How economists can participate in the post-training era
ethayarajh · x · 2026-08-22
Sachin.ethayarajh shares slides from a talk introducing post-training to young economists. He notes that training your own model is more feasible than ever, and economists/social scientists can play a major role in the next era of post-training.
The post-training suite looks roughly like:
- SFT
- Offline Methods (DPO/KTO/SimPO/etc)
- Online Methods (PPO/GRPO/CISPO/etc)
- RL Environments
- Distillation (<--- we are here)
- World Adaptation (<--- the future)
As agents deploy in the real world, the world will adapt to them. This adaptation is largely overlooked by current paradigms, which prefer static environments. We shouldn't assume the real world remains fixed.
More from AGI Musings
- In a world of infinite answers, the question is the scarce part — bratton · 2026-08-22
- In a World of Infinite Answers, the Question is the Scarce Resource — every · 2026-08-22
- TIME op-ed: Formalization is the answer to 'vibe coding' — soumitrashukla9 · 2026-08-22
- US electricians union leader calls for more data centers — Polymarket · 2026-08-22
- Superintelligence hailed as pinnacle of Western tech to lead humanity across star systems — SydSteyerhart · 2026-08-22
- Opinion: Datacenter Sudo Creds Are the 21st Century Keys to Power — curious_vii · 2026-08-22