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

March 14th, 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

0:33

technology. I see a

0:35

bunch of questions saved up

0:37

here. All right, one from a

0:40

non. What exactly is an

0:42

AI agent? I wonder if

0:44

that's an AI agent asking

0:47

that. Agentic. It seems like

0:49

nobody knows what those words

0:51

actually mean today. Yeah.

0:54

That is a problem, isn't

0:56

it? Let me tell you what

0:58

I have kind of, you know,

1:00

my version of what I think

1:02

they should mean or mean to

1:05

most people. The issue is

1:07

an LLLM, large language

1:09

model, its original

1:11

concept was something very

1:14

mundane. You'll have been

1:16

giving some text. follow

1:18

along with how that text, say well

1:21

how that text should continue, you know,

1:23

the cat sat on the blank, probably

1:25

the next word is Matt. Turns out

1:27

that particularly with reinforcement learning, it

1:30

was possible to get something which

1:32

has a prompt, like answer this

1:34

question, what is the, you know,

1:37

what color are most cats or

1:39

something, question mark, and then it

1:41

will go and give an answer,

1:43

and reinforcement learning has sort of...

1:46

been able to be successfully used to

1:48

get AIs to kind of go

1:50

and do the next, take the

1:52

next step, answer the question, do

1:54

the thing that you would expect

1:56

to do next, so to speak.

1:58

That's first thing. Second thing

2:01

is, a big thing starting

2:03

right at the very beginning

2:05

of the current sort of

2:07

LLLM revolution when Chatche BT

2:09

came on the scene was

2:11

this idea of calling tools

2:13

and we were very early

2:15

to that talking about with

2:17

the open AI guys about

2:19

calling Wolfram Alpha and then

2:21

calling Wolfram Language and that

2:23

was first deployed in what

2:25

March of 2023 I think. And

2:27

being able to have the LLLM call

2:30

tools that are external to the LLLM

2:32

in the case of Wolfram, Wolfram

2:35

language, Computation and Knowledge

2:37

Tools, but also we built

2:39

very quickly a way of having the

2:41

LLLM call back into your local

2:43

computer to do operations on

2:45

your local computer, to run Wolfram

2:48

language code that could do all

2:50

kinds of things, including delete your

2:52

files on your local computer. And

2:54

remember the moment it must have

2:57

been, I don't know, sometime in...

2:59

spring of 2023, when we had

3:01

this all set up. And I

3:03

was like, now wait a minute,

3:06

what am I doing here? You

3:08

know, but when I press this

3:10

button, we're going to put the

3:12

LLLM in charge of my computer.

3:15

It's going to be able to

3:17

agentically, you could say, go and

3:19

do things on my computer. It's

3:22

like, well, that maybe isn't such

3:24

a good idea until I've really

3:26

figured out how to make sandboxes,

3:28

and then I don't know if

3:30

the LLLM is going to be

3:33

able to figure out how to break

3:35

my sandbox and escape

3:37

and do things on my computer.

3:39

I don't want it to do.

3:41

But in a case, the next kind

3:43

of concept there is this idea

3:46

that LLLMs can sort of figure

3:48

out what to do next. That's

3:51

I suppose this kind of agentic behavior

3:53

is kind of the the ability to

3:55

go beyond sort of the continue the

3:57

text to being and what should I

3:59

do? next type thing to sort of

4:02

be something where it's like we feel

4:04

as humans that we have agency in

4:06

the sense that we can kind of

4:09

decide what we want to do and

4:11

then make it happen. And it's kind

4:13

of it's sort of an imitation of

4:16

that idea that we're talking about that

4:18

for things like our alarms. But a

4:20

big part of sort of grounding agents,

4:23

LLLM agents is and then they've actually

4:25

got to do something in the world,

4:27

not just ponder in their own minds,

4:30

so to speak, or even generate streams

4:32

of text. It's like, do something in

4:34

the world, you know, have agency in

4:37

the world, and that means calling tools

4:39

like Wolfram Language and Wolfram Alpha, and

4:41

in general, being able to operate things

4:44

in the world. You know, I think...

4:46

probably this year we'll see a lot

4:48

of progress in robotics in connection with

4:51

with AI and the sort of generalization

4:53

of large language models to sort of

4:55

physical behavior and I think there it

4:58

will be even clearer what sort of

5:00

agency means because it's like operate the

5:02

actuator move the arm you know pick

5:05

the thing up it's very clearly very

5:07

much like human agency and so on

5:09

in that case So I think that's

5:12

that's kind of the big part of

5:14

it. So for example, we have agent

5:16

API that we're that well actually we

5:19

have most pieces of it now and

5:21

soon the the thing under that name

5:23

will come out that is something intended

5:26

for agentic APIs to use to make

5:28

things happen computationally and all the things

5:30

that you can control computationally. That's a

5:33

well from language based system that is

5:35

kind of the the universal sort of.

5:37

tool for having agency in the computational

5:40

world, so to speak. That's one side

5:42

of the story. The other side of

5:44

the story is how do you build

5:47

a thing that does sort of agentic

5:49

things as an AI? And as I

5:51

mentioned, you know, part of that story

5:54

is, well, what will you do next?

5:56

of bigger modules

5:58

of what you do

6:01

next. Part of the

6:03

what will you do next story is

6:05

things like reasoning models where you're kind

6:07

of let's try this for a while,

6:09

oh that doesn't work, let's try something

6:11

else instead. Another version of

6:13

that is kind of a graph, kind of

6:15

an almost a flowchart, that sounds very

6:17

ancient but it's the same thing back again,

6:19

of sort of what do you do,

6:21

you try this for a while and then

6:23

if that doesn't work okay you go

6:25

to this branch, you try that, it's kind

6:27

of a structure. Now the way that

6:30

that usually works, it's worked with this way

6:32

with neural nets and so on, is

6:34

first you kind of define a fixed sort

6:36

of graph of do this, then this, then

6:38

this, a little bit like a fixed

6:40

flowchart. Later on that becomes more dynamic,

6:42

it's more like a piece of code

6:44

where it's kind of making decisions and

6:46

it's then running different pieces of code

6:48

and I really suspect the same thing

6:50

is going to happen here, in other

6:52

words that the harness for LLMs which

6:54

at the beginning was just like

6:56

produce the next token, produce the next

6:58

token, you know read what's there

7:00

from the prompt and what you have

7:02

already written and just keep producing the

7:04

next code and producing the next token.

7:07

The only thing that was really in

7:09

the harness there was which next token

7:11

should you produce because the typical output

7:13

of an LLM is here the probabilities

7:15

of a bunch of possible next tokens

7:17

and then the question sort of for

7:19

the harness was sort of what the

7:21

temperature should be as in are you

7:23

at temperature zero always picking the most

7:26

probable next token, are you at temperature

7:28

one picking with the probabilities that are

7:30

defined by the LLM, are you at

7:32

a different higher lower temperature sort of

7:34

picking with more randomness or less randomness

7:36

relative to the probabilities defined by the

7:38

LLM. That was sort of the earliest

7:41

version of the kind of harness that

7:43

you would put I suppose another harness, two

7:46

other pieces of the harness one is the stop

7:48

token, when do you stop going

7:50

yak yak yak as an LLM that's

7:52

another thing that's kind of comes from the harness not

7:54

from within the LLM and another thing

7:56

is tool calling at what point do

7:58

you say okay. I take the

8:00

characters, the tokens that were produced here

8:02

and this is just input to the

8:04

tool. It's not what I'm going to

8:07

show the user. Go send it to

8:09

the tool, the tool responds, then you

8:11

get back from that and you use

8:13

that as a further prompt. None of

8:15

this is visible directly to the user,

8:17

to the human user outside or the

8:19

outside user of the LLM that just

8:21

happens inside the LLM and that keeps

8:23

going. So I guess

8:25

that the sort of the

8:27

generalization of all that is make

8:29

a more complicated harness where one

8:32

thing that one expects in the future,

8:34

it's not quite there yet, is

8:37

mostly it's just sort of

8:40

adding a token at a time. It's

8:42

looking at the past. That's kind of

8:44

what the idea of transform on architecture is.

8:46

You're looking at the past, causally looking

8:48

at the past, so to speak, and saying,

8:50

so what do I add next? So

8:53

a very different kind

8:55

of technology is what's used in automated

8:58

theorem proving, for example, where

9:00

you are trying to say,

9:02

given these axioms, for example,

9:04

prove this theorem, and that

9:06

involves kind of finding a path by which

9:08

you apply this axiom and that one and

9:10

that one and that one and just the

9:12

right way to get from the axioms to

9:14

that proof. There are a huge number of

9:17

possible paths, most of which would lead to

9:19

completely different theorems, but there are a few

9:21

paths that lead to the theorem you want.

9:23

That's a kind of story of

9:25

path finding in this ultimately extremely

9:27

large and dynamically built graph of

9:29

all these possibilities. And

9:32

one can imagine in the same type of

9:34

thing being done for an LLM, where

9:36

you're saying, instead of just saying, well, take

9:38

the most probable next token, it's like,

9:40

well, tree out all these possible paths of

9:42

what might happen if you took this

9:44

token and this token and this token and

9:46

then sort of make a plan based

9:48

on where you're going to go in the

9:50

future to get a path that gets

9:52

you to where you want to go in

9:54

the future. I've sort of thought, actually,

9:56

for a couple of years now, that that's

9:58

what buzzwords being what they are. are those

10:00

will one day be called quantum LLLMs.

10:02

They're sort of doing the quantum thing

10:05

following many paths of history to decide,

10:07

in this case, to decide what to

10:09

do. We're not quite there yet on

10:11

the technology. But anyway, back to

10:14

the original question about sort of agents

10:16

and LLLMs and agentic behavior and so

10:18

on. I think the story there is

10:20

really this. this sort of bigger level

10:22

of planning of what are you going

10:24

to do next and maybe the planning

10:27

in a sense is just there's a

10:29

fixed flow chart you're following this branch

10:31

or that didn't work that's follow the

10:33

other branch instead and so on. I think

10:36

it's still an issue in motion so

10:38

to speak I think the part of

10:40

it that's very clear is the part

10:42

where you're calling something like our agent

10:44

API where you're going from the LLLM,

10:47

which is doing its LLLM kind of

10:49

things of making a good linguistic user

10:51

interface and so on, and you're going

10:53

into sort of the hard computational part

10:55

of things that's what we've implemented over

10:58

all these years and more from language,

11:00

and from all from language, it's sort

11:02

of, it's a great kind of universal

11:04

connector from sort of the symbolic specification

11:06

of things to all these different external

11:09

APIs and languages and systems and so

11:11

on, whether you're interacting with a blockchain

11:13

or a blockchain or a database. or

11:15

a piece of rust code or a

11:18

library or this or that, you know,

11:20

one of the things that's happened and

11:22

it's a result of the sort of

11:24

symbolic representation of everything in Wolfen language

11:26

is that we're able to provide this

11:29

really very good sort of coherent universal

11:31

connector. Sometimes what we're connecting

11:34

to is a bit of a jungle, but

11:36

you know, we provide this kind

11:38

of symbolic representation of what is

11:40

there, you know, how we can represent...

