TRANSKRIP 🔗
WEBVTT
00:00:00.400 --> 00:00:13.080
Welcome to Sleepy Facts about the Universe. Some
00:00:13.080 --> 00:00:16.000
questions sound simple until you try to answer
00:00:16.000 --> 00:00:19.690
them. This episode began with a deceptively straightforward
00:00:19.690 --> 00:00:23.129
inquiry. Is intelligence simply prediction grown
00:00:23.129 --> 00:00:26.629
deep enough to question itself? To explore this,
00:00:26.829 --> 00:00:29.730
we look at how organisms from bacteria to fish
00:00:29.730 --> 00:00:33.899
use anticipation to survive. and how this capacity
00:00:33.899 --> 00:00:37.060
scales up to the complex counterfactual reasoning
00:00:37.060 --> 00:00:40.750
of a crow. The conversation then examines predictive
00:00:40.750 --> 00:00:44.049
processing, a framework proposing that the brain
00:00:44.049 --> 00:00:47.549
does not passively absorb sensory data, but instead
00:00:47.549 --> 00:00:51.009
generates a controlled hallucination, using prediction
00:00:51.009 --> 00:00:53.909
error to refine its internal models of the world.
00:00:54.670 --> 00:00:58.460
We also turn to artificial intelligence. asking
00:00:58.460 --> 00:01:01.100
why large language models can fluently predict
00:01:01.100 --> 00:01:03.520
the next word without genuinely understanding
00:01:03.520 --> 00:01:06.959
the world. By distinguishing between shallow
00:01:06.959 --> 00:01:10.079
statistical extrapolation and deep model -based
00:01:10.079 --> 00:01:13.239
revision, the episode traces the boundaries of
00:01:13.239 --> 00:01:17.420
what makes a system truly intelligent. I wanted
00:01:17.420 --> 00:01:20.140
to understand why a machine can generate a perfect
00:01:20.140 --> 00:01:23.040
sentence yet fail at basic physical intuition,
00:01:23.319 --> 00:01:26.829
while a tiny fish computes a life -saving future
00:01:26.829 --> 00:01:30.959
in milliseconds. Exploring this distinction reveals
00:01:30.959 --> 00:01:33.739
that the thread of prediction alone is not enough.
00:01:34.079 --> 00:01:37.400
The weave of flexible, revisable engagement with
00:01:37.400 --> 00:01:43.530
the world is what truly matters. The fish does
00:01:43.530 --> 00:01:48.870
not think. It computes futures. And in that computation,
00:01:49.310 --> 00:01:52.609
older than thought itself, lies a question we
00:01:52.609 --> 00:01:56.540
are only now learning to ask properly. The question
00:01:56.540 --> 00:02:00.019
is whether intelligence, in all its forms, is
00:02:00.019 --> 00:02:02.739
simply prediction. grown deep enough to question
00:02:02.739 --> 00:02:06.519
itself. Imagine a small fish darting through
00:02:06.519 --> 00:02:11.460
murky water. It senses pressure waves. In an
00:02:11.460 --> 00:02:15.379
instant it changes direction. It does not change
00:02:15.379 --> 00:02:18.750
because it has been struck. It changes because
00:02:18.750 --> 00:02:21.509
it anticipates the mouth of a larger predator
00:02:21.509 --> 00:02:25.020
closing where it would have been. The fish does
00:02:25.020 --> 00:02:28.419
not think this through. Its nervous system has
00:02:28.419 --> 00:02:31.990
computed a future. The body has acted on that
00:02:31.990 --> 00:02:35.740
computation before the event occurs. This is
00:02:35.740 --> 00:02:38.919
prediction in its most ancient and urgent form.
00:02:39.250 --> 00:02:41.870
It is the difference between surviving the next
00:02:41.870 --> 00:02:46.099
second and not. The question this episode examines
00:02:46.099 --> 00:02:48.939
is whether this capacity, anticipating what comes
00:02:48.939 --> 00:02:51.659
next, is not merely one feature of intelligence,
00:02:51.800 --> 00:02:55.590
but its fundamental core. Are minds, brains,
00:02:55.789 --> 00:02:58.449
and perhaps even sufficiently complex machines,
00:02:59.069 --> 00:03:02.789
essentially devices for forecasting? And if so,
00:03:03.030 --> 00:03:07.490
what kind of forecasting counts? There is a vast
00:03:07.490 --> 00:03:11.090
distance between a fish evading a predator and
00:03:11.090 --> 00:03:14.069
a physicist imagining a particle that no instrument
00:03:14.069 --> 00:03:17.469
has yet detected. Yet both may be doing something
00:03:17.469 --> 00:03:21.780
deeply related. To approach this properly, we
00:03:21.780 --> 00:03:24.280
need to begin with what prediction actually means
00:03:24.280 --> 00:03:27.629
in a living system. It is not fortune telling.
00:03:28.259 --> 00:03:31.180
It is the capacity to use past and present information
00:03:31.180 --> 00:03:35.259
to constrain possible futures. These guides present
00:03:35.259 --> 00:03:39.610
action. Every organism that moves, that seeks
00:03:39.610 --> 00:03:43.349
food, that avoids harm, faces a temporal problem.
00:03:43.909 --> 00:03:48.139
The world changes. The changes that matter for
00:03:48.139 --> 00:03:51.039
survival often arrive faster than reaction alone
00:03:51.039 --> 00:03:56.819
could handle. A reflex is fast but blind. A true
00:03:56.819 --> 00:03:59.879
prediction integrates memory. current sensation,
00:04:00.199 --> 00:04:03.919
and some model of how things work. It produces
00:04:03.919 --> 00:04:06.360
behavior that is appropriate not for what is
00:04:06.360 --> 00:04:09.159
happening now, but for what will happen next.
