Latest / Core Memory / Consciousness in the Quantum Realm

Transcript

Consciousness in the Quantum Realm · Core Memory

0:01
I actually have a fun analogy for LLMs that I was just thinking
0:04
about last night.
0:05
SPEAKER_00: You might like this.
0:06
So imagine that kind of like consciousness is something that
0:10
comes first and then
0:11
it generates more and more and more training data for whatever
0:14
large model you want.
0:16
SPEAKER_00: And I think this is actually what happened from evolutionary perspective.
0:20
So we like organisms started with a few neurons and then they
0:24
evolve like a tiny
0:25
brainstem and then they evolve like more brain regions and more
0:27
brain regions.
0:28
SPEAKER_00: And then
0:29
SPEAKER_00: Eventually mammals and us evolved this massive cortex.
0:32
SPEAKER_00: So we've been building this giant cortex,
0:36
SPEAKER_00: which gives us like weirdly sized head compared to other animals.
0:39
And what I think is happening is we're just basically building
0:42
like a bigger and
0:43
SPEAKER_00: bigger and bigger LLM style thing inside our head.
0:48
SPEAKER_00: And that's what evolution has done.
0:50
SPEAKER_00: And we're doing the same thing now with LLMs.
0:51
SPEAKER_00: We're just building this like massive neocortex.
0:55
SPEAKER_00: But everyone thinks the neocortex is the seat of consciousness.
0:58
SPEAKER_00: What if it's not?
0:59
What if it's that really deep thing that started off in
1:02
individual cells that's
1:04
SPEAKER_00: been throwing off training data,
1:05
throwing off training data for millions or billions of years
1:09
and it's grown this
1:10
SPEAKER_00: giant model.
1:25
SPEAKER_02: All right, team, we're going to get a little weird today.
1:28
SPEAKER_02: We are talking to Suzanne Gildert.
1:32
She's a quantum physicist who I've known for a long time and
1:37
has a new startup
1:38
SPEAKER_02: called Nirvanic Consciousness Technologies,
1:41
SPEAKER_02: which is an impressive name and a
1:45
SPEAKER_02: impressive field of research.
1:46
SPEAKER_02: Basically,
1:47
SPEAKER_02: we're going to get into consciousness today,
1:48
SPEAKER_02: obviously,
1:49
SPEAKER_02: and the limits of LLM technology,
1:52
SPEAKER_02: current AI technology,
1:54
SPEAKER_02: whether we're on
1:56
SPEAKER_02: a dead end and,
1:57
SPEAKER_02: and the really interesting stuff is off on some other path that,
2:01
SPEAKER_02: that Suzanne is chasing.
2:02
SPEAKER_02: Um,
2:03
SPEAKER_02: so I think you'll be,
2:05
SPEAKER_02: takes a little minute to get into this,
2:07
but then I think you'll be entertained and your mind will be
2:11
racing as we dive in
2:14
SPEAKER_02: now to explore consciousness in the quantum realm.
2:22
SPEAKER_02: Hello, Suzanne.
2:23
SPEAKER_02: How are you?
2:23
SPEAKER_00: Hi, Ashley.
2:24
SPEAKER_00: I am very good.
2:26
SPEAKER_00: Thank you.
2:26
SPEAKER_00: Loving the new brand.
2:28
SPEAKER_02: Oh, thank you.
2:29
SPEAKER_02: We try.
2:29
SPEAKER_02: We try.
2:30
SPEAKER_02: I saw you.
2:31
SPEAKER_02: You have a Substack as well.
2:32
SPEAKER_00: I know.
2:33
SPEAKER_00: Actually,
2:34
SPEAKER_00: well,
2:34
it's kind of funny because I went on your YouTube channel and I
2:37
saw like your
2:37
banner thing and it's got the core memories with the actual
2:41
wires going through them.
2:42
SPEAKER_00: It's like some computer history geek must have designed that.
2:46
SPEAKER_02: Well, it was me.
2:47
I mean,
2:47
I didn't do the design,
2:49
but I was like,
2:50
we got to have these bits and pieces in there.
2:52
SPEAKER_02: Yeah, we're trying to... That's so cool.
2:54
SPEAKER_02: It was a hard mandate because I was like,
2:55
SPEAKER_02: I want it retro to acknowledge the actual core memory,
3:00
SPEAKER_02: but then kind of futuristic for the stuff we covered.
3:03
SPEAKER_02: That's cool.
3:04
SPEAKER_02: Yeah, I was happy with it.
3:04
SPEAKER_02: And I'm so happy to know somebody who actually knows what core memory is.
3:09
Yeah.
3:11
SPEAKER_02: Where are we catching you?
3:13
SPEAKER_00: Well, I'm in Vancouver at the moment.
3:15
SPEAKER_00: So, yeah.
3:16
SPEAKER_00: Yeah, in my home office.
3:17
SPEAKER_02: Okay.
3:18
SPEAKER_02: All right.
3:18
SPEAKER_02: And I see we've got some of your robots behind you.
3:22
SPEAKER_02: Yeah.
3:22
I thought these might actually be a bit of a throwback because
3:25
some of these robots
3:26
you've sort of seen before and they kind of come back over and
3:29
over again.
3:30
SPEAKER_02: Yes.
3:30
SPEAKER_02: Yeah.
3:30
SPEAKER_02: Well, I'm glad because we're going to be talking robots.
3:34
SPEAKER_02: Well, there's all kinds of places I want to go to chat with you.
3:38
SPEAKER_02: But, you know, I thought...
3:40
SPEAKER_02: There's probably a vague unifying theme in some of this,
3:44
SPEAKER_02: which is as long as I have known you,
3:47
SPEAKER_02: you have been pursuing this idea of where is AI heading,
3:54
SPEAKER_02: where is robotics heading.
3:56
And I mean,
3:57
please obviously correct me if I'm speaking out of turn,
4:00
but I feel like you've...
4:02
have been very open-minded to this idea of what do these new
4:08
forms of intelligence
4:09
SPEAKER_02: and machines look like in terms of consciousness,
4:12
SPEAKER_02: in terms of some of these more philosophical thoughts,
4:16
SPEAKER_02: like these are beings that are coming into this world,
4:19
SPEAKER_02: and do they have rights,
4:22
SPEAKER_02: and how do we interact with them?
4:23
SPEAKER_02: And I know you've been...
4:26
SPEAKER_02: I don't think your view on this particular part has changed.
4:29
I mean,
4:29
you've always been pretty pro-embracing these new life forms
4:34
and living with them.
4:35
SPEAKER_02: Is that fair?
4:37
SPEAKER_00: I think so, yeah.
4:38
SPEAKER_00: I mean,
4:38
I've always been interested in trying to understand what the
4:41
human mind is and how
4:42
SPEAKER_00: it works.
4:43
And as an engineer,
4:44
I've always tried to understand something by building it.
4:47
SPEAKER_00: So the idea of understanding the human mind is like,
4:49
SPEAKER_00: can we build something equivalent in a machine,
4:52
SPEAKER_00: in a robot,
4:53
SPEAKER_00: and get all those pieces put together in the right way,
4:56
SPEAKER_00: that it sort of comes together as a mind?
4:58
SPEAKER_00: So it's really a fascination with the mind, I think.
5:00
SPEAKER_02: I remember I was interviewing you one time,
5:03
SPEAKER_02: and I don't remember the exact quote,
5:04
SPEAKER_02: but it was something like,
5:06
Like you were kind of looking forward to the day that you're
5:08
hanging out with your
5:09
SPEAKER_02: robot buddy at home.
5:12
SPEAKER_02: Yeah.
5:13
And you've been thinking about that type of future since you
5:16
were a kid or that's
5:18
SPEAKER_02: more in your professional light?
5:20
SPEAKER_00: I think I've been thinking about it for quite a long time.
5:25
SPEAKER_00: You know, early on, it was about exploring electronics.
5:28
SPEAKER_00: I'm like, what can I build?
5:29
SPEAKER_00: But as soon as I realized that you can actually make systems move,
5:33
SPEAKER_00: you can make them do things in the world,
5:35
SPEAKER_00: then my mind started to go kind of crazy.
5:37
SPEAKER_00: And, well, where does this end?
5:39
SPEAKER_00: Will we have machines that are eventually like us that are moving around?
5:42
SPEAKER_00: Yeah.
5:43
SPEAKER_00: and have minds like us.
5:45
SPEAKER_00: And at that point,
5:46
I started thinking about interacting with those machines and
5:49
would they think differently,
5:50
SPEAKER_00: would they feel differently to us,
5:52
SPEAKER_00: which I guess brings us to consciousness at some point.
5:55
SPEAKER_00: Yeah.
5:55
SPEAKER_02: I mean, I was going to say, I mean, lots of people have these thoughts.
5:58
SPEAKER_02: You happen to have like dedicated your career to actually seeing if,
6:04
SPEAKER_02: you know,
6:04
SPEAKER_02: exploring the same thing.
6:06
SPEAKER_00: I think I have a weird...
6:07
Way of thinking about this is different than most people
6:10
because I'm very I have a
6:11
SPEAKER_00: very optimistic outlook on most things.
6:13
SPEAKER_00: So when I think about like robots living amongst us or,
6:17
SPEAKER_00: you know,
6:18
SPEAKER_00: machine sentience or what could those systems do and look like,
6:22
SPEAKER_00: I tend to be more utopian and optimistic.
6:25
rather than the dystopian sci-fi kind of horror fancy that a
6:29
lot of people think of.
6:30
So that's generally the first barrier I come across when I talk
6:33
to people.
6:34
SPEAKER_00: They say, oh, they're going to kill us all or this is a horrible future.
6:39
SPEAKER_00: But I don't think that way.
6:40
SPEAKER_00: It's like, why are you thinking like that?
6:42
SPEAKER_02: Right, yeah.
6:43
SPEAKER_02: I mean,
6:43
SPEAKER_02: I tend to go a little more that way as well,
6:46
SPEAKER_02: although,
6:48
SPEAKER_02: gosh,
6:48
I've been reporting this book on AI technology for the last
6:52
couple of years,
6:53
SPEAKER_02: and it gets harder because all the people,
6:56
SPEAKER_02: I'm sitting with the people who are developing this technology,
7:00
SPEAKER_02: and gosh,
7:01
I feel like the vast majority of them seem to be deathly afraid
7:07
of what they're
7:07
SPEAKER_02: making at the same time as they're making.
7:09
SPEAKER_02: I don't know, really, it does get in your head a bit.
7:13
SPEAKER_00: Yeah.
7:13
SPEAKER_00: But then on the other hand,
7:14
people are still buying products that are promoting AI and
7:21
making AI improve.
7:23
And so if we didn't want this as our future,
7:25
I think we wouldn't be voting with our wallets.
7:28
SPEAKER_00: So I do find a little bit of a paradox there.
7:30
SPEAKER_02: Yeah, well, yeah.
7:32
SPEAKER_02: I mean, we've seen this many times with technology.
7:35
SPEAKER_02: It's weird,
7:35
SPEAKER_02: you know,
7:36
something like Facebook where you know you're making this deal
7:39
with the company to
7:40
SPEAKER_02: hand over all your data in exchange for these free services.
7:44
SPEAKER_02: I don't know.
7:44
SPEAKER_02: I go back and forth on whether I know that people are
7:48
are really how aware they are of these decisions they're making
7:51
as this stuff is
7:53
SPEAKER_02: foisted upon them.
7:54
Two things I know for sure is that most people think ChatGPT is
7:59
really helpful and
8:00
SPEAKER_00: useful and they use it all the time.
8:02
SPEAKER_00: And then whenever I ask someone,
8:03
would you want a robot to do all your domestic chores for you
8:06
and do all your
8:07
SPEAKER_00: grocery shopping and cooking and cleaning,
8:08
SPEAKER_00: they're like,
8:09
SPEAKER_00: hell yes.
8:10
SPEAKER_00: So those are two good days points.
8:13
SPEAKER_02: Well, let's go there for a second.
8:17
SPEAKER_02: I know with your new startup,
8:19
SPEAKER_02: which we'll get into,
8:20
SPEAKER_02: you've changed some of your thinking around this stuff.
