The Best Way to Achieve AGI Is to Invent It
38:02 Share

I in nine, took ten years to come up with general relativity, and that was just one equation. So what makes you think we're going to soviet six months? I really do think we need fundamental new ideas.

And what I see, unfortunately, is that there is more research than every bit as ual, less diverse than ever, which is very unfortunate. There is a sprint, but it's a sprint to a local optimum. And there's nothing wrong with sprinting to the local.

So there's a lot of managed to be had, for example, making transform faster and efficient. All that is good at the end of day. That is not to get to human.

I having said that, even if that all we do is make transformers Better, that alone is already going to have an incredible impact, is just that A I is a technology of an impact that people are not used to. So even just getting ten percent to A G I is kind of change the world. dramatic.

Welcome again to sixteen Z A I podcast and dared you just hit play at another great discussion. This one features a sixteenth general partner, our team dad, in discussion with noted machine learning researcher and now multiple times published author page o. Domingos pagers, a professor, merits of the university of washington and has been involved in the world, the machine learning at one thousand nine hundred ninety.

He's also the author of the popular twenty fifteen book of the master algorithm and recently published his first novel, twenty forty, a silicon valley satire. And all of discussion touches on those books and the ideas of them. The focus really is on whether our current path of A I research is the right one to get us to the mountain top with the that's agi, super intelligence or whatever anyone thinks is the ultimate goal.

Among other topics, Martini page of assessment critique that continued to click, ability of scaling laws, the wisdom of building out massive AI data, the importance of alms and the transformer architecture on the quest for the one I the master of algorithm is an insightful conversation, which is Better hurt from them than recapped by me. So here goes starting with page, or giving a very appreciated version of this linker resume. As a reminder, please note that the content here for informational purposes only should not be taken as legal, business tax or investment advice or be used to evaluate any investment or security, and is not directed to any investors or potential investors. Is in any a sixteen fund? For more details, please see eight sixteen s that calm slash disclosures.

I've been a mission leving researcher for thirty years since since my p in the night. So i'm kind of an A I old timer, a professor versions of washing. But i'm probably best known as the author of the mass, which was a book, popular ized machine learning that that went best seller. Most recently another book on its a SATA of A I called twenty four seven a and i'm continuing to the research and right and gently try educ public.

So I noticed that you're a meritis. Can you describe what that means regarding your day to day?

A man like, you know who the during the world winter is beautiful because you do everything you want .

and nothing that you don't. exactly.

My joke to people is that back when I was a professor doing research with my hobby, and now I can do IT full time, you arch. So yeah, definitely work on solving a, of course.

is I could come out when we have this conversation, or is IT a secret.

But let me just put IT. This week I ve solved IT. I just haven't told you that .

is that that's and you're not onna tell .

us that I can't yet. I can I need to patent IT and you I that seems like way to do first and prove that IT were Better. There are many details you've saw.

Okay, everybody, you've heard IT here, pedro doing as a, saw the I and by the way, saw that you mean like A G, I.

no, I don't just mean G, I mean super intelligence. Wow, I see this all the time and and I love you because people don't believe me and believe I will have the last law. I just said this a tedi yesterday.

Love that. Okay, well, this this, this goes to the first topic that I want to discuss. What if you've said many times that A I is a marathon, not a sprint, and we're in what seems to me i've also been following a for quite a while that were in like the dead run phase where you have got innovations happening all the time. And so would love to hear the subtext for you saying that originally and then how you view the current madness when IT comes to progress and pace of like maybe just set context.

absolutely. I really do deeply believe that A, I is a marathon on a sprinter. That is why i'm doing what i'm doing.

I I think he's an annoyed ledge, right? I intime took ten years to come up with general relativity, and that was just one equation. So what makes you think we're going to Sophia in six months? So there's a couple of different views on this.

One is that is just about scaling, if if he just about killing than what i'm doing is wasting my tie, right? But I don't I really do think we need fundamental new ideas. And what I see, unfortunately, is that there is more research than never bit such ual less diverse than ever, which is very unfortunate.

