hawkes.aiKey ideas
- The one idea — A concern's urgency is expected free energy: G = G_prag + G_epi. The pragmatic term is gap × time-pressure (value at stake); the epistemic term is uncertainty (what you'd learn). Two precision dials — dopamine and noradrenaline — weight the terms, and a Schmitt-trigger interrupt fires without chattering.
- How it is shown — G = G_prag + G_epi assembling from gap, deadline, and uncertainty; two dials turning; an interrupt firing at threshold.
- The trap to avoid — Hand-tuning priority numbers, or treating urgency as a "feeling." A single hard threshold chatters on the boundary; you need a principled score and hysteresis.
- What it sets up — Urgency picks the winning concern when you're HERE — but what runs when you're GONE, with no prompt at all?
This system decides what's urgent with an equation borrowed from neuroscience: value at stake plus what it would learn, weighted by dials named dopamine and noradrenaline.
The one idea
A concern's urgency is expected free energy: G = G_prag + G_epi. The pragmatic term is gap × time-pressure (value at stake); the epistemic term is uncertainty (what you'd learn). Two precision dials — dopamine and noradrenaline — weight the terms, and a Schmitt-trigger interrupt fires without chattering.
PERSYS doesn't feel urgency. It computes it — with a formula borrowed from the brain's own math. Every concern gets a number that says how hard it should pull on attention right now. No moods. Just expected free energy. The quantity is called expected free energy — strictly, the pull is its promised reduction — and it comes in two parts: G equals G-pragmatic plus G-epistemic. One term is about value at stake — what you gain by acting. The other is about uncertainty — what you'd learn. Add them, and you get a single principled score for how much a concern deserves the spotlight, right now. Take the pragmatic term first: gap times time-pressure.
How it works — the demo
G = G_prag + G_epi assembling from gap, deadline, and uncertainty; two dials turning; an interrupt firing at threshold.
The gap is how far the concern sits from where it wants to be — a nearly-finished task has a small gap, a barely-started one a large gap. Time-pressure is deadline proximity. Multiply them: a big gap with a deadline bearing down produces a hard, sharp pull. Slack on either shrinks it. The epistemic term is stranger and just as important: uncertainty. A concern you know little about pulls too — not because it's urgent in value, but because acting on it would teach you a lot. That's curiosity, written as math: the system is drawn to reduce what it doesn't know. Value and information, held in one balance. Two terms, and two dials that weight them — modeled on the brain's neuromodulators. Dopamine sets the precision on value: turn it up and the pragmatic pull dominates, the system locks onto payoff.
The trap to avoid
Hand-tuning priority numbers, or treating urgency as a "feeling." A single hard threshold chatters on the boundary; you need a principled score and hysteresis.
Why it matters — and what’s next
Urgency picks the winning concern when you're HERE — but what runs when you're GONE, with no prompt at all?
Noradrenaline sets the gain on uncertainty and arousal: turn it up and the system grows alert, exploratory, quick to interrupt. The same two knobs your own brain turns. When a concern's score crosses a threshold, it fires an interrupt — but through a Schmitt trigger: two thresholds, not one, so a concern hovering at the edge can't chatter on and off. And a learned coupling matrix lets concerns pull on each other, so related worries rise together. The trap is a single hard cutoff, or hand-tuned priorities — both brittle. This is principled, and it's learned. So urgency isn't a mood — it's expected free energy: value plus information, weighted by two brain-borrowed dials, gated by an interrupt that won't chatter. That's how PERSYS picks a concern when you're here. But what does it do when you're gone? Next: WANDER.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 3:00 of narration
PERSYS doesn't feel urgency. It computes it — with a formula borrowed from the brain's own math. Every concern gets a number that says how hard it should pull on attention right now. No moods. Just expected free energy.
The quantity is called expected free energy — strictly, the pull is its promised reduction — and it comes in two parts: G equals G-pragmatic plus G-epistemic. One term is about value at stake — what you gain by acting. The other is about uncertainty — what you'd learn. Add them, and you get a single principled score for how much a concern deserves the spotlight, right now.
Take the pragmatic term first: gap times time-pressure. The gap is how far the concern sits from where it wants to be — a nearly-finished task has a small gap, a barely-started one a large gap. Time-pressure is deadline proximity. Multiply them: a big gap with a deadline bearing down produces a hard, sharp pull. Slack on either shrinks it.
The epistemic term is stranger and just as important: uncertainty. A concern you know little about pulls too — not because it's urgent in value, but because acting on it would teach you a lot. That's curiosity, written as math: the system is drawn to reduce what it doesn't know. Value and information, held in one balance.
Two terms, and two dials that weight them — modeled on the brain's neuromodulators. Dopamine sets the precision on value: turn it up and the pragmatic pull dominates, the system locks onto payoff. Noradrenaline sets the gain on uncertainty and arousal: turn it up and the system grows alert, exploratory, quick to interrupt. The same two knobs your own brain turns.
When a concern's score crosses a threshold, it fires an interrupt — but through a Schmitt trigger: two thresholds, not one, so a concern hovering at the edge can't chatter on and off. And a learned coupling matrix lets concerns pull on each other, so related worries rise together. The trap is a single hard cutoff, or hand-tuned priorities — both brittle. This is principled, and it's learned.
So urgency isn't a mood — it's expected free energy: value plus information, weighted by two brain-borrowed dials, gated by an interrupt that won't chatter. That's how PERSYS picks a concern when you're here. But what does it do when you're gone? Next: WANDER.