Following Employment at AnthropicResearch Insights Made Simple
Research in a company of agents
Research acceleration: The view inside OpenAI · Alexander Polomodov
/ Research Insights Made Simple · Research in a company of agents
Slide contents
1. Research in a company of agents
Research acceleration: The view inside OpenAI · Alexander Polomodov
2. OpenAI keeps a meter toward RSI
3. Four questions for the report
4. Who and what the telemetry counts
5. The median researcher: $600 a day
6. Research grows fastest of all teams
7. Agent workdays now triple human ones
8. Weekly peak: 74% at 4+ agents
9. Code: 7× lines vs pre-2025 average
10. Experiments per experimenter up 2.2× since January
11. Agents build, run and answer; people decide
12. Support-channel posts down almost fivefold
13. 4–8 hour task success nearly tripled
14. Longer tasks need more human steering
15. A model grades success; 25 human checks
16. The breach response sank RL compute
17. Astra −59%, others +17%, total barely moved
18. What is measured, what is claimed
19. Same words as Anthropic, other meters
20. Confidence depends on the claim
21. Count up to the accepted result
22. Acceleration needs a verifiable outcome