8:23
SPEAKER_02: But your previous two startups,
8:25
SPEAKER_02: Kindred and Sanctuary,
8:28
SPEAKER_02: you were sort of building helpers like that in a lot of ways.
8:33
SPEAKER_02: I mean, some of it was evolved into robotic technology.
8:38
And then I think it's fair to say with Sanctuary getting more
8:43
to like a more
8:45
SPEAKER_02: embodied form of AI.
8:46
SPEAKER_02: But that was the question that you were on for a bit, right?
8:49
SPEAKER_02: Was to use, I mean, yeah, can you explain to me?
8:52
SPEAKER_02: I mean, using advancing AI technology and to give it an embodiment.
8:59
Yeah,
9:00
so the original idea of Kindred back in 2014 was actually a
9:04
really simple idea.
9:05
SPEAKER_00: It was,
9:07
we now have this like VR technology and we have like sensors
9:12
and things that we can
9:13
SPEAKER_00: put on our bodies.
9:14
SPEAKER_00: So could we move a humanoid robot using all this gear?
9:18
So I was kind of imagining this like advanced suit that you
9:21
would wear and it would
9:23
SPEAKER_00: map to a humanoid robot.
9:24
SPEAKER_00: A bit like Ready Player One,
9:26
SPEAKER_00: you know,
9:26
that movie where they go into the VR and then they're actually
9:29
controlling avatars,
9:30
SPEAKER_00: but it would be a humanoid robot instead.
9:33
And my thought was if you could do that and the robot could
9:35
then do anything the
9:36
SPEAKER_00: person could do,
9:37
SPEAKER_00: then it's immediately generating training data.
9:40
SPEAKER_00: And with enough training data, you could automate that process.
9:44
So you could have an AI like watching learning in the
9:46
background as a person moved.
9:49
a robot or say thousands of people moved thousands of robots
9:53
and that that was
9:54
really the core idea of the whole thing and then that played
9:57
out over the next
9:58
eight years that idea yeah so i mean you end up selling that
10:01
company for hundreds
10:03
of millions if i recall correctly and then so that's good for
10:07
you and then and then
10:09
and then with sanctuary it evolves it evolves into something um
10:13
yeah that i would
10:15
SPEAKER_02: say is is um
10:18
SPEAKER_02: comes with bigger, deeper questions attached to it, right?
10:23
SPEAKER_00: Yeah,
10:23
SPEAKER_00: so the original idea of Kindred was kind of the same,
10:25
SPEAKER_00: like we wanted to do this in humanoids,
10:27
SPEAKER_00: but at that point,
10:28
SPEAKER_00: humanoids weren't quite advanced enough,
10:30
SPEAKER_00: like the hardware wasn't there.
10:33
SPEAKER_00: So what happened at Kindred was the idea ended up morphing into a
10:37
niche version of itself,
10:39
where the robot ended up being a robot arm in a warehouse.
10:42
SPEAKER_00: But the technology idea was exactly the same,
10:44
that people had trained that robot arm by showing it what to do
10:48
through these VR
10:49
SPEAKER_00: teleoperation gear.
10:50
SPEAKER_00: And then at Sanctuary, it was kind of a reboot.
10:52
It was like,
10:52
okay,
10:53
it's 2018 and humanoid robots are getting better and better.
10:56
So can we try the same thing again,
10:59
but with a fully general purpose robot this time?
11:02
SPEAKER_02: But you,
11:03
SPEAKER_02: but in the background,
11:04
SPEAKER_02: I mean,
11:04
SPEAKER_02: you were always a scientist and a researcher and you were thinking even,
11:07
SPEAKER_02: unless I'm wrong,
11:09
SPEAKER_02: even on sanctuaries,
11:10
SPEAKER_02: like,
11:10
SPEAKER_02: is like,
11:11
SPEAKER_02: how is this evolving towards this thing,
11:13
SPEAKER_02: not just doing tasks,
11:14
SPEAKER_02: but becoming kind of its own being,
11:16
SPEAKER_02: right?
11:17
SPEAKER_00: Yeah, that's right.
11:18
SPEAKER_00: That's right.
11:18
And we designed a whole bunch of different what are called
11:20
cognitive architectures.
11:22
SPEAKER_00: So it's like, how do you put the pieces of a mind together?
11:25
SPEAKER_00: to bring a machine to life.
11:27
And that's the goal of people who design cognitive
11:30
architectures is they're trying to,
11:31
SPEAKER_00: I guess,
11:32
SPEAKER_00: deconstruct the brain into pieces,
11:34
SPEAKER_00: simulate each piece in software,
11:36
put it back together and like watch the thing boot up and see
11:39
how it behaves.
11:40
SPEAKER_00: But what happened over that time,
11:42
SPEAKER_00: like between,
11:43
SPEAKER_00: I guess,
11:43
SPEAKER_00: 2018 and like 2023,
11:46
SPEAKER_00: is deep learning kept getting better and better and better.
11:49
SPEAKER_00: So now in this field,
11:51
everyone thinks you just train an end-to-end deep learning
11:54
model differently.
11:55
on multimodal data and you don't need to build the mind piece
11:59
by piece anymore.
12:00
SPEAKER_00: You just have,
12:01
SPEAKER_00: like with large language models,
12:03
SPEAKER_00: you just have a giant amount of data,
12:05
you train one giant network and then the robot will kind of
12:08
like behave as it's
12:10
SPEAKER_00: been shown with its training data.
12:13
SPEAKER_02: What do you make of that thesis?
12:15
Because I spent some time with physical intelligence and with
12:19
some researchers in
12:20
SPEAKER_02: NVIDIA's group and with figure AI.
12:23
SPEAKER_02: And they're all exactly to your point.
12:25
SPEAKER_02: I mean, it was this idea that we were doing this robotics work.
12:29
SPEAKER_02: Then we solved math, we solved language mathematically.
12:34
SPEAKER_02: And that
12:37
We could piggyback off all this LLM work and infuse it into
12:41
robots and they would
12:42
somehow gain this general purpose knowledge of the world where
12:45
you like set
12:46
SPEAKER_02: something on the table and now you just have to say in language,
12:50
SPEAKER_02: pick up the cup and the robot,
12:53
SPEAKER_02: you know,
12:53
SPEAKER_02: discerns a very different way than the way we used to train this,
12:57
SPEAKER_02: which is like you'd have to show it.
12:59
SPEAKER_02: Thousands of cups have it pick it up tens of thousands of times.
13:02
And now you're kind of using language to get this head start
13:05
about context in the world.
13:06
SPEAKER_02: Do you I mean, this this in the last year.
13:11
SPEAKER_02: Hundreds of millions,
13:11
if not billions of dollars have been pumped into the startups
13:14
taking this approach.
13:16
SPEAKER_02: I was so curious what you make of it.
13:18
SPEAKER_00: Yeah.
13:19
SPEAKER_00: So I have two main thoughts on this.
13:20
SPEAKER_00: I have to be careful not to get into like the hot water here and,
13:24
SPEAKER_00: you know,
13:24
SPEAKER_00: say anything bad about it.
13:25
SPEAKER_00: No, no.
13:25
SPEAKER_01: We welcome the hot water.
13:28
SPEAKER_00: My first thought is it might work.
13:31
SPEAKER_00: I think the amount of data required is going to be ridiculous.
13:35
SPEAKER_00: Yeah.
13:36
If you look at how much data was used to train some of the
13:40
large language models,
13:41
SPEAKER_00: it was already all the data we have on the internet.
13:44
SPEAKER_00: It was just immense amount of data.
13:46
SPEAKER_00: But if you think about a humanoid robot now,
13:48
SPEAKER_00: it's not just one modality like text,
13:50
SPEAKER_00: like letters of the alphabet.
13:52
SPEAKER_00: It's like high definition camera feeds,
13:55
SPEAKER_00: audios,
13:56
SPEAKER_00: touch sensors,
13:58
SPEAKER_00: motor,
13:59
SPEAKER_00: position sensors from all over the body.
14:00
SPEAKER_00: There's
14:01
SPEAKER_00: So the data is much richer.
14:03
So in order to learn a model that like predicts and controls
14:08
all those different
14:08
SPEAKER_00: kinds of modalities,
14:09
SPEAKER_00: not just text,
14:11
it's you need like an exponential amount of data more than what
14:14
we needed for
14:15
SPEAKER_00: language models.
14:16
SPEAKER_00: So I'm not saying it won't work.
14:18
I actually think it will work in a weird way where if you stick
14:22
within the training data,
14:24
the robots will actually be able to learn to do things
14:28
autonomously and they'll
14:29
SPEAKER_00: generalize a little bit.
14:30
SPEAKER_00: But what I'm nervous about is it won't be like a person.
14:35
SPEAKER_00: So in my mind,
14:36
there's still something missing because people can go into a
14:40
general environment
14:41
that they've never seen before and start making smart decisions
14:44
and doing things
14:45
SPEAKER_00: without having seen training data of that type.
14:48
And I've just never seen an example of a humanoid robot
14:52
generalizing in that same way.
14:54
SPEAKER_00: So you could say,
14:55
SPEAKER_00: oh,
14:55
SPEAKER_00: well,
14:55
maybe humans have just seen like trillion gazillion episodes of
15:00
training data from
15:01
SPEAKER_00: robots haven't yet.
15:02
But there's another way to think about it is maybe we're
15:03
actually using a different
15:05
SPEAKER_00: process at that point.
15:06
SPEAKER_00: Maybe we're not just using like trained model.
15:09
SPEAKER_00: Maybe there's something else going on.
15:10
And that's like my second point,
15:11
which is I think there's something else going on.
15:13
SPEAKER_00: And I think it's related to consciousness and our conscious awareness.
15:17
SPEAKER_02: Yeah, which I 100% want to talk about.
15:20
SPEAKER_02: But really quickly,
15:21
SPEAKER_02: I mean,
15:21
maybe we don't have to call out any of the startups by name if
15:25
you don't want to
15:26
SPEAKER_02: badmouth them.
15:27
SPEAKER_02: But generally,
15:28
so I've had the same questions because I've gone into some of
15:32
these labs and I see them.
15:35
SPEAKER_02: A big part of their thesis is,
15:36
SPEAKER_02: you know,
15:37
SPEAKER_02: I can't remember when I went to visit Kindred,
15:39
SPEAKER_02: but it must have been for the first time.
15:41
SPEAKER_02: It was probably like eight years ago or something.
15:43
SPEAKER_01: Yeah, that's right.
15:44
SPEAKER_02: Yeah.
15:45
SPEAKER_02: And you had these arms doing very similar tasks to what I'm seeing now.
15:50
When I walk into these startups,
15:52
they just have arms picking up objects,
15:54
folding shirts.
15:56
SPEAKER_02: Mm-hmm.
15:56
You know,
15:57
just doing these closing microwave doors,
16:00
just repetitive things.
16:02
SPEAKER_02: And when I talk to the companies,
16:03
SPEAKER_02: I mean,
16:04
they're arguing that even though companies like Google and
16:09
Kidred did this kind of
16:10
SPEAKER_02: thing in the past,
16:11
SPEAKER_02: that they are going to do this to your point.
16:15
SPEAKER_02: Yeah.
16:16
SPEAKER_02: on a higher level of volume so that we get more amounts of data.
16:21
SPEAKER_02: But I don't know if I'm sold on that thesis.
16:24
SPEAKER_02: I mean, even that part, is that even possible?
16:28
SPEAKER_02: Like, are they right that
16:30
even though companies doing this type of work existed before
16:34
they weren't doing it
16:35
SPEAKER_02: at kind of like the scale that we could get to now.
16:38
SPEAKER_00: Yeah.
16:38
It's a really interesting question because if you'd asked the
16:41
same thing about like
16:42
SPEAKER_00: language models back in 2013,
16:44
we would have been wrong because language models weren't able
16:48
to do what they do now.
16:49
SPEAKER_00: And so maybe it is just a volume of data.
16:52
SPEAKER_00: problem, but I don't know.
16:54
SPEAKER_00: I'm skeptical.
16:55
SPEAKER_00: I mean, you've mentioned you've been to a lot of these robot companies.
16:59
SPEAKER_00: Have you ever just been seeing the robot demo,
17:02
SPEAKER_00: and then it's moving things from place to place,
17:04
SPEAKER_00: and then it suddenly stops and turns around and looks at you and goes,
17:08
SPEAKER_00: hey,
17:08
SPEAKER_00: Ashley's here.
