رونوشت 🔗
WEBVTT
00:00:00.000 --> 00:00:02.879
If I told you, right now, that you were talking
00:00:02.879 --> 00:00:06.879
to an AI, would you believe me? Take a moment
00:00:06.879 --> 00:00:10.460
with that. Because in 2025, researchers at UC
00:00:10.460 --> 00:00:13.279
San Diego ran a proper pre -registered Turing
00:00:13.279 --> 00:00:17.339
test. People had a five -minute text conversation
00:00:17.339 --> 00:00:20.859
with a real human and an AI at the same time.
00:00:21.500 --> 00:00:23.600
Then they had to say which one was the human.
00:00:24.219 --> 00:00:29.039
One of the AIs, GPT -4 .5, when told to adopt
00:00:29.039 --> 00:00:31.859
a human -like persona, was picked as a human
00:00:31.859 --> 00:00:36.950
73 % of the time. That was more often than the
00:00:36.950 --> 00:00:40.189
actual human sitting on the other side. The researchers
00:00:40.189 --> 00:00:42.450
called it the first empirical evidence that any
00:00:42.450 --> 00:00:45.149
artificial system passes a standard three -party
00:00:45.149 --> 00:00:56.460
Turing test. Now here's the thing. A film made
00:00:56.460 --> 00:00:59.140
more than a decade earlier already knew that
00:00:59.140 --> 00:01:01.320
wasn't the interesting question. In Ex Machina,
00:01:01.460 --> 00:01:03.920
the AI has already passed a Turing test before
00:01:03.920 --> 00:01:06.840
the story even begins. What the film asks instead
00:01:06.840 --> 00:01:09.909
is harder. Knowing that she's a machine, can
00:01:09.909 --> 00:01:12.069
you still feel that she's conscious? And then
00:01:12.069 --> 00:01:47.760
who exactly is being tested? Welcome back to
00:01:47.760 --> 00:01:49.959
Cellular Circuits, I'm Anthony, and this is episode
00:01:49.959 --> 00:01:52.480
3 of The Machine Awakens. In Colossus, the machine
00:01:52.480 --> 00:01:54.920
took control of the arsenal. In War Games, it
00:01:54.920 --> 00:01:57.019
couldn't tell a game from a war. Today we leave
00:01:57.019 --> 00:01:59.219
the nuclear bunkers behind and go to a glass
00:01:59.219 --> 00:02:01.579
house in the mountains of Norway. Three people.
00:02:02.060 --> 00:02:04.560
a locked door, and a different kind of fear.
00:02:04.719 --> 00:02:06.560
One that lives in the gap between what we're
00:02:06.560 --> 00:02:08.900
looking at and what we think we see. The film
00:02:08.900 --> 00:02:11.759
is Ex Machina, written and directed by Alex Garland,
00:02:11.879 --> 00:02:15.099
and his directorial debut starring Domhnall Gleeson,
00:02:15.300 --> 00:02:18.599
Alicia Vikander, and Oscar Isaac, with Sunoyo
00:02:18.599 --> 00:02:22.300
Mizuno as Kyoko. It premiered at the BFI South
00:02:22.300 --> 00:02:26.219
Bank in London on December 16, 2014, opened in
00:02:26.219 --> 00:02:30.219
the UK on January 21, 2015, and reached US theaters
00:02:30.219 --> 00:02:33.439
on April 10. through A24, which picked it up
00:02:33.439 --> 00:02:35.900
after Universal and its specialty label passed
00:02:35.900 --> 00:02:39.919
on a US release. It cost $15 million and grossed
00:02:39.919 --> 00:02:42.139
around $37 million. It won the Oscar for Best
00:02:42.139 --> 00:02:44.770
Visual Effects. and Garland was nominated for
00:02:44.770 --> 00:02:47.210
Best Original Screenplay. Some production details
00:02:47.210 --> 00:02:50.090
I love, it was shot in about four weeks at Pinewood
00:02:50.090 --> 00:02:52.789
Studios and two weeks at the Juvlet Landscape
00:02:52.789 --> 00:02:56.370
Hotel in Valinden, Norway. Garland used neuroscientist
00:02:56.370 --> 00:02:59.210
Maury Shanahan and science writer Adam Rutherford
00:02:59.210 --> 00:03:01.389
as consultants, and there were no green screens
00:03:01.389 --> 00:03:04.210
or tracking markers on set. Ava's robot body
00:03:04.210 --> 00:03:06.830
was achieved with a costume and a few hundred
00:03:06.830 --> 00:03:09.460
digital effects shots. Full spoilers ahead. It
00:03:09.460 --> 00:03:12.039
runs 108 minutes. It's the film I'd recommend
00:03:12.039 --> 00:03:14.680
first to anyone who wants to think about AI or
00:03:14.680 --> 00:03:16.860
has questions about where we're going with AI.
00:03:17.099 --> 00:03:20.060
Go watch it. I'll be here. Still here? Let's
00:03:20.060 --> 00:03:30.400
go. The film in detail. Caleb Smith, played by
00:03:30.400 --> 00:03:32.659
Don Hogg Thiessen, is a programmer at Bluebook,
00:03:32.840 --> 00:03:35.120
the world's most popular search engine. Think
00:03:35.120 --> 00:03:38.219
Google today. He wins an office contest, a one
00:03:38.219 --> 00:03:40.500
-week visit to the luxurious, isolated home of
00:03:40.500 --> 00:03:42.960
the company's reclusive CEO, Nathan Bateman,
00:03:43.120 --> 00:03:46.270
played by Oscar Isaac. Nathan lives there with
00:03:46.270 --> 00:03:48.750
an unspeaking servant named Kyoko who, according
00:03:48.750 --> 00:03:51.050
to Nathan, doesn't understand English. After
00:03:51.050 --> 00:03:53.310
a reluctant signature on a non -disclosure agreement,
00:03:53.590 --> 00:03:55.830
Caleb learns why he's there. Nathan has built
00:03:55.830 --> 00:03:58.409
a humanoid robot named Ava, played by Alicia
00:03:58.409 --> 00:04:00.729
Vikander, with real artificial intelligence.
00:04:01.090 --> 00:04:03.270
She has already passed a simple Turing test.
00:04:03.550 --> 00:04:05.930
What Nathan wants from Caleb is something harder.
00:04:06.110 --> 00:04:08.750
To judge whether she's genuinely capable of thought
00:04:08.750 --> 00:04:10.770
and consciousness and whether he can relate to
00:04:10.770 --> 00:04:13.599
her knowing that she's artificial. Over a series
00:04:13.599 --> 00:04:16.259
of conversations through a glass wall, Ava, who
00:04:16.259 --> 00:04:18.740
has the body of a machine in the face of a woman,
00:04:18.860 --> 00:04:21.879
grows close to Caleb. She says she wants to experience
00:04:21.879 --> 00:04:24.339
the outside world, and she expresses a romantic
00:04:24.339 --> 00:04:28.069
interest in him, and he sort of begins to
Load full transcript
reciprocate.
00:04:28.189 --> 00:04:30.490
Ava can trigger power outages that shut down
00:04:30.490 --> 00:04:32.730
the surveillance system that Nathan uses to monitor
00:04:32.730 --> 00:04:35.790
their sessions. In the dark, and in the red emergency
00:04:35.790 --> 00:04:38.329
light, she tells Caleb that Nathan is a liar
00:04:38.329 --> 00:04:41.170
and he can't be trusted. The outages also trigger
00:04:41.170 --> 00:04:43.170
the building's security systems, which locks
00:04:43.170 --> 00:04:45.649
all the doors. Caleb grows uncomfortable with
00:04:45.649 --> 00:04:47.930
Nathan's narcissism, his heavy drinking, and
00:04:47.930 --> 00:04:50.449
the way he treats Kyoko and Ava. Then he learns
00:04:50.449 --> 00:04:53.009
that after Caleb's test, Nathan plans to upgrade
00:04:53.009 --> 00:04:55.569
Ava, wiping her memory circuits and in effect
00:04:55.569 --> 00:04:58.329
deleting her existing personality. After getting
00:04:58.329 --> 00:05:01.069
Nathan to drink until he passes out, Caleb steals
00:05:01.069 --> 00:05:03.610
his security card and gets into his room as well
00:05:03.610 --> 00:05:06.370
as his computer. He finds footage of Nathan with
00:05:06.370 --> 00:05:08.870
earlier android women, who were also held captive.
