Full episode with Stephen Wolfram (Apr 2020): https://www.youtube.com/watch?v=ez773teNFYA
Clips channel (Lex Clips): https://www.youtube.com/lexclips
Main channel (Lex Fridman): https://www.youtube.com/lexfridman
(more links below)

Podcast full episodes playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4

Podcasts clips playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

Podcast website:
https://lexfridman.com/ai

Podcast on Apple Podcasts (iTunes):
https://apple.co/2lwqZIr

Podcast on Spotify:
https://spoti.fi/2nEwCF8

Podcast RSS:
https://lexfridman.com/category/ai/feed/

Stephen Wolfram is a computer scientist, mathematician, and theoretical physicist who is the founder and CEO of Wolfram Research, a company behind Mathematica, Wolfram Alpha, Wolfram Language, and the new Wolfram Physics project. He is the author of several books including A New Kind of Science, which on a personal note was one of the most influential books in my journey in computer science and artificial intelligence.

Subscribe to this YouTube channel or connect on:
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

I can answer the question for you, but is it basically, not the philosophical, the deep, the profound, the impact of it, I'm talking about in terms of tools, in terms of things you can download, in terms of stuff you can play with, what is it?

What does it fit into the infrastructure?

What are the different ways to interact with it?

Right, so I mean, the 2 big things that people have sort of perhaps heard of that come from Wolfram Language, 1 is Mathematica, the other is Wolfram Alpha.

It's this system that is basically an instance of Wolfram language, and it's used to do computations, particularly in sort of technical areas.

And the typical thing you're doing is you're typing little pieces of computational language and you're getting computations done.

It's very kind of, there's like a symbolic.

So symbolic language, so I mean, I don't know how to cleanly express that, but that makes it very distinct from how we think about sort of, I don't know, programming in a language like Python or something.

Right, so the point is that in a traditional programming language, the raw material of the programming language is just stuff that computers intrinsically do.

And the point of Wolfram Language is that what the language is talking about is things that exist in the world or things that we can imagine and construct.

It's not sort of, it's aimed to be an abstract language from the beginning.

And so for example, 1 feature it has is that it's a symbolic language, which means that the thing called, you'd have an X, just type in X and Wolfram Langelder would just say, oh, that's X.

I don't know what it is, computation, in terms of the internal computer.

Now that X could perfectly well be the city of Boston.

Or it could perfectly well be the trajectory of some spacecraft represented as a symbolic thing.

And that idea that 1 can work with, sort of computationally work with these different, these kinds of things that exist in the world or describe the world, that's really powerful.

And that's what, I mean, you know, when I started designing, well, when I designed the predecessor of what's now Wolfram language, which is a thing called SMP, which was my first computer language.

I kind of wanted to have this sort of infrastructure for computation, which was as fundamental as possible.

I mean, this is what I got for having been a physicist and tried to find, you know, fundamental components of things and wound up with this kind of idea of transformation rules for symbolic expressions as being sort of the underlying stuff from which computation would be built.

And that's what we've been building from and Wolfram Language.

And, you know, operationally what happens, it's, I would say by far the highest level computer language that exists.

And it's really been built in a very different direction from other languages.

So other languages have been about, there is a core language, it really is kind of wrapped around the operations that a computer intrinsically does.

Maybe people add libraries for this or that, but the goal of Wolfram Language is to have the language itself be able to cover this sort of very broad range of things that show up in the world.

And that means that, you know, there are 6,000 primitive functions in the Wolfram Language that cover things, you know, I could probably pick a random here.

I'm going to pick just for fun, I'll pick, let's take a random sample of all the things that we have here.

So let's just say random sample of 10 of them, and let's see what we get.

So these are really different things from-

That's a thing for converting between different types of Boolean expressions.

So for people just listening, Stephen typed in random sample of names, sampling from all functional, how many you said there might be?

6,000, from 6,000, 10 of them, and there's a hilarious variety of them.