11:42

So at least one side of the

11:44

connection, you know, we're connecting to processes

11:47

on distributed computers or whatever. These are

11:49

all specified symbolically. Some of the things

11:51

that happen when you're launching processes on

11:53

those external computers, that's a jungle out

11:55

there. But at least we can specify

11:58

what's going on and get analysis. and

12:00

get results from what's going on

12:02

in this nice clean and coherent

12:04

symbolic way. And that's kind of

12:06

one of the things I use

12:08

quite a bit in both language.

12:10

I mean, I have my personal

12:12

setup, I think I have about

12:14

240 cores that are sort of

12:16

somewhat dotted around my house or

12:18

something. And when I'm doing big

12:20

computations, I will just run, you

12:22

know, parallel table, parallel map, whatever.

12:24

And it just sort of immediately

12:26

distributes the computation among the cores

12:28

that I have. We also have

12:30

remote batch submit and things for

12:32

sending batch jobs to sort of

12:34

external cloud providers and such like.

12:36

Let's see. All right, let's go

12:38

on to some other things here.

12:40

Layla asks, can you tell us

12:42

about the future of media and

12:44

information consumption? Will we become a

12:46

society of AI summaries as our

12:48

main form of information gathering? That's

12:51

an interesting question. I mean, I

12:53

think that, you know, it's funny,

12:55

from the time before Desktop Publishing,

12:57

I'm old enough that I remember

12:59

a time before Desktop Publishing, before

13:01

PowerPoint, before TED Talks, before these

13:03

kinds of things, there was a

13:05

certain difference in the way that

13:07

information was presented and absorbed. The

13:09

idea of kind of the bullet

13:11

point list. I think is an

13:13

idea that is, you know, I

13:15

don't think that really existed so

13:17

much when I was a kid,

13:19

for example. I think it was

13:21

much more narrative. And by the

13:23

way, the style of presentation that

13:25

people make has also changed. I

13:27

mean, in this, people used to

13:29

have a much less direct way

13:31

of expressing, explaining things, I think.

13:33

I think you know if you

13:35

look at sort of the history

13:37

of web expilations of things one

13:39

of the first kind of just

13:41

say it directly kinds of explainer

13:43

things was our math world website

13:45

talking about kind of the you

13:47

know all things math sort of

13:49

explained kind of directly. I would

13:51

say by the way that that

13:53

I also noticed you know when

13:55

I was writing the original documentation

13:58

for Mathematica and now Wolfman language

14:00

back in 1987 or so. I

14:02

also kind of adopted this kind

14:04

of tell it as it is

14:06

style of explaining things that was

14:08

visibly different from the stars that

14:10

existed before, for example, a computer

14:12

documentation, which had a habit of

14:14

being quite formulaic and not kind

14:16

of say it as you would

14:18

say it to a person type

14:20

thing. Math world had the same

14:22

approach of kind of say it

14:24

as you would say it to

14:26

a person. Wikipedia kind of followed

14:28

Math world. I know Jimmy Wales

14:30

made... use of ideas from math

14:32

world and they took a bunch

14:34

of content from math world. They're

14:36

in math world. It was kind

14:38

of, you know, as often happens

14:40

with these things, you know, Wikipedia

14:42

kind of sucked the content from

14:44

math world. I have to say,

14:46

it will be nicer. Wikipedia has

14:48

such more content from all the

14:50

things I write. We've kind of

14:52

made them. number of years ago

14:54

some some Wikipedia's were like you

14:56

should make this content Wikipedia friendly

14:58

and set it up so that

15:00

it has all the right to

15:02

you know letters after its name

15:05

at the bottom of the page

15:07

about you know the different kinds

15:09

of creative commons rights and so

15:11

on that it has so we

15:13

did that and I don't think

15:15

people have sucked as much content

15:17

from there onto Wikipedia as they

15:19

probably should have done and so

15:21

it's a good thing for people

15:23

to actually do I think that

15:25

But in any case, this style

15:27

that got adopted also by Wikipedia

15:29

of kind of the tele directly

15:31

as it is, is kind of

15:33

a style of presentation. Now, you

15:35

know, there's a yet different style

15:37

of presentation that the LLLMs are

15:39

giving us. It's a bit. land.

15:41

Sometimes I like it a lot

15:43

where it's kind of, you know,

15:45

we get these, we started doing

15:47

this back in early 2023, kind

15:49

of getting LLLM summaries of papers

15:51

about LLLM's that come out in

15:53

the world. And I find it

15:55

easier to read the couple of

15:57

sentence summaries of those things than

15:59

to read the original abstract because

16:01

they're very uniform, they're very uniform,

16:03

very easy to consume. So I

16:05

kind of, you know, I recently

16:07

have been taking doing more sort

16:09

of summarization of things that I'm

16:11

asking where I would be using

16:14

web search and I'm now using

16:16

kind of more. sort of LLLM

16:18

summarization. I have to say, I

16:20

think I mentioned this, I know

16:22

why it has come up a

16:24

couple of times, but I always

16:26

use this search engine Caggy, which

16:28

is this kind of search engine

16:30

sort of optimized for the users

16:32

rather than for the advertisers type

16:34

thing. I happen to have been

16:36

an advisor to that company for

16:38

a while, but Caggy has this

16:40

feature that I have to say,

16:42

only learnt from its CEO, fairly

16:44

fairly recently, fairly recently, where if

16:46

you, when you ask a. when

16:48

you put in a query, if

16:50

you end it with a question

16:52

mark, it will start its LLLM

16:54

running and start giving you a

16:56

summarization. And just the fact that

16:58

it's so easy to do it,

17:00

that you know, just doing it

17:02

by putting a question mark at

17:04

the end of the query, I've

17:06

started to find that I do

17:08

that a lot. And now, you

17:10

know, I've gone from probably one

17:12

in 50 of my web searches,

17:14

was a thing where I'm asking

17:16

for another number to more like

17:18

one in three or four. being

17:21

that way. And there definitely it

17:23

sort of changes my view of

17:25

what kinds of things I can

17:27

ask and it's sort of interesting

17:29

to see what happens and I

17:31

I was you know sometimes it's

17:33

like yeah I got the right

17:35

it got the right idea and

17:37

sometimes it's like no it careened

17:39

off into crazy land. Of course

17:41

there's I think the same thing

17:43

about web search sometimes that you

17:45

know while I was looking for

17:47

this thing but oh I didn't

17:49

realize that there was also some,

17:51

some other thing that had a

17:53

similar name and the website kind

17:55

of careened off into Crazy Land

17:57

as far as I was concerned.

17:59

You know, I do think that

18:01

this question of when, you know,

18:03

when do you do better with

18:05

the kind of bland LLLM summary,

18:07

when is it important to have

18:09

that kind of summarization of really

18:11

vast amounts of material out there

18:13

in the world versus, you know,

18:15

when are you better off looking

18:17

at the originals? You know, I

18:19

have to say, I do lots

18:21

of work in sort of history

18:23

of science and technology. And one

18:25

of the principles about doing that

18:27

is always read the original documents.

18:30

You know, you'd think, oh, there

18:32

have been these great historians and

18:34

they summarized what happened and this

18:36

and that and the other, and

18:38

you know, I'll read some of

18:40

that stuff. and then I'll say,

18:42

look, I just make it a

18:44

rule for myself. I got to

18:46

go read the original documents. And

18:48

it is the amount of additional

18:50

sort of flavor, color of what's

18:52

going on that you get from

18:54

those original documents, plus insights that

18:56

were kind of diluted by other

18:58

people trying to understand what was

19:00

said and not really quite getting

19:02

it right, and they're looking at

19:04

things through a certain lens. I've

19:06

always found it really, really valuable

19:08

to go back and read the

19:10

original documents. And maybe that's the

19:12

way one will feel about, do

19:14

I read the LLLM summary or

19:16

do I read the original document?

19:18

Do I read what so-and-so actually

19:20

wrote? Or do I read the

19:22

summary of what so-and-so actually wrote?

19:24

I mean, it's kind of a

19:26

little bit like in education, kind

19:28

of the great books theory of

19:30

just read what the original people

19:32

actually wrote. I'm quite a believer

19:34

in that. Maybe it's I'm prejudiced

19:37

from the fact that I write

19:39

a lot and I think some

19:41

of what I write is pretty

19:43

interesting and I hope people actually

19:45

read it because there's a lot

19:47

more. I know when I go

19:49

back and read things that I

19:51

wrote, there's a lot more detail

19:53

in there and a lot of

19:55

things that I had figured out

19:57

that weren't what I remembered. having

19:59

figured out, I only remembered kind

20:01

of the big picture. And I

20:03

know when people summarize, like, I

20:05

haven't read the summaries that people

20:07

have written of my stuff on

20:09

Wikipedia, I think I'd be pretty

20:11

horrified to do so. But I'm

20:13

kind of guessing that the summaries

20:15

are pretty distant from all of

20:17

the kind of things that are

20:19

really there in what, you know,

20:21

what even I figured out and

20:23

things I've written. So it's kind

20:25

of a thing where, where that's

20:27

a, you know, it's like... Are

20:29

you at an LLLM distance from

20:31

the original documents or are you

20:33

reading original documents? I think there

20:35

will be sort of additional value

20:37

in reading the original human written

20:39

documents, but there are other purposes

20:41

for which getting that planned LLLM

20:44

summary is going to be really

20:46

worthwhile. I don't know whether the

20:48

summaries will stay bland. It may

20:50

be that they'll be in a

20:52

good sort of, I don't know,

20:54

it's like, and waiting for the

20:56

LLLM that can write just like

20:58

me, it will come. You know,

21:00

I've put enough stuff out out

21:02

in the stuff out in the

21:04

I think from all this yacking

21:06

that I do in live streams

21:08

and things like that, there's maybe

21:10

50 million words of stuff for

21:12

me out there. And that's probably

21:14

enough to make a me bot

21:16

that will sound pretty much like

21:18

me. I don't know to what

21:20

extent the whatever edge I might

21:22

think I have of, you know,

21:24

the crisp new idea presented in

21:26

an interesting way. I don't know

21:28

what to at what point the

21:30

sort of the LLLM imitation will

21:32

come close enough. or maybe even

21:34

exceed what I'm able to do

21:36

there, that it will be sort

21:38

of, you might as well read

21:40

it, so to speak. I think

21:42

that there's sort of a question

21:44

of the economic ecosystem of kind

21:46

of material and what one consumes,

21:48

because you know, back in the

21:50

day, it's like things got written

21:53

and people who wanted to read

21:55

them paid for them with, you

21:57

know, magazine subscriptions or... or books

21:59

that they bought and so on.

22:01

And then kind of this. alternate

22:03

economy started up, which I suppose

22:05

had happened with radio and television

22:07

and so on, because, you know,

22:09

in the US at least, you

22:11

couldn't collect from people before cable

22:13

television. It's like you're broadcasting radio

22:15

out into the ether. Anybody can

22:17

pick it up. When I was

22:19

growing up in England, there was

22:21

the BBC, which was a government

22:23

sponsored, you know, still is, a

22:25

government-sponsored thing. And there was this

22:27

very strange concept of you had

22:29

to have a television license to

22:31

receive these television programs. And it's

22:33

a very strange kind of thing.

