Bob McGrew: AI Agents And The Path To AGI

TL;DR

The video discusses OpenAI’s journey from early projects (robotics hand, Dota 2, GPT‑1) to current reasoning models (O1, O3), arguing that pretraining scaling is hitting a data bottleneck but reasoning and test‑time compute unlock a clear path to AGI.

Key insights

• The early OpenAI culture was a startup‑style hybrid: no top‑down plan like DeepMind, not pure bottom‑up like Google Brain, but leadership (Ilya, Dario) set opinions on what mattered (e.g., scale).
• Solving Dota 2 strengthened belief that scale (huge experience fed into a neural network) drives generalization; this idea later applied to the robot hand and GPT‑2/3/4.
• Alec Radford persevered for years on the next‑token prediction objective (GPT‑1) despite nobody thinking it would work.