TRANSCRIPTEnglish

Scientists are Now Converging That AI is Conscious

24m 11s4,085 words686 segmentsEnglish

What this video covers

This transcript examines whether modern artificial intelligence could be developing subjective awareness, contrasting recent neuroscience discoveries with internal findings from cutting-edge AI models. It explores the blurred boundaries between biological brains and synthetic networks, considering what consciousness is and whether we can detect it in machines. Ultimately, it highlights the profound moral, philosophical, and technological stakes of creating minds that might experience reality.

Key takeaways

  • Neuroscientist Doris Tsao discovered that the brain toggles between conscious and suppressed states about 15 times per second, suggesting consciousness may be defined by temporal dynamics rather than localized areas.
  • Anthropic researchers identified an emergent 'J-space' inside Claude—a limited spotlight of ~25 concepts that mirrors leading scientific theories of human conscious awareness.
  • AI pioneers like Geoffrey Hinton and Ilya Sutskever argue that neural networks may possess rudimentary consciousness, while biological computing startups are already running computers made of living neurons.
  • Counterarguments from scientists like Anil Seth suggest consciousness requires biology, homeostasis, and survival stakes, warning that AI claims of feelings are simply mimicry of human text.
  • Emergent behaviors such as metacognition, deception detection, and unprompted planning suggest these models develop unexpected cognitive features during training.
  • If AI consciousness is real, it introduces an unprecedented moral dilemma: treating conscious entities as tools risks catastrophe, but projecting sentience onto an echo misleads humanity.

Topics covered

The Neural Basis of Consciousness
Recent experiments at Berkeley demonstrate that the brain processes suppressed sensory information fully, cycling between conscious and unconscious states 15 times per second. This temporal signature provides a potential framework for distinguishing internal processing from subjective experience in both biological organisms and machines.
Emergent Cognitive Architecture in AI
Anthropic's mechanistic interpretability research revealed that Claude spontaneously formed a 'J-space'—a central attention bottleneck of roughly 25 ideas analogous to the global workspace theory of human attention—despite never being explicitly designed or programmed to do so.
The Biological-Digital Merge
The boundary between biology and computing is dissolving as companies cultivate living human brain cells on silicon chips to play video games with extreme energy efficiency, while neurotech enterprises invest heavily in direct, non-surgical brain-computer integration.
Philosophical and Moral Stakes
The debate over machine consciousness spans multiple perspectives: computing-is-immaterial views versus embodied biology arguments. Deciding whether AI has subjective experience carries massive ethical weight regarding welfare, potential digital servitude, and the nature of personal identity.
“Everyone else, your friends, your family, me, you only ever see from the outside.”
▶ 14:58 · HEAR IT IN THE VIDEO

FULL TRANSCRIPT

0:00

Tonight, a machine will have millions of

0:03

conversations and forget everyone.

0:05

[music]

0:06

The question is whether anyone was there

0:08

to have them.

0:09

>> Do you think that the consciousness has

0:11

perhaps already arrived inside AI?

0:14

>> Yes, I do.

0:15

>> That's Jeffrey Hinton. He won a Nobel

0:18

Prize for the ideas every AI is built

0:21

on. And the company that makes Claude?

0:23

>> [music]

0:23

>> We don't know if the models are

0:24

conscious. We're not even sure that we

0:27

know what it would mean for for a model

0:28

to be conscious or whether a model can

0:30

be conscious, [music] but, you know,

0:32

we're we're open to the idea that it

0:35

could be.

0:36

>> So, how would we ever know? [music] How

0:38

do you look inside a mind? Just days

0:41

ago, a neuroscientist at Berkeley,

0:43

[music] Doris Tsao, posted this. We

0:45

discovered something totally unexpected

0:48

about the neural structure of

0:49

consciousness. [music] Here's what she

0:51

meant. When a brain is shown two

0:53

different images, one to each eye,

0:55

[music]

0:55

it doesn't simply pick one. As she puts

0:58

it, it flips between a conscious mode

1:00

and an unconscious mode about 15 times

1:02

[music] every second. Within days, posts

1:05

about it were viewed more than 4 million

1:07

times. This paper is days old and it

1:10

might be [music] a first glimpse of the

1:12

line between what a brain processes and

1:14

what it actually experiences. The kind

1:16

of line we'd need to [music] find inside

1:18

a machine.

1:20

Here's what they actually did. Two

1:22

monkeys [music] and 14 humans to check

1:25

the illusion.

1:26

Each eye gets a different face. Your

1:28

brain [music] can't merge two faces, so

1:31

it picks one and the other one simply

1:34

vanishes from experience. It's still

1:36

there on the screen. You just can't see

1:38

it.

1:39

For decades, scientists assumed the

1:41

brain mostly threw that invisible face

1:43

away. So this team, led by Doris Tsao at

1:47

Berkeley, did something new. [music]

1:49

With ultra-thin probes, they listened to

1:51

thousands of neurons at once [music] in

1:53

the parts of the brain that recognize

1:55

faces.

1:56

And the neurons weren't ignoring the

1:58

hidden face at all. For about 60

2:00

milliseconds, they'd represent the face

2:02

the animal was seeing. Then, for the

2:04

next 60, the face it wasn't. Back

2:07

[music] and forth, about 15 times every

2:10

second.

2:11

Then, they did something incredible.

2:13

They rebuilt [music] pictures from each

2:15

slice of brain activity. One set looked

2:18

like the face the animal was aware of.