There is a print, but it's sprint to a local optimum and there's nothing wrong with sprinting to the local. So like there's a lot of miles to be had, for example, in making transform is faster and more efficient. All of that is good. But at the end of the day, there is not going to get to human level intellects. I don't think, having said that, even before we do and you know to to get back to the second part of the question, even if that all we do is make transformers Better and that things like, you know what a one is doing and what not, that alone is already going to have an incredible impact, is just that people are a is a technology of an impact that people are not used to. So even just getting ten percent to agi is kind of change the world dramatic.

So I want to talk about both of those things. The first one is, you know, you talk about transformers, you talk about the current path, mean, right now, things have been following scaling laws pretty well.

And if you talk to focus on the big la B2Be lik e, you know, like we can predict if you add more compute and you add more data, like that's going to be so like this holding, I keep hearing the last holding s the nothing in computer science holds forever. So if you were to predict, are the scaling law is gonna hold? And if not, what do you think is the limiting feature?

okay. So I don't know if the scaling are gonna ld, but actually the scale of laws are actually pretty version. There have been a bunch of papers coming out saying that their basic an illusion, they are creature of the metrics.

They ba ba ba and and let me tell you what a formative experience from me was. I actually spent the first several years of my academic career, in the nineties and early two thousands, largely working on skilling up machine learning organs. That was the first way of skillings up.

And then we got to a point where actually could predict mathematically how well the algorithm would learn, give an infinite data, and he was still stupid. And then IT was like, just fine. But like, no, this is not.

I'm looking for IT, you know, intelligence, this is not. And I started on this path where, I mean, right now he is like, we are limitations this way. A human brain is not a scaled up anto brain.

So at any point is it's fine for people to be doing what up. And thus I scale up what you have to limit. I've done that myself.

That's good, but we shouldn't be under the illusion that the skilling was onna take us there. By the way, somebody y's killing was not look as good as you was what you might think. But every time someone tells you does an exponential here, you you need to hear, ask her instead of exponential. And the question is where I going to ask him to?

So to think so clearly, I hear you that we probably need another architecture. And I think folks are. But assuming that given architecture, I mean, that IT seems like get scaling, do you think there are limits in data or computer? You think like we can keep their dead anthem and keep in computer anthem and gonna some sort of gains?

Or do you that running we can get, but it's going to be diminishing returns on all these fronts. Architecture, data, algorithms, the hard work is set to all of these have big problems. So in a way, the local optima that were in is a bunch of different local ultima all in all these aspects.

And know, I think, when we will look tenets from now, the architecture will be different. The harder will be different. The way we use that with the at set.

the whole work do do you think that s are amended part of like code and code A G I R A S I in the future, like it's can be more complicated and maybe there are other aspects. But the fundamental part, do you think it's .

just not even relevant to the ultimate luck? I think that if you lot language models started in the fifties and they were used for speech and machine translation for many decades, and their role, this is, I think what people often forget was, for example, I translate from, you know, english to french. The language models role is to patch up the french after it's been translated, yeah, to make IT look Better.

So the language model has a role, but he has no connection to the real world. Now what has happened is that because we have all these texts and are not in the way we are using the language model as the world models, well, I don't think that's gona survive. That's just a very interesting accident of history as an interface of language model, you know, to then generate the text after you have actually answered red the question using real knowledge and real reasoning. I it'll probably surviving the in that fashion. But large language models is the foundation of A I is, I think if you told me that, or anyone ten years ago, we would have laughed, and maybe we will laugh again ten years now.

So, so as humans, we've got this very highly evolved brain. The brain experiences the world. We make abstractions in that world, like I call you pedro and you're human and yeah, one house that's right.

And then in our head, we figure out connections between these. We come up with kind of models of how, like I think you act. And then we write all of that down, let's say we write that down in text.