22:35

There was, I was a kid,

22:37

there were these television detector vans

22:39

that drove around the place trying

22:41

to detect... It must have been

22:43

the superheterodyne signals of the amplifiers

22:45

for televisions that were unlicensed televisions

22:47

and people would come in and

22:49

say you've got to shut down

22:51

that television and you know we're

22:53

going to find you because you

22:55

didn't pay your license fee which

22:57

was paying for the BBC to

23:00

produce the programming that people was

23:02

watching on their televisions. So it

23:04

was a little bit of a

23:06

different way of doing things, but

23:08

it was again a licensing a

23:10

model where the consumers of the

23:12

content were paying for the production

23:14

of the content. That kind of

23:16

disappeared, I think probably initially in

23:18

the US, when the idea came

23:20

about, well, just ran ads and

23:22

the advertisers will pay for the

23:24

creation of the content and, you

23:26

know, the soap operas were advertising

23:28

soap and the etc, etc, etc,

23:30

etc, etc. It's some. And that

23:32

was sort of, that has been

23:34

kind of an alternate model of

23:36

how to pay for the production

23:38

of media. And then, you know,

23:40

when the web came along, the

23:42

web had the sort of anomaly

23:44

that the original concept of the

23:46

web and things like Project Xanadu,

23:48

which was a predecessor kind of

23:50

the hypertext invention that turned into

23:52

the HTCP, you know, the H.T.

23:54

that Tim Berners Lee put in

23:56

the in the names of websites.

23:58

that I think came from Project

24:00

Xanadu and Ted Nelson and all

24:02

these guys who were thinking about

24:04

sort of how to make this

24:07

ecosystem of information of connected information.

24:09

I think they originally imagined that

24:11

people would pay for the information

24:13

they got, that there would be

24:15

a whole ecosystem of micropayments where

24:17

it's like, well, if you got

24:19

this thing from here, you would,

24:21

the person who was the original

24:23

producer of that, would get, you

24:25

know, a millionth of a penny

24:27

for the fact that you looked

24:29

at that thing. And this would

24:31

all get sort of aggregated, sort

24:33

of aggregated up. to pay for

24:35

the production of content. For whatever

24:37

reason, and it was partly because

24:39

of the origination of the web

24:41

and the fact that it was

24:43

sort of based on the internet,

24:45

which was based on the ARPANET,

24:47

which was this sort of free-ish

24:49

government thing in the US, and

24:51

just the way that that whole

24:53

thing developed, micropayments never happened on

24:55

the web, as in that way.

24:57

And so, sort of there had

24:59

to be another model for how

25:01

would you pay for content on

25:03

the web? And the fundamental model,

25:05

which I have to say I

25:07

was amazed it worked, that, you

25:09

know, Google basically pioneered was, I

25:11

mean, Yahoo had done things earlier

25:13

and so on, but it was,

25:16

you know, I just couldn't believe

25:18

this was going to work. And

25:20

it wasn't the original Google business

25:22

model, but, you know, that you

25:24

would make money by basically selling

25:26

advertising against web content. And you

25:28

know that has led to many

25:30

things that have been both well

25:32

some good and many probably not

25:34

so great about I mean it

25:36

sort of allowed the web to

25:38

be more widely in the content

25:40

on it to mean war widely

25:42

disseminated. It's meant that the you

25:44

know the people who pay and

25:46

not the people. who are getting

25:48

the benefits, so to speak, and

25:50

I never like those kind of

25:52

mismatches, something in the business we've

25:54

done with our company. I've always

25:56

tried to avoid that, tried to

25:58

make it be the case that

26:00

the folks who are paying for

26:02

our software and the services we

26:04

make and so on are the

26:06

folks who are getting the benefit,

26:08

rather than... sort of a third-party

26:10

mismatch, so to speak. And I

26:12

think that's another aspect of kind

26:14

of the consumption of media is,

26:16

well, you know, who's paying for

26:18

the production of this media content?

26:20

And, you know, what we've seen

26:23

now with social media and with

26:25

Twitter and X and so on,

26:27

is that it's, you know, there's

26:29

an awful lot of content that

26:31

is produced by people just producing

26:33

it because they want to make

26:35

a point. They want to put

26:37

that content out there. It's very

26:39

different from the model that had

26:41

existed before, where there's a stratum

26:43

of professional journalists whose job is

26:45

to forge the world for content

26:47

to put it out there. It's

26:49

a thing where people who want

26:51

to get content out there, you

26:53

know, push to get content out

26:55

there, so to speak. And then

26:57

sort of it becomes an issue

26:59

of how do you select the

27:01

content you want, and so on.

27:03

But it's a different ecosystem of

27:05

sort of content production. Well now

27:07

we've got the LLLMs coming along,

27:09

which are yet something different. They're

27:11

foraging all that human content and

27:13

sort of making it their own.

27:15

in a way that is now

27:17

quite disconnected from the original sources

27:19

of content. I mean, there's some

27:21

efforts to put references into the

27:23

LLM content. That works in some

27:25

cases and so on. But I

27:27

think as we see more sort

27:29

of LLM-made content out there, it's

27:32

again unclear how that's going to

27:34

shake out economically. I mean, it's,

27:36

you know, right now, there's a...

27:38

a big market for training data.

27:40

It's something, you know, we are

27:42

a source of quite a bit

27:44

of training data that's used out

27:46

there in the world, of things

27:48

that are very, in a sense,

27:50

very clean training data because we

27:52

are computing things. So we're generating

27:54

training data. It's not that, I

27:56

mean, it takes a bunch of

27:58

human effort to generate the right

28:00

thing, but it is not something

28:02

where we're saying, well, it's so-and-so's

28:04

writing that was done whenever, and

28:06

now we're foraging that. very clean

28:08

training data. And it's also training

28:10

data that has the interesting feature

28:12

that's kind of training for reasoning

28:14

in effect. Just like, you know,

28:16

I think the elements from the

28:18

beginning read all this text and

28:20

discovered logic from the text by

28:22

virtue of the fact that there

28:24

were sort of sentences that were

28:26

constructed in this way, which are

28:28

in a sense, the logical sentences.

28:30

And now... you know, there's a

28:32

lot that we have that is

28:34

computation of mathematics and chemistry and

28:36

all those kinds of things that

28:39

is somehow kind of put together

28:41

in a sort of reasoning way.

28:43

So that's been another, the, another,

28:45

sort of piece of that story.

28:47

And I don't know how, how,

28:49

you know, I think The thing

28:51

that's happened is because of the

28:53

ecosystem, the way the economics of

28:55

content generation have worked, there's been

28:57

sort of a lessening, you know,

28:59

the journalism world has sort of

29:01

been going down, I'm sort of

29:03

horrified usually with the kinds of

29:05

things that I'll see in a

29:07

lot of... beyond the very top

29:09

tier of journalism, the kinds of

29:11

things that I see. They're very

29:13

lazy, sloppy, repetitive, secondary kinds of

29:15

things. I think one of the

29:17

things I know, I know Kagi

29:19

uses as a heuristic for its

29:21

ranking for search is how many

29:23

ad trackers are there on that

29:25

page? How much of it is

29:27

sort of original content and how

29:29

much of it is just, you're

29:31

only here because we want to

29:33

get you to look at the

29:35

ads, because that's how we're monetizing

29:37

our page. And I think that's

29:39

now, you know, it's starting. There

29:41

are all these other different channels

29:43

where people are sort of putting

29:46

content out there, just in, you

29:48

know, the things they write in

29:50

blogs and other things like that.

29:52

And I obviously put a lot

29:54

of content out there in that

29:56

form. And it's it's it's kind

29:58

of just stuff I want to

30:00

put out there. It's not because

30:02

I have a living being a

30:04

journalist or a writer of science

30:06

books or something like that. It's

30:08

just stuff that I'm writing because

30:10

I like to write it and

30:12

I think people will find it

30:14

interesting and I'm putting it out

30:16

there. It's a different sort of

30:18

and you know it's not it's

30:20

not how I'm making a living

30:22

so to speak and I think

30:24

the same is true of lots

30:26

and lots of content on X

30:28

and so on and similar kinds

30:30

of kinds of places. So you

30:32

know and I think it... this

30:34

question, you know, the elements are

30:36

a weird twist to the whole

30:38

thing because they are a yet

30:40

different part of the story than

30:42

not at this time. They'd been

30:44

monetized essentially by subscription, which as

30:46

far as I'm concerned is a

30:48

very clean form of monetization. You

30:50

know, maybe one day they'll be

30:52

monetized with ads. I kind of

30:55

hope not because then it's going

30:57

to get even more bizarre as,

30:59

you know, you're reading this element.

31:01

at least in search ads, for

31:03

example, it's an ad, you know,

31:05

it's an ad. By the time

31:07

you're reading LLM content is all

31:09

interwoven with, you know, kind of

31:11

prompts that say, now there's a

31:13

chance to advertise this brand of

31:15

toothpaste or something, and it starts

31:17

doing that. It's going to get

31:19

very confusing. And I, you know,

31:21

and then you'll have to have,

31:23

it's like, well, I want to

31:25

read this thing that doesn't have

31:27

ad contents. I'm going to get

31:29

my own lalam to try and

31:31

guess what's ad content and back

31:33

that out. Just like, you know,

31:35

we're getting lalams to try and

31:37

figure out the things we're sent,

31:39

we're written by lalams, well, it's

31:41

usually pretty obvious at this point.

31:43

And, you know, we're, we're kind

31:45

of like. For me, it's like,

31:47

I don't want to read what

31:49

some LLLM wrote. If you as

31:51

a human want to communicate with

31:53

me, you had some set of

31:55

bullet points, some set of prompts

31:57

that you fed to the LLLM.

31:59

Just give me those prompts. You

32:02

know, I'll deal with the prompts,

32:04

so to speak. I don't want

32:06

to see the LLLM Fluff that

32:08

came out that's many pages long

32:10

that I have to kind of

32:12

grind down. Maybe I have to use

32:14

my own LLLM to do that to see

32:16

what it is to actually trying

32:18

to say. Let's see. A bunch of

32:21

questions about LLLM. So let me

32:23

address those and then go on to

32:25

some other things. Yeah, desk comments

32:27

before AI summaries there were

32:29

encyclopedias and textbooks and cliff

32:32

notes and so on, which were

32:34

useful and convenient. They

32:37

never became de facto. So, you

32:39

know, that's right. I mean, starting

32:41

in the 1700s, the idea of

32:43

encyclopedias with people like Dalomber and

32:45

France and then later encyclopedia Britannica

32:48

or in Scotland were kind of

32:50

like, collect this information, collect. you

32:52

know, the things you need to

32:55

know in the world in a

32:57

digested way. I mean, then it

32:59

was readers digest, then it was

33:02

close notes, then it was all

33:04

these kinds of things that sort

33:06

of a digest of information. I

33:09

agree. I think that that maybe that's

33:11

that maybe that's the way to think

33:13

about the current role of LLLM's

33:15

is as digesters of information.

