0:00

You're listening to the Stephen

0:02

Wolfram podcast, an exploration of

0:04

thoughts and ideas from the

0:06

founder and CEO of Wolfram

0:08

Research, creator of Wolfram Alpha

0:10

and the Wolfram Language. In

0:12

this ongoing Q&A series, Stephen

0:14

answers questions from his live

0:16

stream audience about the future

0:18

of science and technology. This

0:20

session was originally broadcast on

0:22

January 31st, 2025. Let's have

0:24

a listen. Hello

0:26

everyone, welcome to another

0:28

episode of Q&A about

0:31

future of science and technology.

0:33

And I see a bunch

0:35

of questions saved up here.

0:37

Okay, here's an interesting

0:40

one from Prometheus. Do

0:42

you imagine humanity exploring

0:44

inner space, i.e.

0:46

virtual worlds, more than

0:49

outer space. It's an interesting

0:51

question. It's In a

0:53

sense, one's comparing the physical

0:55

universe and the extent of

0:57

space in the physical

0:59

universe with the computational

1:02

universe. In my way of

1:04

thinking about it, the computational universe

1:06

is ultimately the rule

1:09

ad, the collection of

1:11

all possible computations one

1:13

can run. And in a

1:15

sense, what one's talking about

1:17

then is the difference between

1:20

exploring physical space and exploring

1:22

rural space. Exploring physical space

1:24

by sending out spacecraft. Exploring

1:26

rural space by looking at... different

1:29

kinds of programs that one can

1:31

study in the computational universe. I

1:33

mean, it's one thing to have

1:35

a virtual world that is emulating

1:37

our physical world that's emulating the

1:39

laws of physics as we currently

1:41

perceive them. It's another to think

1:43

about having a virtual world where

1:45

the laws of physics are up

1:47

the grabs and you can have

1:49

any law you want, any rule

1:51

you want to define how your

1:53

virtual universe should work. I mean, my

1:55

own efforts in studying what

1:57

I now call rheoliology have

1:59

been directed for the last I

2:02

don't know 45 years or something

2:04

to this question of exploring the

2:06

computational universe of possible programs and

2:08

what they do. Some corner of

2:10

that is the programs that are

2:12

the way that we perceive the

2:14

physical universe. But there are other

2:17

programs which in a sense one

2:19

could think of as being the

2:21

ways that aliens could perceive even

2:23

our physical universe that is... that

2:25

are sort of different possibilities, different

2:27

computational possibilities. I think nobody has

2:29

yet quite made the Ruliad video

2:32

game, but it's a really cool

2:34

thing to think about doing. That

2:36

is... a video game where instead

2:38

of being exposed to the laws

2:40

of physics as they are, you're

2:42

exposed to laws of physics as

2:44

they could be. Now my guess

2:47

is, from having personally been exposed

2:49

to those kinds of virtual physics

2:51

as for many decades now, that

2:53

it is one's intuition about how

2:55

things work has to be developed.

2:57

one doesn't immediately have intuition about

2:59

that. One has intuition about the

3:01

physical world as we currently perceive

3:04

it because we've all grown up

3:06

in that physical world, but this

3:08

virtual world of computational possibilities is

3:10

something one just has to get

3:12

used to. My guess is if

3:14

you're sort of really having sensory

3:16

experiences in that world that one

3:19

will end up needing to kind

3:21

of have a bit of an

3:23

on-ramp. from kind of the way

3:25

that we have sensory experiences in

3:27

the current physical world to what's

3:29

possible in the computational universe. You

3:31

know, I have to say this,

3:34

the computational universe is vastly bigger

3:36

than our physical universe. The set

3:38

of all possible rules that one

3:40

can attribute to one's universe is

3:42

much bigger than the set of

3:44

rules that we do attribute to

3:46

our physical universe. There's much more

3:48

to explore and ruleial space. across

3:51

the whole Ruliad than there is

3:53

to explore in physical space even

3:55

though our universe is pretty big

3:57

compared to us. So I think

3:59

it's... It's

4:01

almost for certain

4:04

that there's sort of more

4:06

that one can do in that computational

4:08

universe than in physical space. So

4:10

in a sense, as we

4:13

explore more about that

4:15

computational universe, expanding

4:17

our domain in that computational

4:19

universe is kind of

4:21

like expanding our paradigms for

4:23

understanding how things work. It's

4:26

different from physical space

4:28

where by expanding our domain, we

4:30

might learn a little bit more

4:32

about how the physical universe works

4:34

in places other than right where

4:36

we are, although we think the

4:38

physical universe is fairly homogeneous and

4:40

that the laws of, well, gravity

4:43

looks like it's very homogeneous, but

4:45

many aspects of the universe, it's

4:47

pretty much the same here as

4:49

it is in other places. So

4:52

there's sort of less to learn in a

4:54

sense than there is in the in

4:56

the in the rural universe in in the

4:58

Rulliad, where one can sort of

5:00

look at what happens with very different kinds of

5:02

rules. I have to say

5:04

one thing that I thought

5:07

many years ago now is as

5:09

came out of my thinking about

5:11

this thing I call the principle of

5:13

computational equivalence. The idea that once

5:15

you get above a very low

5:17

threshold, essentially every system

5:19

you look at is equivalent in

5:21

the computational sophistication it shows. So

5:24

for example, that means in particular

5:26

that our brains are equivalent in

5:28

their computational sophistication to many kinds

5:30

of systems, either abstract ones that

5:32

we can make up or physical

5:34

ones out in the world. And

5:37

it's kind of like you think

5:39

about what's the relationship between a rock

5:41

which has all of these electrons

5:43

going around in all those complicated

5:45

patterns and our brain which has

5:47

all of these electrical signals going around

5:49

in all these complicated patterns. Can

5:51

you draw a fundamental distinction between

5:53

the electron activity in a rock and

5:56

the electrical activity in our brains

5:58

and the claim of. the principle

6:00

of computational equivalence is that at the

6:02

level of computational sophistication, you can't

6:04

really make a difference between those two

6:06

things. Now, there are many things

6:08

about what happens in brains that are

6:11

more special to us. Things like

6:13

the fact that we aggregate our experiences

6:15

to believe that we sort of

6:17

have a single thread of experience, even

6:19

though all our neurons are

6:21

sending different signals all the time,

6:23

we aggregate that into the single thread

6:25

of experience. There's no reason to

6:27

think that a rock does the same

6:29

thing. And there are other sort

6:32

of special aspects of the particular arrangement

6:34

of computational things that represents us.

6:36

But when it comes to just sort

6:38

of the rating of who's doing

6:40

the most sophisticated computation, I don't think

6:42

there's a fundamental way in which

6:44

we win relative to the rock. So

6:47

then the question is, in that

6:49

point of view, if you think

6:51

about sort of the ultimate future of humanity, and

6:54

you think about, well, we can take all

6:56

these things that we're very proud of,

6:58

which are our processes of thinking and so

7:00

on, we can in principle upload those

7:02

to some digital form. And then in

7:04

the end, kind of the thing

7:06

I talked about maybe 20 years

7:09

ago now is kind of the potential

7:11

end point is a

7:13

box of a trillion souls, so

7:15

to speak, where you have

7:17

sort of a trillion

7:19

disembodied human consciousnesses in this

7:21

digital medium. And the

7:23

question then is kind of, how do

7:25

you distinguish looking from the outside?

7:27

How do you distinguish the box of

7:29

a trillion souls from the rock? Both

7:32

have lots of electrons going around in

7:34

complicated patterns. What's the difference between

7:36

the box of a trillion souls and

7:38

the rock? And of course,

7:40

the answer that is important to

7:42

us is, well, the box of

7:44

a trillion souls has the details

7:46

of our cultural history and all

7:48

of those kinds of things in

7:50

it. It is a special collection

7:52

of electrons moving around, whereas the

7:54

rock is something not special

7:56

to us. But to

7:58

this question of... What will the box

8:01

of a trillion souls do for

8:03

the rest of eternity, so to

8:06

speak? And in a sense, the

8:08

slightly cynical point of view would

8:10

be the future of humanity

8:12

is a box of a

8:14

trillion souls playing video games

8:16

for the rest of eternity.

8:19

And that seems really bad

8:21

from the point of view

8:23

of what we think is

8:25

important today. it's worth realizing

8:27

that in the course of

8:29

history the things that have

8:31

seemed important to people have changed

8:33

quite a bit. And to one

8:36

of those souls, so to speak,

8:38

in that box, playing quotes

8:40

video games for forever, it's

8:42

that may seem as meaningful as

8:44

anything that we do today. I

8:46

mean, in the past, it might

8:48

not have seemed meaningful to do

8:50

things in social media and so

8:52

on, or things that are very

8:54

virtualized today. One would have said,

8:56

why do you care about those

8:58

things in the past? But today,

9:00

we are embedded in an environment

9:03

where we feel like we have

9:05

a reason to care. Anyway, this

9:07

is all to say that one of

9:09

the things that I talked about maybe

9:11

20 years ago now, was the concept

9:13

of, well, what will those souls do?

9:15

in that, you know, they can, in

9:17

a sense, in that virtualized environment,

9:19

they can achieve anything. They can

9:22

make their bits move around in

9:24

any way they want. They're not

9:26

constrained by the physicality of things

9:28

in our perceived physical world and

9:30

so on. And then the question

9:33

is, well, they can explore kind of

9:35

the things about the physical universe,

9:37

but when you finish doing that,

9:39

where are you going to go?

9:41

Well, you're going to... start potentially

9:43

exploring things in the computational universe

9:45

that are not things realized in

9:48

our physical universe in the way

9:50

that we perceive it at least,

9:52

but they're things that are in principle

9:54

possible. And so I suppose the thing

9:56

that kind of made me pause in a

9:59

sense was the thought. that the

10:01

future of humanity is

10:03

those future instances

10:05

of our consciousness

10:07

exploring the computational

10:09

universe for the

10:11

rest of eternity.