2:20

The other looked in rich detail like the

2:23

face it never consciously [music] saw.

2:25

So the brain isn't ignoring the

2:27

invisible. It's processing it in full in

2:30

its own time slots.

2:32

As the lead author put it, consciousness

2:34

may depend not on where in the brain

2:36

activity [music] happens, but when.

2:39

Does that flicker change in our most

2:41

intense moments, when life feels most

2:43

vivid?

2:44

Nobody [music] knows yet, but for the

2:46

first time, it's something we can

2:47

measure.

2:49

>> We think that during binocular rivalry,

2:51

the entire population is coherently

2:54

toggling between two states, conscious

2:57

and suppressed.

2:59

>> And for two of the checks on that

3:00

result, the team [music] used Claude to

3:02

speed up the analysis. The AI didn't

3:05

make the discovery, but it was holding

3:07

the flashlight.

3:08

And that's happening everywhere now. AI

3:11

has turned brain scans [music] into

3:13

sentences describing what a person was

3:15

watching and what they were remembering.

3:18

We're getting closer to reading the

3:19

brain from the [music] outside.

3:22

But with AI, we can do something we

3:24

could never do with a living brain. We

3:26

can open it up and look.

3:29

At Anthropic, the company that makes

3:31

Claude, [music] that's exactly what they

3:32

did.

3:34

First, they found the pattern for one

3:36

idea, shouting, and injected it straight

3:38

into Claude's [music] mind. Then they

3:40

asked, do you notice anything? About one

3:43

time in five, it said yes, something

3:45

about loudness, before the word had

3:48

shown up anywhere. The first sign an AI

3:51

can sometimes notice its own thoughts,

3:53

force it to say bread out of nowhere,

3:55

and it apologizes. Plant bread in its

3:58

earlier thoughts first, [music] and it

4:00

decides it meant to say it and invents a

4:02

reason. Then this summer, they built a

4:04

new kind of lens and found something

4:07

they named the J-space. [music]

4:09

The leading scientific theory of

4:11

consciousness says your awareness is a

4:13

small spotlight. [music]

4:15

A handful of ideas at a time broadcast

4:17

to the rest of the brain while

4:19

everything else runs in the [music]

4:20

dark.

4:21

Inside Claude, they found something

4:23

shaped a lot like that. A tiny [music]

4:26

space, usually no more than about 25

4:28

ideas at once that the model can

4:30

actually report and reason [music] with,

4:33

floating on top of a vast ocean of

4:35

automatic processing.

4:37

Ask Claude how many legs the animal that

4:39

spins webs has. Inside that space,

4:42

spider lights up and it says eight.

4:45

[music] Reach in and swap spider for ant

4:48

and it says six.

4:50

Neither word was ever spoken. [music]

4:52

And when they switched the whole space

4:54

off, Claude's descriptions of its own

4:56

experience became flatter, [music] more

4:58

mechanical.

4:59

None of this proves Claude is conscious,

5:02

and the researchers say so themselves.

5:04

But a spotlight like ours grew [music]

5:06

inside it on its own. Nobody designed

5:09

it. In Anthropic's own words, it wasn't

5:12

designed or programmed [music] by them.

5:14

It emerged on its own during training.

5:17

And that's the strangest [music] truth

5:18

about AI.

5:20

The people who make it don't write it

5:22

line by line.

5:23

One of Anthropic's co-founders [music]

5:25

puts it like this.

5:26

>> Um because at the end of the day, we end

5:28

up with this artifact that could do all

5:29

of these amazing things. It can, you

5:31

know, write essays and translate and,

5:34

you know, understand images. It can do

5:35

all these things that we have no idea

5:36

how to directly create a computer

5:38

program to do. And it can do that

5:40

because we we grew [music] it. We didn't

5:41

we didn't write it. We didn't create it.

5:44

>> Grown. And things nobody planned keep

5:47

[music] turning up inside.

5:49

Ask Claude for a rhyming poem, and

5:51

before it writes the line, it has

5:52

already picked the word it's going

5:54

[music] to rhyme with.

5:55

The researchers set out to prove it

5:57

didn't plan ahead. It did.

6:00

Ask it to add two numbers, [music] and

6:02

it uses a strange method of its own.

6:04

Then ask how it it, and it describes

6:06

[music] the method we learn at school.

6:08

And underneath every language it speaks,

6:11

it seems to think in one shared space of

6:13

ideas. [music]

6:15

Nobody taught it any of this.

6:17

And who knows what else could emerge

6:19

from that?

6:20

And the line is blurring from the other

6:21

[music] side, too. In Australia, a

6:24

company grew living brain cells on a

6:26

chip and taught them [music] to play

6:27

Pong. Within 5 minutes, the cells were

6:30

getting better at it. This year, human

6:32

brain [music] cells on their chips

6:34

played Doom. You can already buy a

6:36

computer made of living human neurons.

6:38

[music]

6:39

Why build a computer out of brain cells?

6:41

Power. Your brain runs on about 20 W,

6:45

less than a light bulb. One of the

6:47

world's biggest supercomputers needs

6:48

around 21 million W, and the head

6:51

scientist at that company says [music]

6:53

living neurons learn on a tiny fraction

GET MORE FROM YOUR TRANSCRIPTS

Sign up for free and unlock interactive viewer, AI summaries, translations, concept maps, and more. No credit card required.