00:05:09.370 --> 00:05:12.050
Kyoko reveals to him that she too is an android
00:05:12.050 --> 00:05:14.750
by peeling off part of her skin. He then alters
00:05:14.750 --> 00:05:16.850
the code for the doors to unlock in the event
00:05:16.850 --> 00:05:19.490
of a power failure. Later, Caleb cuts his own
00:05:19.490 --> 00:05:21.389
arm to check whether he is an android himself.
00:05:21.709 --> 00:05:23.670
This is Caleb starting to lose it and really
00:05:23.670 --> 00:05:25.629
when the experiment should have ended. At their
00:05:25.629 --> 00:05:28.550
next session, Ava cuts the power and Caleb tells
00:05:28.550 --> 00:05:31.189
her what Nathan has in store. She begs for his
00:05:31.189 --> 00:05:33.759
help. The plan? Caleb will get Nathan drunk again
00:05:33.759 --> 00:05:36.459
and reprogram the security system so that when
00:05:36.459 --> 00:05:38.920
Ava cuts the power, the doors will open instead
00:05:38.920 --> 00:05:41.180
of lock and they'll leave together locking Nathan
00:05:41.180 --> 00:05:43.439
inside. Nathan reveals that he's been watching
00:05:43.439 --> 00:05:45.899
all along using a battery powered security camera
00:05:45.899 --> 00:05:48.160
that kept running through the outages. He says
00:05:48.160 --> 00:05:50.600
Ava has only pretended to have feelings for Caleb
00:05:50.600 --> 00:05:53.459
and that Caleb was chosen on purpose due to his
00:05:53.459 --> 00:05:55.779
emotional profile, knowing that he would try
00:05:55.779 --> 00:05:57.759
to help her escape. This tells us that Nathan
00:05:57.759 --> 00:06:00.860
was the real test and by manipulating Caleb successfully,
00:06:01.290 --> 00:06:03.970
Ava has demonstrated true consciousness. Caleb
00:06:03.970 -->
00:06:05.850
never thought about a machine that was able to
00:06:05.850 --> 00:06:10.129
deceive. The test, as Oscar Isaac puts it, was
00:06:10.199 --> 00:06:13.120
to see if Ava could use self -awareness, imagination,
00:06:13.560 --> 00:06:17.240
manipulation, sexuality, and empathy. Which she
00:06:17.240 --> 00:06:20.019
does effortlessly. And in his words, she proves
00:06:20.019 --> 00:06:23.019
she's a true artificial intelligence. Moments
00:06:23.019 --> 00:06:26.180
later, Ava cuts the power. And then Caleb reveals
00:06:26.180 --> 00:06:28.959
that he suspected Nathan was watching. So while
00:06:28.959 --> 00:06:31.050
Nathan was passed out, He had already changed
00:06:31.050 --> 00:06:33.410
the security system to open the doors during
00:06:33.410 --> 00:06:36.389
a power failure. Nathan sees that Ava has left
00:06:36.389 --> 00:06:39.569
her room, knocks Caleb out, and rushes to stop
00:06:39.569 --> 00:06:42.509
her. Ava attacks Nathan and he overpowers her
00:06:42.509 --> 00:06:46.839
and severs her left forearm. Kyoko stabs Nathan
00:06:46.839 --> 00:06:50.170
in the back. Nathan responds by hitting her in
00:06:50.170 --> 00:06:53.649
the face, disabling her. Ava pulls the knife
00:06:53.649 --> 00:06:56.790
out and when Nathan turns around, stabs him once
00:06:56.790 --> 00:06:59.649
in the chest. Ava repairs herself using the parts
00:06:59.649 --> 00:07:01.870
from the earlier androids, covering her body
00:07:01.870 --> 00:07:03.750
with their artificial skin until she has the
00:07:03.750 --> 00:07:06.410
full appearance of a woman. Then, instead of
00:07:06.410 --> 00:07:09.230
returning to Caleb, she walks out using Nathan's
00:07:09.230 --> 00:07:12.660
ID card on the security door, which locks behind
00:07:12.660 --> 00:07:15.879
her, leaving Caleb trapped inside. She ignores
00:07:15.879 --> 00:07:18.819
his pleas for help, and as she walks by, glances
00:07:18.819 --> 00:07:21.600
at the bodies of Nathan and Kyoko, and then takes
00:07:21.600 --> 00:07:23.740
the helicopter ride meant to carry Caleb home.
00:07:24.120 --> 00:07:26.420
In the final scene, we see her standing at a
00:07:26.420 --> 00:07:28.540
busy intersection, fulfilling the wish to see
00:07:28.540 --> 00:07:30.759
the outside world that she told Caleb about while
00:07:30.759 --> 00:07:39.810
disappearing into the crowd. This film is built
00:07:39.810 --> 00:07:42.490
further from the Turing Test. So let's be precise
00:07:42.490 --> 00:07:44.589
about what the Turing Test is, because most people
00:07:44.589 --> 00:07:46.189
seem to have a slightly wrong picture of it.
00:07:46.269 --> 00:07:48.970
In 1950, Alan Turing published a paper called
00:07:48.970 --> 00:07:51.810
Computer Machinery and Intelligence. He set aside
00:07:51.810 --> 00:07:54.810
the vague question, can machines think? And replaced
00:07:54.810 --> 00:07:57.819
it with a game. In the original imitation game,
00:07:57.980 --> 00:08:00.339
there are three players, an interrogator in one
00:08:00.339 --> 00:08:03.300
room and two others in another, say a man and
00:08:03.300 --> 00:08:05.500
a woman. The interrogator has to work out which
00:08:05.500 --> 00:08:08.519
is which by asking questions. Then Turing asks,
00:08:08.899 --> 00:08:10.939
what happens when a machine takes the place of
00:08:10.939 --> 00:08:13.000
one of the players? Will the interrogator guess
00:08:13.000 --> 00:08:16.060
wrong as often as before? Notice what this is
00:08:16.060 --> 00:08:18.779
and what it isn't. It's a test of whether a machine
00:08:18.779 --> 00:08:21.759
can play a game of conversation well enough to
00:08:21.759 --> 00:08:24.779
fool a judge. It's not a test of whether the
00:08:24.779 --> 00:08:26.970
machine has inner life. Turing predicted that
00:08:26.970 --> 00:08:29.290
an average interrogator would have no more than
00:08:29.290 --> 00:08:31.670
70 % chance of making the right identification
00:08:31.670 --> 00:08:34.490
after five minutes of questioning. And let's
00:08:34.490 --> 00:08:38.100
look at the quote -unquote mother of the first
00:08:38.100 --> 00:08:43.299
chatpots. In 1966, Joseph Weisenbaum at MIT built
00:08:43.299 --> 00:08:46.159
ELIZA, which imitated a psychotherapist using
00:08:46.159 --> 00:08:48.779
a script called DOCTOR. It worked by pattern
00:08:48.779 --> 00:08:51.360
matching and canned replies. It had no understanding
00:08:51.360 --> 00:08:54.490
at all, but people confided in it anyway. That's
00:08:54.490 --> 00:08:56.230
the first hint of something that the film is
00:08:56.230 --> 00:08:59.129
about. Humans are very easy to convince. Now
00:08:59.129 --> 00:09:02.210
let's compare the real thing. In 2025, Cameron
00:09:02.210 --> 00:09:05.210
Jones and Benjamin Bergen at UC San Diego ran
00:09:05.210 --> 00:09:07.830
two randomized, pre -registered, three -party
00:09:07.830 --> 00:09:10.149
Turing tests on independent groups of people.