17:10
I've never had that spark of life moment with any autonomous
17:15
robot I've interacted with.
17:16
SPEAKER_00: The only time I've had it is when a robot's been teleoperated by person.
17:21
SPEAKER_00: and the person is hyper aware of what's going on around them,
17:24
SPEAKER_00: they imbue that into the robot,
17:26
SPEAKER_00: and then the robot cat comes to life.
17:28
There's something fundamentally different when a person is
17:31
teleoperating that you
17:32
SPEAKER_00: just don't get.
17:33
SPEAKER_02: Okay, so...
17:35
It sounds like you're skeptical,
17:38
but my sense is you don't see this as total smoke and mirrors.
17:45
It's like possibly a thesis worth exploring because they might
17:51
crack the code on
17:52
SPEAKER_02: some sort of store of data.
17:54
SPEAKER_00: I certainly don't think it's smoke and mirrors.
17:56
SPEAKER_00: So what I think is it's a question of generalization.
17:59
So these robots will always be able to do something if you make
18:02
the task narrow enough.
18:04
SPEAKER_00: And just as a really limiting case of this,
18:07
SPEAKER_00: there's something called pilot replay,
18:09
SPEAKER_00: which is you teleoperate,
18:10
SPEAKER_00: and then if everything is in exactly the same position,
18:14
SPEAKER_00: the robot will do the task perfectly autonomously.
18:17
SPEAKER_00: So that's one extreme limiting case of zero generalization.
18:21
Then what you can do is you can maybe move all the things you
18:23
have to pick up by a
18:24
SPEAKER_00: millimeter and ask,
18:25
SPEAKER_00: okay,
18:26
SPEAKER_00: is the model generalizing?
18:27
Maybe you change the lighting condition slightly and say,
18:29
does it generalize to that?
18:31
But what I think is going to happen is this kind of circle of
18:35
generalization is not
18:37
SPEAKER_00: going to widen as fast as people think it is.
18:40
So the robot will have had to have seen everything perfectly
18:44
before in its training
18:45
SPEAKER_00: data properly.
18:46
SPEAKER_00: to automate it.
18:47
SPEAKER_00: Now,
18:47
that's fine if you're on like a factory assembly line or
18:50
something like that,
18:50
SPEAKER_00: where things don't change very much.
18:52
But you take that robot and you put it out into the street in
18:55
the middle of,
18:56
SPEAKER_00: you know,
18:56
SPEAKER_00: traffic,
18:57
SPEAKER_00: a bustling environment with people,
18:58
SPEAKER_00: you know,
18:59
SPEAKER_00: trying to kick it and all this kind of stuff.
19:01
I just don't think it's going to work in those kind of really
19:05
ultra general environments.
19:06
SPEAKER_02: Okay.
19:07
SPEAKER_02: I mean,
19:07
SPEAKER_02: that's like,
19:08
that would be a problem because that seems to be where the
19:12
Android type robot folks want.
19:16
SPEAKER_00: That's right.
19:16
SPEAKER_00: And the home, I often say the home is the ultimate test of AGI.
19:21
SPEAKER_00: It's like, we're so familiar with our homes and we think things like
19:24
Cleaning is so boring and mundane,
19:26
but these are actually what I call AGI complete tasks.
19:30
Try cleaning an average American household with all the
19:35
completely random things
19:36
SPEAKER_00: you could come across.
19:38
Trying to do that with kids running around and dogs and all
19:41
this kind of stuff.
19:43
SPEAKER_00: It's a very hard problem.
19:44
SPEAKER_00: It's very general.
19:45
I saw a couple of these companies doing the demos where the
19:48
robot's going to the
19:49
SPEAKER_02: fridge and picking things out.
19:51
SPEAKER_02: And then on X,
19:53
SPEAKER_02: people were like,
19:53
SPEAKER_02: oh,
19:53
this is like your stoned roommate trying to make a sandwich
20:00
after a late night out
20:01
SPEAKER_02: or something.
20:02
SPEAKER_02: You must have reached this conclusion in many ways a bit
20:12
SPEAKER_02: Okay, let me come into this a slightly different way.
20:14
Because what you're working on now,
20:16
I think,
20:17
is not just what we were talking about.
20:19
SPEAKER_02: I mean, you've been hunting for consciousness as well.
20:22
So there's two parts to this puzzle,
20:24
which is the robots doing interactions out in the world.
20:31
SPEAKER_02: And then there's been this other idea, which is...
20:35
SPEAKER_02: If we build these neural networks sufficiently large,
20:40
that they start to take on a level of complexity that starts to
20:44
resemble the brain.
20:46
SPEAKER_02: And obviously,
20:47
SPEAKER_02: as far as I'm aware,
20:48
nobody has a super clear answer on where consciousness lights
20:53
up or the soul or
20:54
SPEAKER_02: anything like that.
20:55
SPEAKER_02: But this theory that...
20:57
SPEAKER_02: well,
20:57
maybe we're just heading on this path where we make things
21:00
complex enough with
21:01
SPEAKER_02: classical computers and it does light up one day.
21:06
SPEAKER_02: And your philosophy,
21:08
SPEAKER_02: as I understand it,
21:09
SPEAKER_02: and your new company is kind of a rejection of that idea.
21:14
SPEAKER_02: Is that fair?
21:15
SPEAKER_02: And did you maybe walk me through what you're trying to do?
21:18
SPEAKER_02: And you must have seen limitations on these multiple fronts that led you.
21:26
yeah so okay there's a lot to unpack there so maybe let me
21:30
start by I guess posing
21:32
a question to the audience is so first of all we have to sort
21:35
of define what
21:36
consciousness is because people have different definitions for
21:38
it some people think
21:39
it equates to like thinking or reasoning or something like that
21:42
so my definition of
21:44
consciousness at least from the what is it what does it look
21:47
like to us is it's
21:48
this concept that we have this subjective experience so we feel
21:52
like we're inside a
21:53
SPEAKER_00: movie that's playing out around us it
21:55
SPEAKER_00: It feels like something to be us.
21:58
SPEAKER_00: That's kind of what consciousness is.
21:59
SPEAKER_00: And,
22:00
SPEAKER_00: you know,
22:01
SPEAKER_00: we see it visually,
22:02
SPEAKER_00: but we also have all these feelings inside us,
22:04
SPEAKER_00: like emotions,
22:05
SPEAKER_00: feelings,
22:05
SPEAKER_00: sensations.
22:06
SPEAKER_00: We're aware of things.
22:07
SPEAKER_00: So that's how I think of consciousness.
22:10
SPEAKER_00: And we can do things unconsciously.
22:14
SPEAKER_00: and not be aware that we're doing them.
22:16
SPEAKER_00: So a good example is driving.
22:18
So if you've been driving for 20 years,
22:20
sometimes you even forget that you drove to work.
22:22
SPEAKER_00: You can't even remember being in the car.
22:25
SPEAKER_00: You just somehow got there.
22:26
SPEAKER_00: So that process,
22:28
SPEAKER_00: after having seen a lot of examples of driving,
22:31
SPEAKER_00: has become...
22:32
SPEAKER_00: It's sort of gone into your unconscious.
22:34
SPEAKER_00: It's gone into your muscle memory.
22:36
SPEAKER_00: It's gone into your autopilot.
22:38
So the models I'm working with are...
22:40
It's like the first time we see something...
22:43
SPEAKER_00: Like the first time you ever drive a car,
22:45
SPEAKER_00: you're like hyper aware,
22:46
SPEAKER_00: you're hyper conscious of that.
22:48
SPEAKER_00: But if you do it over and over and over again, it becomes automatic.
22:52
SPEAKER_00: So the question is, why don't we think about AI in the same way?
22:56
We're giving AI all this training data,
22:58
all this training data that comes from people.
23:00
SPEAKER_00: And we're creating this like unconscious automatic routines.
23:04
But we've completely skipped the part where those systems are
23:07
able to do that for
23:08
SPEAKER_00: themselves to like
23:11
SPEAKER_00: Create that training data in the first place in an online way.
23:14
SPEAKER_00: So in my world,
23:15
SPEAKER_00: if I had a conscious robot and I put it in a car,
23:20
SPEAKER_00: it would be like its first driving lesson ever.
23:23
SPEAKER_00: It would be like, oh, my goodness, what am I doing?
23:25
SPEAKER_00: I feel terrified.
23:26
SPEAKER_00: Like it's grabbing the wheel.
23:27
SPEAKER_00: It's trying to learn to drive the car.
23:29
And it's using that conscious decision-making process to kind
23:32
of now create its own
23:33
SPEAKER_00: training data that over time can become these unconscious subroutines.
23:39
SPEAKER_00: And we're sort of doing that part for it.
23:42
SPEAKER_00: So my question is,
23:42
SPEAKER_00: well,
23:43
can we create that conscious module that in the first place
23:46
allows it to create its
23:48
SPEAKER_00: own training data?
23:49
SPEAKER_00: Does that help?
23:50
SPEAKER_02: Does that make sense?
23:52
SPEAKER_02: It does.
23:53
SPEAKER_02: I like the way...
23:56
SPEAKER_02: Yeah, the order of operations that you laid out is really fascinating.
23:59
SPEAKER_02: Clearly,
24:00
SPEAKER_02: you know,
24:00
SPEAKER_02: in the last year or so,
24:01
we've moved from some of what you were talking about with LLMs
24:06
to this reasoning era.
24:08
SPEAKER_02: I mean, my sense is that people seem to think that...
24:14
SPEAKER_02: You know,
24:14
SPEAKER_02: like on one level,
24:15
SPEAKER_02: we could argue that these LLMs,
24:16
they're just these statistical machines and they're just
24:21
prediction machines.
24:23
SPEAKER_02: And, you know, there's not – it's amazing.
24:28
SPEAKER_02: Like there's a lot that goes into that.
24:29
SPEAKER_02: But other than that,
24:30
there's maybe not that much under the hood that at all
24:33
resembles thinking or
24:35
SPEAKER_02: consciousness on any level.
24:37
SPEAKER_02: And then now we've moved to this reasoning era where,
24:40
SPEAKER_02: again,
24:41
SPEAKER_02: it's based on these same principles.
24:45
SPEAKER_02: working through problems in a kind of human-like way?
24:49
SPEAKER_02: I mean,
24:49
SPEAKER_02: even with that stuff,
24:50
you're not seeing hints that convince you otherwise that
24:54
something's changed in the
24:56
SPEAKER_02: last year or so.
24:57
SPEAKER_00: No, I don't think any AI today is conscious.
25:01
SPEAKER_00: That doesn't mean that they can't do useful things.
25:04
SPEAKER_00: And again, this is like where this definition of consciousness comes in.
25:07
SPEAKER_00: I think we can,
25:09
SPEAKER_00: like,
25:10
SPEAKER_00: if you took the consciousness of a person away,
25:13
they'd still be able to do a lot of things because they've
25:15
already learned how to
25:16
SPEAKER_00: do those and they've become,
25:17
SPEAKER_00: those things have become unconscious routines.
25:20
SPEAKER_00: But that system will be kind of like a zombie.
25:24
It will be doing things that it's done before in the way it's
25:28
done them before,
25:28
SPEAKER_00: and it will have no experience.
25:31
SPEAKER_00: It will have no feeling.
25:32
SPEAKER_00: It will have no subjectivity.
25:34
And what I'm saying is I think it's that subjectivity that
25:37
gives us this ability to
25:38
SPEAKER_00: go into a new environment and do something in a different way.
25:42
So this is where consciousness research gets a little kind of
25:45
crazy and interesting,
25:47
SPEAKER_00: if you like the other word for crazy.
25:50
SPEAKER_00: But I think when we're in this conscious awareness mode,
25:52
I think we're using a different kind of decision making process
25:56
that allows us to
25:57
make decisions in places we've never been in before where we
26:00
don't have training data.
26:02
SPEAKER_00: You might call it intuition,
26:04
SPEAKER_00: I think is the closest word we have in the language,
26:06
SPEAKER_00: but it's certainly not prediction.
26:08
SPEAKER_00: It's not memory.
26:10
SPEAKER_00: It's sort of like good feeling or intuition of what to do.