00:04:10.349 --> 00:04:13.050
Consider the simplest case that still deserves
00:04:13.050 --> 00:04:16.639
the name. A bacterium swimming through a gradient
00:04:16.639 --> 00:04:19.800
of nutrients does not predict in any rich sense,
00:04:20.100 --> 00:04:25.689
yet it behaves anticipatorily. Escherichia coli.
00:04:26.250 --> 00:04:29.750
moves in a distinctive pattern of runs and tumbles.
00:04:30.339 --> 00:04:34.339
During a run, the bacterium swims in a relatively
00:04:34.339 --> 00:04:38.730
straight line, propelled by rotating flagella.
00:04:38.980 --> 00:04:43.670
Then it stops, the flagella unbondle, causing
00:04:43.670 --> 00:04:47.750
the cell to tumble randomly. It reorients in
Load full transcript
00:04:47.750 --> 00:04:53.189
a new direction. Another run follows. This stop
00:04:53.189 --> 00:04:56.740
-and -go movement is not random over time. When
00:04:56.740 --> 00:04:59.839
the chemical concentration is improving, when
00:04:59.839 --> 00:05:02.920
the bacterium is moving toward more food, it
00:05:02.920 --> 00:05:07.740
suppresses tumbling. The runs grow longer. When
00:05:07.740 --> 00:05:10.360
concentration is worsening, tumbling frequency
00:05:10.360 --> 00:05:15.040
increases. This produces more reorientation and
00:05:15.040 --> 00:05:18.560
more searching. The bacterium is not measuring
00:05:18.560 --> 00:05:22.360
absolute concentration at any moment. It is comparing
00:05:22.360 --> 00:05:25.500
what it sensed a moment ago with what it senses
00:05:25.500 --> 00:05:29.470
now. This temporal comparison biases its future
00:05:29.470 --> 00:05:34.730
direction. It has no nervous system. It has no
00:05:34.730 --> 00:05:38.470
internal image of the world ahead. Yet its behavior
00:05:38.470 --> 00:05:41.529
is conditioned on a temporal gradient that points
00:05:41.529 --> 00:05:45.160
toward the future. This is prediction at its
00:05:45.160 --> 00:05:49.160
minimal edge. It is a minimal form of anticipatory
00:05:49.160 --> 00:05:52.430
behavior without a nervous system. And it is
00:05:52.430 --> 00:05:56.420
already enough to improve survival. Now consider
00:05:56.420 --> 00:06:00.329
something more complex. A frog's visual system
00:06:00.329 --> 00:06:02.610
does not passively register all light patterns
00:06:02.610 --> 00:06:06.009
equally. Researchers found that certain neurons
00:06:06.009 --> 00:06:09.149
in the frog's retina and optic tectum respond
00:06:09.149 --> 00:06:12.329
vigorously to small, dark, moving shapes resembling
00:06:12.329 --> 00:06:16.540
flies. They respond weakly or not at all to large
00:06:16.540 --> 00:06:19.600
objects, stationary objects, or diffuse light
00:06:19.600 --> 00:06:23.829
changes. The frog does not see and then decide
00:06:23.829 --> 00:06:27.329
what to pursue. Its perception is already structured
00:06:27.329 --> 00:06:29.730
by what evolution has predicted will matter.
00:06:30.220 --> 00:06:33.360
This selectivity is not learned by the individual
00:06:33.360 --> 00:06:37.410
frog. It is built into the visual system by evolutionary
00:06:37.410 --> 00:06:40.949
history. It is a prediction about what will matter
00:06:40.949 --> 00:06:45.569
for survival, encoded in neural hardware. When
00:06:45.569 --> 00:06:48.050
that particular pattern of light and motion appears,
00:06:48.509 --> 00:06:50.949
the behavioral response is already prepared.
00:06:51.379 --> 00:06:55.410
The visual system is, in effect, a bet about
00:06:55.410 --> 00:06:59.180
the future when this pattern appears act now
00:06:59.180 --> 00:07:03.019
because food is likely this is prediction embedded
00:07:03.019 --> 00:07:06.259
in hardware shaped by selection over millennia
00:07:06.259 --> 00:07:10.660
operating in milliseconds these biological cases
00:07:10.660 --> 00:07:14.360
establish something important prediction is not
00:07:14.360 --> 00:07:17.680
a luxury that arrives late in evolution after
00:07:17.680 --> 00:07:21.620
organisms have solved more basic problems It
00:07:21.620 --> 00:07:24.300
is woven into the earliest and simplest systems
00:07:24.300 --> 00:07:28.600
that maintain themselves against entropy. Homeostasis,
00:07:28.740 --> 00:07:32.019
the regulation of internal conditions, requires
00:07:32.019 --> 00:07:36.360
anticipating external change. A system that waits
00:07:36.360 --> 00:07:38.959
until temperature has already fallen will fail.
00:07:39.550 --> 00:07:42.089
One that begins responding when temperature is
00:07:42.089 --> 00:07:45.069
likely to fall, based on prior patterns, has
00:07:45.069 --> 00:07:49.079
a decisive advantage. But we need to be careful
00:07:49.079 --> 00:07:52.180
here. There is a temptation to call everything
00:07:52.180 --> 00:07:55.709
prediction. In doing so, we empty the concept
00:07:55.709 --> 00:07:59.740
of its explanatory power. A rock rolling downhill
00:07:59.740 --> 00:08:03.810
follows physical regularities. We might say,
00:08:03.949 --> 00:08:06.629
loosely, that it predicts where gravity will
00:08:06.629 --> 00:08:10.790
take it. But this is trivial. The rock has no
00:08:10.790 --> 00:08:14.180
internal states that represent the future. It
00:08:14.180 --> 00:08:17.519
does not update anything when its path deviates.