10:13

I personally have spent some part of

10:15

my life exploring the computational universe

10:17

and it felt a little bit like

10:20

if you imagine the future, you

10:22

imagine imprinting your own preferences on the

10:24

future. This is an ultimate version

10:26

of that, of saying that the thing

10:28

that is the inevitable thing for

10:30

people or generalized people to do is

10:32

the thing that I happen to

10:35

have been doing for a large number

10:37

of decades, so to speak. So

10:40

anyway, there are a few thoughts about

10:42

that. The other thing to realize is that

10:45

in just in terms of sort of the

10:47

physical size of things, there's a

10:49

lot more room down than

10:51

there is up. I mean, in

10:54

rough terms, we're about a meter tall,

10:56

give or take. The universe is

10:58

10 to the 26 meters across. The

11:00

physical universe, as we perceive it, is 10

11:02

to the 26 meters across. There's

11:05

a question in our physics project

11:07

and our physics model, how small

11:09

is the elementary length? We're pretty

11:11

sure that space is ultimately discrete, but

11:14

there's a question of what is that

11:16

length scale that is effective at

11:18

the distance between neighboring atoms of space?

11:21

It's not laid out in space

11:23

as it is, that relationship between

11:25

the atoms of space is what

11:27

defines space, but we can still

11:29

translate, we can still say

11:31

what do we perceive? How many sort of

11:33

distances between atoms of space correspond to

11:35

one meter as we perceive it in space

11:37

as we know it right now? We

11:39

don't know the answer to that. It's

11:42

something we'd really like to know

11:44

the answer to. It's something where I'm

11:46

hoping in the next few months,

11:48

actually, to really try and define some

11:50

experimental ways of exploring what is

11:52

the discrete state scale of space. It's

11:55

also my suspicion that we

11:57

probably already know, that

11:59

is, there are already has been an experiment

12:01

done that probably had some weird

12:03

result that nobody understood and probably

12:06

when you untangle what that experiment actually

12:08

did, you'll realize actually that shows what

12:10

the discreteness scale at scale of spaces.

12:12

So it's a weird case a little

12:15

bit like what you're asking about exploring

12:17

outer space versus inner space. It is

12:19

my suspicion that there's a really good

12:22

chance that we'll be able to make

12:24

progress in understanding the experimental implications of

12:26

our physics project without doing any new

12:28

experiment at all. just by saying, well, you

12:31

know, where does what we're talking about

12:33

relate to things that have already been

12:35

studied? Now, it may help a lot.

12:37

It may be a lot clearer if

12:39

we can do a new experiment of

12:41

our own design, so to speak, and

12:43

say, how does this work? But it's

12:45

quite possible that we can explore the

12:47

almost the inner space of what's been

12:49

already studied. But any case, we don't

12:51

know the elementary length, but there's some

12:54

vague reason to think it might be

12:56

around 10 to the minus 100 meters.

12:58

from our size up to the scale

13:01

of the universe is 26 orders of

13:03

magnitude, from our size down to the

13:05

elementary length, is 100 orders of magnitude.

13:07

So there's a lot more room going

13:10

down than there is going up. If

13:12

we're saying, how can we encode a

13:14

lot of computational activity? We could in

13:16

principle encode that in this very microscopic

13:18

stuff more than we could encode it

13:21

by building a computer that's the size

13:23

of a galaxy, for example, much in

13:25

a sense. it should be much easier

13:28

to build that very tiny thing that

13:30

makes use of things that are the

13:32

size of protons or something than to

13:34

build the computer that sort of spans

13:37

the solar system or the galaxy or

13:39

whatever. Now one of the things that's a

13:41

feature of physics is that to probe

13:43

very small length scales you effectively

13:45

need very high energies. That's a

13:48

feature of the uncertainty principle in

13:50

quantum mechanics and so on. So

13:53

there's sort of the question of what

13:55

would it actually take. to build a

13:57

computer that makes use of...

13:59

of things that are happening at

14:01

the scale of things inside a

14:03

proton or something. We certainly

14:05

don't know how to do that

14:08

at this time. And the best

14:10

we can do to sort of

14:12

probe what's inside a proton is

14:14

to bash two protons together at

14:16

very high speed and sort of

14:18

see what happens when they hit

14:20

and see what pieces kind of

14:22

come out when they hit. that's

14:24

kind of like the approach to

14:26

making a clock where you say

14:28

I just take random pieces of

14:30

metal and throw them at each

14:32

other rather than I create something

14:34

with lots of gears and so

14:36

on inside it. So a few

14:38

thoughts on that. I think in the

14:40

nearer term this question of,

14:43

you know, do you want to

14:45

go to the moon or do

14:47

you want to kind of explore

14:49

virtual reality of things that

14:52

might be possible. I mean, I

14:54

think there's just a lot more

14:56

kind of aspirational value to going

14:58

to the moon than there is

15:00

to exploring sort of the inner

15:02

space of what's possible. Personally, I

15:04

have found exploring the inner space

15:06

of what's possible really exciting, but

15:08

I think that's probably a minority

15:10

opinion, at least at this time

15:12

in history, relative to the aspirational

15:14

value of look I can go

15:16

to the thing that I can see far

15:18

away in the sky type thing. But I think

15:20

in... in terms of what one

15:23

can learn from science, there's a

15:25

lot more to explore in rural

15:27

space than there is in physical

15:29

space. And I think it is a

15:31

much, much easier thing to explore

15:33

more broadly in rural space, and

15:36

it's really a worthwhile thing to

15:38

do, and it's what I've spent some

15:40

significant part of my life doing and

15:42

what I've spent a lot of effort

15:44

building the tools to make easy to

15:46

do. It's a lot. From my point

15:48

of view, it's a lot easier to

15:50

build out a giant tower of technology

15:53

and more from language than it would

15:55

be to build a spacecraft to make

15:57

it actually work. At least that's my point

15:59

of view. other people might think

16:01

differently about that. But even though

16:04

the the number of moving parts

16:06

in the kind of large software

16:08

system that is wolf language is

16:10

very very large, actually very large

16:13

compared to anything that you would

16:15

build as a physical object with

16:17

moving parts, but still it seems

16:19

to me a lot easier to

16:21

just you know sit in one's

16:24

chair and type things into one's

16:26

computer and have have CEDUs and

16:28

GPUs and GPUs running things. than

16:30

to be, you know, strapped in

16:33

a rocket and exposed to lots

16:35

of forces that we humans were

16:37

not really evolved for?

16:39

Let's see. As a question

16:41

here from Max, could the

16:43

spin of electrons lead

16:46

to a communication system?

16:48

So, well, there's a whole

16:51

field called spentonics

16:53

that is sort of

16:55

a generalization of electronics.

16:58

Electronics is about moving

17:00

electrons around, having electric currents

17:02

that are flows of electrons,

17:04

or electric potential, which can

17:07

be thought of as sort

17:09

of accumulating electrons in one

17:11

place rather than another. But

17:13

electrons, in addition to being, well,

17:15

what appear to be point particles, I

17:18

don't think they really are point particles,

17:20

but let's say that they have a place,

17:22

and they also have a momentum, but

17:25

they have one other attribute. which is

17:27

they have this thing called spin for

17:29

an electron. And from a large-scale

17:31

point of view, we can think

17:33

about spin as being like, oh,

17:35

the electron is spinning around on

17:37

its axis. That picture of the

17:39

electron spin is very kind of,

17:41

it seems very mechanical. It's only partly

17:44

useful, I think, in terms of the

17:46

intuition about what's really going on. But

17:48

one can say that there is a

17:50

definite direction associated with this attribute that

17:52

we call spin for an electron. what

17:54

it means to have a definite direction

17:57

in a case where we're dealing with

17:59

quantum mechanics. and you have only

18:01

this one moment when you can

18:03

measure, is the spin-up or is

18:05

the spin-up or is the spin-down,

18:07

is the spin-line this way or

18:09

that way, you don't get to

18:12

sort of probe it many times,

18:14

you just get to say, okay,

18:16

I can check, it's this or

18:18

that, and then you measure it.

18:20

and then you've kind of forced

18:22

it to be either this or

18:25

that, whatever you measured it to

18:27

be. And it's not something where

18:29

in the macroscopic world, you can

18:31

just have an object, you can

18:33

say, I'll look at it once,

18:35

I'll look at it twice, I'll

18:38

look at it any number of

18:40

times, and you'll sort of see

18:42

the same thing every time. In

18:44

the sort of quantum world, it

18:46

is by making that measurement from

18:48

the point of view of our

18:51

physics project, you're picking a particular

18:53

location in branch-heel space. You're picking

18:55

one of the possible outcomes of

18:57

what could have happened with that

18:59

electron. You, as the observer, have

19:01

just are at a particular place

19:04

in branch-heel space, and so you're

19:06

seeing the electron in that form

19:08

rather than in a different form.

19:10

But in any case, so there

19:12

is this attribute of spin in

19:14

electrons, and when electrons scatter against

19:17

each other or scatter magnetic magnetic

19:19

field... electric field, magnetic field, whatever,

19:21

it can change the spin. Well,

19:23

this question of whether you can,

19:25

you know, can you communicate using

19:27

spin, I think spin is a

19:30

very quantum mechanical phenomenon. And so

19:32

as soon as you're sort of

19:34

using it to do things, you're

19:36

sort of throwing yourself into quantum

19:38

mechanics. And that might be a

19:40

good thing if you want to

19:43

check that nobody tampered with your

19:45

electron from where it started to

19:47

where it ended up, because if

19:49

they'd tampered they would have made

19:51

this measurement and they would have

19:53

determined things and so on and

19:56

you would have you would be

19:58

able to tell that it had

20:00

been tampered with. But it's also,

20:02

it's the kind of the general

20:04

rules about kind of how you

20:06

think about things don't don't apply

20:09

there. But so I think the

20:11

answer, well, I mean electron spins

20:13

are in the in the... in

20:15

bulk are what lead to things.