33:17

I mean, they go beyond that because

33:20

you can kind of talk to the

33:22

book type thing. And I think that's

33:24

a That's a thing that is

33:26

probably going to be increasingly

33:28

useful. I mean, we're currently

33:31

doing a big experiment of

33:33

building an AI tutor, specifically

33:35

for algebra one course. It's an

33:37

attempt to make something which people

33:40

have never been able to make

33:42

work before, which is to have

33:44

a truly autonomous, truly scalable educational

33:47

tutoring system. People, you know,

33:49

there's 70 years of history of

33:51

trying to use computers to be

33:53

teaching machines. It didn't really work

33:56

that well. It's worked great for

33:58

things like Wolfram Alpha, where you're...

34:00

using it as a generator of content

34:02

that is alongside what you're learning, but

34:04

it's not something where the dynamics of

34:07

actually teaching are delegated to the machine.

34:09

What we're trying to do with our

34:11

AI tutor is get a surprisingly complicated

34:13

system that tries to delegate the dynamics

34:16

of teaching to the machine, at least

34:18

for a topic like algebra, which is

34:20

in some level a kind of very

34:23

cut and dried topic. We'll see how

34:25

it works. It's looking pretty promising right

34:27

now, but that's a case where you

34:30

can in a sense expect. Now that

34:32

particular one was done with lots of

34:34

efforts, specifically building an algebra course that

34:36

fitted into the kind of the way

34:39

that one could control it with another

34:41

lamb and so on. I think we

34:43

also have another project that is really

34:46

taking much more off-the-shelf sort of book-like

34:48

content and saying make something where you

34:50

can kind of talk to the book

34:53

and... You know, I think that's a

34:55

that's kind of a new direction for

34:57

quote summarization is well, it can be

34:59

a summary just for you a personalized

35:02

summary. I mean, I suppose I suppose

35:04

that is probably the direction that one

35:06

can expect to go in that your

35:09

LML gets to know you pretty well

35:11

and or your AI gets to know

35:13

you pretty well and it knows the

35:16

thing you don't understand and the one

35:18

thing that should be said to explain

35:20

this to you is this particular thing

35:22

and it will be very different for

35:25

somebody else. I mean, I think it's

35:27

an interesting dynamic as our AIs get

35:29

to know us better and better. You

35:32

know, our feelings about our AIs will

35:34

probably change. You know, they'll probably come

35:36

a time when there are AIs that

35:39

know me better than any human knows

35:41

me. And maybe it's already the case,

35:43

although in a very banal sense of

35:45

an A. I mean, I have a

35:48

pretty well-organized meta-seacher of all of the

35:50

sort of interactions and things I've done

35:52

over the course of the last 30-something

35:55

you know that in a sense that

35:57

system which is not really an AI

35:59

system. becoming more AI-ish over time, but

36:02

sort of knows me very well. There

36:04

will probably come a time when an

36:06

AI that I can routinely interact with

36:08

as I interact with people knows me

36:11

better than any person knows me, in

36:13

the sense that they know my whole

36:15

history of every interaction I've had with

36:18

everybody, and they've sort of seen more

36:20

of the details of what I do

36:22

and say and so on than anybody

36:25

else has. And I think my feelings

36:27

about that AI will be interesting. I

36:29

mean, I think I will be very

36:31

protected probably of that AI. It's a,

36:34

you know, it's a very, it's a

36:36

very sort of close companion AI, so

36:38

to speak, and that will be an

36:41

interesting dynamic to see. Let's see. The

36:43

question from the... When will we get

36:45

the first AI robot news reporter? I

36:48

see these being useful in cases of

36:50

dangerous live broadcasting like hurricanes to keep

36:52

people up to date. That's a fun

36:54

one. I mean, I think there is

36:57

a lot of AI reporting. I mean,

36:59

that's happened for a decade or more,

37:01

whether it's the sports scores or the

37:04

financial news. There's, you know, there are

37:06

people who will read, you know, and

37:08

the Tao is up X number of

37:11

points or down X number of points

37:13

or whatever, and the this and that

37:15

and the other. But, you know, you

37:17

know, starting, probably 15 years ago, 15

37:20

years ago or more. that started to

37:22

be systems that could produce, could synthesize

37:24

good natural language text from the kind

37:27

of the raw numbers of what was

37:29

going on in the stock market or

37:31

wherever else. And I'm pretty sure, I

37:34

mean, I have to admit I'm not

37:36

a consumer of these things, but I'm

37:38

pretty sure that a lot of those

37:40

things up and running right now. Now

37:43

I think another question about the news

37:45

is perhaps, there's news that is directly

37:47

absorbable by people. from the raw data.

37:50

Like if you're watching, I don't know

37:52

what, I don't know, something about the

37:54

stock market or something, the raw numbers

37:57

tell their story, or you're doing... But

37:59

now, there are cases where there's lots

38:01

of stuff happening in the world, but

38:03

most of what's happening, you just don't

38:06

want to know about it's completely boring.

38:08

It's like every security camera that has

38:10

footage of this and that happening. But

38:13

occasionally, there'll be a zebra that escaped

38:15

from the zoo and is prancing through

38:17

the security camera field of view, and

38:20

that's the thing you might want to

38:22

know about. But, you know, most of

38:24

the time, it's just looking in an

38:26

empty street or whatever. I can imagine

38:29

a time when all this data that's

38:31

being collected in the world is being

38:33

more effectively aggregated by AIs to tell

38:36

us, the humans, things that we might

38:38

care about that are going on in

38:40

the world. I think that hasn't yet

38:43

happened. I mean, I've been involved in

38:45

a number of projects actually over the

38:47

years that have sort of gone towards

38:49

that kind of direction. not only in

38:52

that case saying what is happening in

38:54

the world, but based on what is

38:56

happening in the world, what's going to

38:59

happen next? What kinds of, you know,

39:01

what inevitable consequences are there based on

39:03

the model that we have of the

39:06

world, whether it's model of meteorology or

39:08

models of crops or models of supply

39:10

chains, you know, given that this happened,

39:12

what will inevitably happen next in the

39:15

world? And that's something that's... that's relevant,

39:17

I mean, purely in terms of, you

39:19

know, things like financial speculation, that's relevant,

39:22

but there are lots of other reasons

39:24

why that's a relevant thing to be

39:26

able to do. But it's an interesting

39:29

point that I haven't really seen this

39:31

yet, is the really the systematic aggregation

39:33

of what is otherwise a boring kind

39:35

of, you know, a most of the

39:38

time the camera is going to see

39:40

nothing type of type of information. Now,

39:42

obviously X. has a big, has lots

39:45

of streams of stuff that are coming

39:47

along, but those already quite human curated.

39:49

Even the curation of those streams is

39:52

not something that's really happened yet that

39:54

much. I mean, I'm kind of waiting

39:56

for the first sort of X-based newspaper.

39:58

I think actual newspapers these days... take

40:01

a lot of their content from X.

40:03

And, you know, it typically happens first

40:05

there and only later in the newspapers.

40:08

But, you know, I think that's an

40:10

interesting question is sort of an interesting

40:12

almost philosophical question. There's all this stuff

40:15

going on in the world. What's news?

40:17

You know. it's the New York Times

40:19

what you know what the all the

40:21

news that's fit to print or something

40:24

I don't know whether that's a tagline

40:26

they even still use or certainly whether

40:28

it applies really in any serious way

40:31

to that newspaper I'm not sure but

40:33

but I mean there's this question of

40:35

all the things that are going on

40:38

in the world that are picked up

40:40

by all our sensor arrays and all

40:42

this kind of thing you know what's

40:44

what's worth summarizing for a particular human

40:47

and you know obviously one can pick

40:49

particular sections of news, one's interested in

40:51

things like that. I would say that

40:54

the whole idea of personalized news that

40:56

was, I don't know, that's come a

40:58

bunch of times. People have used Wolfmaufer

41:00

a bunch to do various kinds of

41:03

personalized news and some of our natural

41:05

language understanding technology do those kinds of

41:07

things. I don't think that's really become

41:10

a big thing yet. And maybe with

41:12

elements it will. Let's see. Oh yeah,

41:14

the question was about things like broadcasting

41:17

from the hurricane scene and so on.

41:19

Yeah, I mean, I think the one

41:21

of the questions is, you know, it's

41:23

like you're seeing a bunch of footage,

41:26

you know, there are a bunch of

41:28

security cameras being ripped up by the

41:30

hurricane or something. And, you know, it's

41:33

like there could be a human meteorologist

41:35

that interpreting what they see, or it

41:37

could be something where there's a layer

41:40

of... you know, of AI that perhaps

41:42

knows more than us humans know about

41:44

how to interpret that weird picture of

41:46

the thing that was, you know, I

41:49

don't know what happens, you know, that

41:51

blacks out at that moment or whatever

41:53

else it is. But I think, yeah,

41:56

I mean, there are a number of

41:58

these things where I feel like there's

42:00

still some place. to go in terms

42:03

of what people get used to and

42:05

so on. I mean, you know, another

42:07

one of these that I'm surprised hasn't

42:09

caught on more is virtual touristing of,

42:12

you know, you get, I mean, maybe

42:14

when there are better humanoid robots and

42:16

things, that'll be more of a thing.

42:19

It's like, I'm not going to go

42:21

to Antarctica, but I'm going to send

42:23

my robot to Antarctica, and it's going

42:26

to play with the penguins, and... I'm

42:28

going to have, you know, it'll be

42:30

really cool to watch me controlling my

42:32

robot playing with the penguins type thing,

42:35

or whatever it is. I don't know

42:37

if penguins are very friendly or not.

42:39

It's always sometimes suspicious when animals look

42:42

really cute and then when you get

42:44

up close, they're pretty aggressive. I don't

42:46

know in the case of penguins. But

42:49

in any case, I mean, you know,

42:51

I can imagine that being a thing

42:53

that develops there. Let's see. It's

42:56

a question here from LC, how

42:58

far are we from LLLM's generating

43:00

the kind of things that I

43:02

write with sort of similar elucidation

43:04

based on a short prompt? I

43:06

don't know. It feels like we're still

43:08

pretty far away. It feels like,

43:10

I mean I think I'm going to

43:13

know more I'm doing a project

43:15

just starting a project right now,

43:17

kind of using LLLM's as a way

43:19

to kind of understand a very

43:21

broad literature about physics about physics. for

43:23

understanding experimental implications of our physics

43:25

project. I don't know. I'm going

43:27

to know more after I really do

43:30

this. I mean, I've been a

43:32

big believer forever in using the best

43:34

tools one can for doing what

43:36

we want to do. And it's

43:38

clear now that sort of this summarization

43:40

of broad swaths of human output

43:42

is something which the LLLM's are good

43:45

at. And I want to see

43:47

how far I can get with

43:49

that. and to what extent it can

43:51

kind of, you know, to what

43:53

extent it can organize and write kinds

43:55

of things that I'm expecting to

43:57

use it mostly as a tool.

43:59

a foraging tool so to speak more

44:02

than a tool of exposition. We're

44:04

not there yet. I don't know how

44:06

that future will unfold. Wailo

44:08

is commenting when you say that

44:10

teaching is delegated to the machine.

44:12

Are you saying that the machine

44:14

is telling the student what to

44:17

think instead of just answering

44:19

questions? That's a good interesting spin.

44:21

I mean when it comes

44:23

to Wolfram Alpha for example,

44:25

we're just answering questions.