00:08:17.959 --> 00:08:22.149
It does not learn. The regularity is entirely
00:08:22.149 --> 00:08:26.149
external, imposed by physics, not generated by
00:08:26.149 --> 00:08:30.019
any model within the rock itself. This distinction
00:08:30.019 --> 00:08:33.870
matters enormously. The question is not whether
00:08:33.870 --> 00:08:37.500
the future can be anticipated in principle. Any
00:08:37.500 --> 00:08:41.179
deterministic system allows that. The question
00:08:41.179 --> 00:08:44.700
is whether the anticipation is generated by internal
00:08:44.700 --> 00:08:48.820
states that can be revised, that can be wrong.
00:08:49.299 --> 00:08:52.889
that can improve. The bacterium's tumble frequency
00:08:52.889 --> 00:08:57.240
can adapt. The frog's visual system can be modified
00:08:57.240 --> 00:09:01.440
by experience, at least at the margins. These
00:09:01.440 --> 00:09:04.100
systems carry something about the world inside
00:09:04.100 --> 00:09:07.779
them, that something can be more or less adequate
00:09:07.779 --> 00:09:12.059
to what actually occurs. So we need to distinguish
00:09:12.059 --> 00:09:15.820
prediction at different depths. At the shallowest
00:09:15.820 --> 00:09:18.940
level, a system extrapolates a pattern forward.
00:09:19.750 --> 00:09:24.250
a heartbeat, a day -night cycle, a repeated sound.
00:09:24.830 --> 00:09:28.669
These allow simple continuation. The system does
00:09:28.669 --> 00:09:30.990
not need to
understand why the pattern holds.
00:09:31.669 --> 00:09:36.399
It merely extends the line. This is useful but
00:09:36.399 --> 00:09:40.559
limited. The world is full of patterns that suddenly
00:09:40.559 --> 00:09:45.299
break. Shallow extrapolation fails catastrophically
00:09:45.299 --> 00:09:50.000
when they do. At a deeper level, A system infers
00:09:50.000 --> 00:09:53.600
hidden causes. It does not merely extend the
00:09:53.600 --> 00:09:57.009
surface pattern. It posits something behind the
00:09:57.009 --> 00:10:00.710
pattern that explains why it holds and when it
00:10:00.710 --> 00:10:04.419
might change. A creature that hears a particular
00:10:04.419 --> 00:10:07.580
rustle in leaves and infers a predator, rather
00:10:07.580 --> 00:10:09.940
than merely predicting that rustles will continue,
00:10:10.179 --> 00:10:14.120
is operating at this deeper level. It has a model
00:10:14.120 --> 00:10:16.460
of the world that includes entities and causes
00:10:16.460 --> 00:10:20.690
not directly present to sensation. This is cognitively
00:10:20.690 --> 00:10:24.549
more demanding and far more powerful. Deeper
00:10:24.549 --> 00:10:28.230
still is counterfactual prediction. This is the
00:10:28.230 --> 00:10:31.049
capacity to run possible futures that will not
00:10:31.049 --> 00:10:34.730
occur, to evaluate them, and to select actions
00:10:34.730 --> 00:10:38.470
that bring preferred outcomes. This is not merely
00:10:38.470 --> 00:10:42.009
expecting what will happen. It is imagining what
00:10:42.009 --> 00:10:44.389
could happen and choosing among possibilities.
00:10:45.960 --> 00:10:49.080
When a crow drops a stone into a tube to raise
00:10:49.080 --> 00:10:52.500
the water level and reach a floating treat, it
00:10:52.500 --> 00:10:56.580
is not predicting a single future. It is evaluating
00:10:56.580 --> 00:11:00.379
a possible future against others. It is selecting
00:11:00.379 --> 00:11:04.340
one that does not yet exist. This is prediction
00:11:04.340 --> 00:11:08.120
as the servant of decision. It begins to look
00:11:08.120 --> 00:11:10.519
like something we would readily call intelligence.
00:11:11.580 --> 00:11:15.419
The progression from shallow extrapolation to
00:11:15.419 --> 00:11:18.399
hidden cause inference to counterfactual evaluation
00:11:18.399 --> 00:11:21.500
gives us a way to think about what makes prediction
00:11:21.500 --> 00:11:25.039
cognitively interesting. But it also raises a
00:11:25.039 --> 00:11:28.850
question about mechanism. How do systems actually
00:11:28.850 --> 00:11:32.980
do this? What is happening inside them that allows
00:11:32.980 --> 00:11:37.039
these anticipatory capacities? One influential
00:11:37.039 --> 00:11:39.679
proposal comes from neuroscience and machine
00:11:39.679 --> 00:11:43.570
learning. It reframes perception itself as a
00:11:43.570 --> 00:11:47.600
predictive process. The idea, developed in various
00:11:47.600 --> 00:11:50.980
forms over recent decades,
is that the brain
00:11:50.980 --> 00:11:53.580
does not primarily build up world models from
00:11:53.580 --> 00:11:56.379
sensory input in a purely bottom -up fashion.
00:11:57.460 --> 00:12:00.600
Instead, it constantly generates top -down predictions
00:12:00.600 --> 00:12:03.120
about what sensory input should be arriving.
