AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Andrej Karpathy — AGI is still a decade away

· Dwarkesh Podcast

Andrej Karpathy reframes the industry's 'year of agents' narrative as more accurately 'the decade of agents,' arguing current systems like Claude and Codex lack the intelligence, multimodality, and continual learning needed to function as autonomous 'employees.' He traces the field's path through AlexNet, the Atari/Universe reinforcement learning detour he now considers a misstep, and finally LLMs, which delivered the representations needed before agents could become viable. He contrasts the current paradigm—building 'ghosts' by imitating internet text—with the evolution-based animal intelligence vision championed by Richard Sutton, and argues humans use far less RL for cognition than commonly assumed. A major thread is the inadequacy of current RL: outcome-based supervision 'sucks supervision through a straw' by treating every token in a successful trajectory as equally good, while process-based supervision breaks down because LLM judges are trivially gameable with adversarial examples (he recounts one where the model produced nonsensical 'da da da' continuations that fooled the judge). Karpathy sees this as a research-level problem requiring 3–5 more major algorithmic ideas. He discusses his new 'nanochat' repository and explains why coding models failed to help him build it: novel, intellectually intense code is exactly what models struggle with, and autocomplete remains his preferred mode. He draws an extended analogy to Tesla's self-driving work, where 'march of nines' means each reliability level takes constant effort and the demo-to-product gap is enormous. Karpathy is skeptical of imminent takeoff scenarios, framing AI as a continuous extension of computing rather than a discrete singularity, and predicting the same gradual 2% GDP-diffusion pattern. He flags coding as 'the perfect first task' for LLMs (text-native, rich tooling infrastructure like diffs/IDEs) but finds it surprising that other pure language-in/language-out knowledge work isn't yielding similar gains. On architecture, he expects transformers and gradient descent to persist but with sparse attention, better data, and a much smaller 'cognitive core'—possibly sub-billion parameters—once memorization is stripped away. Missing pieces include continual learning (no analog of sleep-time distillation into weights), culture (a shared scratchpad across LLMs), and self-play.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.