20:17

like magnetism, ferromagnetism, a permanent magnet

20:19

is a bunch of, in the

20:22

iron or cobalt and nickel or

20:24

get a linear, or whatever you're

20:26

using for your magnet, the electrons,

20:28

there are some electrons in there

20:30

that have their spins all aligned,

20:32

and that's what leads to this

20:35

overall magnetic field, because associated with

20:37

the spin of electron, is a

20:39

magnetic field that's kind of like

20:41

a little tiny bar magnet that

20:43

generates that magnetic field. So I

20:45

think. The, the, the, yes, you

20:48

can imagine using electron spins and

20:50

the interaction between electron spins to

20:52

sort of carry information necessarily in

20:54

a very quantum mechanical way. Usually,

20:56

cubits, as discussed in quantum computing

20:58

and quantum circuits, one imagines that

21:01

one of the common sort of,

21:03

well, both ways to implement them

21:05

and ways to think about them

21:07

is each cubit is like a

21:09

spin that can be either, let's

21:11

say up or down. Let's see.

21:14

Let's see, Intense is asking, how

21:16

would we evolve to live in

21:18

space? Would we even evolve without

21:20

going into space? You know, evolution

21:22

by natural selection requires that some

21:24

organisms live, some die, that the

21:27

organisms that are more favorable for

21:29

the environment get to have more

21:31

children who survive, and that means

21:33

that more of the genes that

21:35

more of the genes that lead

21:37

to whatever it is that makes

21:40

those children survive, be produced and

21:42

survive, those genes will become more

21:44

prevalent in the population. You know,

21:46

if we have a Mars colony

21:48

or something at some point, and

21:50

there are people who kind of

21:53

grow up on Mars and so

21:55

on, you know, if we are

21:57

sort of, if it's the case

21:59

that, that I mean, there'll be

22:01

all kinds of no doubt strangely

22:03

different medical issues. to do with

22:06

different gravity levels to do with

22:08

well I'll assume one can shield

22:10

the radiation and so on, but

22:12

I wouldn't doubt that we would

22:14

experience things that the kinds of

22:16

issues we know in microgravity in

22:19

space, they're all in orbit around

22:21

the earth, for example, where you're

22:23

constantly sort of falling and you're

22:25

therefore weightless. You have essentially zero

22:27

gravity. You know, the way us

22:29

humans. were not built for zero

22:32

gravity, so to speak, and all

22:34

kinds of things like our muscle

22:36

mass decreases, I think our immune

22:38

system function decreases, all sorts of

22:40

things like that, for sometimes known

22:42

reasons, with the muscle mass pretty

22:45

obvious, you're not, you're not, you

22:47

know, unless you put effort into

22:49

it, you're not exercising your muscles

22:51

in the same way as you

22:53

do supporting yourself against the force

22:55

of gravity, the other ones maybe

22:57

not so obvious, but this question

23:00

of sort of sort of would

23:02

the humans on Mars evolve, Well,

23:04

you know, only if there is

23:06

sort of preferential survival of ones

23:08

that are more Martian suitable than

23:10

others. I think in that's so

23:13

it's kind of a sort of,

23:15

it's always an irony. I mean,

23:17

even Charles Darwin pointed this out

23:19

that, you know, the struggle for

23:21

life as he called it, that

23:23

this thing that is sort of

23:26

things fighting with tooth and claw,

23:28

I think he talked about, from

23:30

that sort of kind of ugly

23:32

war of creatures, so to speak,

23:34

emerges the thing that we think

23:36

is really cool, which is higher

23:39

organisms and humans and all that

23:41

kind of thing. And if we

23:43

don't allow the tooth and claw,

23:45

so to speak, then we don't

23:47

get evolution by natural selection in

23:49

the same way. If we chose

23:52

to modify our genetics and engineer

23:54

our genetics, then yes, we can

23:56

get evolution, although it's a different

23:58

kind of evolution, you know, in...

24:00

in animals we often do breeding

24:02

or plants, we often do breeding

24:05

where we're picking, you know, what

24:07

gets to mate with what, so

24:09

to speak, from the outside, we

24:11

can sort of go more extreme

24:13

and just say, okay, let's just

24:15

pick the genome. We can't do

24:18

that yet, but one suspects that

24:20

eventually one will be able to

24:22

do that. It's a very complicated

24:24

issue and a complicated ethical issue

24:26

of if you can pick the

24:28

genome, you know, you are creating

24:31

a human like... like all of

24:33

us, so to speak, but you

24:35

are creating a human. I mean,

24:37

we have some control over the

24:39

humans we create, so to speak,

24:41

when we choose to have children,

24:44

not to have children, whatever else.

24:46

But by the time you actually

24:48

are deciding, okay, my child is

24:50

going to have this particular sequence

24:52

of base pairs. that's a higher

24:54

level of control than we've understood

24:57

that we could have before and

24:59

it raises all sorts of very

25:01

thorny ethical issues about sort of,

25:03

given that you have that control,

25:05

what do you do with that

25:07

control? In some sense, it's easier,

25:10

from an ethical point of view,

25:12

in some situations, to just sort

25:14

of say, well, you know, nature

25:16

will take its course. And it's

25:18

not like there's a human choice

25:20

being made that we can be

25:23

ethically concerned about. It's just that's

25:25

what nature is going to do.

25:27

Let's see. Well, somebody's noticing that

25:29

I have probably the same shirt

25:31

as I had last time. I

25:33

think it may be a different

25:36

actual shirt, but... I'm not sure

25:38

this is what I'm not in

25:40

my most natural habitat here, so

25:42

that's likely to change. Okay, let's

25:44

see. Athelson asks, do you think

25:46

that the rules of... human biology

25:49

are computationally reducible so that we

25:51

eventually will be able to understand

25:53

the aging of ourselves. You know,

25:55

that's an interesting question. You know,

25:57

is aging fundamentally a feature of

25:59

computational irreducibility? Is it something where

26:02

we have the rules as we

26:04

run them, they will just do

26:06

certain things and we don't get

26:08

to have sort of an overarching

26:10

theory of what's happening? Is something

26:12

like aging a bit like the

26:15

law of entropy increase that well

26:17

the molecules just bouncing around and

26:19

eventually with the way that they

26:21

bounce around makes makes them seem

26:23

random to us or and that's

26:25

sort of all we can say

26:28

or can we have a can

26:30

we make bigger statements about what's

26:32

happening with with molecules bouncing around

26:34

we know that a large scale

26:36

we can talk about things like

26:38

fluid mechanics. where we can kind

26:41

of describe the large scale motions

26:43

of molecules without having to deal

26:45

with sort of the microscopic details

26:47

of that. But I don't think

26:49

we know. It's a very interesting

26:51

question. I've sort of paid attention

26:54

to it for 40 years or

26:56

so of sort of what causes

26:58

aging and then what might we

27:00

do about it? Because we do

27:02

know that if we restart the

27:04

organism from the same DNA... and

27:07

we just let it run again,

27:09

you'll get a nice young organism

27:11

that will lead another life, so

27:13

to speak. So it's not as

27:15

if in the course of our

27:17

life, everything is lost. It's that

27:20

in the course of our life,

27:22

the particular instance of it that

27:24

is us, sort of gradually ages

27:26

and degrades. I mean, it's kind

27:28

of like with the computer, it's

27:30

running its operating system, it, the

27:33

operating system will gradually, things will

27:35

happen to it, and the computer

27:37

will start running slowly, and the

27:39

memory will fill up, and this

27:41

and that. But if we reboot

27:43

the computer, then we, then it's,

27:46

then it's kind of stopped. from

27:48

the same genetic material again, and

27:50

it's good as new, so to

27:52

speak. So, you know, there have

27:54

been a bunch of theories about

27:56

sort of what leads to aging.

27:59

There was one sort of very,

28:01

very clear theory that said, well,

28:03

the end caps on our DNA,

28:05

the telomeres, the repeating sequences at

28:07

the ends of our DNA, you

28:09

know, we have maybe about 50

28:11

of those typically, and maybe after

28:14

we start from our kind of

28:16

initial... initial sort of single cell

28:18

that we start off as, that

28:20

after we replicate that 50 times,

28:22

those telomeres will just all sort

28:24

of fall off and the DNA

28:27

will just untangle and we won't

28:29

be able to replicate the DNA

28:31

any further. Well, that theory, of

28:33

course, like everything in biology, it's

28:35

more complicated than one thinks and

28:37

there are certainly enzymes that add

28:40

telomeres. and that's what happens when

28:42

you go back to the, you

28:44

know, to the fertilized egg cell

28:46

again. It's, it's got its full

28:48

complement of telomeres, courtesy of, I

28:50

guess what is it, telomeres, I

28:53

think. And, you know, then there

28:55

started to be companies that could

28:57

sort of measure your effective age

28:59

by telling you how many telomeres

29:01

you had left in the particular

29:03

sample of your DNA that they

29:06

got. And, you know, it's kind

29:08

of fun to do those. But

29:10

then you realize that, well, actually,

29:12

you can increase the number of

29:14

telomeres by, you know, good diet

29:16

and exercise and things like this.

29:19

And that doesn't seem like a

29:21

theory of aging. It doesn't seem

29:23

like it's like every time the

29:25

cells replicate one telomere falls off.

29:27

So, but then there's the whole

29:29

question of what, you know, so

29:32

that's sort of one partial theory

29:34

of aging. Other theories of aging

29:36

have to do with oxidative damage.

29:38

Essentially, things burn up. Things are.

29:40

in existing cells just sort of

29:42

gradually in an existing cell they

29:45

gradually sort of there's the equivalent

29:47

of combustion, where from metabolism, that

29:49

causes things to molecules to not

29:51

be in the same sort of

29:53

organized form that they originally were.