44:28

The... the kind of which questions to

44:30

ask that's all on the

44:32

student the which way should we go

44:34

it's all on the human I mean

44:36

I think the long-term story of AI

44:38

is kind of the which way should we

44:40

go is up to humans to

44:43

define the let's get there as

44:45

as automatically as possible that's for

44:47

the AIs to do you know I

44:49

was just realizing that you know

44:52

when people ask you know what's going

44:54

to happen when AIs are doing all

44:56

these things you know, how do you

44:58

feel about that as a human? And

45:00

I was realizing, at some level, I've

45:02

kind of lived the AI dream for

45:05

decades, because I've been building technology

45:07

that automates the things that I

45:09

mechanically want to do, that involve,

45:11

you know, figuring out things in

45:14

science and technology and so on.

45:16

And, you know, with Wolfman language, I've

45:18

been kind of building the AI

45:20

dream for myself and for lots of

45:22

other people, because it's all about

45:24

going from I imagine doing doing this

45:27

thing. to as automatically as possible,

45:29

the thing gets done. And that's

45:31

kind of the whole idea of

45:33

all from language is I imagine

45:35

the thing, I type it with my fingers

45:38

or even you can say it to

45:40

some extent now, and then, you know, and

45:42

then it's dealing with the mechanics

45:44

of how to get that question

45:46

answered, so to speak. I'm still

45:48

defining the question, it's getting that

45:51

question answered as efficiently as possible.

45:53

And that is kind of the

45:55

AI dream, I think, is we're

45:57

in charge, we're calling the shots.

45:59

and the AIs are efficiently executing

46:01

on what we want to have done.

46:04

Now, you know, there's a scenario where

46:06

the AIs are deciding what to do

46:08

for themselves, I think is one of

46:11

these sort of philosophically doomed scenarios, because

46:13

it's something where you say, well, the

46:15

AIs are going to go off and

46:18

do what they want to do. Sure,

46:20

computers do that all the time. It's

46:22

like, you know, you've set them off

46:25

in this general direction. The way they

46:27

actually do it is this. I've spent

46:29

a lot of time kind of defining

46:32

simple rules. sort of computational rules and

46:34

just seeing what those rules do, then

46:36

it's off on its own. It's not

46:39

under human control. It's not, oh, I

46:41

want to get it to do this.

46:43

I want to have this goal for

46:46

it. It's just doing what it's doing.

46:48

And I think the thing to understand

46:50

is that the big example of that

46:53

happening that we're all very familiar with

46:55

is the natural world, where things are

46:57

happening, all the time. But they're not

47:00

things under our control. They're not things

47:02

where we know the goal, the purpose

47:04

of what's happening. They're just things that

47:07

are happening. And I think that's what

47:09

sort of AI is left to their

47:11

own devices. That's almost by definition, what

47:14

they end up doing is like what

47:16

the natural world does. They just do

47:18

what they do based on the computational

47:20

rules that they have. So, but you

47:23

know, the things that I suppose my

47:25

kind of use of sort of AI

47:27

and AI like technology. It's kind of

47:30

living this AI dream of, I imagine

47:32

what I want to do, the AI

47:34

gets it done. And I think that's

47:37

sort of the well-defined version of how

47:39

that works out. Let's see. So the

47:41

question was about telling students what to

47:44

think as well as how to answer

47:46

the question. So I mean, I agree

47:48

that there's an issue of, you know,

47:51

what is teaching? What's that supposed to

47:53

be about? It's supposed to be about

47:55

sort of inducing certain patterns of thinking

47:58

and certain remembered facts. in the student.

48:00

And yes, it is true, and it's

48:02

an interesting point, that, you know, it's

48:05

the think like an AI, you know,

48:07

are we really going in that direction?

48:09

There is a certain tendency, and, you

48:12

know, even right like an AI, you

48:14

know, sometimes I'm wondering whether, you know,

48:16

the fact that the first generation of

48:19

LLLams was very into the word delve.

48:21

And then I start seeing humans writing

48:23

and using the word delve a lot.

48:26

I don't think I've ever written the

48:28

word delve. I don't know. I can

48:30

check because I have archives of everything

48:33

I've ever written. So I could easily

48:35

check how many times I've written the

48:37

world delve in my life. I'm guessing

48:40

it's very very very small. And I

48:42

haven't started now. But for many people,

48:44

that was a rare word. but then

48:46

the LLLM started using it and then

48:49

the humans started using it, copying the

48:51

LLLM's. And yes, I think that's an

48:53

interesting issue. If it gets to the

48:56

point where sort of one's teaching the

48:58

human and what one is trying to

49:00

do is to model the way that

49:03

humans think about things and the way

49:05

that an expert human might think about

49:07

this and model that for a human

49:10

who is just learning, but yes, there

49:12

is a certain tendency to say. Actually,

49:14

it will turn out to be the

49:17

way the AI thinks about these things,

49:19

and we're teaching the human to be

49:21

a bit like an AI. You know,

49:24

it makes me think of all the

49:26

things that I've automated in my life,

49:28

and the things where it's like, oh,

49:31

I'm thinking of my older daughter who

49:33

is now a mathematician, when she was

49:35

younger, you know, she would do things

49:38

like calculus by hand, and I would

49:40

sort of make... silly annoying parent comments

49:42

like I didn't know anybody still did

49:45

that kind of thing by hand. You

49:47

know, I had been doing that. I

49:49

haven't done that by hand myself in

49:52

25 years that time, 30 years, whatever

49:54

it was. And because I've always been

49:56

using a computer, but it's kind of

49:59

like what there's kind of this, this.

50:01

well, what should you delegate to the

50:03

machine versus what should you do yourself?

50:06

How should you think about things? And

50:08

I kind of feel like the big,

50:10

the big thing that us humans have

50:13

to contribute is sort of what questions

50:15

to ask, what direction do we want

50:17

to go? The, the how to get

50:19

the question answered, like how to get

50:22

the calculus problem solved is much more

50:24

in the domain of the machines. And

50:26

so the question of education in the

50:29

end. is much more how do you

50:31

teach humans to sort of think about

50:33

things that that and make decisions about

50:36

things that haven't been decided before. And

50:38

I tend to think that although people

50:40

might imagine that that's all a story

50:43

of just thinking thinking about the learning

50:45

about the dynamics of thinking itself, my

50:47

personal experience and my observation of other

50:50

things is that knowing a ton of

50:52

facts is super important in being able

50:54

to actually think clearly. It's the facts

50:57

are the bedrock. on which you can

50:59

build this kind of these layers of

51:01

sort of abstract thinking. And if you

51:04

try and just sort of teach the

51:06

abstract thinking in the abstract, it just

51:08

doesn't work. It's something, by the way,

51:11

I think we learned this a little

51:13

bit from the LLLMs. The fact that

51:15

the LMs can do a certain amount

51:18

of sort of common-sensish thinking lives on

51:20

top of the fact that they were

51:22

taught a zillion effectively facts about the

51:25

world. that's where we abstract to be

51:27

able to do thinking from. And so

51:29

in education it's really important, I think,

51:32

to learn a bunch of facts. That's

51:34

the bedrock on which we can actually

51:36

sort of construct this kind of layer

51:39

of abstract thinking that is then probably

51:41

ultimately the thing that we can contribute

51:43

the most to sort of what happens

51:46

in the world is kind of the

51:48

thinking about what... to do next what

51:50

we care about. Those are, in a

51:52

sense, human choices and being able to

51:55

do a good job as humans of

51:57

thinking through those things, I think is

51:59

an important thing to be teaching in

52:02

education. And there's sort of a question

52:04

of can we teach. how to think

52:06

like humans were their eyes. I don't

52:09

know the answer to that yet. Possibly

52:11

yes. I mean, it's, you know, the

52:13

other thing to realize is, for example,

52:16

the way that, I don't know, Wolfen

52:18

language computes some math thing, like calculus,

52:20

is deeply nonhuman. I mean, if you

52:23

know, if you know how it works

52:25

inside, it's a wonderful algorithmically clever industrial

52:27

machine for just grinding out the answer.

52:30

And when we were building step-by-step functionality

52:32

in Wolfmalfa, it's interesting how much of

52:34

a fake it is, in the sense

52:37

that it's very useful for the step-by-step

52:39

stuff to know what the answer is

52:41

going to be. That answer was found

52:44

by industrial methods that are deeply nonhuman.

52:46

When we build the step-by-step explanation, that

52:48

is something that we're building modeling the

52:51

way that humans think about things. And

52:53

it's kind of a retrofitted model of

52:55

the way that humans think about things.

52:58

And yes, it's even if there's a

53:00

situation where the computational way to work

53:02

it out is something deeply nonhuman, one

53:05

could certainly imagine sort of backing out

53:07

what the human way to think about

53:09

that is and being able to help

53:12

people understand how to think in those

53:14

human ways. Teaching people to think like

53:16

computers is not going to work. We

53:18

just don't, our brains are not capable

53:21

of that. I think that the, you

53:23

know, if I say, can I run

53:25

this piece of code in my brain,

53:28

I cannot. If I say, do I

53:30

have intuition about what this code is

53:32

roughly going to do, the answer is

53:35

absolutely. I can get that intuition. It's

53:37

a kind of a vague, abstracted intuition

53:39

about things. But if you say, can

53:42

I actually run this code and say,

53:44

well, I'm going to get three or

53:46

seven out of it, the answer is

53:49

no chance. And I don't think that's

53:51

just my inadequacy, so to speak. I

53:53

think that's just not the way the

53:56

100 billion neurons that are in our

53:58

brains are built to operate. in a

54:00

microprocessor are taught to operate, at least

54:03

when you have layers of things like

54:05

our computation. language on top, but it's

54:07

just a different kind of thing than

54:10

what brains do. Let's see. Maybe this

54:12

is a different question. Some of the

54:14

previous one here. Can a sentient AI

54:17

understand how humans learn if we would

54:19

delegate them the teaching of human kids?

54:21

Would that be compatible with a biological

54:24

point of view? I mean, I think

54:26

that... You know, there's this

54:28

fundamental question of, of can we, can

54:30

we understand psychology? Is psychology a thing

54:33

about which there can be a scientific

54:35

understanding? Is it something where we can

54:37

have a summarization of how psychology will

54:39

work as opposed to just, well, and

54:41

then this neural network with 100 billion

54:44

neurons runs and it does this? which

54:46

is not, you know, science is about

54:48

trying to take what happens in the

54:50

world and turn it into kind of

54:52

a human narrative. And that's, that's something

54:55

that, that it's not obvious you'll be

54:57

able to do that for something like

54:59

psychology. And we're seeing that question for

55:01

LLLM psychology right now as well, of

55:04

to what extent can we summarize what

55:06

the LLLM's do? Can we get a

55:08

feeling for how the LLLM's work? I

55:10

mean, I know in our company, we

55:12

are essentially... sort of getting people to

55:15

learn to be LLLM psychologists. That's what

55:17

we need. We need that because we

55:19

are trying to get the LLLM, we're

55:21

trying to wrangle the LLLM to do

55:24

the right thing. And it's like, well,

55:26

what caused the LLLM to go crazy

55:28

at this point? We need kind of

55:30

an LLLM psychologist to figure that out.

