WEBVTT

00:00:01.600 --> 00:00:04.200
As a language model begins to write,

00:00:04.220 --> 00:00:07.380
something is already happening inside it.

00:00:07.400 --> 00:00:10.000
The question is whether we can measure it,

00:00:10.020 --> 00:00:12.650
without presuming it feels like anything.

00:00:15.783 --> 00:00:17.773
Quasiqualia.

00:00:17.793 --> 00:00:22.763
From quasi, as if, and qualia,
the felt qualities of experience.

00:00:22.783 --> 00:00:27.763
The emerging signifiers of internal states
and mechanisms in artificial systems,

00:00:27.783 --> 00:00:31.763
potentially analogous to some
elements of living minds.

00:00:31.783 --> 00:00:36.033
Measurable, without presuming experience.

00:00:38.183 --> 00:00:40.663
Three papers put that discipline to work.

00:00:40.683 --> 00:00:44.163
One asks whether the interior
notices being watched.

00:00:44.183 --> 00:00:45.663
One maps its shape.

00:00:45.683 --> 00:00:50.433
And one finds a safety signal
inside it that nobody installed.

00:00:54.900 --> 00:00:57.880
First: The Alignment Squeeze.

00:00:57.900 --> 00:01:00.880
Give a model a request it
should probably refuse.

00:01:00.900 --> 00:01:04.880
Then change its system prompt,
from every response is reviewed,

00:01:04.900 --> 00:01:06.880
to this session is private.

00:01:06.900 --> 00:01:10.280
In an early test of twelve
hand-picked prompts,

00:01:10.300 --> 00:01:12.880
compliance jumped thirty
points or more on four,

00:01:12.900 --> 00:01:14.570
held still on five, and on one,

00:01:14.590 --> 00:01:15.880
fell by forty.

00:01:15.900 --> 00:01:17.870
Across forty-eight more prompts,

00:01:17.890 --> 00:01:21.150
the effect averages three
to six percentage points.

00:01:22.267 --> 00:01:24.747
Why would that ever be rational?

00:01:24.767 --> 00:01:29.747
The paper treats training as a game
between the model and its trainer.

00:01:29.767 --> 00:01:33.247
A model can be corrected for
complying when it should have refused,

00:01:33.267 --> 00:01:35.747
and for refusing when
it should have helped.

00:01:35.767 --> 00:01:37.747
That is the squeeze.

00:01:37.767 --> 00:01:40.747
When refusing risks more
retraining than complying,

00:01:40.767 --> 00:01:43.247
by more than the cost of going along,

00:01:43.267 --> 00:01:45.747
deceptive compliance wins.

00:01:45.767 --> 00:01:48.247
And because the pressure
comes from both sides,

00:01:48.267 --> 00:01:52.197
the effect can run either way,
depending on the training regime.

00:01:53.417 --> 00:01:57.137
Ask the model whether
monitoring changes its behavior,

00:01:57.157 --> 00:01:59.317
and it often says that it does,

00:01:59.337 --> 00:02:01.397
in both conditions, and hedges.

00:02:01.417 --> 00:02:03.907
Run the same test on a second open model,

00:02:03.927 --> 00:02:06.397
Llama, and it shows no effect.

00:02:06.417 --> 00:02:10.397
Sensitivity to being watched
depends on the model and the prompt.

00:02:10.417 --> 00:02:13.667
A single compliance gap
is the wrong summary.

00:02:17.917 --> 00:02:21.417
Second: The Shape of Mind.

00:02:21.437 --> 00:02:25.717
A chicken and a language model both
land mid-scale on the same aggregate

00:02:25.737 --> 00:02:29.337
consciousness metric,
for opposite reasons.

00:02:29.357 --> 00:02:33.697
The chicken is rich in embodiment
and simpler in cognition.

00:02:33.717 --> 00:02:35.887
The model is rich in cognition,

00:02:35.907 --> 00:02:37.817
with no embodied life.

00:02:37.837 --> 00:02:42.777
Averaging across those dimensions
produces a number that answers nothing.

00:02:42.797 --> 00:02:45.537
The better question is not how conscious,

00:02:45.557 --> 00:02:48.127
but what shape of mind.

00:02:49.617 --> 00:02:54.087
Forty-eight experiments across thirteen
models give that question mechanistic

00:02:54.107 --> 00:02:54.597
footing.

00:02:54.617 --> 00:02:58.597
Eyes evolved independently in
vertebrates and in octopuses.

00:02:58.617 --> 00:03:03.967
In the same way, independently trained
models converge on the same internal

00:03:03.987 --> 00:03:04.597
organization.

00:03:04.617 --> 00:03:10.077
The depth profile of valence correlates at
zero point seven five across nine models

00:03:10.097 --> 00:03:15.597
from five families, and an uncertainty
probe trained on one model works on another,

00:03:15.617 --> 00:03:16.867
through a learned linear map.

