If AI is a verifiable domain, how do you get AIs to construct benchmarks to apply to themselves?
It seems that the only benchmarks that could be automatically constructed are those for tasks that are themselves verifiable (theorem proving, coding, etc).
Non-verifiable tasks require humans to define what exactly a good job is.
The RSI hypothesis is dependent on an assumption that progress on verifiable tasks generalizes to non-verifiable tasks.
There are hundreds of verifiable tasks in AI development as simple benchmarks on the models themselves, such as
1. Performance engineering ability
2. Ability to reconstruct visuals in code alone
3. Ability to orchestrate cheaper models towards a given task
4. Pretty much any verifiable engineering task
The loop is pretty clear, have AI's work on pre and post-training, then benchmark their outputs against these tasks to determine how to steer AI progress. The non-verifiable stuff isn't on the critical path towards these accoutrements anyway.
It seems that the only benchmarks that could be automatically constructed are those for tasks that are themselves verifiable (theorem proving, coding, etc).
Non-verifiable tasks require humans to define what exactly a good job is.
The RSI hypothesis is dependent on an assumption that progress on verifiable tasks generalizes to non-verifiable tasks.