26:15
SPEAKER_02: Yeah.
26:15
SPEAKER_02: Okay.
26:15
SPEAKER_02: So where LLMs and current AI is very brute force,
26:20
SPEAKER_02: it's missing all of this nuance and those funny feelings that we have.
26:25
And you don't see a path even as the current models add
26:31
complexity for that just to
26:33
SPEAKER_02: infuse itself in or to appear out of nowhere.
26:36
SPEAKER_00: I think it's still the way around.
26:37
SPEAKER_00: I think it has to come from the
26:39
So we're building these increasingly complex hierarchies of
26:43
unconscious behaviors
26:44
and expecting consciousness to pop out of them when I think
26:48
it's more of like a
26:50
SPEAKER_00: fundamental layer,
26:51
SPEAKER_00: like consciousness is what creates those behaviors.
26:54
SPEAKER_00: in people and in animals.
26:55
SPEAKER_00: I think consciousness is what creates that triniator in the first place.
26:58
SPEAKER_00: It allows you to create those unconscious routines.
27:00
SPEAKER_00: So it's sort of like flipping it on its head.
27:03
SPEAKER_00: And maybe a fun thing you can think about is,
27:06
so I have like a microscope at home and I love looking at
27:09
little microorganisms.
27:10
SPEAKER_00: Of course you do.
27:13
SPEAKER_00: I know, I thought you might like this, actually.
27:15
SPEAKER_00: But just another one of my weird hobbies.
27:18
SPEAKER_00: Yeah.
27:19
But so you've got to like go to the pond down the road and you
27:22
scoop some pond
27:23
slime and then you put it under the microscope and you see all
27:25
these little
27:26
SPEAKER_00: creatures and you can interact with them.
27:28
You can like poke them with a needle and some of them are
27:30
really small and just
27:32
SPEAKER_00: watch what they do.
27:33
SPEAKER_00: They are very smart.
27:35
SPEAKER_00: So they're like zipping around looking for food.
27:37
SPEAKER_00: But if you start interacting with them, they go into a different mode.
27:40
SPEAKER_00: They start like hiding.
27:41
SPEAKER_00: They'll run away from you.
27:42
SPEAKER_00: You will not be able to catch them.
27:44
SPEAKER_00: It's kind of crazy.
27:44
SPEAKER_00: And they'll like have all these super interesting behaviors.
27:48
SPEAKER_00: They don't have a single neuron.
27:49
SPEAKER_00: They don't have a nervous system.
27:52
SPEAKER_00: Right.
27:53
SPEAKER_00: They're making decisions somehow.
27:55
SPEAKER_00: Right.
27:56
without any of the things we would think are now going into
28:00
like massive trained AI networks.
28:02
SPEAKER_00: So if you think about it,
28:04
it's kind of an interesting question is what inside that little
28:07
dude is making
28:08
SPEAKER_00: those decisions and having that
28:10
little 5 micron guy be hyper aware of what's going around it so
28:14
it can survive I'm
28:16
pretty sure that guy from an evolutionary perspective has never
28:19
had like a pencil
28:21
poked in its face so it's a totally novel situation yeah and so
28:26
this has to be
28:27
built in even at that most basic level and I think
28:31
consciousness is in that level I
28:33
think it's a very deep thing that gets deep into biology and
28:36
physics so I think
28:37
SPEAKER_00: consciousness came first
28:39
SPEAKER_00: And then learning came from consciousness, not the other way around.
28:44
SPEAKER_02: Okay, okay.
28:45
SPEAKER_02: And so let's run with that.
28:48
SPEAKER_02: And you are a quantum physicist,
28:52
and I'm about to enter territory where I'm going to embarrass
28:56
myself terribly.
28:58
SPEAKER_02: But, you know, I've seen you with your new company.
29:00
SPEAKER_02: I've seen you make this argument.
29:02
SPEAKER_02: So why...
29:05
SPEAKER_02: You seem to be saying that quantum properties,
29:08
SPEAKER_02: possibly quantum computing elements,
29:11
SPEAKER_02: lend themselves better to what you're describing.
29:15
SPEAKER_02: Okay,
29:15
SPEAKER_02: so for the rest of us,
29:17
why would quantum properties be at the heart of that
29:22
fundamental consciousness,
29:23
SPEAKER_02: possibly,
29:24
SPEAKER_02: whereas what we do with these classical computers,
29:29
SPEAKER_02: just these transistors going on and off,
29:31
SPEAKER_02: are not?
29:32
SPEAKER_00: Mm-hmm.
29:32
SPEAKER_00: Yeah, okay.
29:34
So there's different ways of tackling this problem,
29:36
different ways of thinking about it.
29:38
SPEAKER_00: What got me interested in it,
29:39
SPEAKER_00: first of all,
29:40
was some work that was done by Sir Roger Penrose and Stuart
29:43
Hameroff back in the day.
29:45
SPEAKER_00: And they jointly came up with a theory,
29:48
like Penrose came up with a theory that maybe consciousness is non-computable.
29:52
It's like not something you could simulate with a digital computer,
29:56
if you like.
29:57
SPEAKER_00: And he has a bunch of mathematical reasons for why that might be.
30:00
SPEAKER_00: So he was thinking about it from a very abstract perspective,
30:03
SPEAKER_00: sort of mathematical,
30:04
SPEAKER_00: philosophical argument.
30:06
SPEAKER_00: And then Stuart Hameroff came along and said,
30:07
SPEAKER_00: hey,
30:08
there are these structures in the brain called microtubules
30:10
that might actually
30:12
SPEAKER_00: have a mechanism
30:13
through which you could have consciousness happening,
30:16
and it relates to quantum physics.
30:18
SPEAKER_00: And so Penrose really liked this,
30:20
and it's related to this not being able to compute things using
30:24
a digital computer.
30:25
Because if you think consciousness is not a digital computable phenomenon,
30:30
SPEAKER_00: then you have to say,
30:30
SPEAKER_00: well,
30:30
SPEAKER_00: what is it?
30:31
SPEAKER_00: It has to be something else.
30:33
And the only other thing we really have right now is something
30:36
that goes deeper
30:37
SPEAKER_00: than digital computation,
30:38
SPEAKER_00: which is quantum information.
30:41
So you have to find a type of computer that you can't simulate
30:44
with a digital
30:44
computer if you're looking for consciousness being something
30:48
that's not digital,
30:48
SPEAKER_00: if that makes sense.
30:50
SPEAKER_02: I mean, I get where you're going.
30:51
I guess,
30:52
I mean,
30:52
couldn't people just argue like,
30:54
okay,
30:54
if it's not this...
30:58
digital phenomenon that we know i mean couldn't it be anything
31:01
why does it have to
31:01
be microtubules and right properties it's kind of too hot two
31:06
different parts to
31:07
the argument so the penrose part is just like forget about the
31:10
microtubules and the
31:11
biology uh penrose thinks that our brain okay i just said
31:15
forget about the biology
31:16
but then the brain like penrose thinks that our mind our
31:21
thoughts are not able to
31:23
SPEAKER_00: be simulated and
31:24
SPEAKER_00: on a digital computer.
31:25
So then the question is,
31:27
what physics are they using that's different than classical physics?
31:32
Because all classical physics can be simulated on a digital computer,
31:36
on a Turing machine.
31:37
So then Penrose thinks that there are certain things in quantum
31:41
physics that might fit the bill.
31:44
SPEAKER_00: And there's one thing in particular,
31:46
which is so there's some weird prophecies of quantum physics
31:48
like superposition,
31:50
SPEAKER_00: which means to a zero and a one that can be in the same state at once.
31:54
SPEAKER_00: So it could be like a zero one combination state.
31:58
SPEAKER_00: So that's one interesting thing that goes beyond standard computation.
32:01
SPEAKER_00: Then there's something called entanglement.
32:04
which basically means if you've got two bits of classical
32:07
information and you put
32:08
SPEAKER_00: them on opposite sides of the galaxy,
32:11
the only way they can communicate is via sending information at
32:14
the speed of light.
32:15
If you have two quantum bits and you put them on opposite ends
32:18
of the galaxy,
32:18
SPEAKER_00: they can still kind of interact with each other,
32:20
SPEAKER_00: or they still have correlations,
32:22
SPEAKER_00: even though they're really far apart.
32:24
SPEAKER_00: So that's entanglement, and that's kind of weird.
32:26
And then the weirdest one of all is called the measurement
32:28
problem or the measurement process.
32:30
So you take a highly quantum mechanical system and then you
32:34
measure it or you observe it.
32:36
SPEAKER_00: It chooses a classical reality to end up in.
32:40
SPEAKER_00: So before you've done this measurement process,
32:42
it's sort of in a weird fuzzy probability of all possible
32:46
things that could happen.
32:48
SPEAKER_00: And then once you observe it, it chooses it.
32:51
one particular reality to end up in so the reason people
32:55
thought this might be
32:55
related to consciousness is that we seem to have this agency
32:59
this ability to
33:01
consider lots of options and then choose like make a choice and
33:05
so the in the
33:07
quantum consciousness quantum agency community that's the
33:10
process by which this
33:11
choice is made is this quantum measurement process so this is
33:14
like this bond of
33:17
SPEAKER_02: Our concept of free will or something like that with computing.
33:22
SPEAKER_00: Yeah.
33:22
So people think that there might be a loophole that lets us
33:26
have our free will if
33:28
SPEAKER_00: we switch to quantum computing.
33:29
SPEAKER_00: Because in classical computing, there's absolutely no room for free will.
33:33
SPEAKER_00: Because everything is just like clockwork, right?
33:36
SPEAKER_00: It's just computation.
33:38
SPEAKER_02: So, like, I'm way out of my depth on the quantum side.
33:41
SPEAKER_02: What you're saying,
33:41
SPEAKER_02: though,
33:42
SPEAKER_02: I mean,
33:42
SPEAKER_02: there is,
33:42
SPEAKER_02: like,
33:43
SPEAKER_02: this intuitive part of me that is,
33:45
SPEAKER_02: you know,
33:45
SPEAKER_02: there's always been something weird to me about LLMs,
33:49
SPEAKER_02: that it is some of the things I was talking about before,
33:52
SPEAKER_02: this brute force statistical...
33:57
SPEAKER_02: Like, it doesn't quite feel like intelligence, right?
34:00
SPEAKER_02: I mean, it is on some level.
34:02
SPEAKER_02: Clearly, it can do very interesting things.
34:04
SPEAKER_02: But it feels like, yeah, you've reversed engineered these properties.
34:09
SPEAKER_02: And it's clearly... Like, what I...
34:14
SPEAKER_02: and I feel like this is borne out when I,
34:16
SPEAKER_02: when I was reading the stories about that,
34:18
SPEAKER_02: that Google guy freaking out when he was talking to the LLMs at first.
34:23
SPEAKER_02: Um, like to me, it was just, it's, I don't know.
34:26
SPEAKER_02: To me, it's clear that this, I feel like it doesn't have a personality.
34:29
I don't think I'm having this,
34:31
this super existential metaphysical experience with these things.
34:36
And so,
34:36
so I don't know,
34:37
just like intuitively I can see,
34:39
see where you're,
34:40
you're going.
34:41
SPEAKER_02: Um, and then,
34:45
SPEAKER_02: And then we're in this,
34:46
we're like in this interesting time where after all these
34:49
decades of quantum
34:51
SPEAKER_02: computing being chased after and sort of overhyped,
34:59
SPEAKER_02: I guess,
35:00
SPEAKER_02: it seems like we've kind of turned a quarter,
35:01
SPEAKER_02: right?
35:02
SPEAKER_02: I mean, so like are the...
35:06
Would you be playing off some of these quantum computing chips
35:11
people are developing right now?
35:12
SPEAKER_02: Or is this you have to follow like a totally different path?
35:16
SPEAKER_00: Yeah, absolutely.
35:17
So I think what's happened recently that sort of made this what
35:21
I'm trying to do
35:21
possible is that quantum computers are now mature enough that
35:24
we can actually run.
35:26
reasonable number of different algorithms on them they're
35:29
they're not there yet
35:30
like the quantum computing industry is still playing catch-up
35:33
with the classical
35:34
and in fact it's kind of a weird race because the classical
35:37
computing industry is
35:38
still still accelerating like with gpus and things so they're
35:41
sort of racing
35:42
against something that's trying to get away from them at the
35:44
same time and
35:46
SPEAKER_00: So that's interesting, right?
