Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI
- Source
- youtube.com
- Author
- Latent Space
- Date
Context
Noam Brown of OpenAI, a reasoning researcher who earlier worked on the Diplomacy-playing bot Cicero, argues on the Latent Space podcast that reasoning models can improve on tasks without an easy right answer. Deep research is the evidence offered: there is no simple score for a good report, yet people can tell a good report from a mediocre one, and Brown says that is enough signal to keep improving the product.
Brown adds that a base model needs enough capability before extra thinking time helps, and doubts the approach would have gained much on GPT-2, while saying the measurements to define this have not been done. The episode was recorded after Brown's 2025 Diplomacy world championship win and treats o3 as a current model. For background on spending compute at answer time, see this review of test-time compute.
- test-time compute — Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
- multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
Checking sign-in…
Loading comments…