29:55

I mean, then there are other

29:58

things like progressive genetic damage that

30:00

builds up, although there are proofreading

30:02

enzymes that try to after every

30:04

time our DNA replicates, try to

30:06

make sure that there aren't at

30:08

least... very small changes that got

30:11

made, though those can be corrected

30:13

from what's around that change. But

30:15

so it's still a bit mysterious

30:17

what the real sort of cause

30:19

of aging is. And I think

30:21

it is something where sort of

30:24

the whole computational irreducibility story is

30:26

potentially quite relevant. I tend to

30:28

think that eventually we will have

30:30

more of an overarching theory. of

30:32

biology than we have right now.

30:34

Biology has been rather allergic to

30:37

theory, because a lot of simple

30:39

theories don't work in biology. Biology

30:41

has this sort of meta feature

30:43

that usually if you try and

30:45

explain something in biology, the most

30:47

complicated possible explanation, but there's many

30:50

footnotes and as many special cases,

30:52

will be what's actually going on.

30:54

Whereas in something like physics, it's

30:56

much more likely that the simplest

30:58

explanation will be the right explanation.

31:00

So I think... That's something to

31:03

realize that, but I think it

31:05

is a, you know, I do

31:07

think that sort of the the

31:09

problem of aging is probably solvable.

31:11

I mean, we certainly know what

31:13

we know that you go back

31:16

to the beginning again with another

31:18

generation of organism and the clock

31:20

starts again. It is very likely

31:22

that Well, so there's a question

31:24

of different kinds of organisms, different

31:26

species, have very different aging characteristics.

31:29

One of the things that's... seems

31:31

a little bit disappointing, if true,

31:33

is that we just evolved to

31:35

die after a certain period of

31:37

time because from the point of

31:39

view of the whole species, that

31:42

was a good thing. It might

31:44

not be a good thing from

31:46

our individual point of view, but

31:48

in a sense, if every member

31:50

of your species is immortal, but

31:52

they all get a bit slower

31:55

and so on, eventually it's kind

31:57

of like forget problems of social

31:59

security, you've got the infinite version

32:01

of that, of the young are

32:03

always sort of the young and

32:05

agile are taking care of the

32:08

old and just sort of hanging

32:10

around, or alternatively another thing to

32:12

say is it could be the

32:14

case, and that, you know, the...

32:16

the organism that learns something when

32:18

it's very young, and then it

32:21

lives for a really, really long

32:23

time, it's like, I know everything

32:25

I need to know, I never

32:27

lead to learn anything new. And

32:29

it's kind of like, there can

32:31

be no progress made. And for

32:34

all we know, sort of aging

32:36

is an evolutionary adaptation that is...

32:38

kind of the thing that makes

32:40

progress possible, and that without that

32:42

it would just be like, you

32:44

know, everybody is, I mean, you

32:47

do see this, I have to

32:49

say, but cynically in some organizations

32:51

where the people there are sort

32:53

of just getting older and doing

32:55

the same thing that they always

32:57

did. and it takes the young

33:00

to come in and change things.

33:02

It isn't always the young who

33:04

changed things. That's a bit of

33:06

an illusion, I think. Maybe I'm

33:08

just speaking as an ancient person,

33:10

but it is not my impression

33:13

that sometimes the young say, I'm

33:15

embedded in this environment, this is

33:17

the only way it can be.

33:19

And the older, a bit more

33:21

like, yeah, I kind of know

33:23

that environment. Actually, there's other things

33:26

one could do type thing. But

33:28

any case. So, you know, there's

33:30

this question of is there a

33:32

way. Is there a way. Is

33:34

there a way. Is there a

33:36

way. If it is indeed something

33:38

that evolution has put upon us

33:41

for the benefit of the species,

33:43

if not the benefit of us

33:45

as individuals, or even our current

33:47

conception of the species, then maybe

33:49

there's just a way to flip

33:51

the switch and say, okay. don't

33:54

do that thing that was evolutionarily

33:56

set up to do. You know,

33:58

people have imagined an elixir of

34:00

eternal youth, so to speak, sometimes

34:02

got called the philosopher's stone back

34:04

in the Middle Ages and so

34:07

on. You know, is it conceivable

34:09

that such a thing could exist?

34:11

I think it's conceivable. I mean,

34:13

an example of something that isn't

34:15

that, but sort of almost maybe

34:17

cut might have been that, was

34:20

this question about stem cells. So,

34:22

you know, in... in a human,

34:24

for example, start from one cell,

34:26

it divides, it divides, it keeps

34:28

on dividing. Eventually, it differentiates into

34:30

cells that will be heart muscle

34:33

and brain cells and skin cells

34:35

and so on. And that process

34:37

of differentiation, I think there's like

34:39

12 or 13 levels of differentiation

34:41

in us humans. We don't have

34:43

that many types of cells in

34:46

the end. Lots of copies of

34:48

each individual type. But the... that

34:50

process of differentiation, it was thought

34:52

that the only way you can

34:54

have a pluripotent stem cell, a

34:56

stem cell that can turn into

34:59

any other kind of cell, was

35:01

to sort of restart and have

35:03

the new generation in the fertilized

35:05

egg cell or fetal cells or

35:07

whatever else. But then it was

35:09

discovered, what was it, maybe 15

35:12

years ago or something now, that...

35:14

there was just a way to

35:16

essentially erase the memory of what

35:18

a cell was supposed to be

35:20

and take something like a skin

35:22

cell and reprogram it to be

35:25

a stem cell. It doesn't always

35:27

work perfectly. It's, you know, it's

35:29

a sort of a question of

35:31

does that erasure of memory, does

35:33

it really work or does the

35:35

cell somehow in some corner sort

35:38

of remember that it was a

35:40

skin cell and does the, is

35:42

it's genetics really stable? or when

35:44

you've done that erasure process, does

35:46

that make the genetic somehow start

35:48

splitting out and making tumors and

35:51

things like this? And you might

35:53

have thought, well, let's just revert

35:55

ourselves all to be being, you

35:57

know, lots of cells to be

35:59

being stem cells. That's a really

36:01

bad idea to do in general

36:04

because basically then you just everything

36:06

would turn into a tumor basically

36:08

But the the possibility that

36:10

one could build stem build

36:12

from stem cells one could

36:14

build Cells that can replace

36:17

cells that exist specific kinds

36:19

of cells that exist That's

36:21

a real thing and you know

36:23

big successes recently with

36:26

pancreatic beta cells for

36:28

curing type one diabetes,

36:30

some success I think

36:32

with cardiomyocytes, muscle cells

36:35

for the heart, some successes, although

36:37

not deployed in brain cells and so

36:39

on, I think it's kind of like

36:41

almost the catalog you can say, well

36:43

what kinds of cells do we now

36:46

know how to take from a created

36:48

stem cell? to guide them through this

36:50

pathway of differentiation and give them this

36:52

thing and feed them that and have

36:54

them do this for five days and

36:57

then have them do that for a

36:59

week and this and that and the

37:01

other to sort of chemically trick them

37:03

into deciding that they're going to turn

37:05

into this particular type of brain

37:08

cell. But so, you know, that's an

37:10

example of something which doesn't make it

37:12

as the elixir of eternal youth, so

37:14

to speak. But it's a, you know,

37:16

it's a thing in that direction. I

37:19

mean, another, another thing that people

37:21

talk about is the idea that,

37:23

you know, blood from young organisms

37:25

has stuff in it that helps,

37:27

that is more energizing. And if

37:29

you give the blood from the

37:31

young organism to an old organism,

37:33

it has to, it has to

37:35

match. if you don't want to

37:38

have to suppress the immune system

37:40

and suppress the immune system, causes

37:42

so many other problems. But, you

37:44

know, then that's a mysterious way

37:46

that one is sort of transporting

37:48

youth into the future. And maybe

37:50

that's something where it would be

37:52

possible to identify what the particular

37:54

factors are that say, you know, I'm

37:56

young blood and so to speak and sort

37:59

of activate things. to make the

38:01

organism feel younger, so to

38:03

speak. Let's see. I mean, just

38:06

to say, I mean, I think there

38:08

are different approaches. The

38:10

thing with stem cells

38:12

right now is using them to

38:14

kind of create specific

38:16

organs. You know, if you build

38:19

the scaffold of a lung and

38:21

then you fill it or the

38:23

scaffold of a kidney. and then

38:26

you fill it with the right

38:28

with stem cells and get those

38:30

stem cells to become kidney cells

38:32

or become lung cells, then there's

38:34

a chance that you can create an

38:37

artificial organ that will be if

38:39

you if the stem cells come

38:41

from you, then your immune system

38:43

will be happy with whatever cells

38:45

came out from those stem cells

38:48

that it when in general when

38:50

you have organ transplants and things

38:52

like that the because we all

38:54

have slightly different sort of bar

38:56

codes for our immune system. Our

38:58

immune system is trying to recognize,

39:00

does that thing that just got

39:03

put into me, is that part

39:05

of me or is that an

39:07

alien thing? And it's, if you take,

39:09

you can get a high degree

39:12

of compatibility, but not perfect, just

39:14

because the commonatorics are too big,

39:16

in, you know, unless you're dealing

39:18

with an identical twin, it's not

39:21

going to be the exact same

39:23

genetics. But if you have a

39:25

stem cell that was created from

39:27

you, then it will have the

39:29

exact same genetics as you, and

39:31

so your immune system will be

39:34

perfectly happy with it. And if

39:36

you can have turned those stem

39:38

cells into something that's useful, you

39:40

know, a new, you know, organ

39:42

of some particular type or something

39:45

like that, then, you know, it's a

39:47

whole effort to sort of, you know,

39:49

piece of surgery to connect it in

39:51

and so on, but... that's how that

39:53

can work in terms of sort of

39:55

the broad scale reversal of aging. It's

39:57

just not at all known how to

39:59

do that. It's not, and you know,

40:01

as I say, just saying, revert everything

40:03

to being a stem cell is absolutely

40:05

not the right thing to do. And

40:08

whether there is a sort of

40:10

a broad, you know, this is

40:12

how aging broadly works, and there's

40:14

some switch that you can turn

40:16

back, something you can turn back

40:19

broadly, it's just not known. And

40:21

it's a very interesting question whether

40:23

sort of thinking about things, computation.

