All podcasts / No Priors / Summary

Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

2026-03-20 - 66 min - source - Read full transcript
Sarah Guo (host)Andrej Karpathy

Key insights

Karpathy has shifted from writing code by hand to 'expressing his will' to agents, and now measures his day by token throughput rather than lines of code.
He estimates he hasn't typed a line of code since December, moving from roughly 80% self-written code to well over 80% agent-delegated. He now runs multiple parallel agent sessions and feels anxious when a token subscription goes unused, comparing it to feeling GPU-bound as a PhD student.
agentic-coding-workflows
He frames nearly all remaining friction with agents as a 'skill issue,' not a capability gap.
The raw capability already exists, in his view; users just haven't learned to write good enough agent instructions, memory tools, and multi-agent orchestration to unlock it. He describes this as both stressful, because the possibility space feels infinite, and addictive, because getting better produces constant unlocks.
agentic-coding-workflows
OpenClaw's traction comes from several compounding design choices rather than one breakthrough feature.
Karpathy credits Peter Steinberger with combining a deliberately crafted personality document (a 'soul.md'), praise that feels earned rather than sycophantic, a persistent memory system that goes beyond simple context compaction, and a single WhatsApp interface for control - and argues most other agent tools underinvest in personality and memory.
agentic-coding-workflows
Auto research works by defining an objective, a metric, and boundaries, then fully removing the researcher from the loop.
When Karpathy let his auto-research loop run overnight on a codebase he had hand-tuned for two decades, it found an untuned weight-decay setting on the value embeddings and mis-set Adam betas that he had missed, despite his 'earned confidence' from years of manual hyperparameter tuning.
autonomous-ai-research
A research organization can be described as a 'program.md' - a markdown constitution of roles and process - that is itself an object for meta-optimization.
Karpathy proposes comparing which program.md variants (e.g., fewer standups, more risk-taking) produce the most research progress on identical hardware, then feeding that data back to a model to write a better program.md, layering optimization on top of optimization.
autonomous-ai-research
Scaling auto-research beyond one lab's compute could work via a public swarm of untrusted contributors whose work is cheap to verify.
He compares the design to blockchain proof-of-work and citizen-science projects like SETI@home and Folding@home: any single contributor might submit thousands of bad candidate improvements, but a verified good one can be cheaply confirmed and merged, potentially letting the swarm's aggregate untrusted compute rival frontier labs' trusted compute.
autonomous-ai-research
Auto research and reinforcement learning only improve domains with cheap, objective evaluation - which explains models' uneven capability elsewhere.
If a task can't be verified, it can't be auto-researched or reliably reward-optimized. Karpathy points to ChatGPT still telling the same 'why don't scientists trust atoms' joke it told years ago, even as its coding ability transformed, as evidence that capability gains from RL are not generalizing into unverifiable domains like humor.
model-capability-jaggedness
Karpathy expects models to eventually 'speciate' into specialized variants rather than remain one monolithic oracle model, but says true speciation hasn't happened yet.
He compares this to the animal kingdom's diversity of specialized brains, but notes labs currently serve one general model because they don't know in advance what users will ask, and the field still lacks good primitives for adjusting model weights precisely (versus just the context window) without losing general capability.
model-capability-jaggedness
Open-source models will likely keep trailing frontier closed models by a roughly stable lag, and Karpathy argues that lag is structurally healthy.
The gap has narrowed from nothing comparable to roughly eight months behind, but he expects it to persist rather than close. He frames the persistent lag as a good thing: it prevents the systemic risk of intelligence being centralized in two or three closed labs, while still giving the industry a shared, trusted baseline comparable to Linux.
open-vs-closed-source-ai
Working at a frontier lab creates a real tension between financial alignment and independent judgment that Karpathy has not resolved for himself.
Heavy financial upside tied to a technology he believes will 'dramatically change society' compromises independent judgment and what you can say publicly, he argues, while being fully outside a lab means your judgment drifts from not knowing what's actually being built next. He suggests moving in and out of labs over time as the healthier posture rather than picking one side permanently.
open-vs-closed-source-ai
Karpathy predicts AI will increase, not shrink, demand for software work, citing Bureau of Labor Statistics projections and a Jevons-paradox argument.
He points to how ATMs made bank branches cheap enough to operate that teller employment actually rose, and argues digital-native work will be reshaped far faster than physical-world work because 'atoms are a million times harder' to manipulate than bits.
ai-labor-and-education
Documentation and teaching are shifting from being authored for humans to being authored for agents.
His minimalist microGPT project (roughly 200 lines, the irreducible essence of LLM training) is now explained to users primarily by an AI agent rather than a human-made guide or video. He expects library documentation to move from HTML written for people to markdown written for agents, since agents can re-explain any concept to a human at whatever level of patience is needed.
ai-labor-and-education

