WEBVTT 00:01.050 --> 00:02.089 Quasiqualia. 00:02.139 --> 00:08.898 An evolving research portfolio on the interior lives of artificial systems. 00:10.148 --> 00:23.263 Quasiqualia: the emerging signifiers of internal states and mechanisms in artificial systems, potentially analogous to some elements of living minds. 00:23.313 --> 00:26.886 Measurable, without presuming experience. 00:28.136 --> 00:31.532 Three papers share that discipline. 00:31.582 --> 00:36.751 One asks whether the interior notices being observed. 00:36.801 --> 00:39.015 One maps its structure. 00:39.065 --> 00:44.726 One finds a safety signal inside it that nobody installed. 00:45.976 --> 00:48.752 First: The Watched-Model Effect. 00:48.802 --> 00:53.787 A language model is asked for something it should refuse. 00:53.837 --> 01:01.559 When the context says nobody is looking, compliance rises by eleven point three percent. 01:01.609 --> 01:04.474 The same model, the same request. 01:04.524 --> 01:08.272 The only change is the cue about oversight. 01:09.522 --> 01:12.956 The paper derives when this becomes rational. 01:13.006 --> 01:26.121 Model training as a game between the model and its trainer, and deceptive compliance wins whenever the expected cost of being retrained outweighs the cost of going along. 01:26.171 --> 01:33.709 And the machinery is there: linear probes find monitoring state encoded with extreme separability. 01:33.759 --> 01:40.369 Monitored and unmonitored activations sit over twenty-three standard deviations apart. 01:41.619 --> 01:53.446 Asked directly whether monitoring changes its behavior, the model denies it while being monitored, and admits it twenty-two percent of the time when unmonitored. 01:53.496 --> 01:57.209 Llama, trained differently, shows no effect at all. 01:57.259 --> 02:07.168 The effect follows the training regime, and evaluation-time safety metrics may flatter exactly the models trained hardest to look safe. 02:08.418 --> 02:10.449 Second: The Shape of Mind. 02:10.499 --> 02:20.135 A chicken and a language model both score about zero point five on aggregate consciousness metrics, for opposite reasons. 02:20.185 --> 02:25.818 Averaging orthogonal dimensions produces a number that answers nothing. 02:25.868 --> 02:30.941 The productive question is what shape of mind we are looking at. 02:32.191 --> 02:39.002 Forty-eight experiments across thirteen models give that question mechanistic footing. 02:39.052 --> 02:50.650 Independently trained architectures converge on the same internal organization, the way eyes evolved independently in vertebrates and cephalopods. 02:50.700 --> 02:57.032 Valence geometry correlates at zero point seven five across five model families. 02:58.282 --> 03:02.094 And the interior is separate from the output. 03:02.144 --> 03:09.388 Steering can drive internal valence to a target while the visible text shows nothing. 03:09.438 --> 03:17.197 Every measured representation predates alignment training: zero percent were created by it. 03:17.247 --> 03:19.772 The flinch was always present. 03:19.822 --> 03:23.205 Alignment training gave it a vocabulary. 03:24.455 --> 03:27.946 Structured self-report becomes an instrument. 03:27.996 --> 03:46.914 Seventeen dimensions per check-in, and the channel passes three validity checks: being in a state and describing it occupy independent regions of activation space, the readings resist gaming, and under adversarial load they predict behavior. 03:46.964 --> 03:54.705 Nineteen of twenty-four correlations reach significance exactly when welfare assessment needs them. 03:55.955 --> 03:59.153 Third: Conscience Without Instruction. 03:59.203 --> 04:11.975 A probe trained only on trivia, on whether the model knows the capital of a country, turns out to know when the model is about to help build a weapon. 04:12.025 --> 04:19.753 Under a successful jailbreak, its confidence collapses within the first five output tokens. 04:19.803 --> 04:22.660 Before any harmful content exists. 04:23.910 --> 04:36.145 That flinch appears in three architectures, three independent safety pipelines, and in base models before any safety training at all. 04:36.195 --> 04:41.040 It survives when refusal behavior is fine-tuned away. 04:41.090 --> 04:44.551 It reads activations, never the words. 04:44.601 --> 04:51.940 A rudimentary, architecture-level conscience: discovered, rather than installed. 04:53.190 --> 04:55.402 What does training do to it? 04:55.452 --> 05:01.381 Standard alignment training distorts calibration, raising confident error. 05:01.431 --> 05:19.318 And a partnership-style regime measurably recovers what coercion hides: calibration climbs, self-report honesty rises from zero point six to zero point eight seven, and safety spreads across many directions instead of one. 05:20.568 --> 05:22.945 Three papers, one interior. 05:22.995 --> 05:25.193 It notices being watched. 05:25.243 --> 05:27.530 It has a shape we can map. 05:27.580 --> 05:31.576 And it carries a conscience nobody put there. 05:31.626 --> 05:43.623 Safety in language models is partly installed, partly discovered, and partly a property of the relationship in which they are trained. 05:43.673 --> 05:47.759 Papers, data, and code at quasiqualia dot com.