrusted internet contributors can propose commits; verification is cheap (just run the code, check the loss). Structurally similar to SETI@home and Folding@home. “A swarm of agents on the internet could run circles around frontier labs.”
On Model Jaggedness
Models trained via RL improve only in verifiable domains. Outside that loop, they’re stuck — the same “why don’t scientists trust atoms?” joke from 5 years ago persists despite massive capability improvements in code. Karpathy’s vivid description: “I feel like I’m talking to an extremely brilliant PhD student who’s been a systems programmer their whole life — and a 10-year-old. Simultaneously.” The jaggedness is an artifact of what has and hasn’t been optimized, not a fundamental ceiling. But until verifiable coverage expands, the jaggedness will persist.
On Model Speciation
Currently, labs ship monoculture models. Karpathy expects more speciation — smaller models with cognitive cores that specialize for specific niches (like a Lean math prover or CUDA kernel optimizer) — once the science of fine-tuning without capability loss matures. The animal kingdom analogy: diverse brains for diverse niches. Today, weight manipulation is harder than context-window manipulation; that gap will narrow.
On Education Through Agents
“I’m not explaining things to people anymore. I’m explaining things to agents.” Karpathy sees education shifting: instead of writing docs for humans or recording videos, you write skills/curricula for agents — hints about what to teach first, what the progression should be. Agents then explain in the learner’s language with infinite patience. Micro GPT (200 lines of complete LLM training) is his final synthesis: the few bits an agent can’t generate are his contribution; everything else the agent can explain better than he can.
On Being Outside Frontier Labs
Karpathy’s case for independence: inside a frontier lab, you’re not a free agent — financial incentives constrain what you can say, organizational pressure shapes your views, and individual employees rarely determine major decisions. Risk of being outside: judgment drifts because you don’t know what’s coming. His ideal: go back and forth — periods inside for ground truth, periods outside for independent voice.
On Open Source vs. Closed Models
Currently ~6–8 months behind frontier; Karpathy expects this dynamic to persist and considers it healthy. Open source will eat simple use cases; closed labs will stay ahead on “Nobel Prize-level” work. Centralization has a poor track record (“Eastern European precedent”) — he wants more labs in the room, not fewer. “By accident we’re actually in an okay spot.”
Connections
- llm-wiki-pattern — his proposed pattern for personal knowledge management with LLMs
- auto-research — originated; demonstrated on nanoGPT
- coding-agent — macro actions, token throughput as bottleneck, claw framing
- agent-first-software — “apps shouldn’t exist; just APIs and agents” — Dobby as demonstration
Sources
- llm-wiki gist — added 2026-04-12
- Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI — No Priors podcast, added 2026-04-13