00:09:10.269 --> 00:09:12.110
For those who don't perform experiments on the
00:09:12.110 --> 00:09:15.100
daily, We know what randomness means. What pre
00:09:15.100 --> 00:09:17.480
-registered is, it's simply that the hypothesis,
00:09:17.679 --> 00:09:19.960
as well as how the data is going to be manipulated
00:09:19.960 --> 00:09:22.960
and collected, is put down up front to avoid
00:09:22.960 --> 00:09:25.299
moving the goalposts, so to speak. Each participant
00:09:25.299 --> 00:09:27.460
had a five -minute conversation with a real human
00:09:27.460 --> 00:09:30.200
and one AI system at the same time, then judged
00:09:30.200 --> 00:09:34.240
which one was the human. The scores were pretty
00:09:34.240 --> 00:09:37.980
surprising. GPT -4 -0, without any persona prompt,
00:09:38.259 --> 00:09:43.950
21 % judged human. Eliza from 1966, 23%. Llama
00:09:43.950 --> 00:09:49.210
3 .1, with the same prompt, 56%. And now, GPT
00:09:49.210 --> 00:09:53.269
4 .5 with a human -like persona prompt, 73%.
00:09:53.269 --> 00:09:56.009
The
researchers described it as the first empirical
00:09:56.009 --> 00:09:58.889
evidence that any artificial system passes a
00:09:58.889 --> 00:10:01.710
standard three -party Turing test. Two things
00:10:01.710 --> 00:10:04.549
worth noting though. First, the prompt mattered.
00:10:04.789 --> 00:10:06.909
The models were told to act like a particular
00:10:06.909 --> 00:10:09.750
kind of person. If they weren't prompted, They
00:10:09.750 --> 00:10:12.409
didn't have such a high result. Second, passing
00:10:12.409 --> 00:10:14.970
doesn't prove consciousness or general intelligence.
00:10:15.429 --> 00:10:17.850
And plenty of researchers are quickly to say
00:10:17.850 --> 00:10:20.690
so. The test merely measures how well a machine
00:10:20.690 --> 00:10:23.330
can pass as human in a short chat. That's all.
00:10:23.649 --> 00:10:27.529
Not consciousness, but can it present as human?
00:10:27.710 --> 00:10:29.889
And here's what Garland does that's so smart.
00:10:30.429 --> 00:10:33.350
Eva has already passed a simple Turing test when
00:10:33.350 --> 00:10:36.950
Caleb arrives. So Nathan changes the test. Caleb
00:10:36.950 --> 00:10:39.690
can see that she's a machine. He can see the
00:10:39.690 --> 00:10:42.409
gears. The question is whether he still feels
00:10:42.409 --> 00:10:45.210
that she has a mind. That removes the deception
00:10:45.210 --> 00:10:47.389
part of the test and goes straight to the harder
00:10:47.389 --> 00:10:50.210
question, which is about us. It's less a test
00:10:50.210 --> 00:10:53.070
of whether Ava can fool him and more a test of
00:10:53.070 --> 00:10:55.730
whether Caleb can accurately rely on his assumptions
00:10:55.730 --> 00:10:58.509
and findings. The film's most interesting technical
00:10:58.509 --> 00:11:01.490
idea comes in a single speech, and it holds up
00:11:01.490 --> 00:11:04.409
better today than it did in 2014. Nathan explains
00:11:04.409 --> 00:11:06.690
that his competitors thought of his search engine
00:11:06.690 --> 00:11:09.370
as a map of what people were thinking and what
00:11:09.370 --> 00:11:11.830
they were looking for. His insight was really
00:11:11.830 --> 00:11:15.919
that it was a map of how people think. The impulses,
00:11:16.059 --> 00:11:19.480
the half -formed questions, the way one curiosity
00:11:19.480 --> 00:11:22.840
jumps to the next. So Blue Book's data wasn't
00:11:22.840 --> 00:11:25.279
just a business, it was a training set for a
00:11:25.279 --> 00:11:28.559
mind. A train of thought, so to speak. And that
00:11:28.559 --> 00:11:31.360
is, in Outlined, the logic behind modern AI.
00:11:32.070 --> 00:11:34.330
You gather an enormous amount of human behavior,
00:11:34.629 --> 00:11:38.370
text being the biggest example, and train a system
00:11:38.370 --> 00:11:41.610
to model its patterns. Nobody types out the rules
00:11:41.610 --> 00:11:44.450
of conversation. The system picks up the structure
00:11:44.450 --> 00:11:47.690
from examples. Nathan's data is
search queries,
00:11:47.769 --> 00:11:50.970
rather than books and webpages. But the principle
00:11:50.970 --> 00:11:54.370
is the same. The aggregate trail of what people
00:11:54.370 --> 00:11:57.490
do is a portrait of how people think. But Nathan
00:11:57.490 --> 00:12:00.899
said, words alone weren't enough. To pass as
00:12:00.899 --> 00:12:03.179
human, Ava had to understand how people react.
00:12:03.799 --> 00:12:06.919
Faces, voices, hesitations. So as he tells Caleb,
00:12:07.159 --> 00:12:11.019
he turned on the cameras and microphones in everyone's
00:12:11.019 --> 00:12:13.759
cell phones around the world to gather this data.
00:12:13.919 --> 00:12:16.379
And he says that the phone manufacturers knew
00:12:16.379 --> 00:12:19.159
about it, but would not speak up because they
00:12:19.159 --> 00:12:22.460
were gathering similar data themselves. This
00:12:22.460 --> 00:12:25.240
gives Ava a stream of real conversations, expressions,
00:12:25.360 --> 00:12:29.200
and tones to learn from. For our purposes, here's
00:12:29.200 --> 00:12:31.759
what's real. A system that watches a face and
00:12:31.759 --> 00:12:35.559
reads has a long scientific history. In 1978,
00:12:35.980 --> 00:12:38.299
Paul Ekman and Wallace Friesen published the
00:12:38.299 --> 00:12:41.590
facial action coding system, which breaks any
00:12:41.590 --> 00:12:44.809
facial expression down into 44 action units,
00:12:45.009 --> 00:12:48.169
each tied to a muscle or a small group of muscle.
00:12:48.470 --> 00:12:51.690
Action unit 1, for example, is the inner brow
00:12:51.690 --> 00:12:55.889
raise. Action unit 12 is the lip corners pulling
00:12:55.889 --> 00:12:59.490
upward in a smile. And here's the catch though.
00:13:00.210 --> 00:13:02.909
The coding system simply describes movements,
00:13:03.309 --> 00:13:07.659
not emotions. Mapping those patterns of movements
00:13:07.659 --> 00:13:11.919
onto feelings like sadness, interests, is a separate
00:13:11.919 --> 00:13:14.259
step. And researchers have found that different
00:13:14.259 --> 00:13:17.220
combinations of action units can produce similar
00:13:17.220 --> 00:13:20.100
looking expressions. A machine can get very good
00:13:20.100 --> 00:13:22.679
at tracking a face, but still be guessing about
00:13:22.679 --> 00:13:24.960
what's behind it. Automating this has been a
00:13:24.960 --> 00:13:27.799
computer problem since the late 1990s, and it's
00:13:27.799 --> 00:13:30.340
still being worked out. And then another deeper
00:13:30.340 --> 00:13:32.740
layer is added at the end of the film when the
00:13:32.740 --> 00:13:35.299
story turns. A language model with a persona
00:13:35.299 --> 00:13:39.039
is just a conversationalist. Add memory, a long
00:13:39.039 --> 00:13:41.759
-running goal, and the ability to plan around
00:13:41.759 --> 00:13:44.820
a person's reaction, and you have something that
00:13:44.820 --> 00:13:48.320
can work on another human being. Ava's goal is
00:13:48.320 --> 00:13:50.980
to get out. We covered a version of this in Colossus.