Well, we've got things about dollar requester address that has to do with interacting with the world of the cloud and so on.

I want to pick another 10 because I think this is some...

Okay, so yeah, there's a lot of infrastructure stuff here that you see if you just start sampling at random, there's a lot of kind of infrastructural things.

If you more, you know, if you more look at the-

Some of the exciting machine learning stuff you showed off, Is that also in this pool?

I mean, so 1 of those functions is like image identify is a function here.

Text in a current image, accessing the webcam, took a picture of yourself.

But anyway, we can say, image identify open square brackets and then we just paste that picture in there.

Image identify function running from the picture.

It says I look like a plunger because I got this great big thing behind my head.

Classify, so this image identify classifies the most likely object in the image.

Okay, well it thinks there's a, oh, it thinks it's pretty unlikely that it's a primate, a hominid, a person.

That hopefully will not give you an existential crisis.

And then 8%, or I shouldn't say percent, but-

No, that's right, 8% that it's a hominid.

And yeah, okay, it's really, I'm gonna do another 1 of these just because I'm embarrassed that it, it didn't even see me at all.

There we go, let's try that, let's see what that did.

We took a picture with a little bit more of your body.

A little bit more of me, and not just my bald head, so to speak.

So then I would, but you know, so this is image identify as an example of 1.

And that's part of the, that's like part of the language.

And I mean, you know, something like, I could say, I don't know, let's find the geo nearest.

Let's try finding the 10 volcanoes nearest here.

So, geo nearest volcano here, 10 nearest volcanoes.

We can now, we got a list of volcanoes out and I can say geo list plot that And hopefully, okay, so there we go.

So there's a map that shows the positions of those 10 volcanoes.

Of the East Coast and the Midwest, and it's the, well, no, we're okay.

But The fact that right in the language, it knows about all the volcanoes in the world, it knows computing what the nearest ones are, it knows all the maps of the world and so on.

It's a fundamentally different idea of what a language is.

That's why I like to talk about it as a full-scale computational language.

And just if you can comment briefly, I mean, this kind of, the Wolfram language along with Wolfram Alpha represents kind of what the dream of what AI is supposed to be.

There's now a sort of a craze of learning, kind of idea that we can take raw data and from that extract the different hierarchies of abstractions in order to be able to, like in order to form the kind of things that Wolfram language operates with.

But we're very far from learning systems being able to form that.

And the context of history of AI, if you could just comment on, there is, you said computation X, And there's just some sense where in the 80s and 90s, sort of expert systems represented a very particular computation X.

And there's a kind of notion that those efforts didn't pan out.

But then out of that emerges kind of wolfram language, wolfram alpha, which is the success, I mean.

Yeah, right, I think those are, in some sense, those efforts were too modest.

That is, they were looking at particular areas and you actually can't do it with a particular area.

I mean, like even a problem like natural language understanding, it's critical to have broad knowledge of the world if you want to do good natural language understanding.

And you kind of have to bite off the whole problem.

If you say we're just gonna do the blocks world over here, so to speak, you don't really, it's actually, it's 1 of these cases where it's easier to do the whole thing than it is to do some piece of it.

You know, 1 comment to make about, so the relationship between what we've tried to do and sort of the learning side of AI, you know, in a sense, if you look at the development of knowledge in our civilization as a whole, there was kind of this notion pre 300 years ago or so now, you wanna figure something out about the world, you can reason it out.

You can do things which are just use raw human thought.

And then along came sort of modern mathematical science.

And we found ways to just sort of blast through that by in that case, writing down equations.

Now we also know we can do that with computation and so on.

And so that was kind of a different thing.

So when we look at how do we sort of encode knowledge and figure things out?

1 way we could do it is start from scratch, learn everything, it's just a neural net figuring everything out.

But in a sense that denies the sort of knowledge-based achievements of our civilization.

Because in our civilization we have learned lots of stuff.

We've surveyed all the volcanoes in the world.

