In this post
This post summarises Bob McGrew: AI Agents And The Path To AGI, a January 2025 conversation between Garry Tan and Bob McGrew, former Chief Research Officer at OpenAI, published on the Y Combinator channel. They discuss the evolution of AI, the path towards Artificial General Intelligence (AGI), and the future of jobs. McGrew talks about his time at OpenAI, scaling laws, and advice for startups. The points below follow the video; the quotes are his words, not ours.
Key Takeaways #
- Reasoning and test-time compute are key to developing reliable AI agents.
- Scaling laws have been crucial in improving AI capabilities.
- Startups should focus on using the best AI models available to succeed.
- The future of jobs will involve collaboration with AI, not just replacement.
The Early Days of OpenAI #
Bob McGrew’s journey into AI began unexpectedly. He left Palantir intending to start a robotics company, but found himself at OpenAI, a place full of very smart people with big ambitions. Early projects included teaching a robot hand to solve a Rubik’s Cube and building an AI to play Dota 2. In his words, Dota 2 “really strengthened our belief that scale was the path to improving artificial intelligence.”
The Role of Scaling Laws #
Scaling laws have been a game-changer in AI development. McGrew explains that the “secret idea” behind Dota 2 was taking huge amounts of experience and feeding it into a neural network, which then learned and generalized from it. The same approach was later applied to the robot hand. For language, GPT-1 showed that a Transformer trained on the simple objective of guessing the next token could generate coherent text. Bringing in Dota’s and the robot hand’s idea of training at larger scale on diverse data led, in his telling, to GPT-2, GPT-3 and GPT-4.
The Shift Towards Reasoning #
As AI technology advanced, McGrew said the pre-training approach is reaching a bottleneck on data. The new mechanism is reasoning and test-time compute: letting a model think longer to get a better answer. He points to OpenAI’s o1 and o3 models, and to Gemini Flash Thinking. In his view this makes more reliable AI agents possible.
The Future of AI Agents #
McGrew believes reasoning will “unlock the possibility of agents to do actions on your behalf.” That has always been possible, he says, but never good enough, because you need a lot of reliability before you will wait five minutes or five hours for a result. His rule of thumb: adding a nine of reliability (90% to 99%, or 99% to 99.9%) is roughly an order of magnitude more compute, and reasoning lets models spend that compute by thinking longer. He says the problem is not solved, but there is a clear path.
Startups and AI #
For founders building AI startups, McGrew’s advice is to start with the very best model you can, because a startup only succeeds if it exploits something at the frontier of AI. Get it working first, then use distillation to cut cost. Your time matters most, so build the product quickly, work out where the value is by iterating with users, and only then think about cost. He also says AI needs a twist that helps people solve the problem they actually have, not just run their existing workflow with AI.
The Future of Work #
As AI continues to evolve, the nature of work will change. McGrew expects two roles: the “lone genius,” such as OpenAI researcher Alec Radford working alone with a computer that leverages him, and the manager, who will be “the CEO of your own firm” that is mostly AI. He compares it to cameras and artists: artists are still around, and more people may learn to paint. He also notes that in the 1880s most people were farmers and about 3% of Americans are now, and that those farmers could not have imagined their grandchildren’s jobs. He calls himself an optimist about humanity.
Robotics and AGI #
Looking ahead, McGrew says robotics companies are now where LLM companies were five years ago. Within the next five years, he expects “the ChatGPT moment for robotics.” He names Skild AI and Physical Intelligence, which are building foundation models for robots, and calls their progress “really dramatic.” He notes it is harder to scale because you have to build physical robots.
Conclusion #
McGrew’s view is that reasoning has opened a path to more reliable agents, and that startups should start from the best available model. He also calls it a “deep mystery” that AI adoption is slower than people expected in 2018. Watch the full conversation for his own words.