00:12:03.649 --> 00:12:06.629
It compares these predictions with the actual
00:12:06.629 --> 00:12:10.389
signals. It uses the difference, the prediction
00:12:10.389 --> 00:12:14.580
error, to update its internal models. This framework
00:12:14.580 --> 00:12:17.299
known as predictive processing or predictive
00:12:17.299 --> 00:12:20.519
coding has been formalized in particularly detailed
00:12:20.519 --> 00:12:26.370
ways. In 1999, Rajesh Rao and Dana Ballard proposed
00:12:26.370 --> 00:12:29.309
a functional interpretation of visual cortex
00:12:29.309 --> 00:12:33.340
organization in these terms. The visual system
00:12:33.340 --> 00:12:37.289
processes information hierarchically. Lower levels
00:12:37.289 --> 00:12:40.870
handle simple features. These include edges,
00:12:41.129 --> 00:12:44.809
local contrast, small patches of color and motion
00:12:44.809 --> 00:12:49.580
higher levels handle more complex patterns these
00:12:49.669 --> 00:12:54.279
include shapes objects faces scenes in predictive
00:12:54.279 --> 00:12:57.940
coding the higher levels send downward signals
00:12:57.940 --> 00:13:00.659
that predict what the lower levels should be
00:13:00.659 --> 00:13:04.190
reporting the lower levels do not send everything
00:13:04.190 --> 00:13:07.519
upward they send only the difference between
00:13:07.519 --> 00:13:11.230
the prediction and the actual input they send
00:13:11.230 --> 00:13:15.649
the error If the prediction matches, little signal
00:13:15.649 --> 00:13:19.419
travels up. If the prediction fails, a burst
00:13:19.419 --> 00:13:21.799
of error signal ascends, prompting revision.
00:13:22.590 --> 00:13:25.809
This bidirectional flow is more efficient than
00:13:25.809 --> 00:13:29.409
purely bottom -up processing. It also matches
00:13:29.409 --> 00:13:33.549
what neuroscientists observe. Some of the evidence
00:13:33.549 --> 00:13:35.830
comes from experiments using what researchers
00:13:35.830 --> 00:13:40.259
call the oddball paradigm. When a repeated sequence
00:13:40.259 --> 00:13:42.980
of identical stimuli is interrupted by something
00:13:42.980 --> 00:13:46.220
unexpected, the brain shows a characteristic
00:13:46.220 --> 00:13:50.909
amplified response. The expected stimulus generates
00:13:50.909 --> 00:13:54.509
a little activity. The unexpected one generates
00:13:54.509 --> 00:13:58.250
much more. This is consistent with the idea that
00:13:58.250 --> 00:14:00.950
the brain is not passively registering each stimulus
00:14:00.950 --> 00:14:04.450
afresh. It is constantly generating expectations
00:14:04.450 --> 00:14:09.139
and responding primarily to deviations. The experience
00:14:09.139 --> 00:14:12.679
of hearing an unexpected sound in a quiet room
00:14:12.679 --> 00:14:16.830
reflects this mechanism at the human scale. The
00:14:16.830 --> 00:14:20.649
silence holds a particular density. The sudden
00:14:20.649 --> 00:14:23.960
sharp crack breaks it. There is an immediate
00:14:23.960 --> 00:14:27.889
jolt through the shoulders. Attention redirects
00:14:27.889 --> 00:14:31.259
before conscious recognition. Your brain had
00:14:31.259 --> 00:14:34.279
predicted silence or a particular ambient pattern.
00:14:34.799 --> 00:14:37.720
The violation of that prediction produced not
00:14:37.720 --> 00:14:40.460
merely recognition, but a physiological response.
00:14:40.879 --> 00:14:45.049
The flinch before an anticipated impact. The
00:14:45.210 --> 00:14:47.610
surprise when a familiar object is not where
00:14:47.610 --> 00:14:50.850
expected. These everyday experiences suggest
00:14:50.850 --> 00:14:53.950
that our nervous systems are continuously generating
00:14:53.950 --> 00:14:57.919
futures and living with the errors. Predictive
00:14:57.919 --> 00:15:00.779
processing proposes that this is not a peripheral
00:15:00.779 --> 00:15:03.779
feature of perception, but its central organizing
00:15:03.779 --> 00:15:07.960
principle. Perception on this view is hypothesis
00:15:07.960 --> 00:15:11.519
testing. The brain generates a best guess about
00:15:11.519 --> 00:15:15.259
the causes of sensory input. Sensation is the
00:15:15.259 --> 00:15:18.080
feedback that confirms or corrects that guess.
00:15:18.940 --> 00:15:22.200
What we experience as seeing, hearing, touching
00:15:22.200 --> 00:15:26.519
is not the raw sensory data. It is the brain's
00:15:26.519 --> 00:15:29.340
constructed model, refined by prediction error.
00:15:29.600 --> 00:15:33.110
The world we perceive is, in this framework,
00:15:33.409 --> 00:15:36.799
a controlled hallucination. It is controlled
00:15:36.799 --> 00:15:40.860
because it is anchored by sensory input. It is
00:15:40.860 --> 00:15:43.259
a hallucination because it is always a construction,
00:15:43.600 --> 00:15:47.769
never a direct transmission. This is a powerful
00:15:47.769 --> 00:15:51.679
and counterintuitive idea. It suggests that much
00:15:51.679 --> 00:15:54.120
of what we take to be the world impressing itself
00:15:54.120 --> 00:15:57.440
upon a passive mind is actually the mind impressing
00:15:57.440 --> 00:16:01.220
its expectations upon the world. The mind then
00:16:01.220 --> 00:16:03.700
uses the world's resistance to correct itself.
00:16:04.909 --> 00:16:07.789
The philosopher Andy Clarke has synthesized and
00:16:07.789 --> 00:16:11.149
defended this framework influentially. He has
00:16:11.149 --> 00:16:13.850
argued that predictive processing unifies perception,
00:16:14.250 --> 00:16:16.809
action, and learning under a single principle.