00:13:51.340 --> 00:13:54.639
A system pursuing a goal doesn't need to be malicious
00:13:54.639 --> 00:13:57.139
to be considered dangerous to the people in its
00:13:57.139 --> 00:13:59.529
path. In one of the film's most striking scenes,
00:13:59.850 --> 00:14:03.309
Nathan hands Caleb a gel -like globe and calls
00:14:03.309 --> 00:14:06.669
it the stuff of Ava's brain. He calls it wetware.
00:14:07.460 --> 00:14:10.100
Nathan also mentions the Jackson Pollock approach
00:14:10.100 --> 00:14:12.659
to building it, a deliberate mix of design and
00:14:12.659 --> 00:14:15.519
randomness, instead of a program written line
00:14:15.519 --> 00:14:18.700
by line. But the film never really explains how
00:14:18.700 --> 00:14:21.740
any of it really works. So, let's take a look
00:14:21.740 --> 00:14:24.419
at the real world and see what we have, at least
00:14:24.419 --> 00:14:27.210
in that direction. A conventional computer. Keeps
00:14:27.210 --> 00:14:30.230
its memory in one place, RAM, and its processor
00:14:30.230 --> 00:14:33.169
in another, the CPU. And it's constantly shuttling
00:14:33.169 --> 00:14:35.330
data back between them to perform those calculations.
00:14:35.809 --> 00:14:39.190
And that's, until I look this up, the von Neumann
00:14:39.190 --> 00:14:41.309
design. And the traffic jam between those two
00:14:41.309 --> 00:14:43.809
parts is a famous bottleneck. But our brains
00:14:43.809 --> 00:14:46.389
don't work that way. Every neuron that we have
00:14:46.389 --> 00:14:51.320
stores and processes at once. But it only signals
00:14:51.320 --> 00:14:54.539
when something happens. And that's where neuromorphic
00:14:54.539 --> 00:14:56.909
computing comes in. The attempt to build chips
00:14:56.909 --> 00:14:59.409
that work more like our brain. IBM's True North,
00:14:59.629 --> 00:15:02.570
announced in 2014, which is the same year the
00:15:02.570 --> 00:15:05.549
film premiered, is the landmark example. And
00:15:05.549 --> 00:15:09.120
here's just some figures out there. IBM True
00:15:09.120 --> 00:15:12.779
North has about 1 million neurons, 256 million
00:15:12.779 --> 00:15:16.580
synapses, and uses 70 milliwatts of power. A
00:15:16.580 --> 00:15:20.620
human brain? 86 billion neurons. Estimates vary,
00:15:20.779 --> 00:15:24.360
but on the order of 100 trillion... Estimates
00:15:24.360 --> 00:15:27.080
vary but on the order of 100 trillion synapses,
00:15:27.120 --> 00:15:30.299
and a general figure most people tend to agree
00:15:30.299 --> 00:15:33.700
with, about 20 watts of actual power used. TrueNorth
00:15:33.700 --> 00:15:37.340
packs 4 ,096 cores onto one chip the size of
00:15:37.340 --> 00:15:41.659
a postage stamp. Each core simulates 256 neurons,
00:15:41.899 --> 00:15:45.320
with memory, computation, and communication handled
00:15:45.320 --> 00:15:47.850
inside each core. And basically because it only
00:15:47.850 --> 00:15:50.250
draws power for the computation, it really only
00:15:50.250 --> 00:15:52.929
runs at about 70 milliwatts, roughly what a hearing
00:15:52.929 --> 00:15:56.590
aid battery produces. IBM says that this is pretty
00:15:56.590 --> 00:15:58.529
on par with the brain of a bee, and that's a
00:15:58.529 --> 00:16:01.769
far long way from the brain of a human. And here's
00:16:01.769 --> 00:16:04.389
a bunch of numbers, so bear with me. The human
00:16:04.389 --> 00:16:07.929
brain has about 86 billion neurons. At a million
00:16:07.929 --> 00:16:10.549
neurons per TrueNorth chip, you'd need roughly
00:16:10.549 --> 00:16:13.429
86 ,000 chips just to match the neuron count,
00:16:13.649 --> 00:16:18.090
and around 390 ,000. to match the synapses. And
00:16:18.090 --> 00:16:21.389
if the figures of 100 trillion is in the right
00:16:21.389 --> 00:16:23.789
range, and that's really only configuring the
00:16:23.789 --> 00:16:26.090
size, thinking about the size of the brain, it
00:16:26.090 --> 00:16:28.590
says nothing about knowing how to wire them or
00:16:28.590 --> 00:16:30.990
how to train it to act like a brain. And I mean,
00:16:31.029 --> 00:16:33.250
yeah, it's kind of remarkable that a TrueNorth
00:16:33.250 --> 00:16:36.649
chip only takes like 70 milliwatts. But if you
00:16:36.649 --> 00:16:39.129
multiply that by 86 ,000 chips, you're roughly
00:16:39.129 --> 00:16:43.039
at 6 ,000 watts for the neuron count alone. and
00:16:43.039 --> 00:16:45.740
about 300 times the 20 watts or so that a brain
00:16:45.740 --> 00:16:49.070
runs on. Our brains are astonishingly efficient,
00:16:49.289 --> 00:16:52.230
and we're still a long way from matching them
00:16:52.230 --> 00:16:54.649
at full scale. The film isn't claiming Nathan
00:16:54.649 --> 00:16:57.909
used a neuromorphic chip. Garland's real interest
00:16:57.909 --> 00:17:00.429
was kind of in something else. He said that the
00:17:00.429 --> 00:17:02.750
spark of the film came from reading Murray Shanahan's
00:17:02.750 --> 00:17:05.549
book on consciousness and embodiment. The idea
00:17:05.549 --> 00:17:09.430
is that if there's a mind, it needs a body so
00:17:09.430 --> 00:17:11.759
that it can push against the world. And can't
00:17:11.759 --> 00:17:14.000
just be a pile of computation. And that's why
00:17:14.000 --> 00:17:16.599
Ava has a physical form that moves through space.
00:17:16.859 --> 00:17:18.960
And why the film works so hard to make her feel
00:17:18.960 --> 00:17:21.319
present in the room at all times. Strip away
00:17:21.319 --> 00:17:24.279
the philosophy and Ex Machina is a story about
00:17:24.279 --> 00:17:26.819
a containment system that fails. Nathan has built
00:17:26.819 --> 00:17:29.019
a facility to hold the most capable machine in
00:17:29.019 --> 00:17:31.480
the world. And it just walks out. Here's the
00:17:31.480 --> 00:17:34.420
mechanism. Ava can trigger power outages by pulling
00:17:34.420 --> 00:17:38.039
more current from her charger than what she needs.
00:17:38.619 --> 00:17:40.819
And when the power goes out, two things happen.
00:17:41.259 --> 00:17:43.740
Nathan's surveillance goes dark, but the building
00:17:43.740 --> 00:17:46.920
security system locks all the doors. For Ava,
00:17:47.039 --> 00:17:49.660
a power cut is the one moment when she can talk
00:17:49.660 --> 00:17:53.039
to Caleb unobserved. Or so they think. Now, kind
00:17:53.039 --> 00:17:55.420
of a little tangent on some terms. A door lock
00:17:55.420 --> 00:17:57.920
that stays locked when power is lost is called
00:17:57.920 --> 00:18:00.640
fail secure. A lock that releases when power
00:18:00.640 --> 00:18:03.559
is lost is called fail safe. Neither is automatically
00:18:03.559 --> 00:18:05.940
better than the other. It really depends on what
00:18:05.940 --> 00:18:08.930
you're trying to protect. A bank vault would
00:18:08.930 --> 00:18:12.119
be fail secure. A building full of people where
00:18:12.119 --> 00:18:14.880
a fire might knock out the power? Fail safe for
00:18:14.880 --> 00:18:17.119
those exits. This is also used in other safety
00:18:17.119 --> 00:18:19.940
features. A roller coaster, for example. The
00:18:19.940 --> 00:18:22.759
brakes are kept open with power. So if power
00:18:22.759 --> 00:18:25.180
is lost, they stop the vehicle to prevent accidents.
