A collection of some wisdom for working with models.
Some lessons may also apply to humans. 😉
Strategy
- Build systems that leverage models; don’t build models.
- Expect models to get smarter and cheaper.
- Create systems that perform better when scaled.
- The human brain is a beautiful starting point for agent design, but also build simple systems that scale with more compute.
Thinking Well
- Novelty requires some randomness.
- Models discern great ideas quicker than they create them.
- Big problems require lots of thought.
- The quality distribution of thought/experience is Gaussian.
Prompting
- Never encourage in prompt when you can enforce with code.
- Make the models’ job as easy as possible.
- Know when not to give the model data.
- If your prompt is long, you’ve done something wrong.