Hill climbing a moving hill doesn't work.
- Hill climbing a moving hill doesn't work.
- LLMs are moving hills.
- The models are still improving rapidly.
- Don't over optimize for their current behavior.
From other episodes
- If it's a hill climbing exercise, first, make sure you're on the right hill.2026-01-06
- Someone's personality could emerge even from very small starting biases.2025-11-04
- Learning happens in feedback loops.2024-12-09
- If you peek into how multi-modal models work, they're a ball of cheap, random hacks that turn out to be terrifically, unreasonably effective.2024-11-11
- A judo move: if you can frame a problem as an optimization problem (without dangerous externalities) you can now use hill climbing techniques on it.2024-10-28
- LLMs make the "good enough" viable zone larger.2024-10-14
- Once you frame it as a smooth problem (differentiable), it can be optimized and use hill climbing.2024-08-19
- An interesting pattern: using LLMs to astroturf content in an ecosystem.2024-07-01