We've done, you know, we've figured out lots of algorithms for this or that.

Those are things that we can encode computationally, and that's what we've tried to do.

And we're not saying just, you don't have to start everything from scratch.

So in a sense, a big part of what we've done is to try and sort of capture the knowledge of the world in computational form, in computable form.

Now, there's also some pieces which were for a long time undoable by computers, like image identification, where there's a really, really useful module that we can add that is those things which actually were pretty easy for humans to do that had been hard for computers to do.

I think the thing that's interesting that's emerging now is the interplay between these things, between this kind of knowledge of the world that is in a sense very symbolic and this kind of sort of much more statistical kind of things like image identification and so on.

And putting those together by having this sort of symbolic representation of image identification, that that's where things get really interesting and where you can kind of symbolically represent patterns of things and images and so on.

I think that's kind of a part of the path forward, so

Yeah, so the dream of, so the machine learning is not, in my view, I think the view of many people is not anywhere close to building the kind of wide world of computable knowledge that will from a language of build.

But because you have a kind of, you've done the incredibly hard work of building this world, now machine learning can be, can serve as tools to help you explore that world.

And that's what you've added with the version 12, right?

You added a few, I was seeing some demos, it looks amazing.

Right, I mean, I think, you know, this, it's sort of interesting to see the, the sort of the, once it's computable, once it's in there, it's running in sort of a very efficient computational way, but then there's sort of things like the interface of how do you get there?

You know, how do you do natural language understanding to get there?

How do you pick out entities in a big piece of text or something?

That's, I mean, actually a good example right now is our NLP NLU loop, which is, we've done a lot of stuff, natural language understanding, using essentially not learning-based methods, using a lot of, you know, little algorithmic methods, human curation methods, and so on.

In terms of when people try to enter a query and then converting, so the process of converting, NLU defined beautifully as converting their query into a computational language, which is a very well, first of all, a super practical definition, a very useful definition, and then also a very clear definition of natural language understanding.

Right, I mean, a different thing is natural language processing, where it's like, here's a big lump of text, go pick out all the cities in that text, for example.

And so a good example of it, so we do that, we're using modern machine learning techniques and it's actually kind of an interesting process that's going on right now, is this loop between what do we pick up with NLP using machine learning versus what do we pick up with our more kind of precise computational methods in natural language understanding.

And so we've got this kind of loop going between those, which is improving both of them.

And I think you have some of the state-of-the-art transformers, like you have BERT in there, I think.

So it's cool, so you have, you're integrating all the models.

I mean, this is the hybrid thing that people have always dreamed about or talking about.

That makes you just surprised, frankly, that Wolfram language is not more popular than it already is.

That's a, it's a complicated issue because it's like it involves, you know, it involves ideas and ideas are absorbed slowly in the world.

And then there's sort of like what we're talking about, there's egos and personalities, and some of the absorption mechanisms of ideas have to do with personalities, and the students of personalities, and then a little social network.

So it's interesting how the spread of ideas works.

You know what's funny with Wolfenlanguage is that we are, if you say, you know, what market sort of market penetration, if you look at the, I would say, very high end of R&D and sort of the people where you say, well, that's a really, you know, impressive, smart person, they're very often users of orphan language, very, very often.

If you look at the more sort of, it's a funny thing, If you look at the more sort of, it's a funny thing, if you look at the more kind of, I would say, people who are like, oh, we're just plodding away doing what we do, they're often not yet, well, from language users and that dynamic, it's kind of odd that there hasn't been more rapid trickle down, because we've really, you know, the high end, we've really been very successful in for a long time.

But with, that's partly, I think, a consequence of, my fault in a sense, because it's kind of, I have a company which is really emphasizes sort of creating products and building a sort of the best possible technical tower we can, rather than sort of doing the commercial side of things and pumping it out in sort of the most effective way.

an interesting idea that, you know, perhaps you can make it more popular by opening everything up, sort of the GitHub model.