35:47
Because people have been trying to cope with quantum algorithms
35:49
and show that
35:49
SPEAKER_00: quantum computers are faster and better.
35:51
But what they're doing is they're applying them to the same
35:53
problems that people
35:54
SPEAKER_00: are solving with classical computers.
35:56
SPEAKER_00: So it's this like arms race constantly.
35:59
And what I'm saying is,
36:01
well,
36:01
maybe this consciousness thing is a problem that...
36:04
SPEAKER_00: just can't be solved by a classical computer at all.
36:07
Because a classical computer doesn't have the ability to choose
36:10
a reality through
36:13
SPEAKER_00: this measurement process.
36:14
SPEAKER_00: There's just no way to simulate that classically.
36:17
SPEAKER_00: So if...
36:19
If we can get these algorithms to a mature enough state where
36:23
they're able to control,
36:24
SPEAKER_00: say,
36:24
SPEAKER_00: physical robots and make choices and make decisions,
36:27
I think the quantum computers are doing something fundamentally
36:30
different at that
36:31
SPEAKER_00: point than the classical computers.
36:33
SPEAKER_00: Now, people might say, oh, well, that's not consciousness.
36:36
But even if you don't believe that,
36:38
they're still doing something radically different.
36:40
SPEAKER_00: They're using some resource to make a decision.
36:44
that might be smarter than a classical computer we can put put
36:48
these algorithms to
36:49
SPEAKER_00: work and so we have chips now that we can run these things on and
36:52
SPEAKER_02: Okay.
36:53
SPEAKER_02: And I mean, I was like trying to imagine, I mean, you have a startup.
36:58
I saw in a couple spots,
37:00
you were giving a talk about like,
37:02
it seemed like this.
37:04
SPEAKER_02: Okay.
37:04
SPEAKER_02: Part of it seemed like a very researchy.
37:07
SPEAKER_02: And then part of it seemed like,
37:09
SPEAKER_02: no,
37:09
SPEAKER_02: yes,
37:09
you actually want to make products that you would sell based
37:12
off all this stuff.
37:13
SPEAKER_02: What on earth is like a day-to-day business?
37:16
SPEAKER_02: What are you working on?
37:17
SPEAKER_02: Okay,
37:18
SPEAKER_02: if we're going to bring this chase,
37:21
SPEAKER_02: this idea of consciousness appearing in these quantum machines,
37:25
SPEAKER_02: what is your day-to-day work like?
37:28
I have no idea what this would even look like or how it would
37:31
manifest itself.
37:32
SPEAKER_00: Yeah.
37:33
SPEAKER_00: So the way I'm testing it is,
37:37
SPEAKER_00: firstly,
37:37
I think you have to put the quantum system in something like
37:40
you have to put it in
37:41
SPEAKER_00: control of an AI or in control of a robot to see this sort of effect.
37:45
You need to set it up so it's making decisions that allow it to
37:50
take action.
37:50
SPEAKER_00: And then those actions affect a feedback loop.
37:53
SPEAKER_00: So it basically needs to be in a feedback loop with its environment.
37:56
SPEAKER_00: And that's what we see in life.
37:58
Like all life we think is conscious or has consciousness is
38:01
always in a perception
38:02
SPEAKER_00: action feedback loop with its environment.
38:05
So what's happened up till now is quantum computing researchers
38:07
have taken quantum
38:08
SPEAKER_00: systems and tried to isolate them from the environment.
38:11
SPEAKER_00: And there you can see these sort of like weird quantum properties.
38:15
The quantum properties in their potential states,
38:19
but they're not actually doing anything.
38:20
SPEAKER_00: So they just sit there and be weird and quantum.
38:24
SPEAKER_00: If you can connect them to a physical system,
38:27
SPEAKER_00: you can put them in this cyclical loop where they take in information,
38:32
SPEAKER_00: put that into a quantum state,
38:33
SPEAKER_00: and then that quantum state collapses and now...
38:36
what reality it chooses becomes like the action action that the
38:39
robot takes.
38:40
SPEAKER_00: So very practically, what does this look like?
38:42
SPEAKER_00: Well,
38:43
I have some small robot systems that I'm working with and I've
38:46
worked with these
38:47
SPEAKER_00: kind of things for many years.
38:48
SPEAKER_00: So it's easy to get the perception data coming into the robot and actions
38:53
SPEAKER_00: out of the robot through the motor system, that works in real time.
38:57
SPEAKER_00: So it's very important that it's like online.
38:59
And then instead of just having say a neural net that takes in
39:04
perception and does
39:05
SPEAKER_00: an action prediction,
39:05
SPEAKER_00: you have a quantum neural net.
39:08
SPEAKER_00: So again, how do you do that?
39:09
SPEAKER_00: Well, there's quantum computers available today.
39:12
SPEAKER_01: That was my question.
39:13
SPEAKER_00: Yeah.
39:14
So there's quantum computers available today that you can use
39:17
as a variety of
39:18
SPEAKER_00: different ones that offer.
39:19
SPEAKER_00: And if you go onto Azure or AWS
39:22
Or sometimes the company's homepage,
39:24
they have their own like private cloud thing.
39:26
SPEAKER_00: You can actually download an SDK of your choosing,
39:30
SPEAKER_00: whichever code,
39:31
SPEAKER_00: you know,
39:31
SPEAKER_00: whichever programming language you like.
39:33
And you can actually program quantum algorithm and then it will
39:36
send the problem or
39:39
SPEAKER_00: program off to the quantum computer and the answer comes back.
39:42
SPEAKER_00: So it's a cloud enabled system.
39:45
SPEAKER_00: Like quantum is not a weird future thing.
39:47
SPEAKER_00: It's here today.
39:48
SPEAKER_00: Now it might be noisy,
39:50
SPEAKER_00: it might be like small numbers of qubits,
39:52
SPEAKER_00: you know,
39:52
SPEAKER_00: it might not have scaled yet,
39:54
SPEAKER_00: but the end-to-end thing is there.
39:56
SPEAKER_00: So I just like code in Python,
39:59
I write my code and then I'll have a quantum algorithm and then
40:04
I'll send off the
40:05
actual part that queries the quantum state and does this
40:08
collapse process.
40:09
I have write some code that does that and then the quantum
40:12
computer in the cloud
40:13
SPEAKER_00: actually does it,
40:13
SPEAKER_00: sends me back the answer and then I use the answer to control the robot.
40:18
SPEAKER_02: Like what kind of robots?
40:19
SPEAKER_02: Like the ones behind you or arms?
40:20
SPEAKER_00: Yeah, this one I can show you.
40:24
these are small small robots this is you may remember this guy
40:28
yeah yeah i remember
40:30
those running around yeah so this is just like um it's kind of
40:35
made from a kit so
40:36
it's not an off-the-shelf robot you can buy but um we have some
40:40
of the similar
40:41
things at kindred and sanctuary they're kind of really good
40:43
test platforms for
40:44
SPEAKER_00: prototyping codon so it's just a small like where do you order that from
40:49
SPEAKER_00: Well, it's just parts, so I build them from scratch.
40:53
SPEAKER_02: When you said a kit, I didn't know if you could buy the whole kit.
40:55
SPEAKER_00: Oh, no, I just buy the parts individually.
40:57
SPEAKER_00: But you can actually buy some similar kind of quadrupeds,
41:01
SPEAKER_00: but they don't have exactly all the right specs that I want.
41:04
SPEAKER_00: So I tend to just build my own.
41:06
Okay,
41:06
so you have this kind of animal,
41:08
like,
41:09
for people who are just listening,
41:10
Suzanne.
41:11
SPEAKER_00: Oh, sorry, I should be more descriptive.
41:13
SPEAKER_02: No,
41:13
SPEAKER_02: no,
41:13
SPEAKER_02: I think most people are probably watching,
41:15
SPEAKER_02: but you held up,
41:17
SPEAKER_02: I mean,
41:17
SPEAKER_02: it vaguely looks like a cat,
41:19
SPEAKER_02: I would say,
41:19
SPEAKER_02: you know,
41:20
SPEAKER_02: and is robotic.
41:22
And so you're sending the quantum operations off to the cloud
41:31
where it's getting...
41:33
SPEAKER_02: calculated on a quantum computer,
41:35
SPEAKER_02: and then it's coming back to this machine,
41:38
SPEAKER_02: and you're just trying to create this...
41:42
And then there's going to be some interaction between that cat
41:45
that has sensors and
41:46
SPEAKER_02: things like that with the quantum code.
41:49
SPEAKER_00: Is that the right idea?
41:50
SPEAKER_00: That's right.
41:51
SPEAKER_00: So think about it if your audience is familiar with neural nets.
41:55
You can think about it as the information comes in through the
41:59
robot sensors.
42:00
Those information like go into the neural net from the sort of
42:04
like low layers.
42:05
SPEAKER_00: Yeah.
42:05
SPEAKER_00: And then you run something like an inference mode.
42:09
SPEAKER_00: So you do what's called a forward pass of the neural net.
42:12
So the neural net,
42:13
some of its inputs,
42:15
units will be fixed by the perception feed.
42:18
SPEAKER_00: And then you run a forward pass.
42:20
SPEAKER_00: And then some other output neurons will be the action.
42:24
SPEAKER_00: And so you're doing exactly the same thing.
42:26
SPEAKER_00: But now the neural net is like quantum neural net.
42:28
SPEAKER_00: It's running on a quantum computer rather than a classical GPU.
42:33
SPEAKER_00: Okay.
42:33
SPEAKER_02: And this is like...
42:35
SPEAKER_02: I mean, this is the research part.
42:37
SPEAKER_02: You're exploring what this looks like.
42:40
SPEAKER_00: Yeah, that's right.
42:41
So just to come to your question about the kind of research
42:44
versus business side of it,
42:45
SPEAKER_00: because one of the big questions I get asked is,
42:48
SPEAKER_00: why are you doing this as a company?
42:50
SPEAKER_02: Yeah, like are you going to be selling consciousness that we can buy it?
42:56
SPEAKER_00: Yeah, well, we like to think it's like consciousness as a service.
42:58
SPEAKER_02: I kept picturing it in a jar, but I don't know what I would do with that.
43:03
I'm kind of half joking here,
43:05
but I think there is...
43:06
Okay,
43:06
it's very researching.
43:07
SPEAKER_00: So what we're doing now is we're in this R&D phase.
43:11
SPEAKER_00: So mostly what we're doing is scientific research.
43:13
SPEAKER_00: But the reason it's a company is because...
43:17
There's a lot of interesting innovations here to be discovered
43:21
along the way.
43:22
SPEAKER_00: So when you're doing something completely new,
43:24
SPEAKER_00: okay,
43:24
SPEAKER_00: you don't have a product on day one,
43:26
SPEAKER_00: but you're like discovering things as you go along.
43:29
SPEAKER_00: And they're not just scientific discoveries.
43:31
SPEAKER_00: They're actually technological or engineering discoveries.
43:34
So you can kind of generate a lot of intellectual property
43:38
around those ideas.
43:39
SPEAKER_00: And because there's literally no one working in the field...
43:43
I think we're the only company,
43:44
like the only quantum consciousness company in the world.
43:47
SPEAKER_00: I think I could say that as of today, February 2025.
43:51
SPEAKER_00: There may be more coming soon, but I think we're like the first one.
43:55
So the weird thing,
43:57
the dangerous thing about that is it's very risky,
43:59
right?
44:00
Because what you're doing is very researchy and there's still
44:02
open science questions.
44:03
SPEAKER_00: But the advantage is that you're in this green field of innovation there.
44:08
SPEAKER_02: Yeah,
44:08
SPEAKER_02: and are you,
44:09
SPEAKER_02: I mean,
44:09
SPEAKER_02: I totally get what you're saying about the innovations along the way.
44:13
SPEAKER_02: Do you feel like,
44:14
SPEAKER_02: is the expectation they come more from the,
44:18
SPEAKER_02: in the near term,
44:19
from pushing the limits of quantum programming and where we are
44:25
with quantum software,
44:26
SPEAKER_02: or is it more advancing the state of neural nets or both?