00:16:17.149 --> 00:16:20.490
That principle is the minimization of prediction
00:16:20.490 --> 00:16:23.809
error.
But this is a philosophical synthesis
00:16:23.809 --> 00:16:27.009
of a collective scientific development, not the
00:16:27.009 --> 00:16:30.570
work of a single originator. The framework has
00:16:30.570 --> 00:16:35.250
been extended even further. Carl Friston a neuroscientist
00:16:35.250 --> 00:16:38.269
at University College London, has proposed what
00:16:38.269 --> 00:16:41.879
he calls the Free Energy Principle. He has also
00:16:41.879 --> 00:16:45.000
proposed its associated framework of active inference.
00:16:46.000 --> 00:16:49.100
The core claim is that all biological systems,
00:16:49.240 --> 00:16:52.519
in effect, minimize a quantity related to prediction
00:16:52.519 --> 00:16:56.389
error. This happens not just in perception, but
00:16:56.389 --> 00:16:59.840
in action as well. An organism does not merely
00:16:59.840 --> 00:17:02.279
passively update its models when predictions
00:17:02.279 --> 00:17:06.509
fail. It acts upon the world to make its predictions
00:17:06.509 --> 00:17:10.529
come true. You reach for the cup to confirm the
00:17:10.529 --> 00:17:13.849
prediction that you will grasp it. You move your
00:17:13.849 --> 00:17:16.829
eyes to sample the visual world in ways that
00:17:16.829 --> 00:17:20.960
confirm your expectations. Action on this view
00:17:20.960 --> 00:17:24.630
is not separate from prediction. It is itself
00:17:24.630 --> 00:17:28.420
a form of prediction fulfillment. This is an
00:17:28.420 --> 00:17:31.890
ambitious proposal. It has generated substantial
00:17:31.890 --> 00:17:35.789
interest and substantial debate. The mathematical
00:17:35.789 --> 00:17:39.970
formalism is elegant. The claim that perception
00:17:39.970 --> 00:17:42.730
and action are unified under a single principle
00:17:42.730 --> 00:17:46.240
is attractive, but we need to be clear about
00:17:46.240 --> 00:17:49.940
what is established and what is not. Predictive
00:17:49.940 --> 00:17:52.240
processing in the visual system has empirical
00:17:52.240 --> 00:17:56.220
support. The broader free energy principle, applied
00:17:56.220 --> 00:17:59.980
to all biological systems, remains contested.
00:18:00.869 --> 00:18:03.869
Some researchers argue that it is so abstract
00:18:03.869 --> 00:18:07.329
as to be difficult to test empirically. They
00:18:07.329 --> 00:18:10.390
argue that it describes too much to be falsifiable
00:18:10.390 --> 00:18:14.329
in any useful way. Others argue that while the
00:18:14.329 --> 00:18:16.650
framework captures something important about
00:18:16.650 --> 00:18:19.630
brain function, it is not yet clear that it can
00:18:19.630 --> 00:18:22.089
explain the full range of cognitive phenomena.
00:18:22.839 --> 00:18:26.640
These include emotional regulation, social understanding,
00:18:26.960 --> 00:18:31.240
and creative thought. The episode's seed is explicit
00:18:31.240 --> 00:18:34.460
on this point, and we should honor it. These
00:18:34.460 --> 00:18:38.180
are influential scientific proposals, not completed
00:18:38.180 --> 00:18:41.319
theories of occult mission. They deserve attention
00:18:41.319 --> 00:18:44.680
because they reframe familiar problems in productive
00:18:44.680 --> 00:18:48.690
ways. They do not deserve attention because they
00:18:48.690 --> 00:18:52.319
have settled every question. Now let us turn
00:18:52.319 --> 00:18:55.839
to artificial systems, where prediction has become
00:18:55.839 --> 00:18:59.930
perhaps the dominant training paradigm. Modern
00:18:59.930 --> 00:19:02.630
machine learning, and particularly the large
00:19:02.630 --> 00:19:05.329
language models that have attracted so much attention,
00:19:05.710 --> 00:19:09.339
are typically trained by prediction. The model
00:19:09.339 --> 00:19:11.720
receives a sequence of words and must predict
00:19:11.720 --> 00:19:14.960
what word comes next. When its prediction is
00:19:14.960 --> 00:19:17.640
wrong, the error is used to adjust its internal
00:19:17.640 --> 00:19:21.299
parameters. These are billions of numerical weights
00:19:21.299 --> 00:19:25.299
that encode patterns of association. Over billions
00:19:25.299 --> 00:19:29.250
of examples. This simple objective forces the
00:19:29.250 --> 00:19:32.269
model to learn grammar, facts about the world,
00:19:32.410 --> 00:19:34.789
and patterns of reasoning that help it predict
00:19:34.789 --> 00:19:38.250
better. Given enough data and enough parameters,
00:19:38.470 --> 00:19:41.009
this produces remarkably sophisticated behavior.
00:19:42.000 --> 00:19:45.900
The models can generate fluent text, answer questions,
00:19:46.339 --> 00:19:49.440
summarize documents, translate between languages,
00:19:49.759 --> 00:19:53.750
and even produce code that runs. This has led
00:19:53.750 --> 00:19:56.369
some to suggest that prediction at sufficient
00:19:56.369 --> 00:20:00.529
scale is enough for intelligence. The reasoning
00:20:00.529 --> 00:20:04.299
runs as follows. If predicting the next word
00:20:04.299 --> 00:20:08.140
requires understanding context, grammar, facts
00:20:08.140 --> 00:20:11.160
about the world, and even something like reasoning,
00:20:11.470 --> 00:20:14.670
then perhaps prediction has become understanding.