40:25

thinking about biology in this kind

40:27

of broad computational way might lead

40:30

one to a way to think

40:32

about that question. I think it's

40:34

very interesting. And as I get

40:36

older, I get more and more

40:39

sort of interested in that

40:41

question. Well, a bunch of questions

40:43

here about AI and LLLMs. Let

40:45

me see what, well, those questions

40:47

here, one from Waffle about

40:50

latest LLLMs, they say, are doing

40:52

very advanced mathematics mathematics. Do

40:54

you think we can get

40:57

AI to the point that

40:59

it's solving open problems and

41:01

creating new mathematics? I've kind

41:04

of written a fair amount about

41:06

that actually. Doing very

41:08

advanced mathematics, yes, they can

41:10

write something that seems like

41:12

a math paper. When you

41:15

start to dissect it, it often

41:17

falls apart. You have to be pretty

41:19

lucky to end up something that

41:21

really follows through. and means the

41:23

right thing at the end. It's,

41:25

you know, my analogy for this,

41:27

what's happening in machine learning is

41:30

kind of like it's building something like

41:32

a stone wall. It's got a

41:34

bunch of rocks that are certain

41:36

shapes that it can sort of

41:38

pluck out of the computational universe

41:40

and it's assembling those to try

41:42

and build up what you ask

41:44

it to build. And that's something

41:46

that works up to a point,

41:48

but it's hard to build a

41:50

skyscraper without of random-shaped rocks and

41:52

so on. the thing that in

41:54

general sort of can you

41:57

build sort of this whole chain

41:59

of of mathematical reasoning and so

42:01

on and expect the skyscraper not

42:03

to fall over, well only if

42:05

you do it in a precise

42:07

formal way. And what we've built with

42:10

Wolfram language is that story of

42:12

being able to do computation for

42:14

mathematics and many other things in

42:16

that sort of precise formal way

42:18

and being able to be able

42:20

to build that tower sort of as

42:23

tall as you want. And I think this

42:25

question of whether... sort of how AI

42:27

helps with that. One thing that actually

42:29

I'm about to start to really trying

42:31

to push on is, well, it read

42:33

the literature. It read a million papers.

42:36

And so it has sort of a

42:38

broad idea, a broad sort of vague

42:40

understanding of sort of how things fit

42:42

together that can be very useful

42:44

to us. You know, I feel like... I have a

42:47

broad understanding of how things fit together,

42:49

but it knows a lot more detail

42:51

than I do, and I think it

42:53

could potentially help in sort of defining,

42:55

if I give it a rough direction,

42:57

being able to fill in a little

43:00

bit more detail so that one can

43:02

know in what direction to go. I think...

43:04

There are questions about sort of, can

43:06

you, if you're proving films in some

43:09

formal way, can you get the AI

43:11

to sort of pick the steps you

43:13

choose? I'm skeptical about that one. I

43:15

think that it is, AIs do well

43:17

at sort of human-like tasks, like

43:19

saying what's a plausible next word

43:21

or next sentence for this essay

43:23

or what's a plausible thing that

43:25

somebody writing this math paper would

43:27

say next. When it comes to

43:30

these much more austere and very

43:32

nonhuman things, like a long automated

43:34

proof. they're really kind of

43:36

lost in that case, at least

43:38

in what I've been able to

43:40

figure out so far. And I

43:43

think the thing to say in

43:45

general is, you know, the current

43:47

generation of sort of chain of

43:49

reasoning AI models of which Deep

43:52

Seek is the is the big

43:54

excitement of the last couple of

43:56

weeks, is what's happening there. What's

43:59

happening there? sort of the thing

44:01

that is interesting and surprising is

44:03

that it seems like the AI is

44:05

kind of planning and then it's going

44:08

through and it's trying various things and

44:10

if something doesn't work it backs up

44:12

and it tries something different and it

44:14

seems like it's really doing a very

44:16

human-like exploration thinking exploration so to speak

44:19

it even has a little tag called

44:21

think that means that the thing that

44:23

it's you know producing now is it's

44:25

in our thoughts so to speak but You

44:27

know, one of the things that's true

44:29

right now is that right now, the

44:31

way the system works is it's making

44:34

a plan and then it's trying to

44:36

execute on that plan and maybe it'll

44:38

go back and forth in parts of

44:41

that plan, but it's sort of already

44:43

made the plan. It isn't kind of

44:45

looking around as it progresses through the

44:47

plan and saying, well, let me... look

44:50

at this piece of computation here and

44:52

change the plan based on what happened

44:54

from that piece of computation. I mean,

44:56

I will say that I think that

44:59

the idea of making sort of a

45:01

chain of reasoning that is many steps

45:03

long, if you can't turn those steps,

45:05

into something hard and computational, it just

45:08

won't work to have a large number

45:10

of steps assembled one after another, because

45:12

as soon as something's gone wrong at

45:14

one step because it's a bit mushy,

45:16

the whole of the rest of the tower

45:19

is going to topple over. And

45:21

something we've been looking at quite

45:23

a bit is using our computational

45:25

language as the thing into which

45:27

to crisp and kind of the

45:29

pure LLLM side of things, so

45:31

that at every step... you're saying,

45:33

I want to, the LLLM, is

45:35

forcing itself to represent what it's

45:37

talking about in a precise computational

45:39

way that can be expressed in

45:41

our language and where we can

45:43

do all kinds of computations from

45:45

it, and more importantly, it's something

45:47

precise where it kind of, there's a

45:49

definite brick that got assembled there that

45:51

you can then go on and continue

45:54

the chain from there knowing that that

45:56

part of the chain is not, is

45:58

not mushy. So I think that that

46:00

That's some, now, you know, in, in, I think

46:02

in the future, sort of this

46:04

interleaving of kind of LLLM,

46:07

neural net type activity with

46:09

computation is really sort of

46:11

the winning combination, but the

46:14

thing that it's been very

46:16

difficult to figure out is

46:18

how to do fine-grained interleaving.

46:20

The typical interleaving that one

46:22

does right now is that the

46:24

LLM will keep going and eventually

46:26

it will decide okay I need

46:28

to call Wolfram language or Wolfram

46:30

Alpha or something and I will

46:32

generate some text together with a

46:34

stop token and it will just

46:37

the text will say okay send

46:39

this stuff to to Wolfram language

46:41

and then the the LLM will stop

46:43

and some external harness will pick that

46:45

up and say okay I see what

46:48

I should send to Wolfram language it

46:50

will go and get a result back.

46:52

It'll then, and then it'll read

46:54

that result, and then it'll treat that

46:56

as part of what was in its

46:58

context from before, and then it

47:01

will keep going from there and

47:03

make a conclusion. I mean, there are

47:05

a couple of technologies. One is

47:07

Rags, retrieval augmented generation, where you're

47:10

saying, basically do a search in

47:12

a collection of documents for things

47:14

that roughly match some particular query.

47:16

that comes into the system, and

47:18

then you're saying, okay, I know

47:21

based on that query and based

47:23

on the search I've done, here

47:25

are 100 things that you, the

47:27

LLLM, might like to think about.

47:29

And that's then very useful

47:31

for what the LLLM produces after

47:34

that. It doesn't have to know

47:36

everything itself. It just got essentially

47:39

prompted with 100 things it might

47:41

think about. And for example, our

47:43

Wolfen notebook assistant that came out

47:45

a month or so ago now,

47:47

is a... It has a lot of

47:50

technology around that kind of thing

47:52

built in where depending on what

47:54

you've asked about it will be

47:56

sort of hinted that it will

47:59

essentially have the documentation and so

48:01

on, and it will get hints

48:03

about what you might want to

48:06

think about this or that. But

48:08

so this notion of retrieval augmented

48:10

generation, where you're retrieving things from

48:12

existing documents, that's one thing. The

48:15

thing that we're really able to

48:17

do a lot with is computational

48:19

augmented generation, where instead of there

48:21

being a fixed thing where you're

48:24

looking up... something that matches a

48:26

fixed document, you're instead saying, I've

48:28

got this thing that I produced

48:30

as an other lamb, now going

48:33

to compute from that, and sort

48:35

of an infinite universe of things

48:37

you could compute, you get back

48:40

the result of the computation, and

48:42

you use that to augment the

48:44

future generation that you make. But

48:46

I think that the emerging story

48:49

will be one of sort of

48:51

computation and AI kind of hand

48:53

in hand in hand. the sort

48:55

of neural net approach with training

48:58

and so on, together with the

49:00

kind of computational approach that I've

49:02

spent, well, the last 45 years

49:04

developing, of representing the world computationally.

49:07

It's different from a programming language.

49:09

A programming language is, you know,

49:11

C, Java, Python, whatever, is about

49:14

kind of representing what goes on

49:16

inside a computer. and letting us

49:18

sort of tell the computer in

49:20

its terms what to do. The

49:23

whole idea of computational language and

49:25

our world language is to represent

49:27

the world computationally, to represent, you

49:29

know, cities and chemicals and graphs

49:32

and images and so on in

49:34

a computational way in sort of

49:36

the way that that so that

49:38

we can represent things in the

49:41

world computationally and manipulate those things

49:43

in some in some precise. sort

49:45

of abstract way, not just writing

49:48

programs so to speak for the

49:50

innards of the computer, but representing

49:52

the world in a precise way.

49:54

And I think that's the thing

49:57

that really opens up the possibility

49:59

of kind of sort of going

50:01

back and forth between the neural

50:03

net that has its kind of

50:06

rough kind of pattern matching based

50:08

way of thinking about the world

50:10

and the sort of formalized way

50:12

of thinking about the world where

50:15

one can sort of build whole

50:17

towers of reasoning and so on.