44:31
SPEAKER_02: Mm-hmm.
44:33
SPEAKER_00: Yeah, that's a really good question.
44:35
SPEAKER_00: I think initially it will be quantum algorithms discoveries.
44:38
SPEAKER_00: So if you go to,
44:39
SPEAKER_00: I actually did this last night,
44:41
if you go to AWS and like get an AWS account and sign into
44:45
Amazon Bracket,
44:46
SPEAKER_00: it's called,
44:47
SPEAKER_00: it's like their quantum computing platform.
44:50
SPEAKER_00: And you can actually,
44:51
there's like a tab on the side that's like quantum algorithms
44:54
and there's like five.
44:56
SPEAKER_02: Right.
44:56
SPEAKER_02: Yeah.
44:56
SPEAKER_02: It's early days.
44:58
There's really not very many quantum algorithms that have
45:01
actually been innovated
45:03
SPEAKER_00: and people have run them on quantum computers.
45:05
SPEAKER_00: And there's even fewer that are useful.
45:07
Like most of them are just mathematical curiosities like random
45:10
circuit sampling
45:11
SPEAKER_00: and things like that.
45:12
SPEAKER_00: It's not actually useful for anything.
45:14
So I think there's a huge opportunity here to come up with
45:18
really new innovative
45:19
SPEAKER_00: quantum algorithms that no one's thought of before in this whole area.
45:25
SPEAKER_00: And then how does that tie into the kind of more classical AI stuff?
45:30
I don't think we're going to innovate in any classical neural
45:33
nets anytime soon,
45:34
but there might be innovations in how to connect classical AI
45:38
into this quantum
45:39
SPEAKER_00: conscious decision-making system.
45:41
SPEAKER_00: Yeah.
45:42
One of the things I'm really interested in is how do you get
45:44
the best of both worlds?
45:46
So we've trained all these classical AI models to do
45:49
unconscious things really well
45:51
SPEAKER_00: from training data.
45:52
But what if the consciousness module could actually generate
45:55
that training data in
45:56
SPEAKER_00: the first place?
45:56
SPEAKER_00: So when the robot suddenly finds itself in a completely novel situation,
46:01
SPEAKER_00: very complex,
46:02
SPEAKER_00: very dynamic,
46:02
SPEAKER_00: it switches into conscious decision-making mode.
46:05
SPEAKER_00: And then it switches over to the cloud platform I'm offering.
46:09
SPEAKER_00: It starts making decisions tentatively,
46:12
SPEAKER_00: finding its way around,
46:13
SPEAKER_00: and it's throwing off training data at the same time.
46:16
SPEAKER_00: That can now be improving or training the classical AI model.
46:21
So then the next time it's in that environment,
46:22
it can just use the unconscious module.
46:26
SPEAKER_00: So that's the kind of idea of the two halves of it.
46:30
It must be an interesting experience to make a slide deck where
46:34
you're like,
46:36
we're going to pursue quantum consciousness and nobody else is
46:40
doing this.
46:40
SPEAKER_02: Well,
46:41
SPEAKER_00: it's actually fascinating,
46:42
SPEAKER_00: though,
46:42
SPEAKER_00: because the kind of people I talk to,
46:45
they fall into very different camps depending on what they've
46:48
heard about this before.
46:49
Some people have heard about Penrose's work,
46:51
and so they're kind of already open to it.
46:55
SPEAKER_00: Some people believe in strong AI.
46:57
SPEAKER_00: They're like, oh, AI is already conscious.
46:59
SPEAKER_00: Some people are like, it'll never be conscious.
47:01
SPEAKER_00: And so you have a very different starting point depending on people's,
47:05
SPEAKER_00: not just their background,
47:06
SPEAKER_00: but their actual philosophical perspective on reality.
47:11
SPEAKER_02: Yeah.
47:11
SPEAKER_02: Well,
47:12
SPEAKER_02: and I,
47:13
SPEAKER_02: I mean,
47:13
SPEAKER_02: it's not like,
47:13
SPEAKER_02: I'm sure you've,
47:14
I know you've thought about these things and I'm sure you've
47:16
heard them before.
47:17
SPEAKER_02: I mean, a couple of things went through my head.
47:19
Like when you were talking about the example of the organisms
47:24
on the slide under
47:25
SPEAKER_02: your microscope and it having this
47:28
SPEAKER_02: Like,
47:28
SPEAKER_02: I get your argument that maybe you want to start this from this,
47:33
SPEAKER_02: some of this stuff is built in at this fundamental level,
47:36
SPEAKER_02: but wouldn't other people just argue,
47:38
SPEAKER_02: I mean,
47:39
SPEAKER_02: you know,
47:39
SPEAKER_02: if that creature doesn't even have a neuron,
47:42
SPEAKER_02: I mean,
47:42
SPEAKER_02: this is like,
47:44
SPEAKER_02: whatever you're chasing is so basic and just,
47:48
SPEAKER_02: you know,
47:49
SPEAKER_02: it's just like a...
47:51
SPEAKER_02: like chlorophyll reacting to light.
47:53
SPEAKER_02: I mean,
47:53
SPEAKER_02: there's,
47:53
SPEAKER_02: you know,
47:54
SPEAKER_02: there's just nothing that interesting going on there that you're chasing.
47:59
SPEAKER_02: And yeah,
48:01
SPEAKER_02: I mean,
48:03
SPEAKER_02: and well,
48:03
SPEAKER_02: God,
48:04
SPEAKER_02: maybe I should start with that one bucket and then,
48:06
SPEAKER_02: you know,
48:06
SPEAKER_02: whether,
48:07
whether something like free will and consciousness are just
48:11
illusions for us anyway.
48:13
SPEAKER_02: I mean,
48:13
I guess you have a thesis and you're exploring it and you've
48:16
thought about all
48:16
SPEAKER_02: these things.
48:18
SPEAKER_00: Well,
48:18
the interesting thing about consciousness research is I agree
48:21
most of it is very
48:22
SPEAKER_00: vague and hand-wavy and fuzzy.
48:23
And the reason I like the quantum consciousness hypothesis is
48:27
it sounds like the
48:28
SPEAKER_00: weirdest of all,
48:30
SPEAKER_00: but there's actually a way you can test it experimentally.
48:34
So what I don't like about other models of consciousness like
48:37
IIT or global
48:38
workspace and stuff like that is they don't really make very
48:41
good testable
48:41
SPEAKER_00: predictions that I could turn into a technology.
48:44
SPEAKER_00: Whereas the quantum consciousness hypothesis, it does.
48:46
It says like if information that's coming in from an organism
48:51
shapes a quantum
48:52
SPEAKER_00: state and then the collapse of that is what enacts a decision,
48:56
SPEAKER_00: that's something I can actually build and test.
48:58
SPEAKER_00: So although on the face of it, quantum consciousness sounds like
49:02
SPEAKER_00: totally wacky, even amongst consciousness theories.
49:05
SPEAKER_00: In a way, it's the most analytical.
49:07
SPEAKER_00: It's the most scientific.
49:08
SPEAKER_00: It's the closest we have to something we can actually engineer and test.
49:13
SPEAKER_00: So that's why I'm interested in it.
49:15
So it's like this idea of this interaction between the basic
49:23
particles of the
49:24
SPEAKER_02: universe and life and that when you...
49:28
SPEAKER_02: It's the Schrodinger's cat kind of thing that then...
49:32
us as an organism are influencing how those particles are
49:38
interacting with the universe.
49:40
So this is the part where it's this union of the physics of the
49:48
universe and the organism
49:50
SPEAKER_00: That's right.
49:51
SPEAKER_02: Living in that universe.
49:52
SPEAKER_00: Yeah.
49:53
SPEAKER_00: And this is not a new thing.
49:55
SPEAKER_00: Like the early founding fathers of quantum mechanics thought about this.
49:58
SPEAKER_00: Like Erwin Schrödinger wrote a book called What is Life,
50:02
SPEAKER_00: which I would definitely recommend anyone watching reads.
50:05
And it was all about how life harnesses quantum mechanics to
50:08
make decisions and
50:09
SPEAKER_00: even in things like DNA molecules.
50:13
SPEAKER_00: Do we think of these things as random mutations or are they actually like
50:16
SPEAKER_00: purposeful quantum decisions that are being made at the molecular level?
50:20
So this stuff goes really deep and there's people now like
50:23
Michael Levin who are
50:24
working on uncovering really interesting properties of
50:27
biological systems where
50:28
they're sort of somehow orchestrating and acting as a whole in
50:32
ways we can't really explain.
50:34
SPEAKER_00: So yeah, I think there's a lot.
50:35
SPEAKER_00: There's just a lot to pick out here.
50:38
SPEAKER_02: I know you said there's going to be discoveries along the way,
50:43
SPEAKER_02: but for you to somehow be sitting in your lab,
50:45
SPEAKER_02: in your office one day,
50:46
and now your robotic cat is doing things we've never seen a
50:52
robotic cat do before,
50:56
SPEAKER_02: where does the field of quantum compute?
50:59
SPEAKER_02: Some people are predicting, oh, by 2030, we're going to have the first...
51:05
truly usable quantum computers in in kind of general purpose
51:10
like you know they're
51:11
gonna have enough qubits to do much more interesting things i
51:14
mean do we have to
51:15
reach that point do we have to be at like something far beyond
51:20
that or do you have
51:21
a sense as to to what you need to do this yeah i think we
51:26
already have quantum
51:27
computers that can do interesting things i think maybe they're
51:30
just being applied
51:31
SPEAKER_00: to the wrong algorithms so
51:34
If you look at what quantum computing companies have been doing
51:37
over the past few years,
51:38
SPEAKER_00: they've been trying to prove something called quantum supremacy.
51:40
And the idea is,
51:42
can the quantum computer do a calculation that cannot be done classically?
51:47
So that Google and D-Wave and a few others have claimed now
51:50
that they're able to
51:51
compute things that would take longer than the lifetime of the
51:53
universe to compute
51:55
SPEAKER_00: on a classical computer.
51:56
SPEAKER_00: So we already have quantum computers that are doing...
52:00
that appear to have an amazing computational resource inside
52:04
them it's just that no
52:05
one's applied that to a real world problem that actually
52:09
matters so this idea that
52:12
could this be put into a robot and then that robot is able to
52:16
make intuitive
52:18
decisions that say increase its learning rate or have it learn
52:22
faster than a pure
52:24
reinforcement learning system that's a way you could now maybe
52:26
harness that
52:27
quantum supremacy resource and actually put it to use so i
52:31
think we have systems
52:32
today that can do that but in 2030 if we're up to like a
52:36
million cubits then it'll
52:38
be doing that on steroids basically okay so i think we can see
52:42
what i call a spark
52:43
of life today if we get the experiment set up just so but then
52:47
yeah we're going to
52:48
SPEAKER_00: need the scaling to get that to to something that could be like usefully
52:52
SPEAKER_00: implemented in robots
52:54
SPEAKER_02: Okay, so that's fascinating.
52:56
You're just pushing on all kinds of things you feel like other
52:59
people are ignoring
53:00
SPEAKER_02: or going about the wrong way at the moment.
53:03
Yeah,
53:04
I kind of feel like the quantum computing industry needs
53:09
something new.
53:10
injected into it right like why is everyone just constantly
53:13
trying to break
53:13
cryptography it's so boring like why don't you do something
53:16
different right do
53:18
something that has never been done before and i think that's
53:21
where quantum
53:21
computers are going to find their uh like the the kind of like
53:25
the magic bullet or
53:26
like where they really fit it's going to be in something we
53:28
haven't thought of yet
53:29
let's just not do computation faster let's do a different kind
53:33
of computation
53:34
SPEAKER_02: completely and i'm again way out of my depth i mean roger penrose
53:39
is like you know such a legend um if i remember right i mean
53:45
but this is kind of
53:46
one of his more controversial it is yeah opinions right and and
53:51
um can you i don't
53:52
know enough to like explain to people why but i mean he's a
53:55
he's a genius didn't he
53:56
didn't stephen hawking study under penrose or yeah i think
54:01
there was a link there
54:02
yeah yeah and they're both like black hole aficionados yeah
54:06
like regardless penrose
54:07
SPEAKER_02: is as
54:08
SPEAKER_02: done pioneering, all kinds of pioneering work.
