Karpathy: No AGI Until 2035, Expect Augmentation

20 Oct 2025

Andrej Karpathy, speaking on the Dwarkesh Patel podcast and summarized by The Neuron (Oct 19, 2025), pushed back on short‑term AGI hype and estimated truly general AI is roughly a decade away — “no AGI until 2035.” Karpathy, who led AI at Tesla and was an early OpenAI member, argued today’s large language models suffer fundamental cognitive limitations: reinforcement learning can reward lucky guesses rather than genuine reasoning (he called it “terrible” and “stupid”); coding agents get confused by custom code and add defensive bloat; and near‑perfect memorization prevents the kind of forgetting that helps humans generalize.

He emphasized the long, hard path from useful but brittle systems to the highly reliable systems people assume (getting from 99% to 99.9% reliability is extremely costly), and warned the market’s expectations — roughly $14 trillion added since ChatGPT — may be overambitious. The practical takeaway: expect a decade of “augmented intelligence” improvements and better scaffolding, not a sudden handoff to autonomous AGI. The newsletter links to the full 2.5‑hour interview and deeper analysis for readers who want the full context.

Source