But there's an interesting, I think I've heard you discuss this, that that turns out not to work in a lot of cases, like in this particular case.

That you want it, that when you deeply care about the integrity, the quality of the knowledge that you're building, that unfortunately, you can't distribute that effort.

Yeah, it's not the nature of how things work.

I mean, what we're trying to do is a thing that for better or worse requires leadership and it requires kind of maintaining a coherent vision over a long period of time and doing not only the cool vision-related work, but also the kind of mundane in the trenches make the thing actually work well work.

That's the mundane, the fascinating and the mundane.

It's building the knowledge, the adding, integrating more data.

Yeah, I mean, that's probably not the most, I mean, the things like get it to work in all these different cloud environments and so on.

That's pretty, you know, that's very practical stuff.

You know, have the user interface be smooth and, you know, have there be take only, you know, a fraction of a millisecond to do this or that.

And It's, but, you know, I think my, it's an interesting thing over the period of time, you know, orphan language has existed basically for more than half of the total amount of time that any language, any computer language has existed.

That is, the computer language is maybe 60 years old, you know, give or take, and orphan Language is 33 years old.

So it's kind of a, and I think I was realizing recently there's been more innovation in the distribution of software than probably than in the structure of programming languages over that period of time.

And we've been sort of trying to do our best to adapt to it.

And the good news is that we have, because I have a simple private company and so on that doesn't have a bunch of investors telling us we've got to do this or that, they have lots of freedom in what we can do.

And so, for example, we're able to, oh, I don't know, we have this free Wolfram Engine for developers, which is a free version for developers.

And we've been, you know, we've, there are site licenses for Mathematica and Wolfram Language at basically all major universities, certainly in the US by now.

So it's effectively free to people and all universities in effect.

And we've been doing a progression of things.

I mean, different things like Wolfram Alpha, for example, the main website is just a free website.

Okay, Wolfram Alpha is a system for answering questions where you ask a question with natural language and it'll try and generate a report telling you the answer to that question.

So the question could be something like, you know, what's the population of Boston divided by New York, compared to New York.

And it'll take those words and give you an answer.

And that have been- Converts the words into computable, into- Into Wolfram language, actually.

And then- Computational language, and then computes the answer.

the underlying knowledge belongs to Wolfram Alpha, to the Wolfram language?

We just call it the Wolfram knowledge base.

I mean, it's been a, that's been a big effort over the decades to collect all that stuff.

And you know, more of it flows in every second.

Like, that's 1 of the most incredible things.

Of course, in the longterm, Wolfram language itself is the fundamental thing, but in the amazing sort of short-term, the knowledge base is kind of incredible.

So what's the process of building that knowledge base?

The fact that you, first of all, from the very beginning, that you're brave enough to start, to take on the general knowledge base.

And how do you go from 0 to the incredible knowledge base that you have now?

Well, yeah, it was kind of scary at some level.

I mean, I had wondered about doing something like this since I was a kid, so it wasn't like I hadn't thought about it for a while.

But most of us, most of the brilliant dreamers give up such a difficult engineering notion at some point.

Well, the thing that happened with me, which was kind of, it's a live your own paradigm kind of theory.

So basically what happened is, I had assumed that to build something like Wolfram Alpha would require sort of solving the general AI problem.

And so I kept on thinking about that, and I thought I don't really know how to do that, so I don't do anything.

Then I worked on my new kind of science project and sort of exploring the computational universe and came up with things like this principle of computational equivalence, which say there is no bright line between the intelligent and the merely computational.

So I thought, look, that's this paradigm I've built.

Now I have to eat that dog food myself, so to speak.

I've been thinking about doing this thing with computable knowledge forever, and let me actually try and do it.

And so it was, if my paradigm is right, then this should be possible.

But the beginning was certainly, it was a bit daunting.

I remember I took the early team to a big reference library and we're like looking at this reference library and it's like, you know, my basic statement is our goal over the next year or 2 is to ingest everything that's in here.