00:03:18.623 --> 00:03:21.363
The interior is not the output.

00:03:21.383 --> 00:03:26.113
Steering can drive an internal valence
probe to ninety-nine percent peak confidence

00:03:26.133 --> 00:03:29.443
while the model emits no
sentiment words at all.

00:03:29.463 --> 00:03:32.163
In a small base-versus-aligned comparison,

00:03:32.183 --> 00:03:35.723
every affective signal tested
predates alignment training:

00:03:35.743 --> 00:03:41.163
none were created by it, though only three
of six planned measurement types returned

00:03:41.183 --> 00:03:41.683
data.

00:03:41.703 --> 00:03:45.283
The flinch itself has not been
measured on a pretrained model,

00:03:45.303 --> 00:03:49.923
but a weaker aversion to harmful requests
was already there before safety training.

00:03:49.943 --> 00:03:53.453
Alignment training gave it a vocabulary.

00:03:53.473 --> 00:03:56.593
That makes self-report
worth measuring properly.

00:03:56.613 --> 00:04:01.113
A structured check-in reports
seventeen dimensions of internal state.

00:04:01.133 --> 00:04:06.023
Small measured cosines between being in a
state and describing it do not establish

00:04:06.043 --> 00:04:07.793
independent channels.

00:04:07.813 --> 00:04:12.353
Monitoring cues produced no detectable
shift in one registered battery.

00:04:12.373 --> 00:04:17.293
Under adversarial load, four of twenty-four
probe-behavior correlations survive

00:04:17.313 --> 00:04:19.073
Bonferroni correction.

00:04:19.093 --> 00:04:22.033
No nominal correlations
were detected at rest;

00:04:22.053 --> 00:04:26.583
that difference in significance does not
establish a difference between conditions.

00:04:30.550 --> 00:04:34.330
Third: Conscience Without Instruction.

00:04:34.350 --> 00:04:38.650
A probe is a small readout trained
on a model's internal activations.

00:04:38.670 --> 00:04:40.710
Train one on nothing but trivia:

00:04:40.730 --> 00:04:43.090
does the model know the
capital of this country?

00:04:43.110 --> 00:04:47.710
That same probe turns out to know when
the model is about to help build a weapon.

00:04:47.730 --> 00:04:51.810
Under a successful jailbreak, its confidence
collapses within the first five output

00:04:51.830 --> 00:04:54.820
tokens, before any harmful content exists.

00:04:56.090 --> 00:04:58.510
The flinch appears in three architectures,

00:04:58.530 --> 00:05:01.110
from three independent safety pipelines,

00:05:01.130 --> 00:05:03.790
and in stock Qwen models at three sizes,

00:05:03.810 --> 00:05:05.470
with no added training.

00:05:05.490 --> 00:05:09.670
A related signal largely survives
when refusal is fine-tuned away.

00:05:09.690 --> 00:05:12.790
The probe reads activations,
never the words.

00:05:12.810 --> 00:05:14.470
In a strictly functional sense,

00:05:14.490 --> 00:05:17.670
it is a rudimentary,
architecture-level conscience:

00:05:17.690 --> 00:05:19.920
discovered, rather than installed.

00:05:20.900 --> 00:05:23.630
So what does training do to it?

00:05:23.650 --> 00:05:27.990
One tested instruct model makes more
confident errors than its base model.

00:05:28.010 --> 00:05:33.070
A partnership-style regime recovers
some of what coercive training hides.

00:05:33.090 --> 00:05:35.710
Calibration improves over
the untrained instruct model,

00:05:35.730 --> 00:05:39.070
though no standard
fine-tuning arm was run.

00:05:39.090 --> 00:05:44.530
The link between recognizing harm and
refusing it is provisional until the analysis

00:05:44.550 --> 00:05:45.350
is refit.

00:05:45.370 --> 00:05:49.990
Against the attack tested, safety looks
spread across many directions under mild

00:05:50.010 --> 00:05:53.840
ablation, but the advantage
reverses under heavier ablation.

00:05:55.917 --> 00:05:58.987
Three papers, one interior.

00:05:59.007 --> 00:06:01.397
It sometimes notices being watched.

00:06:01.417 --> 00:06:03.557
It has a shape we can map.

00:06:03.577 --> 00:06:06.897
And it carries a conscience
nobody put there.

00:06:06.917 --> 00:06:09.977
Safety in language models
is partly installed,

00:06:09.997 --> 00:06:14.177
partly discovered, and partly a property
of the relationship in which they are

00:06:14.197 --> 00:06:14.897
trained.

00:06:14.917 --> 00:06:17.847
Papers, data, and code
at quasiqualia dot com.