Bottom-Up Research
A research mode where you start with what seems interesting or tractable — a technique, a surprising model behavior, a quick experiment — and figure out the problem it solves later or in parallel.
Contrast with top-down: identify the problem first, then find the approach. Most projects are a mix; the balance shifts based on context.
Bottom-up is most productive when:
- The field has unexplored territory (new model generations, new capabilities = tons of low-hanging fruit)
- Experiments are cheap and fast (hours to a day)
- You can get rapid feedback from reality
Bottom-up is less appropriate when:
- The work is conceptually heavy and can't get empirical feedback (AI control theory, policy-facing demos)
- Specific problems need to be solved in a sequence before further scaling (timelines pressure pushes top-down)
Ethan Perez's thesis: bottom-up is underrated in alignment research, where the culture encourages reading and theorizing first. The empirical loop — run something, see what happens, adjust — produces more confident knowledge faster. Many junior researchers spend a month reading alignment critiques and feel demotivated before trying anything. The opposite attitude — "this seems plausible, I can test it in an hour" — is more productive on the margin.
Red flags that override bottom-up: experiment loop > 1 day; fixed cost > 1 month; "even if successful, this wouldn't matter."