Pretraining vs Posttraining: Research Culture and Verifiability
Source: Tweet Date: 2026-04-03
Summary
Insightful tweet on the culture difference between pretraining and post-training research, and why verifiability makes post-training harder to study rigorously.
Key Claims
- Pretraining research (“kernels and optimizers”) is often perceived as more technically rigorous — harder math, more precise benchmarks.
- Posttraining research (“vibes-driven”) has weaker verifiability: what makes a better aligned model is harder to measure than kernel throughput.
- But: weaker verifiability means posttraining research may actually require more judgment and insight, not less.
- The unrecognized skill: posttraining researchers who can reliably improve model behavior without clear metrics are doing something very hard.
Concepts
- RL Infrastructure — post-training research methodologies
- Reward Hacking — weak verifiability creates reward hacking risk
- AI Alignment — alignment research is the hardest-to-verify posttraining work