What should a persistent system ignore?
Framing the question behind Teach AI What to Ignore — and the tension even a toy model makes obvious.
The question
A persistent AI system — one that stays on, watching a stream of input — can't afford to think hard about everything. Most input is irrelevant; a little of it matters a lot. The project asks how such a system can learn to ignore the former and spend expensive reasoning only on the latter.
What the toy model shows
The figure on the projects page is deliberately simple. Each input carries a hidden relevance and a noisy estimate of it; a gate escalates anything whose estimate crosses a threshold. Even this toy makes the central tension visible:
- lower the threshold and the gate escalates more noise, so compute climbs;
- raise it and meaningful inputs start slipping past unnoticed.
While the estimate is noisy, no threshold removes both costs at once. The interesting work is in making the estimate better and the gate smarter — not in picking a magic number.
// the entire decision rule of the toy gate
const escalate = (estimate: number, threshold: number) => estimate >= threshold;Open questions
- What makes a signal meaningful before the system knows what happens next?
- Should the gate adapt its threshold to context, budget or time?
- How do you evaluate a system whose main job is to do nothing, most of the time?
Results will be written up here once they exist.