And that's, you know, It seemed very daunting, but in a sense, I was well aware of the fact that it's finite.

You know, the fact that you can walk into the reference library, it's a big, big thing with lots of reference books all over the place.

You know, this is not an infinite, you know, it's not the infinite corridor of, so to speak, of reference libraries, not truly infinite, so to speak.

But no, I mean, and then what happened, was sort of interesting there, was from a methodology point of view, was I didn't start off saying, let me have a grand theory for how all this knowledge works.

It was like, let's, you know, implement this area, this area, this area, a few hundred areas and so on.

I also found that, you know, I've been fortunate in that our products get used by sort of the world's experts in lots of areas.

And so that really helped because we were able to ask people, the world expert in this or that, and we're able to ask them for input and so on.

And I found that my general principle was that any area where there wasn't some expert who helped us figure out what to do wouldn't be right.

Because our goal was to kind of get to the point where we had sort of true expert level knowledge about everything.

And so that the ultimate goal is if there's a question that can be answered on the basis of general knowledge in our civilization, make it be automatic to be able to answer that question.

And, you know, and now, well, WolfMalfa got used in Siri from the very beginning and it's now also used in Alexa.

And so it's people are kind of getting more of the, you know, they get more of the sense of, this is what should be possible to do.

I mean, in a sense, the question answering problem was viewed as 1 of the sort of core AI problems for a long time.

I had a friend, Marvin Minsky, who was a well-known AI person from right around here.

And I remember when Wolfman Alpha was coming out, it was a few weeks before it came out, I think.

I happened to see Marvin and I said, I should show you this thing we have.

You know, it's a question answering system.

And he was like, okay, types something in, it's like, okay, fine.

And then he's talking about something different.

I said, no Marvin, this time it actually works.

Of course we have a record of what he typed in, which is kind of interesting.

Can you share where his mind was in the testing space?

He was just trying random stuff, you know, medical stuff and, you know, chemistry stuff and, you know, astronomy and so on.

And it was like, Like, you know, after a few minutes, he was like, oh my God, it actually works.

But that was kind of told you something about the state, you know, what happened in AI, because people had, you know, in a sense, by trying to solve the bigger problem, we were able to actually make something that would work.

Now, to be fair, we had a bunch of completely unfair advantages.

For example, we already built a bunch of orphan language, which was very high level symbolic language.

I had the practical experience of building big systems.

I have the sort of intellectual confidence to not just sort of give up in doing something like this.

I think that the, It's always a funny thing.

I've worked on a bunch of big projects in my life.

And I would say that you mentioned ego, I would also mention optimism, so to speak.

I mean, if somebody said, this project is gonna take 30 years, it would be hard to sell me on that.

I'm always in the, well, I can kind of see a few years, something's gonna happen in a few years.

And Usually it does, something happens in a few years, but the whole, the tale can be decades long.

And that's, you know, and from a personal point of view, always the challenge is you end up with these projects that have infinite tales.

And the question is, do the tales kind of, do you just drown in kind of dealing with all of the tales of these projects?

And that's an interesting sort of personal challenge and like my efforts now to work on fundamental theory of physics, which I've just started doing and I'm having a lot of fun with it, but it's kind of making a bet that I can do that as well as doing the incredibly energetic things that I'm trying to do with orphan language and so on.

And underlying that, I mean, I just talked for the second time with Elon Musk, and you 2 share that quality a little bit of that optimism of taking on basically the daunting, what most people call impossible, and he and you take it on out of, you can call it ego, you can call it naivety, you can call it optimism, whatever the heck it is, but that's how you solve the impossible things.

Yeah, I mean, look, what happens, and I don't know, you know, in my own case, you know, it's been, I progressively got a bit more confident and progressively able to decide that these projects aren't crazy.

See all Lex Fridman transcripts on Youtube

What is Wolfram Language? (Stephen Wolfram) | AI Podcast Clips