50:19

Let's see. What's the next step

50:21

for LLLM's to advance? So

50:24

that's an interesting question. I mean,

50:26

I think that, well, there's several

50:28

different things. I mean, there are

50:31

different kinds of training data. There

50:33

are things about, you know, video

50:35

is starting to come online. There'll

50:37

be robotic training data, and that

50:39

will allow one to have kind

50:42

of things that... are for sort

50:44

of modeling what happens with robots

50:46

and probably humanoid robots because that's

50:48

the ones where we're going to

50:50

have sort of as much more

50:53

knowledge than anywhere else about sort

50:55

of what happens in the world

50:57

but the thing that's humanoid shaped.

50:59

I think the that's one kind

51:01

of thing. Another kind of thing

51:04

is well We were all surprised

51:06

by Chat GBT, we were all

51:08

surprised that it turned out to

51:10

be easier than we expected to

51:12

produce fluent essays that kind of

51:15

coherently made sense over quite long,

51:17

long, long stretches. And, you know,

51:19

we kind of thought that human

51:21

language was something more special than

51:23

that. We knew that human language

51:26

had certain grammatical structure. People who

51:28

have been doing computational linguistics since

51:30

the 1960s have known that, you

51:32

know, It's sort of thought about

51:34

language as, you know, noun, verb,

51:37

noun, etc, etc, etc. subject, verb,

51:39

object, in English, all those kinds

51:41

of things. What we sort of

51:43

realize now is in addition to

51:45

the rules of grammar that tell

51:48

us sort of syntactically how to

51:50

put sentences together what part of

51:52

speech follows, what part of speech.

51:54

There's also kind of... the idea

51:56

of the semantic grammar of language,

51:59

what concepts fit with what other

52:01

concepts. And I think, you know,

52:03

my sort of theory of chatGBT

52:05

is that, and LLLMs in general,

52:07

is that what they sort of

52:10

learnt, they inferred from kind of

52:12

large chunks, a trillion words of

52:14

language or something, they were able

52:16

to sort of statistically infer in

52:18

some sense, kind of rules of

52:21

semantic of semantic grammar. that they

52:23

can then use to produce meaningful

52:25

sentences. Now an interesting question that

52:27

arises when you look at things

52:29

like deep seek that seem to

52:32

be sort of emerging where there

52:34

seems to be emergent reasoning happening.

52:36

It's not quite as deep perhaps

52:38

as one might have hoped, but

52:40

there's some emergent way in which

52:43

arguments are being created and executed.

52:45

And the question is, what's the

52:47

sort of formal representation of those

52:49

plans? At some level... when it

52:51

comes to human language, we sort

52:54

of have the idea from grammar,

52:56

syntactic grammar, of what it might

52:58

look like to have a sort

53:00

of formalized structure to represent language.

53:03

There's a question, what does it

53:05

look like to have a formalized

53:07

structure for reasoning? And I'm pretty

53:09

sure there's an easy answer to

53:11

that. And it may be very

53:14

close to programs that we know

53:16

very well from even programming languages,

53:18

let alone computational language. I think

53:20

the... It's kind of in a

53:22

sense logic is a sort of

53:25

base level of kind of a

53:27

structuring for sentences that leads to

53:29

reasoning. It's, and maybe you can

53:31

extend it a bit further than

53:33

that, but I guess the question

53:36

is when you have these much

53:38

larger scale plans, what are the,

53:40

you know, what, what, what, what

53:42

kind of formal thing do you,

53:44

can you use to describe that

53:47

chain of reasoning? And I think

53:49

that will help us to understand,

53:51

you know, just how significant, is

53:53

the fact that this emerges. Maybe

53:55

it tells us something about science

53:58

or something about philosophy, more so...

54:00

than, and maybe we can then

54:02

just take what the LLLM essentially

54:04

discovered for us, and then take

54:06

it and use it in a

54:09

much more formal way, rather than

54:11

having to make the LLLM rediscover

54:13

it every time, so to speak.

54:15

I noticed a question, the comment

54:17

from Taki saying, I should do

54:20

a conversation with Yoshabak. I know

54:22

Josha quite well. I think we've

54:24

done, I think we did a

54:26

conversation on his podcast once, but

54:28

I know him quite well. He's

54:31

been trying to get launched with

54:33

this California Institute for Machine Consciousness,

54:35

which I've been helping a bit

54:37

with. If nothing else, it's a

54:39

good science fiction name. And one

54:42

sort of imagines the, well. It's

54:44

I kind of I kind of

54:46

I kind of imagine it at

54:48

some level as a place of

54:50

experimental philosophy so to speak where

54:53

it's kind of like you've got

54:55

this you know you potentially have

54:57

sort of the artificial brain on

54:59

which you can do philosophical experiments

55:01

philosophy has generally not been an

55:04

experimental science it's been a theoretical

55:06

science where you just sort of

55:08

have to think about things but

55:10

it's something which in modern times

55:12

you know kind of the philosophy

55:15

is something you can imagine kind

55:17

of experimentally exploring with things like

55:19

LLLMs. You know, I had fun

55:21

a few months ago with a

55:23

humanoid robot that had an LLLM

55:26

inside, so to speak. I did

55:28

a live stream, which I'm sure

55:30

you can find on the web

55:32

of a surprisingly long conversation with

55:34

a humanoid robot sort of powered

55:37

by an LLLM, which was kind

55:39

of... to me sort of viscerally

55:41

philosophically interesting in the sense that

55:43

it really was a strange feeling

55:45

to talk to this humanoid thing

55:48

with eye motions that were fairly

55:50

realistic, although I think it stared

55:52

a lot more than humans stared.

55:54

which got me very, kind of,

55:56

was very disquieting after a while.

55:59

But it was sort of an

56:01

interesting experience in kind of experimental

56:03

philosophy, so to speak. Memes asks,

56:05

have any LLLM agents been trained

56:08

on my big book, New kind

56:10

of science, or maybe on the

56:12

science, that's in New kind of

56:14

science? We have tried to train

56:16

some LLLMs on that. It's not

56:19

terribly successful so far, because... In

56:21

a sense, well, it's not been

56:23

successful in doing the thing that

56:25

is really a big reach, which

56:27

is to sort of break computational

56:30

irreducibility, to be able to say,

56:32

given just that you feed the

56:34

rules in, have the AI jump

56:36

ahead and say what will happen.

56:38

We don't think that's theoretically possible

56:41

in general. But there is a

56:43

question of sort of the things

56:45

of the sort of pockets of

56:47

reducibility that exist, to what extent

56:49

can those be found by an

56:52

AI? And if you feed it

56:54

enough different examples of different kinds

56:56

of things, where we humans notice

56:58

that, oh, there are some regularities,

57:00

like, for example, back 40 years

57:03

ago now, I kind of noticed

57:05

that there were four basic classes

57:07

of behavior in these things called

57:09

cellular automata, these simple rules, computational

57:11

rules that you can run, there

57:14

were four basic classes of behavior

57:16

that you could kind of busially

57:18

identify. And that's something that I

57:20

would think. is perfectly accessible to

57:22

kind of neural net investigation. So

57:25

it's an interesting question, kind of

57:27

what, if you feed the LLLMs

57:29

or AIs enough kind of NKS

57:31

material, do they start kind of

57:33

inventing a language about it? Do

57:36

they start having kind of effectively

57:38

having discovered a description? It will

57:40

be very alien to us, perhaps

57:42

less alien to me than to

57:44

anybody else, but still very alien

57:47

even to me, even to me,

57:49

even to me. if one starts

57:51

sort of being able to reason

57:53

in terms of these structures that

57:55

exist in this very abstract space.

58:00

Let's see. Well, there's questions here about

58:02

reinforcement learning. Justin is asking about external

58:04

reasoning paradigm, more generally reinforcement learning in

58:06

LLLams. Yeah, I mean, just to explain

58:09

a little bit about that, in LLLam,

58:11

kind of what you're trying to do

58:13

is to make this whole neural net

58:15

that's going to that's going to take

58:18

kind of a representation of a piece

58:20

of text up to some point. And

58:22

it's then going to predict, how does

58:24

that piece of text go on? What's

58:26

the probability that the next word is

58:29

the versus a versus cat versus dog?

58:31

And the way it's trained to do

58:33

that is you've got a whole piece

58:35

of text, and you basically cover up

58:38

the end of the text, and you

58:40

say, OK, I want to tweak this

58:42

neural on that, so that it will

58:44

correctly reproduce what was there if I

58:46

take off the cover. And that's kind

58:49

of the incremental approach to learning. that

58:51

it's both incremental in the way that

58:53

you tweak the neural net and it's

58:55

also it's kind of it's it's learning

58:58

to produce a token at a time

59:00

as it produces that that stream of

59:02

text and typical LLMs you just explicitly

59:04

see them writing out one word at

59:06

a time. That's not just done for

59:09

effect. It's done, sometimes it is a

59:11

little bit done for effect, but a

59:13

large part of that is just, it

59:15

is generating those words sequentially. And just

59:18

like we do probably, you know, I'm

59:20

not sure that I have a plan

59:22

for my next word is going to

59:24

be, it's just that the mechanics of

59:26

my neural nets successfully produces the next

59:29

word. So is it with an L11.

59:31

Okay, so the idea of the enforcement

59:33

learning is not... is to do something

59:35

a bit more global, to say, okay,

59:38

let's look at the whole thing that

59:40

came out, the whole essay that came

59:42

out, the whole answer to that question,

59:44

the whole math problem it did, whatever

59:46

else, and say, is that right or

59:49

wrong? And then say, okay, given that

59:51

that was right or wrong, how does

59:53

that feedback to tweaking? this neural net.

