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

Noam Brown, a researcher at OpenAI, joins the podcast to discuss the rapid advancements in AI, particularly in multi-agent systems and their implications for the future. Brown was a foundational contributor to OpenAI's reasoning models and is now focused on multi-agent systems.

Multi-Agent Systems and the Millennium Prize Problem

Brown begins by explaining the concept of multi-agent systems, which were recently used by OpenAI to solve one of the Millennium Prize Problems. He clarifies that while the achievement is significant, the core reason for success lies in the power of the base model, not solely the multi-agent approach. Brown explains that reasoning models improve with more 'think time,' similar to humans. However, this leads to a latency bottleneck. To overcome this, multi-agent systems parallelize the computation, allowing multiple agents to work on a problem simultaneously, thus speeding up the process. This is less efficient than a single agent with full context, but it's a very effective way to scale test-time compute if done well.

Brown addresses the 'parallelization penalty,' noting that while multi-agent systems offer speedups, they are often sub-linear. For example, four agents might solve a problem twice as fast, meaning the cost is 2x more for a 2x speedup. The degree of parallelization depends on the problem; math problems are highly parallelizable, while creative tasks like writing a novel are not. He emphasizes that while OpenAI has demonstrated multi-agent systems with up to 16 agents in blog posts, scaling to 10,000 agents for rigorous scientific evaluation is prohibitively expensive. The Millennium Prize Problem solution was a single data point, and a full ablation study to compare against a single agent has not yet been performed.