And so like there's all of this energy going into kind of understanding the world, coming a great relations and running that down. There's one view that the l ams basically take what all of the work that we've done and just use IT as a cash and reggina ated. And there's another one that I can actually derby deeper notions of like Pearls and Martins and whatever in these. And so are you saying it's more of a cash for the worth that humans have done? No.

no, it's actually neither of voice. So there's this notion that allowance are just a cash parents. And on this, that is a completely stupid notion like, by definition, machine learning about generalizing.

If all we wanted to do was memorized the data, doing back property would just be a ridiculous waste of time. Now at the same time, I don't think that what they actually doing is going deeper. There is the problems that they are currently able to go deeper.

What they're doing is they generalizing some distance from the data, which because you have an enormous amount of data, lets you look really good. So in a way, they do look more intelligent than they are. And we're still for good god, exactly what is that they can and can do.

But clearly know there's something you say that I think Carries a big global, which is we do not write down everything we know. We have learned that over many years of trying to extract, acknowledge the website. The reason these models make sometimes rely to be there, that they don't have common sense about how the world works that two year old has, and that party still missing.

I see. So do you think in the whole world, models keep peace to this, at least conceptually?

I, I, I agree with you on on a lot of things, disagree on some the vision. Folks, of course, like han and fae and attended one not like hell. No, language isn't gonna get us there. I ve actually worked .

most exactly, say right first .

and you know it's fine for people to place that.

Can you actually just can you see the suffer people? So I don't think the. Lay person, listen to this, understands this rift. But IT is so huge, which is like almost people come from two different backgrounds to this point, and is a massive skis on that we don't talk enough about. So if you could you do that.

that would be like history. The rift that goes back to the beginning of A I, and it's the embodiment. Rift is to soft.

A I do I. Neither our is to be embodied or no. And there are very strong and very good arguments on both sides in, in fact, traditional, symbolic.

I was liking not to help with all that. John mccarthy, one of the funders of I, was like, i'm going to solve A I by writing down a bunch of formas in first theologic. Why I thought was ridiculous.

But I have actually spent most of my life working in the mining the web mode because I think this is the first test path to progress because we do let let me put this way the mistake that maybe sometimes people like yeah and someone make, which is is the select but this is you know how human intelligence he does not have to voltage the same way. IT has different resources, different constraints, IT said. IT, like IT, has an amazing fossil fuel, like fossil fuels, part the industrial pollution, but the location is part by the internet, which this data we have at the end of the day.

However, you know, this is born of our experience, like mining the web. And I think L L. Are just a continuation of this. We are still gonna be missing a lot of things because the text that we write actually recipes and writing text for other people. So I presuppose that they know a bunch of stuff that isn't even we're talking about.

And then in fact, this actually one of the most beautiful things about machine learning is that, you know, I know how to ride a bike. I don't know how to explain how to ride a bike. I don't know how to ride a program to ride a bike, but the machine learning can learn to ride a bike if I make IT ride a bike, not from text about .

how to right a bike. So are you on roughly on the side of the ebola ed folks, roughly on the side of the text boxes are somewhere totally different?

No, very good. So I think there's a doctor jet and mr, high quality. This is like the stuff that i'm working on right now, right again, that i've worked on for has been i've done send vision stuff and were not, but overwhelmed. What I have done is work. So like, you know, I know know, if you will, because I do think, like we have all this data, why not make progress there? I do think, however, at the end of the day, we are gonna need, you know, robots and video understanding and interaction and what not, or something equivalent that not simulation, to really, truly get human level intelligence.

I mean, this may be a noncircular question, but that seems to me that any time you have the cameras and trying to reconstruct the world, you're dealing with a very long tail of like exception bases and such as I mean historical, every time we've gone down that path with like autonomous vehicles just been wait way more expensive, way, way, way longer and we ve got to wait worse results. That's not the case. The elements like every like eighteen months we have this market proveth you think going to go back into the long tail of like the universes.

a complex place? Oh, interesting. You should say that because first of all, I do with you, with your vision and robotics are much harder than language.