59:55

It's much easier to tweak a neural

59:58

net if you say, well, just made

1:00:00

a small mistake here. I can sort

1:00:02

of propagate back that error and tweak

1:00:04

the weights and the neural net, the

1:00:06

numbers and the neural net to make

1:00:09

it a bit closer to having got

1:00:11

the right answer instead of the wrong

1:00:13

answer there. When you're looking at the

1:00:15

sort of the whole thing, it's more

1:00:17

difficult, it's more complicated, to decide how

1:00:20

you make changes the neural net. to

1:00:22

make it get closer to the thing

1:00:24

you want to be the answer, so

1:00:26

to speak. And it's also the case

1:00:29

that you can expect the neural net

1:00:31

to kind of surface its own questions

1:00:33

to be asking, and then if you

1:00:35

have an outside arbiter and outside critic

1:00:37

of what's happening, and outside, you know,

1:00:40

you're running it against Wolfen language or

1:00:42

something, and it's, Wolfen language is telling

1:00:44

you, that answer is wrong, and then

1:00:46

you can go and sort of feed

1:00:49

that back. and make changes so you

1:00:51

won't get that same wrong answer the

1:00:53

next time. It's sort of the idea

1:00:55

of reinforcement learning as opposed to traditional

1:00:57

LLLM training is that in traditional LLLM

1:01:00

training it's kind of like you're trying

1:01:02

to get a token at a time

1:01:04

and you're sort of training for that,

1:01:06

whereas reinforcement learning you're really training for

1:01:09

whole answers so to speak. And there's

1:01:11

a bit more arbitrariness in how that's

1:01:13

done and there's sort of gradual engineering

1:01:15

advances in how to make that work.

1:01:17

I mean, I think the thing to

1:01:20

realize is that there's another level which

1:01:22

is kind of the harness inside of

1:01:24

which you're operating the AI and doing

1:01:26

things like falling out to computational tools

1:01:29

or whatever else. That's something that happens

1:01:31

kind of at a higher level of

1:01:33

the thing that is sort of watching

1:01:35

what the AI is doing and is

1:01:37

noticing that the AI is saying, hey,

1:01:40

I want to ask, you know, Wolfram

1:01:42

Al for some question or something like

1:01:44

that, and then it's getting that question

1:01:46

picked up, answered, and fed back to

1:01:49

the LLL. As I said, I think

1:01:51

the ultimate future is a much more

1:01:53

fine-grained sort of arrangement of LLLM-type operations

1:01:55

with computational operations, but we don't yet

1:01:57

really know how to do that. There's

1:02:02

a question here from butchering asking

1:02:04

how many years away do you

1:02:06

think we are from gray goo?

1:02:08

Self-replicating nano machines so Was a

1:02:11

scenario from what must have been

1:02:13

the 1990s? People were talked about

1:02:15

before before any worried about the

1:02:18

AIs actually they were already worrying

1:02:20

about the AIs taking over now

1:02:22

that I think about it people

1:02:24

have been worrying about the AIs

1:02:27

taking over since long before can

1:02:29

actually from pretty much about the

1:02:31

time electronic computers started to be

1:02:34

to be deployed. But another kind

1:02:36

of evolution goes past us, the

1:02:38

world passes us by, beyond the

1:02:40

AIs takeover, is a much more

1:02:43

extreme version of that, which is

1:02:45

self-replicating nano machines takeover. Right now,

1:02:47

if you say, sort of how

1:02:49

do you make a machine? that

1:02:52

has components of a molecular scale?

1:02:54

The answer is we, biology, is

1:02:56

sort of the best example we

1:02:59

know of a molecular scale orchestrated

1:03:01

machine where we've got all these

1:03:03

little pieces that fit together in

1:03:05

this very precise molecular way and

1:03:08

one molecule does this and interacts

1:03:10

with another molecule and so on.

1:03:12

We have a bunch of our

1:03:14

operation is at a molecular scale.

1:03:17

There are much less... much sort

1:03:19

of simpler molecular scale things that

1:03:21

can happen like when a crystal

1:03:24

grows it's arranging molecules in this

1:03:26

very precise array and that it's

1:03:28

sort of a molecular scale molecular

1:03:30

scale precision but it only produces

1:03:33

this repeating crystal let's say I

1:03:35

kind of am thinking about crystals

1:03:37

that that have more computation even

1:03:39

than the way that they're formed

1:03:42

but that's a that's a quite

1:03:44

different footnote to the story. But

1:03:46

the main thing is that. Right

1:03:49

now, we are the only self-replicating

1:03:51

molecular scale things that exist. But

1:03:53

you could imagine that we could

1:03:55

simplify the whole process of self-replication.

1:03:58

and we can end up with

1:04:00

just a molecule that just, you

1:04:02

know, or, you know, two or

1:04:04

three molecules that forget all this

1:04:07

stuff with, you know, metabolism and

1:04:09

RNA and cell membranes, all these

1:04:11

kinds of things, and water and

1:04:14

everything, and it could just be

1:04:16

this arrangement of two or three

1:04:18

molecules that have the feature that

1:04:20

those molecules will just pick up

1:04:23

atoms from everything else and just

1:04:25

keep... you know, making more and more

1:04:27

of those atoms. It's kind of like

1:04:29

a polymer, like, you know, plastics or

1:04:32

polymers, where they have, you know, hydrocarbon

1:04:34

pieces that just, you're just adding on

1:04:36

longer and longer and longer chain to

1:04:38

make that polymer molecule. And the question

1:04:40

would be, could you make kind of

1:04:43

a thing that is a more complicated

1:04:45

thing that would sort of polymerize the

1:04:47

whole world? where you would turn

1:04:49

everything in the world, all the

1:04:51

atoms in the world, all the

1:04:53

carbon and silicon and oxygen and

1:04:56

so on that exists in the

1:04:58

world, could you have this little

1:05:00

nano machine that would just, it

1:05:02

could be just a few molecules

1:05:05

that would somehow just sort of

1:05:07

eat up the world and turn

1:05:09

it into itself. Well, we don't know

1:05:11

even close to how to do that.

1:05:13

I mean, when you see, for

1:05:15

example, You know, some crystals,

1:05:17

if you have a super

1:05:19

saturated solution of some particular

1:05:22

material and you suddenly cool

1:05:24

it down, a crystal will

1:05:26

very rapidly form. You know,

1:05:28

you can have situations where it's

1:05:30

just immediately, I don't know, a

1:05:32

nice one that does this is

1:05:34

bismuth, which melts at some very

1:05:37

modest temperature. You can easily melt

1:05:39

it on a regular stove. And

1:05:41

when you let it cool down

1:05:43

quickly, it immediately forms into these

1:05:45

very lovely. kind of square, it

1:05:47

was a very nice looking stuff,

1:05:50

that forms into these little square

1:05:52

pieces and so on, and that

1:05:54

happens very quickly. It kind of

1:05:56

forms that spontaneously just by the way

1:05:58

the atoms are bismuth. fit together. And

1:06:01

so the question would be, could

1:06:03

something like that happen? Could you

1:06:05

have sort of a gray goo?

1:06:07

Its very name of gray goo

1:06:09

kind of says, well, it's all

1:06:11

over if the world starts getting

1:06:13

eaten by the sort of polymerizing

1:06:15

kind of critter that's just replicating

1:06:17

copies of itself. You know, We're

1:06:19

not close to that. In fact,

1:06:21

the whole enterprise of nanotechnology that

1:06:23

was pretty popular in the late

1:06:25

80s, early 1990s, unfortunately, came upon

1:06:27

hard times. I mean, the idea

1:06:29

was, can one take... sort of

1:06:31

machinery that we know exists at

1:06:34

a large scale with clockwork and

1:06:36

levers and all these kinds of

1:06:38

things and can one shrink it

1:06:40

down to a molecular scale and

1:06:42

quite a lot was figured out

1:06:44

about how to do that and

1:06:46

a lot of different sort of

1:06:48

mechanical engineering issues about you know

1:06:50

how do you lubricate something that's

1:06:52

the size of a few molecules

1:06:54

those kinds of questions and and

1:06:56

how much do things just sort

1:06:58

of stick together and how do

1:07:00

you how do you deal with

1:07:02

that. kind of thing. Well a

1:07:04

fair amount was figured out about

1:07:07

that and there was some actually

1:07:09

there was a in the US

1:07:11

there was some government initiatives of

1:07:13

you know let's fund nanotechnology and

1:07:15

really make that a big thing.

1:07:17

Unfortunately I mean what really happened

1:07:19

maybe I'm being a bit too

1:07:21

cynical here is that there was

1:07:23

sort of an interscientist kind of

1:07:25

rivalry between folks like chemists and

1:07:27

material scientists who after all have

1:07:29

been dealing with molecules in their

1:07:31

ways for a long time and

1:07:33

the nanotechnology crowd that was dealing

1:07:35

with molecules their way and it

1:07:38

kind of it and in the

1:07:40

end it seems like what happened

1:07:42

was the nanotechnology direction just sort

1:07:44

of got stomped on by these

1:07:46

existing fields and it really hasn't

1:07:48

been pursued. My own feeling is

1:07:50

that there's a lot of promise

1:07:52

there. I think one thing that's

1:07:54

probably not correct is the idea

1:07:56

of let's take machinery that we

1:07:58

know how it operates on the

1:08:00

scale of centimetres and let's shrink

1:08:02

it down to the scale of

1:08:04

nanometers and have it. be the

1:08:06

same kind of machinery, but just

1:08:08

on a much smaller scale, my

1:08:11

guess is that's not really the

1:08:13

right way to do it. A

1:08:15

better way to do it is

1:08:17

to think about sort of how

1:08:19

do you take the components that

1:08:21

exist of molecules we have and

1:08:23

how do you assemble them to

1:08:25

sort of compile up to a

1:08:27

thing that's useful to us. It's

1:08:29

kind of like saying, well, you

1:08:31

know, we could, if we wanted

1:08:33

to arithmetic arithmetic. we could do

1:08:35

what people like Charles Babbage did

1:08:37

when they made mechanical computers back

1:08:39

in the early 1800s and they

1:08:41

had, you know, a wheel that

1:08:44

had, you know, digit zero through

1:08:46

nine and they had cogs that

1:08:48

connected that to carry bits and

1:08:50

so on. Well, turns out... that

1:08:52

just having a bunch of nand

1:08:54

gates in a microprocessor and arranging

1:08:56

them in the right way, you

1:08:58

can achieve the same thing. You

1:09:00

don't need to build in the

1:09:02

structure of decimal arithmetic with carry

1:09:04

bits and so on to the

1:09:06

generic computer. And my guess is

1:09:08

the same is true with molecular

1:09:10

computation that one can start with

1:09:12

very much more mundane components and

1:09:15

essentially purely in software in effect.

