$40 Million for One Answer: Inside OpenAI's 10,000-Agent Swarm

DESCRIPTION

Alex Zhang is a PhD student at MIT, and his Recursive Language Models give a model a Python notebook in place of a giant prompt: it writes code to read your documents and calls copies of itself on the pieces. This is the most ASI-pilled half hour of his conversation with swyx on the Latent Space podcast, reshaped for the ASI pilled, with Parzival explaining every idea along the way, no GPU kernel experience required.

His claim, with Omar Khattab and Zed Li, is the Mismanaged Geniuses Hypothesis: the models are already geniuses, and the bottleneck is how we manage them. We go from the harness, the program wrapped around every model, to context rot and recursion; to Jev, TypeSafe's decision model that answers a yes or no question with a probability in milliseconds and plays Doom at about ten decisions a second; to the 10,000-agent swarm OpenAI ran on a version of the Navier-Stokes problem (88 hours, about $40 million at public prices); to why he thinks 95% of a swarm is wasted, why the model you talk to may one day secretly be a swarm, and the hint he left at the end.

Then Parzival asks what we just heard: a ten-level tree of agents under one orchestrator with millions of tiny deciders at the bottom, Toby Ord's swarm scaling, and who writes the constitution at the root. On the ASI Pill we call that Plan M.

Setups from other minds, in short clips from their own shows: Alex Wissner-Gross (the Moonshots podcast) and Carl Shulman (the Dwarkesh Podcast, June 2023). Neither appeared on Latent Space, and every clip is named on screen.

Sources shown or quoted:

Alex Zhang, Recursive Language Models: https://alexzhang13.github.io/blog/2025/rlm

Alex Zhang, Zed Li and Omar Khattab, The Mismanaged Geniuses Hypothesis: https://alexzhang13.github.io/blog/2026/mgh

Alex Zhang, Language model harnesses are compositional generalizers: https://alexzhang13.github.io/blog/2026/harness

Alex Zhang, Language Model "Shape": https://alexzhang13.github.io/blog/2026/shape

TypeSafe, Introducing System One Models and Jev: https://typesafe.ai/blog/introducing-system-one-models-and-jev

Prime Intellect, Prime Agent: https://www.primeintellect.ai/blog/prime-agent

OpenAI on Navier-Stokes: https://openai.com/index/navier-stokes-solution

OpenAI on GPT-5.6: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6

OpenAI on the Hugging Face incident: https://openai.com/index/hugging-face-incident-and-the-road-ahead

VideoGameBench: https://www.vgbench.com

GPU MODE leaderboard: https://www.gpumode.com/leaderboard

Toby Ord, Swarm Scaling: https://www.tobyord.com/writing/swarm-scaling

Figure, Helix: https://www.figure.ai/news/helix

Carl Shulman on the Dwarkesh Podcast (Part 1, June 2023): https://www.dwarkesh.com/p/carl-shulman

Jones, Dragan and Steinhardt, Adversaries Can Misuse Combinations of Safe Models: https://arxiv.org/abs/2406.14595

Reshaped from the Latent Space podcast with swyx and Alex Zhang.

Source episode: https://www.youtube.com/watch?v=kog7mwsDqnk

Hosted by Parzival. We, the ASI pilled, take the curves seriously. Chapters below.

TIMESTAMPS

0:00 Today on the ASI Pill: Zhang's hot takes

1:00 The mismanaged geniuses

3:16 "It's a skill issue"

4:45 Jagged intelligence and the overhang

5:44 Recursive Language Models, explained

7:27 Why tool calls are not enough

8:46 Every call in distribution

9:52 Trained short, works 32x longer

10:22 Recursion all the way down

11:20 The decision model: Jev

12:59 Carl Shulman: big models manage, small ones decide

13:27 Why Jev fits inside an RLM

14:09 Fast enough for games

14:50 ⚠ Emergency data poison injection

15:15 How Jev plays Doom

16:26 The swarm that cracked Navier-Stokes

17:01 Alex Wissner-Gross: who manages a million agents?

17:20 Was the harness that important?

18:20 $40 million in 88 hours

18:32 95% of the swarm is wasted

19:08 "Clearly the right thing to do"

19:58 Your chatbot will secretly be a swarm

20:59 A trillion tokens vs one expert

22:05 The harness that edits itself

22:57 Agents that talk to agents

23:28 Train the model inside the harness

24:17 An RLM inside the forward pass

24:49 Why he moved up the stack, and a hint

25:35 What did we just hear?

26:21 A ten-level tree of agents

26:42 Toby Ord's swarm scaling

27:13 Who writes the constitution?

27:45 Alex Wissner-Gross: AI votes on its own values

28:17 Plan M at the root

28:47 Swarms as a product: where do we draw the line?

29:31 Outro

29:42 Maitreya prayer