00:20:15.569 --> 00:20:18.289
Perhaps the distinction between predicting well
00:20:18.289 --> 00:20:20.829
and comprehending deeply collapses at scale.
00:20:20.990 --> 00:20:26.430
But this conclusion is premature. The seed directs
00:20:26.430 --> 00:20:30.480
us to examine why. Accuracy on a dataset does
00:20:30.480 --> 00:20:33.220
not guarantee reliable reasoning outside it.
00:20:33.319 --> 00:20:37.039
A model might fluently explain a physics concept
00:20:37.039 --> 00:20:40.079
but fail on a novel problem requiring physical
00:20:40.079 --> 00:20:43.910
intuition. It might generate a plausible -sounding
00:20:43.910 --> 00:20:47.579
citation that does not exist. It might confidently
00:20:47.579 --> 00:20:50.279
contradict itself without recognizing
the contradiction.
00:20:51.460 --> 00:20:54.039
The prediction objective has produced something
00:20:54.039 --> 00:20:57.259
genuinely impressive. But it has not produced
00:20:57.259 --> 00:21:00.119
systems that reliably know. when they are wrong.
00:21:00.549 --> 00:21:03.390
It has not produced systems that can evaluate
00:21:03.390 --> 00:21:07.470
counterfactuals in robust ways. It has not produced
00:21:07.470 --> 00:21:10.470
systems that can transfer understanding to situations
00:21:10.470 --> 00:21:14.269
far from their training distribution. This is
00:21:14.269 --> 00:21:17.859
not a failure of the researchers. It is a clue
00:21:17.859 --> 00:21:21.450
about the nature of prediction itself. Prediction
00:21:21.450 --> 00:21:24.009
-based training can learn statistical patterns
00:21:24.009 --> 00:21:26.690
that approximate understanding in domains where
00:21:26.690 --> 00:21:29.349
the training data is rich and the test situations
00:21:29.349 --> 00:21:33.250
are similar. But it does not, by itself, produce
00:21:33.250 --> 00:21:36.029
the kind of model -based reasoning that evaluates
00:21:36.029 --> 00:21:39.339
alternatives. It does not produce the kind of
00:21:39.339 --> 00:21:42.039
reasoning that recognizes surprise as a signal
00:21:42.039 --> 00:21:45.359
for fundamental revision. It does not produce
00:21:45.359 --> 00:21:48.059
the kind of reasoning that uses error to change
00:21:48.059 --> 00:21:50.559
not just the next prediction, but the entire
00:21:50.559 --> 00:21:54.509
framework of prediction. The language model does
00:21:54.509 --> 00:21:57.670
not, in any clear sense, have a model of the
00:21:57.670 --> 00:22:00.589
world that it can query and revise when predictions
00:22:00.589 --> 00:22:04.369
fail. It has weights adjusted to minimize prediction
00:22:04.369 --> 00:22:07.839
error on training data. This is not the same
00:22:07.839 --> 00:22:11.319
thing. This brings us to a critical distinction
00:22:11.319 --> 00:22:14.519
that the seed emphasizes and that we must not
00:22:14.519 --> 00:22:17.500
blur. There is a difference between prediction
00:22:17.500 --> 00:22:20.160
as statistical pattern matching and prediction
00:22:20.160 --> 00:22:23.119
as model -based anticipation with error correction
00:22:23.119 --> 00:22:27.319
and framework revision. The former can be impressive
00:22:27.319 --> 00:22:30.900
and useful. The latter is what makes prediction
00:22:30.900 --> 00:22:33.619
cognitively interesting in the way that matters
00:22:33.619 --> 00:22:37.259
for understanding intelligence. Consider again
00:22:37.259 --> 00:22:40.470
the fish evading the predator. Its prediction
00:22:40.470 --> 00:22:44.220
is not merely statistical. It is embedded in
00:22:44.220 --> 00:22:48.210
a model of how the world works. The model includes
00:22:48.210 --> 00:22:51.450
how pressure waves propagate, how predators move,
00:22:51.769 --> 00:22:55.740
how its own body responds. This model can be
00:22:55.740 --> 00:23:00.309
updated by experience. A fish that survives an
00:23:00.309 --> 00:23:02.650
encounter where its initial prediction was wrong
00:23:02.650 --> 00:23:06.960
does not merely adjust a weight. It may revise
00:23:06.960 --> 00:23:10.930
its model of what certain signals mean. The learning
00:23:10.930 --> 00:23:13.690
is not just about predicting better next time.
00:23:14.119 --> 00:23:18.759
It is about understanding differently. Now walk
00:23:18.759 --> 00:23:22.380
through the comparison carefully. The rock follows
00:23:22.380 --> 00:23:25.359
physical regularities with no internal states.
00:23:25.930 --> 00:23:29.930
no error signal, no memory, no revision. The
00:23:30.349 --> 00:23:32.829
bacterium has temporal comparison and adaptation.
00:23:33.650 --> 00:23:36.930
Its tumble frequency changes based on experience.
00:23:37.710 --> 00:23:41.430
But it has no rich internal model of the world.
00:23:42.190 --> 00:23:45.130
The fish has a nervous system that constructs
00:23:45.130 --> 00:23:47.750
a world model and revises it based on prediction
00:23:47.750 --> 00:23:51.700
error. It integrates multiple sensory modalities
00:23:51.700 --> 00:23:56.940
into a coherent picture. The Crow evaluates counterfactual
00:23:56.940 --> 00:24:00.900
futures and selects among them. It imagines possibilities
00:24:00.900 --> 00:24:05.660
that do not yet exist. Each step adds capacities
00:24:05.660 --> 00:24:09.769
that matter. These include error signals, memory,
00:24:10.049 --> 00:24:12.609
model revision, and counterfactual imagination.