54:11
SPEAKER_02: But so why,
54:14
SPEAKER_02: as I remember from hearing somebody else talk about this,
54:16
isn't this one of these things where they thought Penrose was
54:19
like a little kind of
54:20
SPEAKER_02: on the kookier side of things he was chasing after it?
54:24
SPEAKER_00: Yeah,
54:24
SPEAKER_00: it's very,
54:25
SPEAKER_00: I mean,
54:25
SPEAKER_00: it's like probably one of the most controversial theories in science,
54:29
SPEAKER_00: but that doesn't mean it's like a bad theory.
54:33
SPEAKER_00: It just means it goes against everyone's intuitions.
54:36
And I think the reason it's so controversial and so polarizing
54:40
is it's not just a
54:42
SPEAKER_00: scientific theory.
54:43
hypothesis it actually changes your metaphysical understanding
54:47
of reality so if you
54:49
fall on the penrose kind of camp you have to you basically have
54:53
to let go of
54:54
materialism so what you're saying is consciousness doesn't just
54:58
arise from like
54:59
certain types of networks of brain electrical signals they've
55:04
just put them all
55:04
SPEAKER_00: together in the right way
55:06
boom consciousness emerges this theory is saying no
55:09
consciousness is like the
55:10
fundamental thing that everything else is built out of so
55:14
you've got to flip your
55:16
philosophy on its head to take this seriously and most people
55:20
don't want to
55:21
completely change their philosophical worldview right so that's
55:25
why it gets super
55:26
SPEAKER_00: controversial
55:27
SPEAKER_02: Okay.
55:28
SPEAKER_02: And if we buy into what,
55:30
SPEAKER_02: if what you're,
55:31
SPEAKER_02: this path that you're down and what you're laying out is true,
55:35
SPEAKER_02: I mean,
55:35
SPEAKER_02: then the,
55:38
SPEAKER_02: yeah,
55:39
SPEAKER_02: I've been talking to people all year who are like,
55:41
SPEAKER_02: oh,
55:42
SPEAKER_02: well,
55:42
SPEAKER_02: now that reasoning's here fully by the end of 2025,
55:44
SPEAKER_02: we're at,
55:46
SPEAKER_02: we're at super intelligence.
55:48
SPEAKER_02: And,
55:48
SPEAKER_02: and I mean,
55:49
SPEAKER_02: based on what you're saying,
55:50
SPEAKER_02: you,
55:51
you have no expectation that something like that would happen
55:54
and that people should,
55:56
SPEAKER_02: um,
55:57
At least kind of like pump the brakes on how far LLMs can even
56:01
get us towards truly
56:02
SPEAKER_02: interesting things.
56:04
I actually have a fun analogy for LLMs that I was just thinking
56:07
about last night.
56:08
SPEAKER_00: You might like this.
56:09
So imagine that kind of like consciousness is something that
56:13
comes first and then
56:15
it generates more and more and more training data for whatever
56:18
large model you want.
56:20
And I think this is actually what happened from evolutionary perspective.
56:24
SPEAKER_00: So we,
56:24
like organisms started with a few neurons and then they evolved
56:28
like a tiny brain
56:28
stem and then they evolved like more brain regions and more
56:31
brain regions.
56:31
SPEAKER_00: And then eventually mammals and us evolved this massive cortex.
56:36
SPEAKER_00: So we've been building this giant cortex,
56:39
SPEAKER_00: which gives us like weirdly sized head compared to other animals.
56:43
And what I think is happening is we're just basically building
56:46
like a bigger and
56:46
SPEAKER_00: bigger and bigger LLM style thing inside our head.
56:52
SPEAKER_00: And that's what evolution has done.
56:53
SPEAKER_00: And we're doing the same thing now with LLMs.
56:55
SPEAKER_00: We're just building this like massive neocortex.
56:58
SPEAKER_00: But everyone thinks the neocortex is the seat of consciousness.
57:02
SPEAKER_00: What if it's not?
57:03
What if it's that really deep thing that started off in
57:06
individual cells that's
57:08
SPEAKER_00: been throwing off training data,
57:09
SPEAKER_00: throwing off training data for millions or billions of years,
57:12
SPEAKER_00: and it's grown this giant model?
57:14
SPEAKER_00: I think that's more like what's happening.
57:16
SPEAKER_00: So we are building this weirdly disconnected cortex.
57:20
SPEAKER_00: It's disconnected from a consciousness module.
57:23
So it will be able to do things,
57:25
but it will never be able to do anything new.
57:28
SPEAKER_00: It doesn't have that kind of,
57:30
SPEAKER_00: whether you think of it as free will or true creativity,
57:34
SPEAKER_00: I don't think it'll have that.
57:36
SPEAKER_00: It'll like be stuck wherever you've finished training it.
57:40
forevermore it'll never go beyond that and so here's a really
57:45
you might like this
57:45
actually so uh space right so if we want to send machines out
57:51
into space because
57:52
it's going to be hard for us to send human bodies and human
57:56
biology and things
57:57
right out there we're just not designed to live in space so
57:59
we'll have to end up
58:00
having machine civilizations to at least support us if not be
58:05
our legacy i think
58:07
SPEAKER_00: those things are going to need to be conscious
58:09
SPEAKER_00: Because what we're going to do,
58:10
we're just going to train like GPT-7 and then fix the model and
58:14
then send the model
58:15
SPEAKER_00: out there.
58:16
How is it going to ever generate its own training data and come
58:21
across a completely
58:22
SPEAKER_00: new environment,
58:23
SPEAKER_00: like the edge of a black hole or something?
58:26
It's going to need that same intuition and creativity and
58:30
decision-making we have
58:31
SPEAKER_00: when we're highly conscious.
58:33
SPEAKER_00: So I think in order for us to become a space-faring civilization,
58:37
SPEAKER_00: we need to solve consciousness.
58:39
SPEAKER_02: You said you've talked to people in Silicon Valley,
58:43
SPEAKER_02: in the field,
58:44
SPEAKER_02: and that it kind of depends on their perspective,
58:47
SPEAKER_02: how open they are to this.
58:49
SPEAKER_02: But when you're going to the top five or six companies that have...
58:54
SPEAKER_02: hundreds of billions,
58:56
trillions of dollars at stake at this point on this arms race
59:01
towards LLMs.
59:02
SPEAKER_02: I mean,
59:03
SPEAKER_02: when you're talking to the top researchers from those companies,
59:07
SPEAKER_02: you know,
59:07
SPEAKER_02: are they open to this idea or are you getting more pushback?
59:11
SPEAKER_02: They feel like they're very strong in their conviction.
59:14
SPEAKER_00: I think it's a mixed bag.
59:16
So I've talked to some deep learning researchers who say that
59:20
they don't think it's
59:21
SPEAKER_00: going to scale to full AGI.
59:24
And by AGI,
59:25
I mean something that could do anything a person could do in an embodiment.
59:28
SPEAKER_00: So I should be very clear.
59:30
I think language models are going to solve all of our
59:33
non-embodied problems.
59:37
SPEAKER_00: problems, like all of our research and programming.
59:41
SPEAKER_00: We're still using them as tools on our computer.
59:44
I think when we start trying to put them into robot bodies and
59:47
then have those
59:47
SPEAKER_00: robots go out there and say,
59:49
SPEAKER_00: do original science research,
59:50
SPEAKER_00: I just don't think that's going to work.
59:52
SPEAKER_00: And I've spoken to quite a few people who agree with me.
59:56
They just don't see how these robots are going to ever get that
1:00:00
kind of capability
1:00:01
SPEAKER_00: without all the training data in the multiverse.
1:00:06
SPEAKER_00: But I do think some people are then very bullish on it, the other side.
1:00:10
SPEAKER_00: They think the reward is just enough and all this kind of stuff.
1:00:15
SPEAKER_02: Yeah.
1:00:17
What about the idea of,
1:00:19
like,
1:00:19
where do you fall on reverse engineering the brain?
1:00:23
SPEAKER_02: Like,
1:00:23
SPEAKER_02: if,
1:00:23
SPEAKER_02: you know,
1:00:24
there's efforts underway now to put record numbers of electrode
1:00:29
sensors into brains
1:00:31
to record out how the neurons are firing in such precise detail
1:00:37
that you're,
1:00:39
SPEAKER_02: you know,
1:00:39
SPEAKER_02: we've got a long way to go,
1:00:40
SPEAKER_02: but that maybe you're replicating the exact...
1:00:43
SPEAKER_02: electrical computation behavior of a physical brain,
1:00:47
and it's at a high enough fidelity that you can put that into a
1:00:50
computer and get
1:00:51
SPEAKER_02: intelligence out.
1:00:52
I mean,
1:00:53
obviously,
1:00:53
according to your theory,
1:00:55
that would still be lacking this other element.
1:00:57
SPEAKER_00: Yeah.
1:00:58
So maybe I can make a prediction or kind of a fun scientific
1:01:02
experiment here.
1:01:03
so um you know when you're just in everyday waking life you can
1:01:07
switch your
1:01:08
conscious attention so you can be like conscious of something
1:01:13
and then you're not
1:01:14
conscious of a whole bunch of other things so if I focus on
1:01:17
like I don't know the
1:01:18
feeling in my left hand I'm not focusing on the feeling in my
1:01:21
right foot so I kind
1:01:22
of become I just don't even notice it's not that so what I
1:01:26
predict is say we're
1:01:27
SPEAKER_00: able to
1:01:28
open up the brain and have all these like connections to every
1:01:31
single neuron and
1:01:32
we're looking at the activity of all the neurons what I think
1:01:36
is when you switch
1:01:37
conscious focus to different parts of like your nervous system
1:01:41
or the environment
1:01:42
the neurons that are involved in that will become unpredictable
1:01:46
and the neurons
1:01:47
that are involved in the stuff you're not focusing on will be
1:01:50
predictable I see
1:01:52
So the way I think of this model of consciousness in the brain
1:01:57
is that we have two
1:01:58
SPEAKER_00: parallel processing paths running at all times.
1:02:01
SPEAKER_00: We have a classical computer,
1:02:02
SPEAKER_00: which is just neurons firing in a way that we can model and predict,
1:02:07
SPEAKER_00: like,
1:02:07
SPEAKER_00: you know,
1:02:07
SPEAKER_00: neurons three and four fire,
1:02:09
SPEAKER_00: and then the neuron,
1:02:10
SPEAKER_00: like,
1:02:10
SPEAKER_00: fires 60% probability or something.
1:02:12
SPEAKER_00: So we can model that.
1:02:14
SPEAKER_00: But when that process goes under conscious control...
1:02:18
SPEAKER_00: I think things are going to start changing,
1:02:20
SPEAKER_00: things like maybe the firing rates we see in neurons,
1:02:22
SPEAKER_00: and the model no longer works,
1:02:25
SPEAKER_00: right?
1:02:26
And I think what's happening is we have this other parallel
1:02:29
quantum computing
1:02:30
SPEAKER_00: process that's going on,
1:02:31
SPEAKER_00: and it's probably inside the neurons.
1:02:33
SPEAKER_00: or inside the synapses or something.
1:02:35
SPEAKER_00: It might not even be microtubules.
1:02:37
SPEAKER_00: It could be something else.
1:02:38
But I think that takes over when you go into conscious control
1:02:41
mode and it subtly
1:02:42
SPEAKER_00: affects the final output firing of the neuron.
1:02:46
So I think all this brain-computer interface stuff is strongly
1:02:49
related to this.
1:02:50
SPEAKER_00: We'll actually be able to test this scientifically.
1:02:53
SPEAKER_00: If a neuron is doing something consciously,
1:02:55
maybe it just starts firing in a different pattern than if it's
1:02:58
doing it unconsciously.
1:03:01
SPEAKER_02: I love this.
1:03:02
SPEAKER_02: I feel like microtubules needs a rebranding.
1:03:05
SPEAKER_02: I mean, that definitely sounds like some conspiracy theory.
1:03:10
SPEAKER_02: That's right.
1:03:12
SPEAKER_00: Let's call them quantum cores or something.
1:03:16
SPEAKER_02: That's where you need to go.
1:03:17
SPEAKER_02: I think you should definitely work on that.
1:03:21
SPEAKER_02: Just a couple last questions for you, just being mindful of your time.