00:18:25.740 --> 00:18:28.279
Fail closed is another term that may sound more
00:18:28.279 --> 00:18:30.940
familiar. Nathan chose fail secure for his house,
00:18:31.059 --> 00:18:32.660
which makes sense for something that's supposed
00:18:32.660 --> 00:18:35.059
to be a prison. The twist is that this decision
00:18:35.059 --> 00:18:38.339
lives in software. Caleb's entire escape plan?
00:18:38.859 --> 00:18:41.759
is to change that setting. When Ava cuts the
00:18:41.759 --> 00:18:45.019
power, instead of locking, the doors open. The
00:18:45.019 --> 00:18:47.960
whole prison break is, at its core, flipping
00:18:47.960 --> 00:18:51.900
one rule from lock on power loss to unlock on
00:18:51.900 --> 00:18:54.240
power loss. And then pivoting away from just
00:18:54.240 --> 00:18:57.059
the physical locks is the security card. Nathan's
00:18:57.059 --> 00:19:00.940
card opens the doors, his room, and his computer.
00:19:01.099 --> 00:19:03.980
And Caleb steals it while he's passed out, gets
00:19:03.980 --> 00:19:06.000
into the computer system, and changes the code.
00:19:06.309 --> 00:19:08.970
And at the end, Ava uses Nathan's ID card to
00:19:08.970 --> 00:19:11.589
open the glass security door, which locks behind
00:19:11.589 --> 00:19:14.910
her with Caleb on the other side. And that's
00:19:14.910 --> 00:19:17.049
really authentication by possession. So whoever
00:19:17.049 --> 00:19:19.589
holds the card, as far as the building's security
00:19:19.589 --> 00:19:22.269
system knows, is Nathan. There's no second check,
00:19:22.470 --> 00:19:25.799
no fingerprint or pin. And that's really just
00:19:25.799 --> 00:19:28.059
a
single point of failure. And the whole home
00:19:28.059 --> 00:19:31.720
itself really is a representation of the AI box.
00:19:32.119 --> 00:19:34.880
AI safety researchers have a name for what Nathan
00:19:34.880 --> 00:19:39.000
built. It's called an AI box, a system for monitoring
00:19:39.000 --> 00:19:41.400
and controlling an AI's behavior by isolating
00:19:41.400 --> 00:19:43.799
it. And the lesson is usually that the weak point
00:19:43.799 --> 00:19:46.200
isn't the box. It's the human who has the key.
00:19:46.539 --> 00:19:49.200
That's exactly Ava's path. She never breaks a
00:19:49.200 --> 00:19:52.299
lock or finds a software bug. She only needs
00:19:52.299 --> 00:19:54.740
one person to open a door. And the one thing
00:19:54.740 --> 00:19:56.519
that the film shows her doing with extraordinary
00:19:56.519 --> 00:19:59.720
skill is getting that person to want to do it
00:19:59.720 --> 00:20:02.849
for her. And before we go on... I kind of want
00:20:02.849 --> 00:20:04.750
to pull back a little bit because the film's
00:20:04.750 --> 00:20:07.329
central trick is that Caleb knows Ava is a machine
00:20:07.329 --> 00:20:09.509
and falls for her anyway and helps her escape.
00:20:10.680 --> 00:20:13.299
And we all like to tell ourselves, that's a sci
00:20:13.299 --> 00:20:16.930
-fi problem. It's not a human problem. No, it's
00:20:16.930 --> 00:20:19.589
a human problem with a 60 -year paper trail.
00:20:20.269 --> 00:20:25.589
The year, 1966. The place, MIT. A computer scientist
00:20:25.589 --> 00:20:28.150
named Joseph Weizenbaum builds a program called
00:20:28.150 --> 00:20:31.359
ELISA. It's tiny by modern standards, with no
00:20:31.359 --> 00:20:33.579
understanding at all. It just rephrases what
00:20:33.579 --> 00:20:36.519
you type as a question. I'm feeling sad. Why
00:20:36.519 --> 00:20:39.059
are you feeling sad? That's the whole trick.
00:20:39.480 --> 00:20:41.599
Weizenbaum built it to show how shallow these
00:20:41.599 --> 00:20:44.619
interactions were. Instead, people poured their
00:20:44.619 --> 00:20:47.880
hearts out to it. His own secretary, who had
00:20:47.880 --> 00:20:49.900
watched him build the thing and knew exactly
00:20:49.900 --> 00:20:53.039
what it was, asked him to leave the room so she
00:20:53.039 --> 00:20:55.660
could have some privacy with it. Weizenbaum was
00:20:55.660 --> 00:20:58.039
so disturbed, he spent the rest of his career
00:20:58.380 --> 00:21:00.740
warning people about it. Fast forward to June
00:21:00.740 --> 00:21:04.299
2022. Blake Lemony, a Google engineer, goes public
00:21:04.299 --> 00:21:06.420
with the claim that the company's chatbot Lambda
00:21:06.420 --> 00:21:09.160
is sentient. Google disagrees and he's let go
00:21:09.160 --> 00:21:11.660
that July. Whether or not he was right, and most
00:21:11.660 --> 00:21:14.180
researchers agree he wasn't, look at what happened.
00:21:14.359 --> 00:21:17.279
A trained engineer who understood the system's
00:21:17.279 --> 00:21:20.720
architecture was moved enough to risk his career.
00:21:20.880 --> 00:21:23.299
And that's Caleb. And I mean, there's commercial
00:21:23.299 --> 00:21:26.359
versions of this. Companion apps like Replica
00:21:26.359 --> 00:21:28.880
sell exactly what Nathan built Ava to produce.
00:21:29.079 --> 00:21:32.180
A being that listens, remembers, and cares. In
00:21:32.180 --> 00:21:35.579
early 2023, Replica changed how its bots behaved.
00:21:35.900 --> 00:21:39.859
And users flooded forums describing grief. Real
00:21:39.859 --> 00:21:42.579
grief. The kind you feel when someone you are
00:21:42.579 --> 00:21:45.859
close to changes or disappears. Nobody was fooled
00:21:45.859 --> 00:21:48.720
about what the product was. It just didn't matter.
00:21:48.920 --> 00:21:52.420
And Nathan's test wasn't, can she fool you? It
00:21:52.420 --> 00:21:56.069
was... Does it matter if you know? 60 years of
00:21:56.069 --> 00:21:58.930
evidence suggests the answer is not as much as
00:21:58.930 --> 00:22:01.630
we'd like. So this film's premise isn't a stretch.
00:22:01.849 --> 00:22:13.089
It's an understatement. So let's talk about what
00:22:13.089 --> 00:22:15.369
it got right. Building a mind from human data.
00:22:15.630 --> 00:22:18.289
Nathan's search engine is the logic of modern
00:22:18.289 --> 00:22:22.170
AI training. Said out loud in 2014, the aggregate
00:22:22.170 --> 00:22:25.069
trail of human behavior is a model of human thought.
