Noam Brown – Agent swarms, alignment, & recursive self-improvement

Noam Brown is a researcher at OpenAI who has made significant contributions to the development of the reasoning models and multi-agent systems. Last week, OpenAI announced that they had solved a Millennium Prize Problem using a system consisting of 10,000 AI agents that worked together for 88 hours and used 130 billion tokens. Brown has been interested in how reasoning models can help predict future capabilities of AI systems as test-time compute increases.

As the reasoning models scale up, they demonstrate a clear pattern where longer thinking time leads to better performance, similar to how humans perform better on tasks like the SATs when given more time. However, as the thinking time increases, there is a latency bottleneck that can impact the efficiency of the system. One solution to this problem is to utilize multi-agent systems, where multiple agents work in parallel to accelerate the process.

Although the agents have less context compared to a single agent, they can still collaborate effectively and produce results much faster than a human could.