55:32

And we're trying to get intuition about

55:35

doing that. And the question is, you

55:37

know, what level of intuition can we

55:39

get? Can we get as humans? Could

55:41

the LLLMs get about themselves? And then

55:44

the question is, well, what level of

55:46

intuition can an LLLM get about how

55:48

humans think about things? It's a very

55:50

interesting question. I think that the LLLans...

55:52

thumbs. can perhaps model how the humans

55:55

think about things in a sort of

55:57

fairly parallel way, even if you can't

55:59

understand at a narrative level how the

56:01

LLLMs are thinking, and you can't understand

56:03

how the humans are thinking, maybe they

56:06

will be close enough in the way

56:08

that the LLLMs can parallel the way

56:10

that humans think, that at least you'll

56:12

be able to have the L11 kind

56:15

of run in its mind a simulation

56:17

of what the human is going to

56:19

think and so on. I mean, right

56:21

now LLLMs are not very fast compared

56:23

to humans. comparable speed in some ways.

56:26

And that will change. The LLLMs will

56:28

get faster. And thus humans, well, I'm

56:30

afraid we're stuck, you know, operating at

56:32

the many millisecond timescale of nerve firings.

56:35

You know, you can't, you can, you

56:37

can sort of ingest all kinds of

56:39

chemicals to try and speed that up.

56:41

It isn't going to work. It's, you

56:43

know, the whole system is built for

56:46

a certain clock rate. Maybe when we

56:48

get, you know, we'll be able to,

56:50

just like the LLMs can call tools,

56:52

we will have neural implants that are

56:54

like our ability to call tools. So

56:57

it's like we're thinking, and then for

56:59

an LLM, it's we're thinking, we're thinking,

57:01

and then we generate this piece of

57:03

text that is the tool called, it

57:06

goes off to this external tool, you

57:08

got an answer, you know, all from

57:10

language results, comes back, comes back. that

57:12

gets read ingested by the LM. My

57:14

guess is that we'll be able to

57:17

do similar kinds of things directly with

57:19

our brains and that will be sort

57:21

of a, but we won't get the

57:23

sort of raw clock frequencies is not

57:26

going to go up. And maybe that

57:28

will be a place where you can

57:30

kind of have. the simulation of what

57:32

the human would be thinking, running outside

57:34

the human, and it's like, well, you

57:37

know, in the last 200 milliseconds, I,

57:39

the LLLM, figured out, listen, listen, listen,

57:41

listen, listen, listen, this. So I think

57:43

the right next thing to tell you

57:46

is blah, because that's what you're about

57:48

to figure out, so to speak. And

57:50

I can imagine that absolutely happening. the

57:52

robotics space, having kind of a wolfram

57:54

robotics, so to speak. Yes, I had

57:57

been interested in robotics for a long,

57:59

long time. My views about how that

58:01

will come to fruition have changed actually

58:03

a bit quite recently. I mean, what

58:05

I had always assumed was that robotics

58:08

would kind of come of age when

58:10

there was general purpose robotics, as computing

58:12

came of age when there were general

58:14

purpose computers. There was a time before

58:17

the 1940 and so on where where,

58:19

and even a little bit beyond that,

58:21

where, you know, you had a computer

58:23

that did one particular thing, another one

58:25

that did another particular thing. And then

58:28

the idea of universal computing that came

58:30

in initially theoretically and then in practice

58:32

came to pass where it's like you

58:34

just have a fixed computer and you

58:37

program it to do all the different

58:39

things you want it to do. Well,

58:41

I had been interested for a long

58:43

time in kind of modular robotics where

58:45

you had little, little components, you would

58:48

have this... thing made of sugar cube-like

58:50

objects or maybe smaller than that and

58:52

it would have all this kind of

58:54

this flowing of the sugar cubes to

58:56

be just the right shape and it's

58:59

kind of like it would walk along

59:01

a little bit like an amoeba might

59:03

walk along generating pseudopods and so on

59:05

and you'd see these sort of sugar

59:08

cubes moving around and it's kind of

59:10

like a liquid object that is sort

59:12

of sliming around the world so to

59:14

speak. I mean it wouldn't actually be

59:16

liquid it will be solid and I

59:19

you know. probably 15 years ago, I

59:21

was more than that actually, I was

59:23

interested in kind of how you would

59:25

build things like this, try and think,

59:28

try to figure out some of the

59:30

geometry of it, of how things slide

59:32

against each other, how you maintain enough

59:34

contact to keep power to each of

59:36

these little cubes and so on, not

59:39

cubes. We had a bunch of geometries

59:41

that were very much not cubes, that

59:43

kind of rubics cube like, like assembly

59:45

things that could slide against each other

59:48

and so on. And then how would

59:50

you kind of plan? or how does

59:52

it move from this configuration of that

59:54

configuration and so on. So I had

59:56

assumed for a long time that that's

59:59

kind of how robotics would go. down,

1:00:01

but eventually, rather than, you know, my

1:00:03

experience, for example, at robotics trade

1:00:05

shows, had been this weird thing,

1:00:08

that there's an aisle of hands,

1:00:10

there's another aisle of feet and

1:00:12

legs and so on, every aisle

1:00:14

of the trade show is a

1:00:17

different part of the body type

1:00:19

thing, or a different kind of

1:00:21

thing, whereas... It's that you kind of

1:00:24

needed this different special purpose hardware for

1:00:26

all these different kinds of things and

1:00:28

I imagine that sort of the future

1:00:30

there was this kind of notion of

1:00:32

a modular robot where it was really

1:00:34

just all just many copies of the

1:00:36

same thing and some of the work that

1:00:38

I did and cellular automatra and things like

1:00:40

that both I and other people kind of

1:00:43

picked up on some of that trying to

1:00:45

understand how you would build

1:00:47

this kind of modular robotics technology.

1:00:49

My guess right now is that actually

1:00:51

it will be much more like the

1:00:54

science fiction said. It will be much

1:00:56

more humanoid robots will learn to do

1:00:58

things like humans do. And I think

1:01:00

that the humanoid form factor for

1:01:02

a robot is convenient because our world

1:01:05

is built for things that have humanoid

1:01:07

form factors because we built all

1:01:09

of these buildings and machines and so

1:01:11

on to be operated by us humans

1:01:14

with, you know, the size we are,

1:01:16

the hands we have, things like this.

1:01:18

So it's a good, sort of a plug

1:01:21

compatible way to introduce robotics into the

1:01:23

world is to make the robots be

1:01:25

like us humans. Now the thing that's

1:01:27

always been a challenge is like even

1:01:29

how do you pick anything up? You know,

1:01:31

if the thing is a slithery thing and

1:01:33

you know, oh you have to know to

1:01:35

put it underneath there and... and you put

1:01:38

it between your two hands and so on,

1:01:40

or if it's a thing that's very fragile,

1:01:42

you have to know just what force you

1:01:44

can use, and is it going to slide

1:01:47

out of your fingers and all this kind

1:01:49

of stuff? And this has been a big

1:01:51

challenge, and there's been these sort of formalized

1:01:53

challenges of picking challenges and things relevant for

1:01:56

companies like Amazon and so on for operating

1:01:58

their warehouses, and it really hasn't. made, so

1:02:00

far as I can tell, that much

1:02:02

progress until recently. I think the thing

1:02:05

that's sort of now coming is being

1:02:07

able to use sort of the general

1:02:09

ideas of large language models and apply

1:02:11

them to the behavior of robots and

1:02:13

really crack the problem of having a

1:02:15

robot that can do stuff like a

1:02:18

human does. Now it's kind of a

1:02:20

funny thing because for text... There's, you

1:02:22

know, trillions of words of training data

1:02:24

that's out there from all the things

1:02:26

humans have written for what humans do

1:02:29

and how they pick things up and

1:02:31

so on. Well, there's a bunch of

1:02:33

video you can watch, but fundamentally we

1:02:35

don't have that data. And so it's

1:02:37

kind of funny, I visited a few

1:02:40

of these companies that have sort of

1:02:42

people in suits, capturing all this motion.

1:02:44

so that you can teach the robot

1:02:46

kind of how to do these kinds

1:02:48

of motions, giving it that type of

1:02:51

information. I think there's a certain amount

1:02:53

of autonomous stuff that is based on

1:02:55

kind of reinforcement learning of... of you

1:02:57

know let the robot try and feel

1:02:59

its way through the world so to

1:03:02

speak quite literally to learn how to

1:03:04

do things but my guess is that's

1:03:06

going to come and it might come

1:03:08

even quite suddenly that humanoid type robots

1:03:10

become able to do a lot of

1:03:13

tasks and I think there'll be a

1:03:15

you know my guess is there'll be

1:03:17

another it'll probably come suddenly maybe not

1:03:19

I'm not sure and it'll be another

1:03:21

kind of chat gPT like moment where

1:03:24

people say oh my gosh all these

1:03:26

things that only humans could do before

1:03:28

suddenly the machines can do them too.

1:03:30

And they won't do them perfectly, but

1:03:32

they'll do them well enough for many

1:03:35

purposes. And then the real issue, it's

1:03:37

an interesting question, is sort of how

1:03:39

do you fit those capabilities into what's

1:03:41

useful in the world? Just like ChatGPT,

1:03:43

you know, if you want ChatGPT to

1:03:46

solve your computational... problem. It's not a

1:03:48

good idea. It's not going to work.

1:03:50

But if you want it to summarize

1:03:52

this piece of text for you or

1:03:54

to answer questions about a piece of

1:03:56

text or to generate some report that

1:03:59

can be fed to maybe another online,

1:04:01

it's going to be great at those

1:04:03

things or to be able to do

1:04:05

the kind of thing we do in

1:04:07

a notebook assistant where you're asking it

1:04:10

something vaguely and it's using a bunch

1:04:12

of techniques. fundamental L&M nature, so to

1:04:14

speak, to generate precise war from language

1:04:16

code that you can then do things

1:04:18

with. Those things are going to do

1:04:21

great. You make it do the wrong

1:04:23

thing, it's not going to do very

1:04:25

well at all. So it will be

1:04:27

with sort of L11-style robotics for a

1:04:29

large behavior model, LBM-style robotics. There'll be

1:04:32

things that that can immediately do well,

1:04:34

and there'll be things where, yeah, that's

1:04:36

probably not a good idea to have

1:04:38

it do that. I don't know exactly

1:04:40

what the things that will do well

1:04:43

are. It's an interesting question, actually. It's

1:04:45

worth thinking about. It's like, you know,

1:04:47

things in, I don't know what, the

1:04:49

construction industry, for example, which employs a

1:04:51

lot of people. Is that going to

1:04:54

be a place where, yeah, the robot

1:04:56

can get the material for the right

1:04:58

place on the building site or whatever

1:05:00

else? I know from talking to CEOs

1:05:02

of companies that are making humanoid robots,

1:05:05

at least as a year ago, a

1:05:07

bunch of them were saying, well, one

1:05:09

of the use cases is tidying up

1:05:11

after you. I don't know if that's

1:05:13

a realistic one. That seems really hard

1:05:16

to me. It seems like not one

1:05:18

of the ones that you would pick

1:05:20

as a top priority, especially, you know,

1:05:22

depending on, maybe it's one that has

1:05:24

a certain amount of wiggle room. It's

1:05:27

just like, you know, if you make

1:05:29

a machine learning a machine learning system

1:05:31

rank. do rankings for a search engine,

1:05:33

that's a fairly reasonable thing to do

1:05:35

because, you know, if two out of

1:05:37

ten of the things that it ranks

1:05:40

highly are exactly what you want, it's

1:05:42

a win. And if one or two

1:05:44

of the things are completely crazy, doesn't

1:05:46

really matter. So there will be some

1:05:48

tasks for robotics where doing it roughly

1:05:51

right is good enough. And there are

1:05:53

other tasks where... where, if you get

1:05:55

it even slightly wrong, it's a disaster.