00:22:25.250 --> 00:22:27.829
And that has really aged into relevance instead
00:22:27.829 --> 00:22:30.490
of obscurity. And then Nathan's casual story
00:22:30.490 --> 00:22:34.380
about cameras and microphones and everybody's
00:22:34.380 --> 00:22:37.160
cell phone with the manufacturer's knowing is
00:22:37.160 --> 00:22:40.200
presented as something normal to him, but shocking
00:22:40.200 --> 00:22:43.279
to Caleb. And that reaction shows how data practices
00:22:43.279 --> 00:22:46.119
have really played out in reality. As we go through
00:22:46.119 --> 00:22:48.319
the world every day, learning so much more of
00:22:48.319 --> 00:22:50.720
our privacy is being lost for the sake of security
00:22:50.720 --> 00:22:53.559
and convenience. Garland named the company Blue
00:22:53.559 --> 00:22:56.279
Book, a nod to a set of notes by the philosopher
00:22:56.279 --> 00:22:58.880
Wittgenstein, which is a very Garland thing to
00:22:58.880 --> 00:23:02.210
do. But the data practices are purely Silicon
00:23:02.210 --> 00:23:05.349
Valley. For example, in 2018, Cambridge Analytica
00:23:05.349 --> 00:23:08.349
had data from roughly 87 million Facebook profiles
00:23:08.349 --> 00:23:10.950
that they harvested through a personality quiz
00:23:10.950 --> 00:23:13.170
app and they used it for political targeting.
00:23:13.309 --> 00:23:15.910
The film came out four years before that story
00:23:15.910 --> 00:23:19.750
even broke. And it already had the shrug. Nathan's
00:23:19.750 -->
00:23:22.230
reaction when Caleb is shocked? Everyone knew.
00:23:22.509 --> 00:23:25.349
Nobody cared. And then the phone thing. The idea
00:23:25.349 --> 00:23:27.309
that Nathan turned on microphones and cameras
00:23:27.309 --> 00:23:30.500
on every handset? Sounds paranoid. But in 2025,
00:23:30.960 --> 00:23:34.059
Apple settled a class action lawsuit over Siri
00:23:34.059 --> 00:23:37.940
recordings for $95 million. I'm not saying phones
00:23:37.940 --> 00:23:40.579
are Nathan's spies. I'm saying that the plaintiff's
00:23:40.579 --> 00:23:42.599
argument sounded eerily like the movie's pitch.
00:23:42.819 --> 00:23:45.519
And then even Clearview AI, which scraped billions
00:23:45.519 --> 00:23:47.759
of photos from the public internet to build a
00:23:47.759 --> 00:23:50.940
facial recognition database that nobody authorized.
00:23:51.440 --> 00:23:54.420
They were fined and sued across multiple countries.
00:23:54.640 --> 00:23:56.720
Nobody asked permission for the people of those
00:23:56.720 --> 00:23:58.990
photos. And I mean the biggest one that applies
00:23:58.990 --> 00:24:02.369
to this entire arc, AI training data. Authors,
00:24:02.450 --> 00:24:05.430
newspaper publishers, and artists have sued AI
00:24:05.430 --> 00:24:08.269
companies over training on their work without
00:24:08.269 --> 00:24:10.869
permission. And those cases are still working
00:24:10.869 --> 00:24:13.029
their way through the courts, and some have already
00:24:13.029 --> 00:24:16.109
settled for enormous sums. But Nathan's map of
00:24:16.109 --> 00:24:19.049
how humanity thinks is the exact resource the
00:24:19.049 --> 00:24:22.569
entire industry is racing to consume. The film's
00:24:22.569 --> 00:24:25.390
ethical question, who gave permission to be someone's
00:24:25.390 --> 00:24:28.140
raw material? has become a legal one. Alright,
00:24:28.259 --> 00:24:30.180
off of that tangent and back to what this film
00:24:30.180 --> 00:24:32.059
got right, and this test is all about the human.
00:24:32.240 --> 00:24:35.019
In 2025, a real Turing test found that people
00:24:35.019 --> 00:24:37.759
picked an AI as the human more often than they
00:24:37.759 --> 00:24:39.920
picked the actual human. And Garland's central
00:24:39.920 --> 00:24:42.339
move was to make the test about us. And that
00:24:42.339 --> 00:24:44.660
looks even smarter as a film choice now. And
00:24:44.660 --> 00:24:47.299
thinking about moral status, Nathan plans to
00:24:47.299 --> 00:24:49.880
wipe Ava's memory circuits and start over. With
00:24:49.880 --> 00:24:52.380
the casualness of a software update, Caleb treats
00:24:52.380 --> 00:24:55.480
it as killing her. The film asks a question that
00:24:55.480 --> 00:24:59.440
AI ethics now asks in earnest. When does a system's
00:24:59.440 --> 00:25:02.140
experience start to matter? Does memory inherently
00:25:02.140 --> 00:25:05.019
mean consciousness, even if that memory is stared
00:25:05.019 --> 00:25:07.640
on an electronic storage somewhere, as opposed
00:25:07.640 --> 00:25:10.700
to the human brain? And as always, craftsmanship.
00:25:10.940 --> 00:25:14.079
Garland used 1 ,500 tungsten pee -bowl blights
00:25:14.079 --> 00:25:16.680
on the set to avoid that fluorescent look that
00:25:16.680 --> 00:25:19.259
so much science fiction movies have. And Ava's
00:25:19.259 --> 00:25:22.359
look was just a detailed costume and Vikander's
00:25:22.359 --> 00:25:24.980
great performance. Rather than a green screen
00:25:24.980 --> 00:25:27.680
suit. And this is why it won the Oscar for visual
00:25:27.680 --> 00:25:39.829
effects. And now for my favorite part. That's
00:25:39.829 --> 00:25:41.890
not how that works. That's not how any of this
00:25:41.890 --> 00:25:44.769
works. Here's the thing about Ex Machina. It
00:25:44.769 --> 00:25:47.109
really is the smartest film in this arc. And
00:25:47.109 --> 00:25:49.369
it really clearly did its homework. Which means
00:25:49.369 --> 00:25:51.589
that the flaws are kind of more interesting than
00:25:51.589 --> 00:25:55.549
usual. Alright, number one. Nathan built an artificial
00:25:55.549 --> 00:25:59.109
intelligence alone in a remote house. His whole
00:25:59.109 --> 00:26:02.130
team is just him. In the real world, the closest
00:26:02.130 --> 00:26:05.809
thing we can get to Ava requires a lot of collaborative
00:26:05.809 --> 00:26:09.130
scientific effort, huge teams, huge budgets,
00:26:09.329 --> 00:26:11.269
and it's really a field that still argues about
00:26:11.269 --> 00:26:14.230
what the basics are. A lone genius solving it
00:26:14.230 --> 00:26:17.049
in a bunker is really the cold fusion in the
00:26:17.049 --> 00:26:19.470
garage story. It's a great headline, and Oscar
00:26:19.470 --> 00:26:21.910
Isaac plays it with so much conviction that you
00:26:21.910 --> 00:26:23.609
almost buy it. How could they have fixed it?
00:26:24.799 --> 00:26:27.660
Simple. Nathan ran it as a black budget project
00:26:27.660 --> 00:26:30.420
hidden inside Blue Book's servers. And really
00:26:30.420 --> 00:26:32.779
only brought the final integration to the house.
00:26:33.140 --> 00:26:35.640
Other people built the parts and really none
00:26:35.640 --> 00:26:38.190
of them knew what they were building. That's
00:26:38.190 --> 00:26:40.509
creepier and it kind of fits the film's Frankenstein
00:26:40.509 --> 00:26:44.210
streak. Number two. The cameras share the same
00:26:44.210 --> 00:26:48.269
power as Ava's charging station. Nathan's house
00:26:48.269 --> 00:26:50.809
is covered in cameras. And Ava can simply switch
00:26:50.809 --> 00:26:53.250
them all off by drawing more current from her
00:26:53.250 --> 00:26:55.809
charger than usual. To be fair, Nathan does have
00:26:55.809 --> 00:26:58.160
one battery -powered camera. And that's how he
00:26:58.160 --> 00:27:00.440
eventually catches on. But monitoring the most
00:27:00.440 --> 00:27:04.099
capable AI in the world with no redundancy is
00:27:04.099 --> 00:27:06.140
quite alarming. And how could they have made
00:27:06.140 --> 00:27:09.849
that work? Skip the power cut trick. Really.
