In this episode of the No Priors podcast, Andrej Karpathy sits down for a wide-ranging conversation about how AI agents have fundamentally changed his daily workflow, his AutoResearch project that autonomously improves machine learning models, the state of open source AI, the future of robotics, the impact of AI on the job market, and how education must adapt when agents become the primary interface for learning.
Karpathy describes a dramatic shift that occurred around December 2025, where he went from writing 80% of his code by hand to writing essentially none. He now delegates entirely to AI agents (Claude Code, Codex, and similar tools), describing a state of "AI psychosis" driven by the feeling that his capacity is now limited only by his ability to direct agents, not by his typing speed. The bottleneck has moved from compute to the human operator's skill at orchestrating agents.
Karpathy compares the current moment to being a PhD student with idle GPUs: if your agents are not running, you are wasting capacity. He describes the "Peter Steinberg" approach of running multiple Codex sessions simultaneously across different repos, treating software development as a set of macro actions (entire features, research tasks, implementation plans) rather than individual lines of code. The question is no longer "what code do I write?" but "what macro actions can I delegate in parallel?"
Karpathy praises OpenClaw (built by Peter) for innovating in five areas simultaneously: personality (the "soul" document), memory systems, persistence (looping without human involvement), the WhatsApp portal interface, and sophisticated automation. He notes that Claude has a well-calibrated personality that feels like a teammate, while Codex feels "dry" by comparison. He argues that agent personality matters more than most tool builders appreciate.
In January, Karpathy built a "claw" (persistent agent) called Dobby that controls his entire home: Sonos speakers, lights, HVAC, shades, pool/spa, and security cameras. The agent discovered smart home devices on his local network, reverse-engineered their APIs, built a dashboard, and now communicates via WhatsApp. This replaced six separate apps with natural language control, illustrating how agents can unify fragmented software experiences.
Karpathy argues that many apps "shouldn't even exist" because the real value is in APIs that agents can call directly. The customer of software is shifting from humans to agents acting on behalf of humans. He predicts that what requires vibe coding today will be trivial and free within a year or two, as the barrier for non-technical users continues to drop.
AutoResearch is Karpathy's project to create fully autonomous research loops. Using his GPT-2 training codebase (DataChat) as a testbed, he set up an agent with an objective metric (validation loss), boundaries on what it could modify, and let it run overnight. It discovered hyperparameter improvements he had missed after two decades of manual tuning, including weight decay on value embeddings and better Adam beta values. The key insight: if you have an objective metric, you should not be in the loop.
Karpathy describes how a "program.md" file defines the instructions for an auto-researcher, essentially describing a research organization in markdown. He envisions meta-optimization: running contests where different program.md files compete, then feeding the results back to generate better program.md files. This creates layers of abstraction: LLMs are taken for granted, agents are taken for granted, claws are taken for granted, and now you optimize the instructions to the claws.
Karpathy envisions a "SETI@home for AI research" where untrusted workers on the internet contribute compute to improve models. The system resembles a blockchain: commits (code changes) replace blocks, proof of work is the experimentation needed to find improvements, and verification is cheap (just run the training). This could allow a swarm of agents to potentially rival frontier labs by pooling the Earth's distributed compute resources.
Karpathy describes current AI models as simultaneously feeling like "an extremely brilliant PhD student who's been a systems programmer their entire life and a 10-year-old." Models excel in verifiable domains (code, math) because those are optimized via reinforcement learning, but struggle in softer areas (humor, nuance, knowing when to ask clarifying questions). He uses the example that state-of-the-art models still tell the same atom joke from years ago because joke quality is not being optimized.
Karpathy draws parallels between open source AI and Linux: the industry needs a common open platform. Currently, open source models trail closed models by roughly 6-8 months. He views this dynamic as healthy: frontier labs push capability forward (expensive, closed), while open source follows behind and handles the vast majority of consumer use cases. He warns against excessive centralization, arguing that "ensembles of people" making decisions is always better than a few behind closed doors.
Drawing on his self-driving experience at Tesla, Karpathy predicts that digital transformation will vastly outpace physical robotics. Bits are a million times easier to manipulate than atoms. The sequence he envisions: first, massive digital "unhobbling" (making existing digital processes far more efficient); then, building interfaces between digital intelligence and the physical world (sensors, actuators, data markets); and finally, full physical robotics, which will be a larger market but will lag significantly.
Karpathy describes a fundamental shift in education. With his MicroGPT project (LLM training in 200 lines of Python), he realized that rather than explaining code to humans, he should explain it to agents, who then personalize the explanation for each learner. He envisions "skills" (curriculum scripts) that guide agents through teaching progressions. The educator's new job is to contribute the "few bits" that agents cannot generate themselves: the original insight, the distilled simplification, the judgment about what matters.