But, but is the thing, they all have a long tail. In fact, a brill, a famous NLP researcher, a used to speak of walking the zip flatlands. He says, when zip 了， that's what all this is.

What is exactly? This is what has been killing people in NLP since beginning. Yeah, aliens do not actually solve this problem that make you look like IT would solve, because they have so much that they can work for the out.

But you still get corner .

kisses all the time. We try the hallucinations, the abba blah, right? They worth in vision, but and they can hidden.

And and I mean, part of the debate is like, well, and if we didn't get enough for these cases and in power at some point, we just know we have everything, but we will never have everything. The space is so large, you will always be able to say something. This surprises and l surprises for this.

I was talking to mark last night, mark Andrews, and he got of something I had forgotten about, which is, in the early days of the internet, there was this mean that the amount traffic was doubling, and everybody just kind of repeated. This is like in the midst nineties. And so everybodys I go, the mountain traffic is double.

World comes that amount traffic is doubling. And then everybody started doing basically the build out of fiber, saying the amount traffic was doubling. And like all the sudden, if your traffic wasn't doubling, you'd say IT was double because then you're not deleting edge.

But IT turns out the traffic wasn't doubling and actually hit the minister returns very, very quickly, right? However, the rest of the industry kept doing all of this build out and then we ended up with, like the fibre glass, and I was like this almost mass delusion, where everybody kind of agreed that this the scaling was happen. When is one of happening? Do you think there's, you know, any of that going now i'm not talking about like not being able to predict IT, but like it's actually slow down. But we want to acknowledge that is slow down.

This is a great analogy, right? That comes up very often and because like you cannot resist if you know that part of history, I like in like the whole a eda singable is just really reminiscent of that, right? In fact, I even joked on twitter at some point.

Like in the eighties, IT was my super collide is bigger than yours. And then two thousands, like my fiber network is bigger than us and it's my edit thing, is bigger than years, right? So in some ways, I actually think that the companies are pouring billions of dollars into this, and they making a something that I interesting are not true.

But there there's an interesting twist to this, which is that AI has infinite appetites for computing. You will never satisfy us. So I think what's going to happen is like they're making this build out on Better assumptions, but in the end, we'll actually have a use for all those that is in is now they building in these GPU, which in away will not be optimal. Those GPU are not going to going to use. I guarantee you will have used this, brother.

I think in the history of the industry, we've never maintained a supply of hands of with your computer longer than, say, about four and a half years. That was kind of the worst case, which is the dog congress.

So in one sense, I I totally agree with, we will use the computer, and we always have historical for the last seven years interested in is like, is there like a mass delusion where weed it's not like we don't predict the future enough is just we can't actually see when it's not happening. What i'm saying is IT became a self fufu ling profession because nobody could admit that IT wasn't doubling anymore. And so do you think any of that happening? Is IT too early like like is have things slowed down but nobody is admitting IT?

Or is IT just too early to know of so down? I think illusions, honestly, I think there's A A certain amount of they on the ground that I don't have and maybe nobody does, but we can examine the assumptions behind this, right? And and the assistance really two parts.

One is, I mean, I recently found, for example, Marks like to talk about this like like he's putting billions into this. Of course, some of them was for other reasons which are not being used for this, but that's fine. Is like if you believe the scaling laws hold and the skillings laws will take us to human level intelligence, it's worth a lot of investment, right? That's one part, but that may be wrong.

The other part is that, however, to do that, we do need, you know, exploiting amounts of computer. What I, if I had to predict, like what going to happen, is that we do not need the trillion dollars to reach A G I. At all. So if you need a trillion.

this is a very bad investment.

I, you know, one hundred millions, so save the other. You know, one hundred and ninety. yeah.

okay. So the air race has been going on for decades. Jeff hinton just got the nobel prize.

