hawkes.aiKey ideas
- The one idea — In idle time PERSYS runs WANDER — it Thompson-samples a concern, walks short associative chains through its memory graph, and surfaces only non-obvious, actionable finds, delivered later in context. Exploration temperature is set by noradrenaline.
- How it is shown — An association-graph walk sparking a non-obvious connection (Tuesday's note ↔ Friday's deadline), queued for the morning.
- The trap to avoid — Two failure modes — go fully dark when idle (discover nothing), or blurt every idle thought (become noise). WANDER explores freely but speaks selectively.
- What it sets up — WANDER keeps finding and storing connections — but a store that keeps EVERYTHING drowns. what should it forget?
There's an AI system that daydreams. With no prompt, it wanders its own memory graph, chaining associations — and only speaks up if it finds something genuinely useful.
The one idea
In idle time PERSYS runs WANDER — it Thompson-samples a concern, walks short associative chains through its memory graph, and surfaces only non-obvious, actionable finds, delivered later in context. Exploration temperature is set by noradrenaline.
It's three in the morning. No prompt, no user, nothing asked. And PERSYS is quietly connecting the note you wrote Tuesday to the deadline you set for Friday. This is what it thinks about when you're gone. Idle doesn't mean off. When you step away, PERSYS doesn't halt and wait — it drops into a background mode called WANDER, the system's version of a mind wandering. No task in front of it, so it goes looking: through its own concerns and memories, for something it hasn't noticed yet. The lights stay on, and they're pointed inward. It doesn't just grind on the most urgent thing. WANDER Thompson-samples a concern — picks one at random, but weighted by how promising it is, so high-value and high-uncertainty concerns come up more often without the rest being starved.
How it works — the demo
An association-graph walk sparking a non-obvious connection (Tuesday's note ↔ Friday's deadline), queued for the morning.
That randomness is deliberate: it's exploration. And how wide it roams is set by noradrenaline, the same arousal dial from before. With a concern in hand, it walks. Memory in PERSYS is a graph of associations, and WANDER takes a few short hops — from the sampled concern to a linked memory, to another — hunting for a path you wouldn't have drawn yourself. Tuesday's offhand note sits two hops from Friday's deadline. Nobody queried that link. The walk found it. Most of what it finds is junk, and WANDER knows it. So there's a filter: surface only the non-obvious and the actionable. That you have a deadline Friday — obvious, dropped.
The trap to avoid
Two failure modes — go fully dark when idle (discover nothing), or blurt every idle thought (become noise). WANDER explores freely but speaks selectively.
Why it matters — and what’s next
WANDER keeps finding and storing connections — but a store that keeps EVERYTHING drowns. what should it forget?
That a detail from Tuesday quietly changes how you'll hit Friday — non-obvious, and useful. Only a real, usable connection survives the cut. And it doesn't wake you. WANDER never pings at 3am; it queues the find and surfaces it later, in context — when you're back and the moment fits. The trap sits on both sides: a system that goes fully dark when idle discovers nothing, and one that blurts every idle thought becomes noise you'll mute. WANDER explores freely, but speaks selectively. So when you're gone, PERSYS wanders: sampling a concern, walking its memory, saving the rare connection for the right moment. But all this remembering and connecting fills a store — and a store that keeps everything eventually drowns. So what should it forget? Next group: forgetting is a feature.
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 2:45 of narration
It's three in the morning. No prompt, no user, nothing asked. And PERSYS is quietly connecting the note you wrote Tuesday to the deadline you set for Friday. This is what it thinks about when you're gone.
Idle doesn't mean off. When you step away, PERSYS doesn't halt and wait — it drops into a background mode called WANDER, the system's version of a mind wandering. No task in front of it, so it goes looking: through its own concerns and memories, for something it hasn't noticed yet. The lights stay on, and they're pointed inward.
It doesn't just grind on the most urgent thing. WANDER Thompson-samples a concern — picks one at random, but weighted by how promising it is, so high-value and high-uncertainty concerns come up more often without the rest being starved. That randomness is deliberate: it's exploration. And how wide it roams is set by noradrenaline, the same arousal dial from before.
With a concern in hand, it walks. Memory in PERSYS is a graph of associations, and WANDER takes a few short hops — from the sampled concern to a linked memory, to another — hunting for a path you wouldn't have drawn yourself. Tuesday's offhand note sits two hops from Friday's deadline. Nobody queried that link. The walk found it.
Most of what it finds is junk, and WANDER knows it. So there's a filter: surface only the non-obvious and the actionable. That you have a deadline Friday — obvious, dropped. That a detail from Tuesday quietly changes how you'll hit Friday — non-obvious, and useful. Only a real, usable connection survives the cut.
And it doesn't wake you. WANDER never pings at 3am; it queues the find and surfaces it later, in context — when you're back and the moment fits. The trap sits on both sides: a system that goes fully dark when idle discovers nothing, and one that blurts every idle thought becomes noise you'll mute. WANDER explores freely, but speaks selectively.
So when you're gone, PERSYS wanders: sampling a concern, walking its memory, saving the rare connection for the right moment. But all this remembering and connecting fills a store — and a store that keeps everything eventually drowns. So what should it forget? Next group: forgetting is a feature.