Starting Point
Prediction-first work can be useful when it compresses many observations into a manipulable model. It becomes weaker when accuracy is treated as explanation.
Main Idea
The interesting zone is not mechanism versus prediction as a binary. It is the point where a predictive model changes what experiment should be done next.
What I Agree With
Prediction can reveal structure that is hard to see from isolated experiments. A good model can make uncertainty more concrete.
What I Question
I question claims that treat benchmark performance as biological understanding. Accuracy can hide shortcut learning, weak perturbation labels, or an evaluation task that is easier than the scientific question.
Current Model
Mechanism-first work identifies causal handles. Prediction-first work can organize high-dimensional observations. The stronger scientific form appears when prediction proposes a perturbation that tests or changes a mechanism.
Connections
- Virtual Cell and Perturbation Prediction
- AI tools as pressure tests for scientific judgment
- Indirect measurement and evidence quality
Open Questions
- What makes a model prediction experimentally useful?
- When should predictive failure be treated as biological signal?
- How can benchmark design reward mechanistic usefulness?
Revision History
- 2026-07-07: Converted into the Thinking Note structure.