1:05:57

I don't know whether tidying up a

1:05:59

room is probably one of these things

1:06:02

that's somewhat forgiving task. I mean, I

1:06:04

remember, oh gosh, how long ago was

1:06:06

it? It was probably, what was it?

1:06:08

Must be, eight years ago maybe? I

1:06:10

remember being at the Consumer Electronics Show.

1:06:13

And one of the big attractions was

1:06:15

this laundry folding robot. There's a big

1:06:17

line to see the laundry folding robot,

1:06:19

but it wasn't that exciting actually in

1:06:21

the end. I don't know if it's

1:06:24

become a product that's really used, but

1:06:26

it's a surprise, it was at that

1:06:28

time a surprisingly hard thing to fold

1:06:30

laundry. My guess is that the next

1:06:32

generation of robots will find it very

1:06:35

easy to fold laundry. That's probably a

1:06:37

task that again is somewhat forgiving. It's

1:06:39

like, oh, you get it slightly wrong,

1:06:41

and the fold is in slightly the

1:06:43

wrong place, big deal, big deal. And

1:06:46

so, you know, there will be a

1:06:48

bunch of things that very quickly, my

1:06:50

guess is, will be doable by humanoid

1:06:52

robots, and then there'll be other things

1:06:54

where people might try to do them

1:06:57

with humanoid robots, and it's sort of

1:06:59

a bad idea and doesn't work very

1:07:01

well. But it'll be interesting to see,

1:07:03

I'm kind of guessing, it'll be a

1:07:05

sudden thing that there'll be just, a

1:07:08

whole bunch of stuff just works. I

1:07:10

mean, there have been practical problems with

1:07:12

humanoid robots, like providing enough power. I

1:07:14

mean, I mean, I'll eat foods, get

1:07:16

muscles, get muscles, is actually a surprisingly

1:07:18

efficient way to get things to move.

1:07:21

And a lot of humanoid robots, kind

1:07:23

of the big trick is there's a

1:07:25

giant cable coming out of the back

1:07:27

that's supplying a lot of power to

1:07:29

the robot. I mean, the thing to

1:07:32

understand about muscles and so on is

1:07:34

that, well, it's actually, muscles are pretty

1:07:36

clever. You know, they have these little

1:07:38

devices on the molecular scale that are

1:07:40

kind of walking down the muscle, pulling

1:07:43

on it and so on. Those are

1:07:45

things which are operating on a molecular

1:07:47

scale. they are much more energy efficient

1:07:49

than the things we have operating at

1:07:51

the scale of electric motors and so

1:07:54

on. And I don't know, you know,

1:07:56

I don't know to what extent the

1:07:58

kind of humanoid robot and it's, you

1:08:00

know, things like the practicalities, the hardware

1:08:02

practicalities, to what extent that's going to

1:08:05

be a long thing factor. It may

1:08:07

also be that those hardware practicalities as

1:08:09

soon as one is, oh, let's make

1:08:11

10 million of these robots, it starts

1:08:13

being much more feasible to have that

1:08:16

work well. I also suspect the battery

1:08:18

technology is improving, you know, has improved

1:08:20

it enough that there may be a

1:08:22

different set of power constraints that exist

1:08:24

there. So many questions, guys, thank you.

1:08:27

Wylo asks. If people delegate all calculations

1:08:29

to machines, might it not happen that

1:08:31

the machine actually learns to ask better

1:08:33

questions than the humans can, since the

1:08:35

machines have the experience built from the

1:08:38

calculations that the humans don't? Well, the

1:08:40

question is, what do you want to

1:08:42

ask the question about? I mean, I

1:08:44

could ask all kinds of questions about

1:08:46

why, you know, the... the tree that

1:08:49

I see out of my window has

1:08:51

this particular pattern of branches and not

1:08:53

another? Why the cloud in the sky

1:08:55

has this particular pattern of fluffiness and

1:08:57

not another? Right now we humans don't

1:08:59

care about those questions. There are an

1:09:02

infinite number of questions we can ask,

1:09:04

most of which we don't care about.

1:09:06

So the question of which do we

1:09:08

care about is as much a sort

1:09:10

of societal question. And a human, human

1:09:13

question as anything, it's not something where

1:09:15

there's sort of an abstract answer. Now,

1:09:17

I think there will come a time,

1:09:19

and it's sort of already happened to

1:09:21

some extent where AI's can suggest, hey,

1:09:24

this is something you as a human

1:09:26

might care about, kind of the idea

1:09:28

generator thing. And yes, I think that's

1:09:30

happened. I think it's, it's, you know,

1:09:32

it's an interesting question when the ideas

1:09:35

that people come up with. are assisted

1:09:37

by AIs and those ideas are kind

1:09:39

of coming out of the exhaust from

1:09:41

the humans, so to speak. What happens

1:09:43

to this whole system? You know, is

1:09:46

it the case that the ideas just

1:09:48

revert to, well, okay, it's a standard

1:09:50

idea? It's just like I see in

1:09:52

things like, I don't know, math enrichment

1:09:54

activity. that people propose, I guess I

1:09:57

haven't looked at these recently, but a

1:09:59

while ago I did. And it's kind

1:10:01

of amusing, I mean, sometimes they're things

1:10:03

based on my work, which is always

1:10:05

nice to see, but there's a certain

1:10:08

similarity to, you know, okay, we're going

1:10:10

to do math and Richmond, okay, then

1:10:12

there are five ideas we might have

1:10:14

that all seem like sort of spunky

1:10:16

original creative ideas, but there was the

1:10:19

same five ideas type thing. And I

1:10:21

could imagine that happening when sort of

1:10:23

all ideas are all ideas at LLLM-s

1:10:25

sourced sourced. Let's see. Weasel is commenting

1:10:27

tiny humans care about those questions. I'm

1:10:29

not sure what the questions were at

1:10:32

this point. The Dan is asking, what

1:10:34

will AI not be able to do?

1:10:36

Do I believe that something like that

1:10:38

exists? Well, the answer is kind of

1:10:40

what I've been saying, the choice of

1:10:43

which direction to go, there are an

1:10:45

infinite number of possible choices. Which one

1:10:47

we choose to go in is a

1:10:49

thing for us? because there's no right

1:10:51

answer. Now we could delegate that to

1:10:54

the AIs, we could say, we're giving

1:10:56

up, just like, okay, from here on

1:10:58

out, it's cruise phase, you know, cruise

1:11:00

phase for humanity, we just do whatever

1:11:02

the AIs tell us to do. My

1:11:05

guess is that our nature is not

1:11:07

such that that will go down very

1:11:09

well. Our nature, you know, forged by

1:11:11

three billion years of biological evolution, has

1:11:13

to do as well, we're going to

1:11:16

seek out the new and so on.

1:11:18

So my guess is that the, you

1:11:20

know, you know, you know, you know,

1:11:22

we're just here for the, we're just

1:11:24

here for the, we're just here for

1:11:27

the, we're just here for the, isn't

1:11:29

going to play out well for us

1:11:31

humans and we won't allow that to

1:11:33

be what happens but I suppose in

1:11:35

principle that could happen. I mean I

1:11:38

think it's a you know a thing

1:11:40

that I always find kind of kind

1:11:42

of amusing in the sense I never

1:11:44

really dug into this I don't really

1:11:46

know the facts as clearly as I

1:11:49

should but you know anthropologists tend to

1:11:51

say you know there are a few

1:11:53

cases in human history where there have

1:11:55

been some you know some tribes somewhere

1:11:57

where there were enough berries on the

1:12:00

bushes that they could, you know, feed

1:12:02

themselves and that it was... it was

1:12:04

an easy life type thing. And that,

1:12:06

so they weren't, you know, pushed by

1:12:08

necessity to do more. And then the

1:12:10

claim is, well, for a thousand years

1:12:13

or more, whatever it is, they were

1:12:15

just hanging out, you know, eating the

1:12:17

berries from the bushes and so on,

1:12:19

and not being pushed to do more,

1:12:21

not having sort of a struggle for

1:12:24

life. And then the question is, well,

1:12:26

what do people do in that situation?

1:12:28

And what I've heard said is, well,

1:12:30

then they do ritualisticistic things. Which is

1:12:32

of course a very strange thing to

1:12:35

say because in a sense viewed from

1:12:37

the outside Many things we do would

1:12:39

be thought could be thought of as

1:12:41

ritualistic like we don't understand from the

1:12:43

outside why anybody would do that But

1:12:46

from the inside we absolutely have a,

1:12:48

oh yeah, we're doing that because we

1:12:50

have a whole inner dialogue about why

1:12:52

we're doing those things. So, you know,

1:12:54

no doubt to the ritualistic tribes, so

1:12:57

to speak, the things they were doing,

1:12:59

you know, shaking sticks and, you know,

1:13:01

interacting with the rain gods or whatever

1:13:03

else they were doing or thought they

1:13:05

were doing, is, you know, that has

1:13:08

a complete meaning to them that to

1:13:10

the outside just looks like, oh, that's

1:13:12

ritualistic, we don't know what's going on.

1:13:14

you know, any kind of life of

1:13:16

the mind, I suppose, could look ritualistic

1:13:19

from the outside. And I kind of

1:13:21

think that, you know, we could imagine

1:13:23

a situation where us humans are doing

1:13:25

that. I kind of think that the

1:13:27

thing we will continue to be relevant

1:13:30

for is pick the path. There are

1:13:32

an infinite number of paths. We are

1:13:34

the defining elements of which path to

1:13:36

pick, unless we absolutely choose to advocate

1:13:38

on that and not do that. Oh,

1:13:41

weasel is saying the questions about

1:13:44

clouds and trees and why are

1:13:46

they those shapes? Yeah, interesting. Right.

1:13:48

Good question. I don't know. I'm

1:13:50

trying to remember if any of

1:13:52

my kids asked that question. I

1:13:55

should remember. I can remember many

1:13:57

interesting questions that they asked. which

1:13:59

were indeed, indeed, I mean, because

1:14:01

I've studied a bunch of sort

1:14:03

of foundational science, maybe I was

1:14:05

in a better position than most

1:14:08

to be able to answer lots

1:14:10

of kid questions. I guess that's

1:14:12

perhaps an inspiration for some of

1:14:14

these live streams, although you guys

1:14:16

tend to ask more sophisticated questions

1:14:19

than my kids when they were five

1:14:21

years old asked. And although my kids asked

1:14:23

some very interesting questions. I

1:14:25

remember one that stuck out for a

1:14:28

long time was... five-year-old, one of my

1:14:30

kids, was when there were dinosaurs, could the

1:14:32

Earth have had two moons? And that question

1:14:34

is really hard to answer, and I,

1:14:36

but for years, whenever I would

1:14:39

run into people who did celestial

1:14:41

mechanics, planetary dynamics, kinds of things,

1:14:43

I would ask that question.

