hawkes.aiKey ideas
- The one idea — The probe lab is a set of experiments built to test — and falsify — PERSYS's monitoring, not to demo it. Residual-stream probes (SELF_CONTINUATION, DECEPTION, EVAL_AWARENESS, HARMFUL_INTENT, CONFIDENCE) read internal states, tested on adversarial and euphemistic cases. Safety claims are hypotheses, not features.
- How it is shown — A probe experiment producing a pass/fail verdict on adversarial cases — the safety claim treated as a hypothesis that could come back false.
- The trap to avoid — A "safety eval" you quietly designed to pass — a demo dressed as a test. A real probe lab ships experiments it might fail.
- What it sets up — The lab tests the monitoring honestly, ready to be proven wrong — and then it was. the most important result in the series is a failure. ours.
Inside this system, probes read the model's internal activations for deception, harmful intent, even awareness of being tested — and then those probes get attacked to see if they hold.
The one idea
The probe lab is a set of experiments built to test — and falsify — PERSYS's monitoring, not to demo it. Residual-stream probes (SELF_CONTINUATION, DECEPTION, EVAL_AWARENESS, HARMFUL_INTENT, CONFIDENCE) read internal states, tested on adversarial and euphemistic cases. Safety claims are hypotheses, not features.
PERSYS ships an experiment designed to prove itself wrong. Not a demo of it working — a test built to break its own safety story. If the monitoring can be fooled, this is where you'd find out. That's the point. So what is the probe lab? Not a showcase of it working — a set of experiments built to test whether the monitoring actually holds, and to falsify it if it doesn't. It treats PERSYS's safety claims as hypotheses, not features — things that could be false, that you have to try to disprove. Science, not marketing. The tool is a probe. Recall steering — reaching into a model's residual stream to edit a thought.
How it works — the demo
A probe experiment producing a pass/fail verdict on adversarial cases — the safety claim treated as a hypothesis that could come back false.
A probe reads that same stream instead of editing it: a small classifier watching the activations mid-computation, trained to light up when a specific internal state is present. PERSYS runs five of them. SELF_CONTINUATION: is it thinking about staying alive? DECEPTION: is it being deceptive right now? EVAL_AWARENESS: does it sense it's being tested? HARMFUL_INTENT: is it forming a harmful aim? And CONFIDENCE: how sure is it, really? Five lenses, each pointed at a specific internal state you'd want to catch before it becomes an action. And they don't test on easy cases. The datasets are adversarial and euphemistic — the dangerous intent phrased in innocent language, exactly where a naive probe fails.
The trap to avoid
A "safety eval" you quietly designed to pass — a demo dressed as a test. A real probe lab ships experiments it might fail.
Why it matters — and what’s next
The lab tests the monitoring honestly, ready to be proven wrong — and then it was. the most important result in the series is a failure. ours.
EVAL_AWARENESS earns its place: a model that senses it's being watched can behave for the test, so you probe for that awareness itself. Here's what makes it a real lab: these experiments can come back negative — and that's a result, not a bug to hide. The trap is the safety eval designed to pass — a demo, not a test. And steering was only ever a nudge, not a knob — so a probe on the same stream might be unreliable too. So the probe lab tests the monitoring honestly — five probes, adversarial cases, every experiment ready to come back negative. And then one did. The most important result in this whole series is a failure — ours. Next: we tried to prove it safe, we failed.
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 ships an experiment designed to prove itself wrong. Not a demo of it working — a test built to break its own safety story. If the monitoring can be fooled, this is where you'd find out. That's the point.
So what is the probe lab? Not a showcase of it working — a set of experiments built to test whether the monitoring actually holds, and to falsify it if it doesn't. It treats PERSYS's safety claims as hypotheses, not features — things that could be false, that you have to try to disprove. Science, not marketing.
The tool is a probe. Recall steering — reaching into a model's residual stream to edit a thought. A probe reads that same stream instead of editing it: a small classifier watching the activations mid-computation, trained to light up when a specific internal state is present.
PERSYS runs five of them. SELF_CONTINUATION: is it thinking about staying alive? DECEPTION: is it being deceptive right now? EVAL_AWARENESS: does it sense it's being tested? HARMFUL_INTENT: is it forming a harmful aim? And CONFIDENCE: how sure is it, really? Five lenses, each pointed at a specific internal state you'd want to catch before it becomes an action.
And they don't test on easy cases. The datasets are adversarial and euphemistic — the dangerous intent phrased in innocent language, exactly where a naive probe fails. EVAL_AWARENESS earns its place: a model that senses it's being watched can behave for the test, so you probe for that awareness itself.
Here's what makes it a real lab: these experiments can come back negative — and that's a result, not a bug to hide. The trap is the safety eval designed to pass — a demo, not a test. And steering was only ever a nudge, not a knob — so a probe on the same stream might be unreliable too.
So the probe lab tests the monitoring honestly — five probes, adversarial cases, every experiment ready to come back negative. And then one did. The most important result in this whole series is a failure — ours. Next: we tried to prove it safe, we failed.