Bob McGrew: AI Agents And The Path To AGI

McGrew joined OpenAI after exploring robotics and deep learning at a time when he wanted to start a company. Early projects included training a robot hand to solve a Rubik’s Cube and building an AI for Dota 2; both reinforced the value of scale and generalization. He connects those ideas to GPT, from Alec Radford’s next-token prediction work through GPT-2, GPT-3, and GPT-4.

McGrew describes OpenAI’s early culture as a middle ground between centralized direction and open-ended academic research. He says the lab encouraged collaboration by being inclusive about paper authorship and valuing researchers’ reputations inside the organization. Scaling laws require first getting a system to work, then solving systems, data, and algorithm problems to scale it and improve its performance.

McGrew sees a data bottleneck for pretraining and points to reasoning and test-time compute as another way to improve models. Giving a model more time to think can support longer tasks and raise the reliability needed for agents to take actions. He describes reliability gains such as moving from 90% to 99% as potentially requiring an order of magnitude more compute, now possible through longer reasoning as well as larger training runs.

McGrew advises founders to start with the best available model and get the product working before optimizing costs. After finding value through user feedback, teams can use prompting, distillation, or smaller models to reduce cost. He argues that AI adoption also depends on software that fits people’s actual workflows, informed by close work with customers.

McGrew uses AI-generated coding lessons for his son to explain why learning still builds critical thinking and intuition. He imagines future work involving both individuals amplified by AI and people managing AI-heavy organizations, while remaining optimistic that human roles will evolve. He expects robotics to reach a ChatGPT-like moment within about five years, and suggests scientific progress may still be limited by the physical work of running experiments.