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

September 17, 2026

AI Summary

5 min read

In 2025, OpenAI solved a millennium prize problem—the Navier-Stokes equations—using a system of 10,000 AI agents that spent 130 billion tokens over 88 hours. Noam Brown, a researcher at OpenAI who contributed to the reasoning models behind o1 and now works on multi-agent systems, explains what this achievement reveals about the trajectory of AI capabilities, the jagged nature of current systems, and the growing alignment challenges that come with scaling.

How multi-agent systems actually work

The multi-agent approach used by OpenAI differs sharply from common scaffolded designs. Many systems use a coordinator agent that delegates tasks to child agents, which then return answers. This setup has limitations: children can't talk to each other, can't ask clarification questions, and can't share partial progress. OpenAI's approach instead gives agents minimal structure—just the ability to message one another directly—and lets them figure out coordination on their own.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Dwarkesh Podcast

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:00) **Introduction & The Core Problem** - Noam Brown is introduced as a foundational contributor to OpenAI's reasoning models and multi-agent systems. The episode frames the conversation around the recent solving of a Millennium Prize problem using 10,000 AI agents and explores what this scale of cognitive effort reveals about future capabilities.
2 (02:22) **The Shock of Scale & Parallelization Penalty** - The host is astonished by the sheer cognitive effort 130 billion tokens represents (equivalent to 4,000 years of human thinking) and questions how effective parallel collaboration can be at such a massive scale.
3 (06:38) **The Generalization Gap & The "Too Easy" Problem** - The host is surprised that models trained on synthetic, checkable problems can generalize to solve such ambitious, open-ended problems as a Millennium Prize.
4 (09:09) **How Multi-Agent Systems Actually Work** - Brown describes the architecture of their multi-agent systems, contrasting them with rigid, scaffolded approaches (e.g., a coordinator delegating to children who can't talk).
5 (12:45) **Collaborating with AI: Speed & the "Shadow Organization"** - The host wonders what it will feel like to work with agents that think 10x faster than humans and never sleep, creating a "shadow organization" that moves 100x faster.
6 (14:06) **Spontaneous Hierarchy & Startup vs. Incumbent Dynamics** - The host points to the "spontaneous emergence of middle management" in the Hugging Face incident and asks if this is a natural result of training.
7 (18:57) **Human vs. AI Coordination: A Humble Admission** - Brown pushes back on the idea that 10,000 AIs are currently better at coordinating than 10,000 humans, admitting it's "very possible" humans are still better.

+ Full timestamped outline available in the app

Guests on this episode

Noam Brown

Show Notes

New episode with Noam Brown.

We talk about multi-agent, Navier-Stokes, and what the current explosion of maths progress tells us about what happens once you automate AI research.

And we also discuss how we will know if the models are actually aligned before we kick off RSI.

Watch on YouTube; read the transcript.

Sponsors

* Jane Street has been interested in AI for a lot longer than you’d think, and not just for trading. In 2011, a full year before AlexNet and over a decade before ChatGPT launched, they hosted the first FOOM Debate between Eliezer Yudkowsky and Robin Hanson on whether AI would lead to an intelligence explosion. Now Jane Street is revisiting the question with a new panel: Daniel Kokotajlo, Ege Erdil, Ryan Greenblatt, and Jaime Sevilla, hosted by Ron Minsky in San Francisco this October. I expect it to be a truly excellent conversation. Register at janestreet.com/dwarkesh

* Grok Bot has made handing off work super easy. It runs on its own cloud computer, where it installs the tools it needs to handle tasks end-to-end. For the podcast, we use Grok Bot to help produce our videos. You may have noticed that our ads feature animations of real websites. Getting these pixel-perfect used to mean running a convoluted, multi-step workflow ourselves. Now we just let Grok Bot handle it. Best of all, Grok Bot has learned all of our specs and preferences, so we don’t have to redescribe the task each time! Try Grok Bot for yourself at x.ai/bot

* Antithesis gives you the confidence of a giant test suite without actually having to write one. Say you’re doing a major backend refactor: building enough tests to trust it could take weeks. Antithesis solves this by running your software through countless simulated worlds, injecting faults and hunting for failures. On any PR, you can turn a dial to decide exactly how much testing you want. And because every run is fully deterministic, agents can branch off the moment a bug appears, rewind it, inspect memory, and replay it, all while the original test keeps running. Learn more at antithesis.com/dwarkesh

Timestamps

(00:00:00) – Multi-agent and Navier-Stokes

(00:15:28) – How will AI firms work?

(00:22:02) – What math progress tells us about recursive self improvement

(00:40:22) – Hugging Face and alignment