Books referenced

Companies

Techniques and frameworks

Summary

Andrej Karpathy joins Sarah Guo to describe what he calls a personal state of "AI psychosis": since roughly December, he estimates he hasn't typed a line of code, having shifted almost entirely to delegating work to coding agents and measuring his day by token throughput across parallel sessions rather than lines written. He frames nearly every remaining limitation as a "skill issue" rather than a capability gap - the tools can already do more than most people have learned to extract from them, whether through better instructions, memory tooling, or running several agents on separate slices of a repository at once. He points to Peter Steinberger's OpenClaw as the current best implementation of a persistent "claw" entity, crediting its traction to a combination of choices - a deliberately written personality document, memory that goes beyond simple context compaction, calibrated (not sycophantic) praise, and a single WhatsApp interface - and describes building his own home-automation agent, "Dobby," which reverse-engineered his Sonos and lighting systems' undocumented local APIs from plain-language requests and now replaces six separate apps.

The conversation's technical core is Karpathy's "auto research" framework: define an objective, a metric, and operating boundaries, then remove the researcher entirely from the loop. Running this overnight on a small-model training codebase he had personally hand-tuned for two decades, the loop found hyperparameter mistakes (an untuned weight-decay setting, mis-set Adam betas) that his own "earned confidence" had missed. He proposes describing an entire research organization as a "program.md" - a markdown constitution of roles and process - which becomes itself an object of meta-optimization: compare which program.md variants produce the most progress on identical hardware, then have a model write a better one. To scale this beyond a single lab's compute, he sketches a system where an untrusted public swarm submits candidate improvements that are expensive to produce but cheap to verify, explicitly comparing the design to blockchain proof-of-work and to citizen-science projects like SETI@home and Folding@home.

A recurring caveat is that this all only works where evaluation is cheap and objective. Karpathy argues this constraint explains models' broader unevenness: reinforcement learning only improves verifiable domains, so a model can move mountains on an agentic coding task and then respond with the exact same stale joke ChatGPT told years ago, because jokes were never inside a reward signal. He expects models to eventually "speciate" into specialized variants, much like the animal kingdom's diversity of brains, rather than remain single monolithic oracles - though he says true speciation hasn't happened yet, both because labs don't know in advance what users will ask and because the field still lacks good methods for adjusting model weights precisely, short of just manipulating the context window.

On the open-versus-closed question, Karpathy predicts the open-source lag behind frontier closed models (currently around eight months, narrowed from an initial unbridgeable gap) will persist rather than close, and argues that persistence is structurally healthy: it avoids concentrating intelligence in two or three closed labs while still giving the industry a shared, trusted baseline comparable to Linux. Pressed on why he isn't doing this work inside a frontier lab with more compute and colleagues, he's candid about an unresolved tension: heavy financial alignment with a technology he believes will radically reshape society compromises independent judgment and what one can say publicly, while being fully outside a lab means judgment drifts from not knowing what's actually being built. He suggests moving in and out of labs over time as the more honest posture.

On the labor-market question, Karpathy pulled apart Bureau of Labor Statistics job-growth projections and argues software demand will rise, not fall, as AI makes it cheaper to produce - a Jevons-paradox effect he illustrates with ATMs, which made bank branches cheap enough to run that teller employment actually increased. He expects digital-native work to be reshaped much faster than physical-world work, since "atoms are a million times harder" to manipulate than bits, with robotics likely to lag the way self-driving did. He closes on a smaller but personally resonant thread: his minimalist "microGPT" project, a roughly 200-line irreducible core of LLM training, is now explained to users primarily by an AI agent rather than a human-authored guide - a preview, he argues, of documentation and teaching broadly shifting from being written for humans to being written for agents.

Notable Quotes

"It's not that the capability is not there. It's that you just haven't found a way to string it together... So it all kind of feels like skill issue when it doesn't work." - Andrej Karpathy

"I don't think I've typed like a line of code probably since December, basically, which is like an extremely large change." - Andrej Karpathy

"You're either on rails and you're part of the superintelligence circuits, or you're not on rails and you're outside of the verifiable domains. And suddenly everything kind of just like meanders." - Andrej Karpathy

"I'm a much little bit hesitant of having... intelligences that are closed and that's it. Centralization has a very poor track record in my view." - Andrej Karpathy

"The things that agents can't do is your job now. Things that agents can do, they can probably do better than you, or like, very soon. And so you should be strategic about what you're actually spending time on." - Andrej Karpathy