hawkes.aiKey ideas
- The one idea — PERSYS gates proactive surfacing with conformal risk control (Angelopoulos–Bates): it sets a threshold from a small calibration set so the dismiss-rate among surfaced items is provably bounded, on average, at or below α = 0.25, with a finite-sample, distribution-free guarantee after ≥20 samples.
- How it is shown — A conformal gate filtering low-value candidates; a calibration set setting the threshold; a dismiss-rate meter holding under a quarter.
- The trap to avoid — Gating on a raw, uncalibrated confidence score — no guarantee, just vibes. Conformal control gives a stated, checkable bound instead.
- What it sets up — PERSYS acts on its own but bounded — the mind is complete. but is its reasoning FAITHFUL? and can you turn it OFF?
Proactive AI has a spam problem: interrupt too often and users tune out. The fix here is math — a provable cap on how often surfaced items get dismissed.
The one idea
PERSYS gates proactive surfacing with conformal risk control (Angelopoulos–Bates): it sets a threshold from a small calibration set so the dismiss-rate among surfaced items is provably bounded, on average, at or below α = 0.25, with a finite-sample, distribution-free guarantee after ≥20 samples.
PERSYS won't surface something unless the math says it's worth your attention. Not a vibe, not a guess — a bound. This is the last gate every proactive thought has to clear: the usefulness gate. The pressure it relieves: concerns and WANDER generate a stream of candidates — far more than you'd want to see. Show them all and PERSYS becomes notification spam. So it needs a bar: only surface what's likely useful. But "likely useful" can't be a feeling — it has to be a guarantee you can check. The tool is conformal risk control — the same machinery from Act 09 that turned a confidence score into a real guarantee, now aimed at usefulness. Give it a calibration set and it picks a threshold so a chosen error rate is provably bounded, distribution-free and valid in finite samples.
How it works — the demo
A conformal gate filtering low-value candidates; a calibration set setting the threshold; a dismiss-rate meter holding under a quarter.
Here's the specific promise. The dismiss-rate among surfaced items is held, on average, at or below alpha — and alpha is set to zero-point-two-five. In plain terms: no more than a quarter of what PERSYS proactively brings you should be stuff you'd wave away. A stated, on-average ceiling on how often it's allowed to waste your attention. And it earns that bound honestly. It needs a modest calibration set — around twenty of your real dismissals — and from those it sets the threshold, after which the guarantee holds with finite-sample validity. This isn't "trust me, that's useful"; it's a cutoff calibrated against how you actually react. Your dismissals are the training signal. So the loop is simple: WANDER proposes, the gate disposes.
The trap to avoid
Gating on a raw, uncalibrated confidence score — no guarantee, just vibes. Conformal control gives a stated, checkable bound instead.
Why it matters — and what’s next
PERSYS acts on its own but bounded — the mind is complete. but is its reasoning FAITHFUL? and can you turn it OFF?
Every candidate only reaches you if it clears the conformal bar. That's what turns WANDER's "explore freely, speak selectively" from a nice intention into a number you can hold it to. The trap is gating on a raw confidence score instead: uncalibrated, no guarantee, straight back to vibes. So PERSYS acts on its own, but bounded — a guaranteed ceiling on how often it wastes your attention. That closes the mind: concerns, urgency, WANDER, forgetting, provenance, a usefulness gate. But a system that acts unasked raises sharper questions. Is its reasoning faithful? Can you turn it off? Next group: faithfulness, shutdown, tripwires.
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
PERSYS won't surface something unless the math says it's worth your attention. Not a vibe, not a guess — a bound. This is the last gate every proactive thought has to clear: the usefulness gate.
The pressure it relieves: concerns and WANDER generate a stream of candidates — far more than you'd want to see. Show them all and PERSYS becomes notification spam. So it needs a bar: only surface what's likely useful. But "likely useful" can't be a feeling — it has to be a guarantee you can check.
The tool is conformal risk control — the same machinery from Act 09 that turned a confidence score into a real guarantee, now aimed at usefulness. Give it a calibration set and it picks a threshold so a chosen error rate is provably bounded, distribution-free and valid in finite samples.
Here's the specific promise. The dismiss-rate among surfaced items is held, on average, at or below alpha — and alpha is set to zero-point-two-five. In plain terms: no more than a quarter of what PERSYS proactively brings you should be stuff you'd wave away. A stated, on-average ceiling on how often it's allowed to waste your attention.
And it earns that bound honestly. It needs a modest calibration set — around twenty of your real dismissals — and from those it sets the threshold, after which the guarantee holds with finite-sample validity. This isn't "trust me, that's useful"; it's a cutoff calibrated against how you actually react. Your dismissals are the training signal.
So the loop is simple: WANDER proposes, the gate disposes. Every candidate only reaches you if it clears the conformal bar. That's what turns WANDER's "explore freely, speak selectively" from a nice intention into a number you can hold it to. The trap is gating on a raw confidence score instead: uncalibrated, no guarantee, straight back to vibes.
So PERSYS acts on its own, but bounded — a guaranteed ceiling on how often it wastes your attention. That closes the mind: concerns, urgency, WANDER, forgetting, provenance, a usefulness gate. But a system that acts unasked raises sharper questions. Is its reasoning faithful? Can you turn it off? Next group: faithfulness, shutdown, tripwires.