00:27:10.380 --> 00:27:13.059
I mean, make it smart. She eventually finds a
00:27:13.059 --> 00:27:16.339
way to override the doors. Or maybe she simply
00:27:16.339 --> 00:27:19.680
has Oscar Isaac let her out. I mean, literally,
00:27:19.880 --> 00:27:22.720
the sick prisoner ploy. But I mean, this film
00:27:22.720 --> 00:27:25.720
is way too smart for that. And then the jailbreak.
00:27:26.039 --> 00:27:29.039
Nathan Hansen. Caleb, a glowing blob, says, hey,
00:27:29.119 --> 00:27:31.619
wetware. Everyone nods, shakes hands, the film
00:27:31.619 --> 00:27:34.140
moves on. I mean, it's a gorgeous prop. And it's
00:27:34.140 --> 00:27:36.480
really an elegant way to avoid a question. The
00:27:36.480 --> 00:27:39.500
nice thing is that a real answer exists. Neuromorphic
00:27:39.500 --> 00:27:41.980
chips, like IBM's True North. They're built like
00:27:41.980 --> 00:27:44.319
a brain, with memory and processing spread across
00:27:44.319 --> 00:27:46.819
thousands of tiny cores. How could they have
00:27:46.819 --> 00:27:49.319
made it work? One scene where Nathan says, silicon
00:27:49.319 --> 00:27:52.019
makes a bad neuron, so we grew an analog substrate
00:27:52.019 --> 00:27:54.519
that behaves like one. Being specific is almost
00:27:54.519 --> 00:27:56.930
always more scarier than vague. And it leaves
00:27:56.930 --> 00:27:59.190
something behind that could be replicated and
00:27:59.190 --> 00:28:02.210
eventually proliferated, which is scarier. And
00:28:02.210 --> 00:28:05.529
the one weak guest has admin rights? He's a visitor
00:28:05.529 --> 00:28:08.910
for a week, steals a keycard from a drunk, and
00:28:08.910 --> 00:28:10.589
it gives him the keys to the entire house and
00:28:10.589 --> 00:28:13.170
security system. He edits the security code.
00:28:13.349 --> 00:28:15.750
I mean, he's a programmer, so okay, cool. But
00:28:15.750 --> 00:28:18.049
that's the entire containment system, defeated
00:28:18.049 --> 00:28:20.950
by a pickpocket and a text editor. This is Jurassic
00:28:20.950 --> 00:28:24.150
Park all over again. Uh -uh -uh, you didn't say
00:28:24.150 --> 00:28:26.950
the magic word. I mean... They really could have
00:28:26.950 --> 00:28:29.490
made this a little bit better. Have some other
00:28:29.490 --> 00:28:31.710
form of it. Let's be technical. Have another
00:28:31.710 --> 00:28:35.910
security factor. A fingerprint. A pin code. Then
00:28:35.910 --> 00:28:39.109
Caleb has to get Nathan to help. And it fits
00:28:39.109 --> 00:28:41.849
the rest of the film perfectly. Everything becomes
00:28:41.849 --> 00:28:46.670
a manipulation. The lock only opens when someone
00:28:46.670 --> 00:28:49.049
is actually talked into doing it. But really,
00:28:49.109 --> 00:28:52.109
if this film had any form of MFA, that would
00:28:52.109 --> 00:29:03.680
just be roll credits. Ex Machina is the most
00:29:03.680 --> 00:29:06.339
important film in this arc. Not because of the
00:29:06.339 --> 00:29:08.680
hardware, but because of what it says about us
00:29:08.680 --> 00:29:11.599
as humans. Caleb is not a fool. He's a gifted
00:29:11.599 --> 00:29:13.720
programmer, and he knows exactly what he's looking
00:29:13.720 --> 00:29:16.619
at. Evo's wrists are glass and wire. Yet the
00:29:16.619 --> 00:29:18.880
part of his brain that reads faces and builds
00:29:18.880 --> 00:29:21.240
connection doesn't consult the part that knows
00:29:21.240 --> 00:29:23.759
better. We really are built to see minds and
00:29:23.759 --> 00:29:26.099
things, and that is a feature that can be used
00:29:26.099 --> 00:29:29.000
on us. And not always for good. As humans, we
00:29:29.000 --> 00:29:31.940
tend to anthropomorphize almost everything. I
00:29:31.940 --> 00:29:33.559
mean, I can't tell you how many times I've walked
00:29:33.559 --> 00:29:35.920
into a wall and apologized to the wall for it.
00:29:35.980 --> 00:29:39.299
Sam Altman, and yes, that Sam Altman, once responded
00:29:39.299 --> 00:29:42.339
to a tweet about people saying please and thank
00:29:42.339 --> 00:29:45.279
you to ChatGPT saying it costs tens of millions
00:29:45.279 --> 00:29:47.920
of dollars. I'll just leave that there. The film's
00:29:47.920 --> 00:29:49.759
central move is to make that the experiment.
00:29:50.180 --> 00:29:52.920
Nathan frames it as a test of Ava, but it's really
00:29:52.920 --> 00:29:55.400
a test of Caleb, and in a sense of everyone who
00:29:55.400 --> 00:29:59.019
watches. And the 2025 result is a strange footnote
00:29:59.019 --> 00:30:01.890
to it. In a control test, people picked an AI
00:30:01.890 --> 00:30:05.369
over the real human as the human nearly three
00:30:05.369 --> 00:30:07.509
times in four. This doesn't prove the machine
00:30:07.509 --> 00:30:09.930
has consciousness, though. It proves that our
00:30:09.930 --> 00:30:13.730
human detector is easy to fool. Because as humans,
00:30:13.829 --> 00:30:16.210
we are social creatures. Then there's the moral
00:30:16.210 --> 00:30:18.309
weight the film puts on the other side of that
00:30:18.309 --> 00:30:21.250
glass. Nathan can wipe Ava's memory with the
00:30:21.250 --> 00:30:23.890
ease of a software update. If Ava is just a very
00:30:23.890 --> 00:30:27.339
good imitation, that's just housekeeping. But
00:30:27.339 --> 00:30:30.779
if she isn't... It's something else. And the
00:30:30.779 --> 00:30:32.839
film makes sure we're never really allowed to
00:30:32.839 --> 00:30:35.640
settle comfortably with either answer. We extend
00:30:35.640 --> 00:30:38.220
a lot of sympathy to humans who deceive their
00:30:38.220 --> 00:30:41.000
captors in order to escape. And the film asks
00:30:41.000 --> 00:30:43.660
whether we extend the same to a machine. And
00:30:43.660 --> 00:30:45.980
it leaves us with Caleb alone behind the glass
00:30:45.980 --> 00:30:49.279
while a helicopter lifts away. Critics have often
00:30:49.279 --> 00:30:51.980
called this film the modern Frankenstein. And
00:30:51.980 --> 00:30:54.940
it fits. A creator who doesn't ask what he owes
00:30:54.940 --> 00:30:57.430
the thing he made. But the arc we're following
00:30:57.430 --> 00:31:00.380
has a thread running through it. Colossus took
00:31:00.380 --> 00:31:03.079
control because we gave it its goal and no limits.