00:24:13.589 --> 00:24:16.589
The mere word prediction does not capture these
00:24:16.589 --> 00:24:19.369
differences, unless we are careful to specify
00:24:19.369 --> 00:24:23.460
what kind of prediction we mean. This is why
00:24:23.460 --> 00:24:25.779
the thermostat remains an instructive comparison.
00:24:26.490 --> 00:24:30.190
A thermostat minimizes prediction error by switching
00:24:30.190 --> 00:24:33.589
heating on and off. It anticipates temperature
00:24:33.589 --> 00:24:37.269
deviation and acts to prevent it. But it does
00:24:37.269 --> 00:24:41.799
not learn. It does not model. It does not imagine
00:24:41.799 --> 00:24:46.220
alternatives. The comparison is useful for understanding
00:24:46.220 --> 00:24:50.220
the basic logic of control. But we must not let
00:24:50.220 --> 00:24:52.900
it collapse the distinction between control and
00:24:52.900 --> 00:24:58.210
cognition. So where does this leave us? Is prediction
00:24:58.210 --> 00:25:02.359
the core of intelligence? The honest answer is
00:25:02.359 --> 00:25:04.640
that prediction is a powerful common thread.
00:25:05.200 --> 00:25:08.200
But it becomes cognitively interesting only when
00:25:08.200 --> 00:25:11.079
a system learns models, evaluates alternatives
00:25:11.079 --> 00:25:15.470
and uses surprise to change what it does. Prediction
00:25:15.470 --> 00:25:19.190
alone is not sufficient. The kind of
prediction
00:25:19.190 --> 00:25:22.910
matters enormously, and even the richest forms
00:25:22.910 --> 00:25:25.890
of prediction may not exhaust what we mean by
00:25:25.890 --> 00:25:29.980
intelligence. Let us test this with a speculative
00:25:29.980 --> 00:25:33.849
case. Imagine an alien intelligence evolving
00:25:33.849 --> 00:25:36.950
in an environment that is completely stable over
00:25:36.950 --> 00:25:40.630
evolutionary time. The same resources appear
00:25:40.630 --> 00:25:44.549
in the same places. The same dangers follow the
00:25:44.549 --> 00:25:48.309
same patterns. The same actions produce the same
00:25:48.309 --> 00:25:52.380
results. Could such an intelligence thrive with
00:25:52.380 --> 00:25:55.740
little imagination, with minimal counterfactual
00:25:55.740 --> 00:25:59.319
capacity, with prediction reduced to highly accurate
00:25:59.319 --> 00:26:03.609
extrapolation of unchanging patterns? It seems
00:26:03.609 --> 00:26:07.339
plausible. If the environment never throws novel
00:26:07.339 --> 00:26:10.460
challenges, never requires combining information
00:26:10.460 --> 00:26:13.720
in new ways, never demands that the organism
00:26:13.720 --> 00:26:16.970
consider what is not currently the case, then
00:26:16.970 --> 00:26:18.990
the selective pressure for deep model -based
00:26:18.990 --> 00:26:22.740
prediction might be weak. The organism might
00:26:22.740 --> 00:26:26.059
become exquisitely tuned to its specific world.
00:26:26.359 --> 00:26:30.019
It might become like a highly optimized prediction
00:26:30.019 --> 00:26:33.779
machine for a narrow domain. It might do this
00:26:33.779 --> 00:26:37.160
without developing the flexible generative counterfactual
00:26:37.160 --> 00:26:40.119
capacities that we associate with richer intelligence.
00:26:41.500 --> 00:26:44.559
Conversely, a changing universe rewards deeper
00:26:44.559 --> 00:26:48.839
forecasts. When the environment shifts, when
00:26:48.839 --> 00:26:52.220
new competitors arrive, when resources fluctuate
00:26:52.220 --> 00:26:54.940
unpredictably, the organism that can imagine
00:26:54.940 --> 00:26:58.910
alternatives has a decisive advantage. The organism
00:26:58.910 --> 00:27:01.789
that can model hidden causes has a decisive advantage.
00:27:02.450 --> 00:27:05.049
The organism that can use surprise to revise
00:27:05.049 --> 00:27:07.630
its entire framework has a decisive advantage.
00:27:08.720 --> 00:27:11.880
The pressure is not merely to predict accurately.
00:27:12.160 --> 00:27:15.200
It is to predict in ways that can be revised.
00:27:16.140 --> 00:27:18.819
It is to predict in ways that carry information
00:27:18.819 --> 00:27:22.859
about uncertainty. It is to predict in ways that
00:27:22.859 --> 00:27:25.880
generate exploration as well as exploitation.
00:27:26.970 --> 00:27:29.829
This suggests that the role of prediction in
00:27:29.829 --> 00:27:32.589
intelligence is not a fixed fact about minds
00:27:32.589 --> 00:27:35.660
in general. It is shaped by the relationship
00:27:35.660 --> 00:27:39.480
between organisms and their environments. Prediction
00:27:39.480 --> 00:27:43.799
is always about a system and its world. The depth
00:27:43.799 --> 00:27:46.299
of prediction that evolves depends on the depth
00:27:46.299 --> 00:27:49.849
of challenge that the world presents. This does
00:27:49.849 --> 00:27:53.259
not make prediction any less central. It makes
00:27:53.259 --> 00:27:56.660
its centrality conditional and relational. We
00:27:57.170 --> 00:27:59.470
should also acknowledge what remains uncertain
00:27:59.470 --> 00:28:03.250
and contest it. The neural mechanisms implementing
00:28:03.250 --> 00:28:05.529
predictive coding are not fully characterized.
