The Curious Incident of Developing Artificial General Intelligence
summary
Argues that AGI progress mainly comes from engineering and compute, not new theory, and that xAI builds and runs a 200K GPUs cluster with 50× faster chips but faces reliability, cosmic‑ray bit flips and one‑rack‑per‑day burn rates; shows next‑token pretraining scales on compute but yields “alien” cognition (Raven images score ~IQ 40 while parsed JSON scores ~IQ 80); proposes reallocating compute toward sampling/RL and new abstraction layers for huge, heterogeneous workloads.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:41From PhD to xAI co-founder
- 9:47Secrets at xAI
- 22:45Next token prediction
- 30:50Measuring machine intelligence
- 37:00Beyond next token prediction
- 43:00Open questions
related talks
31mMyth & Truth behind Artificial General Intelligence
13mArtificial General Intelligence: Why aren't we there yet?
01h 19mThe Al Revolution: The Road to Superintelligence
09mSmale’s 18th Problem and Intelligence Science 2
41m#define AI Engineer
01h 03mDecoding AI Landscape: Business, Market Trends, and Risks
29mWhen to Gen AI? A strategic framework for real-world AI adoption
48mArtificial Intelligence. Our journey towards singularity