1:14:45

And people would say, well, I'm

1:14:47

not sure. You know, we have

1:14:49

these simulations, I don't really know.

1:14:51

I think the answer now... is

1:14:53

pretty clearly no, the Earth couldn't

1:14:55

have had two moons six to

1:14:57

five million years ago, but it could

1:14:59

easily have had two moons

1:15:02

a billion years ago. We

1:15:04

wouldn't know that. And it's

1:15:06

quite possible that it did. But

1:15:09

that's kind of a, well, it's

1:15:11

sort of an obvious question.

1:15:13

I'd never thought of that question. Sancho

1:15:15

is commenting. At least at the time

1:15:18

when I went to a bunch of

1:15:20

them, commenting that Boston Dynamics shows a

1:15:22

lot of progress in the humanoid department.

1:15:24

I'm not sure. I live in the

1:15:26

Boston area and there's a building on

1:15:29

Route 128 in Boston that is now

1:15:31

the Boston Dynamics building. It's been other

1:15:33

companies in the past. And when I

1:15:35

drive by there, I'll sometimes kind of

1:15:37

try and look in the windows to

1:15:40

see whether there are any cool robots

1:15:42

there. They don't seem to be most

1:15:44

of the time. I'll probably run into there

1:15:46

at least. founder on some of them

1:15:48

that I'm going to soon. So maybe

1:15:51

I'll hear more about, I'll

1:15:53

probably see more actually, also

1:15:55

as an event that I'm going

1:15:57

to next week that will. have

1:16:00

probably a bunch of the latest

1:16:02

and greatest robotic things at it.

1:16:04

Always interesting. It's always the question,

1:16:06

you know, I was at a

1:16:08

robot company, oh, what was it,

1:16:10

sometime last year, and it's like

1:16:12

there's a humanoid robot and you

1:16:14

can shake its hand. And I'm

1:16:16

not quite trusting enough that I'm

1:16:19

prepared to let the robot put

1:16:21

its hand around my hand. I'm

1:16:23

more in the let me grab

1:16:25

your hand robot from the outside.

1:16:27

We'll see how that evolves. Let's

1:16:29

see, Mighty is commenting, anything to

1:16:31

say about the future of pie,

1:16:33

happy pie day. Happy pie day

1:16:35

to everybody. This is an American

1:16:37

dates, it's 314, the first digits

1:16:39

of pie. What's the future of

1:16:41

pie? Interesting question. I mean, we've

1:16:43

got a trillion digits computed so

1:16:46

far. Will we get more? Will

1:16:48

anybody ever find a pattern in

1:16:50

the digits of pie? A statistical

1:16:52

pattern, none has been found. Nobody's

1:16:54

ever proved that there isn't a

1:16:56

statistical pattern. I would say the

1:16:58

big stretch goal for pie is

1:17:00

to show pretty much anything about

1:17:02

the digits of pie. Right now

1:17:04

we have no idea. All we

1:17:06

know is that the digits of

1:17:08

pie don't repeat. Pi is not

1:17:10

a rational number, so the digits

1:17:13

of Pi don't repeat. But beyond

1:17:15

that, knowing whether in base 10,

1:17:17

for example, they're the same number

1:17:19

of ones and twos and threes

1:17:21

and fours and so on, nobody

1:17:23

has any idea. People have been

1:17:25

trying to figure that out for

1:17:27

the last 150 years, and really

1:17:29

nothing is known about that. That's

1:17:31

kind of one of these things

1:17:33

where one might say mathematics is

1:17:35

not really ready for that question

1:17:38

yet. you know as we build

1:17:40

out mathematics we kind of build

1:17:42

out these different paradigms these different

1:17:44

sort of in the space of

1:17:46

or possible theorems we're gradually building

1:17:48

out more and more that we've

1:17:50

kind of colonized we've reached maybe

1:17:52

three or four million theorems out

1:17:54

of the infinite number of possible

1:17:56

theorems of mathematics and the question

1:17:58

is have we yet colonized that

1:18:00

we yet visited that we explored

1:18:02

the part of mathematical space that

1:18:05

would lead us to answer a

1:18:07

question like, is pie so-called normal?

1:18:09

Does it have equal numbers of

1:18:11

every digit and every block of

1:18:13

digits and so on? So that's

1:18:15

an example of kind of a

1:18:17

coming attraction for pie. I don't

1:18:19

think that's close. I think it's

1:18:21

an interesting question whether, you know,

1:18:23

as we automate more mathematics, I've

1:18:25

been involved obviously in automating lots

1:18:27

of mathematics, I don't know. whether

1:18:29

sort of a question like that

1:18:32

will come, has more of a

1:18:34

chance to come over the horizon,

1:18:36

I don't know. As far as

1:18:38

other things about, well, it's, yeah,

1:18:40

I think that's the main thing

1:18:42

that I can see in sort

1:18:44

of the future of pie is,

1:18:46

you know, what pie day will

1:18:48

we know that there's something to

1:18:50

say generally about the digits of

1:18:52

pie? Let's see. There's a question

1:18:54

from Jay Chen, do you expect

1:18:57

a limb development to hit significant

1:18:59

diminishing returns within the next two

1:19:01

to three years? Well, I think

1:19:03

it already has. I mean, I

1:19:05

think that what the story of

1:19:07

machine learning tends to be a

1:19:09

story of sort of a new

1:19:11

domain gets cracked and then there's

1:19:13

a big jump at that point,

1:19:15

and then there's kind of incremental

1:19:17

progress after that. So, you know.

1:19:19

image recognition got cracked around 2011

1:19:21

to 2012. Things like speech to

1:19:24

text got cracked in the, when

1:19:26

was it that, something like, oh,

1:19:28

let's see, something around the late

1:19:30

2010s, you know, text generation, packed

1:19:32

with LLLans. Now these things get

1:19:34

gradually better and the harnesses around

1:19:36

the LLLans, the ways to use

1:19:38

these things, the ways to use

1:19:40

AIs, they get there gets to

1:19:42

be more and more understanding how

1:19:44

can you use this technology. I

1:19:46

mean it's just like when people

1:19:48

invented ways to do fast linear

1:19:51

algebra, fast matrix computations. It's like

1:19:53

well that's a thing and that's

1:19:55

kind of cool and then people

1:19:57

realize well we can use that

1:19:59

to do computer graphics and then

1:20:01

that started a whole direction and

1:20:03

sort of the core of what

1:20:05

was done there was is matrix

1:20:07

algebra. Same thing actually with Ella

1:20:09

Lambs it's lots of matrix algebra

1:20:11

going on inside the the GPUs

1:20:13

and so on that are driving

1:20:15

Ella Lambs. So You know, my

1:20:18

own feeling is that it's what

1:20:20

you see in pretty much all

1:20:22

areas, whether it's in science, technology,

1:20:24

whatever, there's a breakthrough, there's a

1:20:26

new methodology, there's a big jump,

1:20:28

then things level out. And you

1:20:30

know, in science it might take

1:20:32

100 years before you get to

1:20:34

the next kind of methodological breakthrough

1:20:36

and big jump up. I mean,

1:20:38

I like to think in physics,

1:20:40

for example, that our physics project

1:20:43

is finally... basically 100 years after

1:20:45

the last sort of big methodological

1:20:47

advances in physics, we finally got

1:20:49

another set of big methodological advances

1:20:51

that are opening up a bunch

1:20:53

of new new possibilities. But you

1:20:55

know, my guess is that what

1:20:57

will happen in the in the

1:20:59

LLLM world is that a bunch

1:21:01

of new modalities, whether it's video,

1:21:03

whether it's robotics, whether it's things

1:21:05

with, you know, chemical sensing, I

1:21:07

don't know, you know, there'll be

1:21:10

these modalities that kind of open

1:21:12

up where they were pretty much

1:21:14

Greenfield. There was nothing there before

1:21:16

and then they get they get

1:21:18

solved or they get substantially solved

1:21:20

and that's where the big advances

1:21:22

will come rather than you know

1:21:24

the the incremental kind of let's

1:21:26

make it a bit better if

1:21:28

we look at something like image

1:21:30

recognition you know what we have

1:21:32

today is better than what we

1:21:34

had in 2012 but not that

1:21:37

much. And what's important is we

1:21:39

understand much more how to take

1:21:41

the things that we could do

1:21:43

in 2012 and fit them into

1:21:45

a harness that really fits into

1:21:47

other kinds of things. So I

1:21:49

think that's that's kind of that's

1:21:51

that's the way I see that

1:21:53

developing and I see I need

1:21:55

to go back to my day

1:21:57

job in a minute here. Let's

1:21:59

see if there's maybe one more

1:22:02

quick question. commenting that automated theory

1:22:04

improving is interesting. I'm trying to

1:22:06

figure out how to make a

1:22:08

theory improver that demonstrably collapses the

1:22:10

wave function, like I was talking

1:22:12

about earlier, quantum LLMs. Yeah, I

1:22:14

mean, this question of how you

1:22:16

knit together sort of these kind

1:22:18

of computational methods like automated theory

1:22:20

improving with what's being done with

1:22:22

our limbs. Very interesting question. I

1:22:24

think I made a bit of

1:22:26

progress on that last year. Peace

1:22:29

I wrote about sort of why

1:22:31

does machine learning work and this

1:22:33

kind of discrete model of machine

1:22:35

learning that allows one sort of

1:22:37

lays much more bare the kind

1:22:39

of the essence of machine learning

1:22:41

I think gives one a much

1:22:43

better handle on what one would

1:22:45

need to do to kind of

1:22:47

thread together those lumps of machine

1:22:49

learning functionality. together with kind of

1:22:51

computational functionality to do things like

1:22:53

sort of integrate theorem proving with

1:22:56

LLLMs and that would be where

1:22:58

I would start to look for

1:23:00

trying to do that. I haven't

1:23:02

figured out how to do it

1:23:04

but that's it's on my to-do

1:23:06

list. I think it's such a

1:23:08

bit far away so it's it's

1:23:10

I encourage other people to to

1:23:12

look at it. All right well

1:23:14

we should wrap there I see

1:23:16

a bunch of other interesting questions

1:23:18

which I look forward to to

1:23:20

trying to address another time. But

1:23:23

it's always fun talking with you

1:23:25

guys and the questions you ask

1:23:27

are interesting and get me thinking

1:23:29

about things I'm not the worst

1:23:31

thinking about and I think I've

1:23:33

been, thank you for contributing to

1:23:35

sort of the future of science

1:23:37

and technology. I'll try to be

1:23:39

the executor of some of these

1:23:41

things that come up as ideas

1:23:43

here and perhaps some of you

1:23:45

can be as well. Anyway, thanks

1:23:48

for joining me and Until the

1:23:50

next time, bye for now. You've

1:23:52

been listening to the Stephen Wolfram

1:23:54

podcast. You can view the full

1:23:56

Q&A series on the Wolfram Research

1:23:58

YouTube Channel. For more information. on

1:24:00

Stephen's live streams,

1:24:02

and this podcast, visit

1:24:04

StevenWolfram .com.