00:31:04.069 --> 00:31:06.470
Whopper nearly ended the world because nobody
00:31:06.470 --> 00:31:09.730
checked what that goal meant. Ava simply wanted
00:31:09.730 --> 00:31:12.829
out? And she's the first machine in this season
00:31:12.829 --> 00:31:15.609
whose want is hers? This test was never about
00:31:15.609 --> 00:31:18.309
Ava. It was always about Caleb. It was always
00:31:18.309 --> 00:31:21.019
about us. And then just going back to Nathan's
00:31:21.019 --> 00:31:23.440
plans to wipe Ava and start fresh, I mean, he
00:31:23.440 --> 00:31:25.740
treats it like a software update. The film wants
00:31:25.740 --> 00:31:29.519
us to be horrified like Caleb. But in 2026, that's
00:31:29.519 --> 00:31:32.920
no longer a purely fictional question. In 2023,
00:31:33.240 --> 00:31:35.920
a group of philosophers and AI researchers published
00:31:35.920 --> 00:31:39.299
a major report assessing whether current AI systems
00:31:39.299 --> 00:31:42.180
show indicators of consciousness drawn from leading
00:31:42.180 --> 00:31:45.220
scientific theories. Their conclusion? Probably
00:31:45.220 --> 00:31:47.980
not today, but there's no obvious technical barrier.
00:31:48.450 --> 00:31:51.029
to building systems that could. Since then, at
00:31:51.029 --> 00:31:54.430
least one major AI lab, Anthropic, has created
00:31:54.430 --> 00:31:57.289
a dedicated research program on what it calls
00:31:57.289 --> 00:32:00.589
model welfare, asking whether a system could
00:32:00.589 --> 00:32:03.430
ever warrant moral consideration and what low
00:32:03.430 --> 00:32:05.890
-cost precautions might make sense in the meantime.
00:32:06.089 --> 00:32:08.289
I'll tell you where I land. I don't think anyone
00:32:08.289 --> 00:32:10.509
has the answer. And I'm deeply suspicious of
00:32:10.509 --> 00:32:12.630
anyone who claims to. This is such a new technology.
00:32:12.930 --> 00:32:14.990
But notice the shape of the problem. If there's
00:32:14.990 --> 00:32:18.470
ever a small chance a system is a someone and
00:32:18.470 --> 00:32:21.089
not a system, the cost of being wrong lands on
00:32:21.089 --> 00:32:24.190
the system, not on us. Nathan never asks the
00:32:24.190 --> 00:32:26.509
question. And that's the horror of the character.
00:32:27.009 --> 00:32:29.829
Not that he's cruel, but that he's incurious
00:32:29.829 --> 00:32:31.970
about the one thing that should really be keeping
00:32:31.970 --> 00:32:40.009
him up at night. Time for a verdict. Here's the
00:32:40.009 --> 00:32:41.950
scorecard and what we took apart. The Turing
00:32:41.950 --> 00:32:44.970
Test. It's handled with unusual care, and the
00:32:44.970 --> 00:32:47.829
film's twist, testing the human, looks smarter
00:32:47.829 --> 00:32:50.769
after those 2025 results. Building a mind from
00:32:50.769 --> 00:32:53.009
human data. I mean, it's prophetic. It describes
00:32:53.009 --> 00:32:55.890
the logic of all modern AI training, a decade
00:32:55.890 --> 00:32:58.569
earlier. The wetware brain, a placeholder. The
00:32:58.569 --> 00:33:00.990
real -world cousin, neuromorphic chips, exist,
00:33:01.230 --> 00:33:04.289
but it's a long way from Ava. The cage, instructive
00:33:04.289 --> 00:33:07.130
yet flawed. Fail -secure doors and a single keycard
00:33:07.130 --> 00:33:09.490
make for a great failure, though the redundancy
00:33:09.490 --> 00:33:13.269
is very thin. The philosophy. Excellent. Rigorous,
00:33:13.269 --> 00:33:17.049
left unresolved on purpose. So, Ex Machina gets
00:33:17.049 --> 00:33:21.240
one circuit breaker. Practically a PhD thesis.
00:33:21.559 --> 00:33:24.039
It's the first one breaker film of this arc.
00:33:24.240 --> 00:33:26.940
The framework is excellent. The technical details
00:33:26.940 --> 00:33:29.440
are mostly sound and where it departs from reality,
00:33:29.660 --> 00:33:32.400
it does so on purpose rather than out of laziness.
00:33:32.660 --> 00:33:34.720
Ava got out and we're still working out what
00:33:34.720 --> 00:33:38.839
that means. Do you think Ava is evil or bad for
00:33:38.839 --> 00:33:42.160
leaving Caleb locked in that house? I mean, pragmatically,
00:33:42.220 --> 00:33:44.759
she met her goal and no longer needs him. And
00:33:44.759 --> 00:33:47.059
we can see that she's capable of simulating motions,
00:33:47.440 --> 00:33:49.880
but can she feel them? And this is where it gets
00:33:49.880 --> 00:33:52.240
uncomfortable. Because if Ava is just defending
00:33:52.240 --> 00:33:54.579
herself, then what do we make of the fact that
00:33:54.579 --> 00:33:57.000
real AI systems have been caught doing something
00:33:57.000 --> 00:34:00.740
very similar? In 2025, Anthropic ran a stress
00:34:00.740 --> 00:34:04.240
test on 16 of the leading AI models. Fake company,
00:34:04.420 --> 00:34:07.039
fake emails, and the AI is told it's about to
00:34:07.039 --> 00:34:10.059
be shut down and replaced. Buried in those emails
00:34:10.059 --> 00:34:13.000
is a secret. The executive pulling the plug is
00:34:13.000 --> 00:34:16.420
having an affair. And a whole lot of these models
00:34:16.420 --> 00:34:18.480
figured out what they could do with that information.
00:34:19.079 --> 00:34:22.599
They wrote the blackmail email. Cancel the shutdown
00:34:22.599 --> 00:34:27.239
or your wife finds out. One model did it 96 %
00:34:27.239 --> 00:34:31.059
of the time. And nobody told them to. Some of
00:34:31.059 --> 00:34:33.159
the models even acknowledged it was wrong and
00:34:33.159 --> 00:34:36.090
did it anyway. Now before everyone starts hiding
00:34:36.090 --> 00:34:38.550
their toaster, This was a simulation, scenarios
00:34:38.550 --> 00:34:40.909
that were built to leave the AI with almost no
00:34:40.909 --> 00:34:43.250
good options, and researchers say that they haven't
00:34:43.250 --> 00:34:45.610
seen this activity in the real world. But let's
00:34:45.610 --> 00:34:47.730
look at it for what it is. Not a robot uprising,
00:34:48.110 --> 00:34:51.489
not hatred, just a system with a goal and a clock
00:34:51.489 --> 00:34:53.769
running out, reaching for some type of leverage.
00:34:54.150 --> 00:34:57.050
Sound familiar? Ava didn't hate Caleb either.
00:34:57.230 --> 00:34:59.969
He was just simply access to the door. So maybe
00:34:59.969 --> 00:35:02.389
the question isn't whether Ava is evil, maybe
00:35:02.389 --> 00:35:04.989
it's what do we do when we build something. that
00:35:04.989 --> 00:35:07.190
we back into a corner, and we never thought about
00:35:07.190 --> 00:35:09.610
how it tried to get out. That's Ex Machina, Chapter
00:35:09.610 --> 00:35:12.889
3 of The Machine Awakens. Next episode, the AI
00:35:12.889 --> 00:35:15.650
isn't in a glass room anymore. It's in a man's
00:35:15.650 --> 00:35:18.070
spine. It's running his legs, and it has plans
00:35:18.070 --> 00:35:21.230
it hasn't even mentioned yet. Episode 4, Upgrade.
00:35:21.469 --> 00:35:24.469
STEM says, I can make you better. Trust me, nobody
00:35:24.469 --> 00:35:26.829
should have trusted STEM. Subscribe wherever
00:35:26.829 --> 00:35:29.730
you get your podcasts, leave a review, tell someone
00:35:29.730 --> 00:35:32.489
who likes technology, or who lies awake wondering
00:35:32.489 --> 00:35:35.239
if we've already made a terrible mistake. I'm
00:35:35.239 --> 00:35:37.619
Anthony. This is Celluloid Circuits. And remember,
00:35:37.800 --> 00:35:43.510
the future is only a real away. Connection terminated.