1:09:17

build up to something that is

1:09:19

practically useful. And I think in

1:09:21

a sense that's what chemistry has

1:09:23

been doing forever in chemical synthesis,

1:09:25

synthetic chemistry, and so on, but

1:09:27

chemistry is much less orchestrated than

1:09:29

one imagines molecular nanotechnology to be

1:09:31

in chemistry. the sort of the

1:09:33

main thing is well how are

1:09:35

you going to get molecules to

1:09:37

interact well well liquids are a

1:09:39

really good case because there are

1:09:41

lots of molecules bouncing around but

1:09:43

they bounce around so if two

1:09:45

molecules are going to sort of

1:09:48

fit into each other they'll probably

1:09:50

find each other by in the

1:09:52

liquid and they'll stick and but

1:09:54

they'll go on and find other

1:09:56

molecules if they don't stick whereas

1:09:58

in a solid the molecules are

1:10:00

nice and close together but they

1:10:02

don't move around in the gas

1:10:04

that just aren't enough molecules but

1:10:06

in liquid. you have all this

1:10:08

sort of randomness of motion of

1:10:10

things jiggling around, it's not like

1:10:12

what seems to happen in biology,

1:10:14

where it seems like there's much

1:10:16

more... orchestration of this molecule is

1:10:18

actively moved by this chain of

1:10:21

other molecules to go to this

1:10:23

place and so on. And I

1:10:25

think that's the thing that one

1:10:27

sort of imagines could be engineered

1:10:29

in nanotechnology is this kind of

1:10:31

orchestrated arrangement of molecules. And I

1:10:33

think there's a lot of wonderful

1:10:35

things that can be done with

1:10:37

that. I think it hasn't been

1:10:39

investigated as much as it could

1:10:41

be. In fact, it makes me

1:10:43

wonder whether there are sort of

1:10:45

machine learning type plays that could

1:10:47

be made. sort of both mining

1:10:49

the existing literature and knowing the

1:10:52

certain amounts about chemistry, so to

1:10:54

speak, that would allow you to

1:10:56

make progress there. I mean, in

1:10:58

a sense, people have tried to

1:11:00

do this with protein engineering and

1:11:02

sort of the hope of text

1:11:04

to protein, so to speak. You

1:11:06

just say what you want the

1:11:08

protein to do, you know, you

1:11:10

say, I want to make a

1:11:12

cage for a molybdenum atom or

1:11:14

something, and then a protein will

1:11:16

get specified that will curl itself

1:11:18

up and make a cage just

1:11:20

the right size for a molybdenum

1:11:22

ion or something. But, you know,

1:11:25

I think, I think, I think

1:11:27

there's sort of great promise in

1:11:29

this sort of molecular scale computation,

1:11:31

molecular scale activity, where the molecules

1:11:33

are carefully arranged to what they

1:11:35

do, much like in life, but

1:11:37

without the immense baggage of life,

1:11:39

that there are probably much simpler

1:11:41

mechanisms that can be set up

1:11:43

for just doing the nanotechnology part

1:11:45

without having the whole organism that's

1:11:47

eating things and so on. It's

1:11:49

just not something that's been explored

1:11:51

very much. It was explored at

1:11:53

the end of the 80s, beginning

1:11:55

of the 90s, and it really

1:11:58

really stopped being explored. And I

1:12:00

suppose I suppose protein engineering, which

1:12:02

is a recent thing that's that's

1:12:04

been discussed in context to machine

1:12:06

learning, there's a little bit an

1:12:08

outgrowth of that, although I think

1:12:10

that proteins are quite floppy and

1:12:12

quite, it's not obvious that proteins

1:12:14

are the right way to do

1:12:16

it. I mean, it's like saying,

1:12:18

let's build a giant sort of

1:12:20

transportation ecosystem. But it's all going

1:12:22

to be based on horses. because

1:12:24

horses succeed in moving from here

1:12:26

to there. And how could you

1:12:28

imagine doing something different? Well, it

1:12:31

turns out we did in our

1:12:33

civilization have a transportation ecosystem based

1:12:35

on horses, but turns out cars

1:12:37

were a better idea in pretty

1:12:39

much every way I can think

1:12:41

of. But it's some. So I

1:12:43

think it's the same thing with

1:12:45

nanotechnology. We can kind of try

1:12:47

and piggyback. on biology, scientific papers.

1:12:49

And... The answer is yes, I

1:12:51

mean there have been a bunch

1:12:53

of efforts to make models that,

1:12:55

well, okay, one of the questions

1:12:57

is when you make a sort

1:12:59

of generic LLLM, one of the

1:13:02

discoveries is that knowing about lots

1:13:04

of stuff is useful even when

1:13:06

you're asking it about specific stuff.

1:13:08

You might have thought that knowing

1:13:10

a little bit about astronomy was

1:13:12

irrelevant if all you wanted to

1:13:14

talk about was ocean life. But

1:13:16

it turns out, and it's true

1:13:18

of us humans, that somehow having

1:13:20

that little bit of common sense

1:13:22

that comes from knowing something about

1:13:24

astronomy, is important everywhere somehow. We

1:13:26

don't completely know how, but the

1:13:28

idea that the LLLM can be,

1:13:30

oh, I'm a specialized LLLM, I

1:13:32

just do this, that tends to

1:13:35

be very narrow and tends to

1:13:37

be important to have this sort

1:13:39

of breadth of knowledge. Now, when

1:13:41

it comes to, did you ever

1:13:43

feed it? sort of the fanciest,

1:13:45

the best scientific papers, whatever that

1:13:47

means, the, there been efforts to

1:13:49

do that. It's all tied up

1:13:51

with, you know, well, who really

1:13:53

owns the rights to these things?

1:13:55

Can the LLLM ingest it? You

1:13:57

know, what happens if you get

1:13:59

a result from the LLLM and

1:14:01

how much of the undigested original

1:14:03

shows through? And there's a lot

1:14:05

of just practical ecosystem of the

1:14:08

world type questions there. But yes,

1:14:10

there's been a fair amount done

1:14:12

on this. So that was a

1:14:14

period of time maybe a year

1:14:16

ago when everybody seemed to be

1:14:18

talking about we're going to make

1:14:20

a special LLLM that's going to

1:14:22

be trained just on science. And

1:14:24

it didn't seem like that was

1:14:26

very promising, because as I say,

1:14:28

it seems like these shards of

1:14:30

common sense that come from different

1:14:32

areas are important. But just having

1:14:34

it know about lots of kinds

1:14:36

of scientific papers, a lot of

1:14:39

the high-end LLMs already know that

1:14:41

and have been trained on large

1:14:43

corpuses of those things. And that's,

1:14:45

you know, that is an interesting

1:14:47

question. I mean, the thing that

1:14:49

I'm about to really try seriously

1:14:51

with the latest models... is this

1:14:53

question of, okay, I'm looking for

1:14:55

experimental implications about physics project, and

1:14:57

that requires a certain amount of

1:14:59

putting together of different things of

1:15:01

sort of, of sort of trawling

1:15:03

out of the literature certain things,

1:15:05

where I can't just search for

1:15:07

it. It's a much vague question

1:15:09

that I have, you know, what

1:15:12

effect does it have on a

1:15:14

quasar if there's, if the dimensionality

1:15:16

of space changes? That's a... That's

1:15:18

a slightly, there are little shards

1:15:20

of information about that from different

1:15:22

places. Can you kind of aggregate

1:15:24

those together? And can you use

1:15:26

that as sort of a tool

1:15:28

for figuring out research? I do

1:15:30

think that my experience so far,

1:15:32

but not much effort at doing

1:15:34

this and it's become progressively more

1:15:36

plausible to do this as the

1:15:38

models have become more sophisticated and

1:15:40

trained on better material and so

1:15:42

on. But my initial take. is

1:15:45

that if you don't have any

1:15:47

idea where you're going, you won't

1:15:49

be led anywhere useful. In other

1:15:51

words, if you try and pull

1:15:53

the thing by the nose in

1:15:55

some direction, it will walk in

1:15:57

an interesting way, so to speak.

1:15:59

If you just say... hey, go

1:16:01

find me an interesting direction. It

1:16:03

will only tell you things that

1:16:05

you kind of already knew. It will

1:16:07

just sort of be, be, you know,

1:16:10

it will, it will revert to kind

1:16:12

of the mean of what people have

1:16:14

said before, so to speak.

1:16:16

But we'll see. I'll probably know

1:16:19

more about this in, well, fairly

1:16:21

short amount of time. All right, I

1:16:23

think it's time for me to go to

1:16:25

my day job here. But thank you

1:16:28

for a lot of interesting

1:16:30

questions and I'm sorry I

1:16:33

didn't get to lots of

1:16:35

other interesting questions that

1:16:37

were here. And I

1:16:39

clearly need to have a

1:16:41

wardrobe update and I will

1:16:43

perhaps do that by the time

1:16:45

of my next live stream. I

1:16:47

think I get to do another

1:16:50

one next Wednesday. So

1:16:52

we'll see. Check out the shirt

1:16:54

for next Wednesday. All

1:16:56

right, well, thanks for joining me and bye

1:16:58

for now. You've been listening to the

1:17:01

Stephen Wolfram podcast. You can

1:17:03

view the full Q&A series

1:17:05

on the Wolfram Research YouTube

1:17:07

channel. For more information on

1:17:10

Stephen's publications, live coding streams,

1:17:12

and this podcast, visit Stephen

1:17:14

Wolfram.com.