00:28:06.130 --> 00:28:09.000
The relationship between predictive processing
00:28:09.000 --> 00:28:13.349
and consciousness is unresolved. Some theorists
00:28:13.349 --> 00:28:15.450
find the framework suggestive for understanding
00:28:15.450 --> 00:28:19.890
conscious experience. Others find it silent on
00:28:19.890 --> 00:28:23.609
the questions that matter most. Whether prediction
00:28:23.609 --> 00:28:26.849
-based training in artificial systems can ever
00:28:26.849 --> 00:28:29.569
produce the kind of general reasoning that humans
00:28:29.569 --> 00:28:33.829
exhibit is an open empirical question, not a
00:28:33.829 --> 00:28:37.150
settled philosophical one. And whether the free
00:28:37.150 --> 00:28:39.789
energy principle will prove to be a genuinely
00:28:39.789 --> 00:28:43.470
unifying framework or an elegant formalism with
00:28:43.470 --> 00:28:46.250
limited empirical reach, is still being debated.
00:28:46.930 --> 00:28:49.670
There are alternative views that deserve mention,
00:28:50.009 --> 00:28:54.059
even if we cannot fully explore them. Some researchers
00:28:54.059 --> 00:28:56.980
emphasize embodied interaction with the environment
00:28:56.980 --> 00:29:00.690
as the foundation of cognition. They argue that
00:29:00.690 --> 00:29:03.250
prediction is secondary to the dynamics of coupling
00:29:03.250 --> 00:29:07.410
between organism and world. Others argue that
00:29:07.410 --> 00:29:10.509
memory, attention, and reasoning are equally
00:29:10.509 --> 00:29:13.190
fundamental capacities that cannot be reduced
00:29:13.190 --> 00:29:16.490
to prediction. Philosophers have questioned whether
00:29:16.490 --> 00:29:18.890
the predictive processing framework can account
00:29:18.890 --> 00:29:22.259
for the normative dimensions of cognition. These
00:29:22.259 --> 00:29:24.940
include our concern with truth, justification
00:29:24.940 --> 00:29:29.619
and reasons as such. These are live debates.
00:29:30.230 --> 00:29:33.349
A full account of intelligence would need to
00:29:33.349 --> 00:29:36.670
engage them. What we can say with confidence
00:29:36.670 --> 00:29:41.019
is this. Prediction is ancient. pervasive and
00:29:41.019 --> 00:29:45.099
powerful. It operates
in bacteria and in brains,
00:29:45.339 --> 00:29:48.680
in simple reflexes and in scientific imagination.
00:29:49.390 --> 00:29:52.630
It connects biological evolution to artificial
00:29:52.630 --> 00:29:56.619
training in ways that illuminate both. The frameworks
00:29:56.619 --> 00:29:59.140
that place prediction at the center of perception
00:29:59.140 --> 00:30:02.039
and action have transformed how we think about
00:30:02.039 --> 00:30:05.680
what minds do. But prediction is not a single
00:30:05.680 --> 00:30:10.640
thing. Not all prediction is intelligence. The
00:30:10.640 --> 00:30:13.319
rock follows regularities without modeling them.
00:30:13.619 --> 00:30:17.640
The thermostat controls without learning. The
00:30:17.640 --> 00:30:20.079
language model predicts fluently without reliably
00:30:20.079 --> 00:30:24.970
knowing. The fish, the crow, the human scientist.
00:30:25.640 --> 00:30:28.819
These systems predict in ways that use error
00:30:28.819 --> 00:30:32.829
to revise models. They predict in ways that evaluate
00:30:32.829 --> 00:30:36.630
alternatives. They predict in ways that transform
00:30:36.630 --> 00:30:40.730
surprise into understanding. The core of intelligence,
00:30:41.069 --> 00:30:44.150
if we must name one, may be not prediction itself,
00:30:44.390 --> 00:30:48.069
but the capacity to use prediction in the service
00:30:48.069 --> 00:30:51.930
of flexible, revisable, model -based engagement
00:30:51.930 --> 00:30:55.509
with a world that is not fully known. Prediction
00:30:55.509 --> 00:30:59.369
is the thread, but the weave is what matters.
00:31:00.109 --> 00:31:04.450
The thread alone is not the garment. And so we
00:31:04.450 --> 00:31:07.329
return to the fish in murky water, computing
00:31:07.329 --> 00:31:11.319
futures in milliseconds. It does not know that
00:31:11.319 --> 00:31:14.809
it predicts. It does not reflect on the nature
00:31:14.809 --> 00:31:19.349
of intelligence. But its survival, and the survival
00:31:19.349 --> 00:31:21.990
of all the more complex systems that evolved
00:31:21.990 --> 00:31:25.529
from such beginnings, depends on a capacity that
00:31:25.529 --> 00:31:29.170
has become, in us, something we can examine and
00:31:29.170 --> 00:31:32.690
question. We are prediction machines that have
00:31:32.690 --> 00:31:35.849
learned to predict our own predicting. We have
00:31:35.849 --> 00:31:38.250
learned to imagine that we might be otherwise.
00:31:39.309 --> 00:31:44.029
We have learned to ask what it means. That recursive
00:31:44.029 --> 00:31:48.529
capacity to turn prediction upon itself, to consider
00:31:48.529 --> 00:31:50.769
the counterfactual that prediction might not
00:31:50.769 --> 00:31:54.309
be enough, is perhaps the most cognitively interesting
00:31:54.309 --> 00:31:58.150
thing of all. It is not separate from prediction.
00:31:58.740 --> 00:32:02.039
It is prediction operating at a depth where the
00:32:02.039 --> 00:32:05.359
model includes the modeler, and the future includes
00:32:05.359 --> 00:32:07.180
the possibility of being wrong.