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

<p>DESCRIPTION</p><p>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.</p><p><br></p><p>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.</p><p><br></p><p>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.</p><p><br></p><p>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.</p><p><br></p><p>Sources shown or quoted:</p><p>Alex Zhang, Recursive Language Models: alexzhang13.github.io/blog/2025/rlm</p>&… Zhang, Zed Li and Omar Khattab, The Mismanaged Geniuses Hypothesis: alexzhang13.github.io/blog/2026/mgh</p>&… Zhang, Language model harnesses are compositional generalizers: alexzhang13.github.io/blog/2026/harness</p&… Zhang, Language Model "Shape": alexzhang13.github.io/blog/2026/shape</p>…, Introducing System One Models and Jev: typesafe.ai/blog/introducing-system-one-models… Intellect, Prime Agent: primeintellect.ai/blog/prime-agent</p>&l… on Navier-Stokes: openai.com/index/navier-stokes-solution</p&… on GPT-5.6: openai.com/index/advancing-the-price-performan… on the Hugging Face incident: openai.com/index/hugging-face-incident-and-the…: vgbench.com</p><p>GPU MODE leaderboard: gpumode.com/leaderboard</p><p>Toby Ord, Swarm Scaling: tobyord.com/writing/swarm-scaling</p><…, Helix: figure.ai/news/helix</p><p>Carl Shulman on the Dwarkesh Podcast (Part 1, June 2023): dwarkesh.com/p/carl-shulman</p><p>…, Dragan and Steinhardt, Adversaries Can Misuse Combinations of Safe Models: arxiv.org/abs/2406.14595</p><p><… from the Latent Space podcast with swyx and Alex Zhang.</p><p>Source episode: youtube.com/watch?v=kog7mwsDqnk</p><p… by Parzival. We, the ASI pilled, take the curves seriously. Chapters below.</p><p><br></p><p>TIMESTAMPS</p><p>0:00 Today on the ASI Pill: Zhang's hot takes</p><p>1:00 The mismanaged geniuses</p><p>3:16 "It's a skill issue"</p><p>4:45 Jagged intelligence and the overhang</p><p>5:44 Recursive Language Models, explained</p><p>7:27 Why tool calls are not enough</p><p>8:46 Every call in distribution</p><p>9:52 Trained short, works 32x longer</p><p>10:22 Recursion all the way down</p><p>11:20 The decision model: Jev</p><p>12:59 Carl Shulman: big models manage, small ones decide</p><p>13:27 Why Jev fits inside an RLM</p><p>14:09 Fast enough for games</p><p>14:50 ⚠ Emergency data poison injection</p><p>15:15 How Jev plays Doom</p><p>16:26 The swarm that cracked Navier-Stokes</p><p>17:01 Alex Wissner-Gross: who manages a million agents?</p><p>17:20 Was the harness that important?</p><p>18:20 $40 million in 88 hours</p><p>18:32 95% of the swarm is wasted</p><p>19:08 "Clearly the right thing to do"</p><p>19:58 Your chatbot will secretly be a swarm</p><p>20:59 A trillion tokens vs one expert</p><p>22:05 The harness that edits itself</p><p>22:57 Agents that talk to agents</p><p>23:28 Train the model inside the harness</p><p>24:17 An RLM inside the forward pass</p><p>24:49 Why he moved up the stack, and a hint</p><p>25:35 What did we just hear?</p><p>26:21 A ten-level tree of agents</p><p>26:42 Toby Ord's swarm scaling</p><p>27:13 Who writes the constitution?</p><p>27:45 Alex Wissner-Gross: AI votes on its own values</p><p>28:17 Plan M at the root</p><p>28:47 Swarms as a product: where do we draw the line?</p><p>29:31 Outro</p><p>29:42 Maitreya prayer</p>