1:03:23
SPEAKER_02: I mean, I think the answer to this might be yes, but
1:03:30
SPEAKER_02: Yeah,
1:03:31
I know you were art and science inclined as a kid from the
1:03:36
influence of your dad on
1:03:38
SPEAKER_02: one side and your mom on another and this blending of the two.
1:03:41
SPEAKER_02: I know you've been pursuing very complex,
1:03:46
sophisticated scientific research for a long time and you
1:03:49
worked in quantum
1:03:50
SPEAKER_02: computing and then all these other fields.
1:03:51
But did you ever imagine you'd be like sitting here in 2025
1:03:55
trying to build quantum consciousness?
1:04:01
SPEAKER_00: It's funny you should say that because I think the answer might be yes.
1:04:04
SPEAKER_01: That's what I thought it would be.
1:04:06
SPEAKER_00: Okay.
1:04:06
So I don't know if any of your audience like believes in
1:04:09
synchronicities or anything like that.
1:04:11
SPEAKER_00: And trust me,
1:04:11
SPEAKER_00: when you start researching consciousness,
1:04:13
SPEAKER_00: you find all these like weird things going on.
1:04:16
SPEAKER_02: Carl Jung.
1:04:17
SPEAKER_02: I'm here for it.
1:04:17
SPEAKER_00: Carl Jung, man.
1:04:18
SPEAKER_00: I'm deep on it.
1:04:19
SPEAKER_00: But anyway,
1:04:20
SPEAKER_00: there's a bunch of,
1:04:21
SPEAKER_00: now I like look back,
1:04:22
there's a bunch of synchronicities I think happened in my life
1:04:25
that sort of ended
1:04:26
SPEAKER_00: up
1:04:27
SPEAKER_00: With me taking this path,
1:04:28
SPEAKER_00: like back in 2010,
1:04:30
I think I was talking already about a quantum brain because I
1:04:33
was doing quantum
1:04:34
SPEAKER_00: computing at the time and thinking about quantum neural nets and stuff.
1:04:37
And I just read On Intelligence by Jeff Hawkins and thought it
1:04:39
was like the best thing ever.
1:04:41
SPEAKER_00: So I was like, well, could we build a quantum brain?
1:04:43
SPEAKER_00: And back then it was just like a kind of silly idea,
1:04:45
SPEAKER_00: but it actually ended up being what I'm doing now.
1:04:48
SPEAKER_00: So yeah.
1:04:49
SPEAKER_00: There's a bunch of things like that.
1:04:50
And I also think that I finally figured out the connection
1:04:53
between art and science.
1:04:55
SPEAKER_00: And it comes through this consciousness angle.
1:04:58
SPEAKER_00: So if you think what is art, it's like difficult to define.
1:05:02
SPEAKER_00: But I think art is a way of thinking through problems and expressing and
1:05:05
SPEAKER_00: representing things in a non-language way,
1:05:09
SPEAKER_00: like a non-symbolic way.
1:05:11
SPEAKER_00: I mean, okay, some people think like poetry and stuff is art as well.
1:05:14
SPEAKER_00: So maybe...
1:05:15
SPEAKER_00: But it's kind of like a weird use of language.
1:05:17
So I think art is a way of trying to solve problems or have
1:05:19
intuitive breakthroughs,
1:05:21
creative breakthroughs that sort of can't be expressed using
1:05:24
normal thinking,
1:05:26
SPEAKER_00: normal words,
1:05:27
SPEAKER_00: normal logic.
1:05:28
So I think art might actually be a deeper expression of our
1:05:32
consciousness or we're
1:05:33
like tapping into those deeper conscious intuitive processes to
1:05:37
solve a problem.
1:05:39
So I think that way art and science are quite synergistic with
1:05:43
each other.
1:05:45
SPEAKER_02: No, I like that.
1:05:47
SPEAKER_02: You were just talking about Unintelligence by Jeff Hawkins.
1:05:51
SPEAKER_02: That must be like one of the most,
1:05:55
like among the circle of people that I run with and probably
1:05:58
you run with,
1:05:59
SPEAKER_02: you know,
1:05:59
SPEAKER_02: like obviously influential book,
1:06:02
SPEAKER_02: but,
1:06:02
SPEAKER_02: you know,
1:06:02
SPEAKER_02: not.
1:06:04
SPEAKER_02: I think it's sold well, but it's not like a totally mainstream book.
1:06:07
SPEAKER_02: But I just feel like it must have impacted.
1:06:10
SPEAKER_02: I hear it come up again and again and again for people.
1:06:12
SPEAKER_02: It's definitely helped shape the last 15 years or so.
1:06:16
SPEAKER_02: And then I haven't heard a lot.
1:06:18
from Jeff lately and we're like in the thick of like all these
1:06:22
things that he was,
1:06:23
SPEAKER_02: he was kind of pursuing.
1:06:25
SPEAKER_00: Oh, they're still going.
1:06:26
SPEAKER_00: And they actually,
1:06:27
SPEAKER_00: um,
1:06:27
new mentor recently announced like a spinoff nonprofit
1:06:31
organization called the
1:06:32
SPEAKER_00: thousand brains project.
1:06:33
SPEAKER_00: So they're open sourcing.
1:06:35
SPEAKER_00: Oh, okay.
1:06:36
SPEAKER_00: So it's all of Jeff's theoretical work.
1:06:39
They're turning it into an open source code project so you can
1:06:43
go and code these
1:06:44
SPEAKER_00: sort of online neocortical-inspired algorithms and things.
1:06:47
SPEAKER_00: So I'm definitely keeping an eye on that as well.
1:06:49
SPEAKER_02: Oh, okay.
1:06:50
SPEAKER_02: So the Palm, it's crazy.
1:06:52
SPEAKER_02: He was like the Palm Pilot guy, right?
1:06:54
SPEAKER_02: Yeah, yeah.
1:06:56
I actually think I know like deep learning and transformers and
1:06:59
big neural nets are
1:07:00
SPEAKER_00: the big thing today,
1:07:01
SPEAKER_00: but I actually think in the long run,
1:07:03
it's going to be these more like online cortically inspired
1:07:06
algorithms that are
1:07:07
SPEAKER_00: going to take over,
1:07:08
SPEAKER_00: especially for embodied systems.
1:07:10
SPEAKER_00: Because when you put something into a robot body,
1:07:12
you need something more like a hierarchy of cortical
1:07:16
controllers that all work
1:07:17
SPEAKER_00: together online.
1:07:19
SPEAKER_00: And you don't get that from modern transformer networks.
1:07:22
SPEAKER_00: Yeah.
1:07:22
Yeah,
1:07:22
you're just making me rewire my brain and my timeline on all
1:07:26
kinds of things.
1:07:26
SPEAKER_02: I think as I've been reporting this book,
1:07:28
SPEAKER_02: I've been so...
1:07:30
SPEAKER_02: San Francisco is like an LLM cult right now.
1:07:33
SPEAKER_02: And if you're talking to everyone in this field every day,
1:07:39
SPEAKER_02: all day,
1:07:41
SPEAKER_02: even though I'm skeptical,
1:07:42
SPEAKER_02: you start to be convinced.
1:07:43
You're like,
1:07:43
oh,
1:07:43
by the end of this year,
1:07:44
I guess this is just solved and we're off on some...
1:07:49
SPEAKER_02: crazy super intelligence singularity timeline.
1:07:52
SPEAKER_02: And,
1:07:52
SPEAKER_02: and even though part of me is like,
1:07:54
SPEAKER_02: okay,
1:07:54
SPEAKER_02: you know,
1:07:55
SPEAKER_02: it's probably not possibly not going to happen.
1:07:58
SPEAKER_02: And,
1:07:58
SPEAKER_02: and,
1:07:58
SPEAKER_02: um,
1:07:59
SPEAKER_02: but now,
1:07:59
SPEAKER_02: I mean,
1:08:00
SPEAKER_02: this is,
1:08:01
SPEAKER_02: you've opened up all these other avenues that,
1:08:03
SPEAKER_02: that,
1:08:04
SPEAKER_02: um,
1:08:05
SPEAKER_02: now I'm going back to my,
1:08:07
Where I started out on all this was like,
1:08:09
oh,
1:08:09
we mathematically solve language.
1:08:11
That's,
1:08:11
like,
1:08:12
interesting,
1:08:12
but it's very limited in how interesting it is.
1:08:15
SPEAKER_02: I don't know.
1:08:16
SPEAKER_02: This is helpful.
1:08:16
SPEAKER_02: You're making me recalibrate.
1:08:18
SPEAKER_02: I appreciate it.
1:08:18
SPEAKER_00: Well,
1:08:19
SPEAKER_00: one piece of advice I'd give to any,
1:08:20
SPEAKER_00: I guess,
1:08:20
SPEAKER_00: budding startup entrepreneurs watching your channel is,
1:08:25
SPEAKER_00: like,
1:08:25
SPEAKER_00: just don't go into what is really hot right now.
1:08:29
SPEAKER_00: Like, don't go into airlines.
1:08:31
SPEAKER_00: Don't go into Transformers because...
1:08:33
You've got your cool idea for a startup,
1:08:35
you're going to start pitching it.
1:08:36
And then Google will just release some like open source side
1:08:38
project next week that
1:08:40
SPEAKER_00: totally wipes you out of the water.
1:08:42
SPEAKER_00: It's crazy to go into that.
1:08:43
SPEAKER_00: It's like a giant like shark bath at the moment.
1:08:46
SPEAKER_00: You've got to pick something.
1:08:47
SPEAKER_00: If you want to innovate, you have to pick something completely different.
1:08:52
You need to pick something that people are like,
1:08:53
that's crazy and ridiculous and insane.
1:08:56
because that it means no one's going to be working on it right
1:08:59
and and so i give
1:09:01
people this advice i'm with you if you're going to take all
1:09:03
this risk and do all
1:09:04
this work you might as well make it worth it i i feel like
1:09:06
quantum consciousness
1:09:07
you've done it let's i'm excited to see where you go um and and
1:09:13
um thank you it's
1:09:14
SPEAKER_02: always so fun talking you always make me think um a lot and and um
1:09:20
SPEAKER_02: Thank you so much for, what's the best thing for people to check out?
1:09:24
SPEAKER_02: You guys' website or your Substack or?
1:09:26
SPEAKER_00: Yeah, the website's got all our links to our other socials on.
1:09:29
SPEAKER_00: So like nirvanic.ai is probably the best place to go at the moment.
1:09:33
We have,
1:09:34
yeah,
1:09:34
Substack newsletter and we also have,
1:09:37
we're quite active on X as well.
1:09:39
SPEAKER_00: Okay.
1:09:40
SPEAKER_00: Yeah.
1:09:41
SPEAKER_00: And I promise you,
1:09:42
SPEAKER_00: Ashley,
1:09:42
SPEAKER_00: like next time you come to visit,
1:09:43
SPEAKER_00: I will have a robot look you in the eye and you'll be like,
1:09:46
SPEAKER_00: holy crap,
1:09:47
SPEAKER_00: that robot's alive.
1:09:49
SPEAKER_01: Okay, I'm here for it.
1:09:50
SPEAKER_01: What does that mean?
1:09:52
SPEAKER_01: When would this visit take place?
1:09:54
SPEAKER_00: Oh, my goodness.
1:09:55
SPEAKER_00: Okay, I'll start that out.
1:09:57
SPEAKER_00: I can't promise any timelines right now, but that's my dream.
1:10:01
SPEAKER_02: All right, well, I will show up with a camera if that can happen.
1:10:05
SPEAKER_02: I'm here for it.
1:10:07
SPEAKER_00: Awesome, thanks.
1:10:08
SPEAKER_02: Thank you.
1:10:08
SPEAKER_02: Thanks so much, Suzanne.
1:10:09
SPEAKER_02: It was great talking.
1:10:11
Thanks.
1:10:16
SPEAKER_02: Today's episode was hosted by me, Ashley Vance.
1:10:18
SPEAKER_02: Our guest was Suzanne Gildert.
1:10:22
SPEAKER_02: Core Memory is produced by me and David Nicholson.
1:10:25
SPEAKER_02: Our theme song is by James Mercer and John Sortland.
1:10:30
SPEAKER_02: And the show is edited by the great John Sortland.