And so can you talk about the significance of that? Both what that means, the computer science and the physicists would love at us. So we are real scientists now. We have a nobel and then be like the actual.

So the work that was a very strAngely noble Price for multiple reasons. And then given that, you can try to understand why the noble committee gave them that Price. The strange thing here is this, and noble pricing physics is supposed to be for great physics, and hole networks and bosman machines are neither great nor physics.

Physics, right? I mean, yes, both some machines use the box and distribution, which, but like I made, who cares? That doesn't them? Physics, right? No one.

no. To pretend that this is and hoping that was with physics there, right? And I think that there several factors.

So IT could be that they're clueless sometimes because this is a different field, right? I think no one three of this is that like they were depressed like, oh, you give able Prices to all these relevant things and there's all the important stuff happening here. So maybe that's what happened.

And maybe they think that that i've seen things, things that happened before, for example, been on committed, where a lot of people can. I have the wrong idea? Maybe they think, hope that was the machines are really important, right? If you look at what the world, they don't seem to be making that decision. So maybe there are little confused.

Another another view, which you know, I think may also be real, they are not mutually love, is is that this is that in some with a very political move to channel some of the A I hype towards physics because it's like putting down A G like, hey, you know, a bunch of this came from physics and everybody there, I know people are really doing this. Like, hey, let's have more people in physics to do deep learning and give us more grant money. And baba, so maybe there is that .

as to what what about there's a theory that is also political and that hinton talks often about safety and you know, like the need regulate and so forth. You think that that had anything to do with the decision? Or do you think that just fear monger .

ing from like I think I think I think maybe that play the part. But my guess is that not because the nobel I mean like is not that the nobel committee is a set of real people and they do have their political leanings, and for example, fear right wing veris, a left of the economist. Instinctively, they sympathy more with the left twin one.

And and and I think they are probably are concerned more than this should be with the dom and that that helps them to like hinton. So maybe they play the role. I don't think this cynically said we're going to for all I know, right? I don't think like we're going to give this guy nobel prize because when to give here, there has as much good and set these needs .

maybe maybe just sticking into kind of the science thing. So here's an observation. I'm kind of a lay person on this like I don't have a deep background and I the background system, however and IT seems to me that like one of the primary use cases right now of A I and alams and gena, I is actually creative.

It's like if you actually like dollar waited, like where people make money from, it's like I will create an image. I would do Better than human. I create a story, I create a video. And in one level, like that kind of was working.

But if you actually look at where many of the advances are happening right now, especially with synthetic data, it's like science, is like actimel systems, is like math, is like sciences, and it's like code, right? And so one belief would be that, well, in axim tic systems you can actually create synthetic data. Therefore we can of push that dat the ARP actually getting away from like the creativity language step.

When we do that, then there's another view that like the more we look at solving science and the more we look at solving math, like we will get Better generally at everything. Do you have a view on this? This is ism. This is actually pretty market right now.

Well, I do believe that there is a general foundation for this way where a book called the master algy's. I do think there is a master along. The goal of my research is is to find that.

But now you have to realize that these things are actually quite different. The creative uses, the math, the protein folding and sea center. How much you do this is the path of generally tells is an interesting question.

I agree with that. The killer APP right now is things like generation. Those are not.

Which is very ironic, because I used to have a slight explaining to people, you know, more of x paradox. What's harder is easy for A I right, which is the opposite of expert. And one of the lines that got the most resistance was creativity, reliability.

And I said it's easier even like a couple years ago, right before, you know, ChatGPT, I said creativity, these in reliability side. And people were like, what are you talking about? Creativity is this unfathomable human thing.

So we were that's actually working out great right now, right? But it's not going to know the economic impact of that will be significant, will be in a world transform. On the math side, the the property that math has and go has and what not is that it's an artificial universe separate from the real one, which makes IT way easier for AI.

In fact, games and math and what not, if you get back to the fifties and sixties, they were the natural model organism for you, I, to try things on, because he was used to start there. And in particularly, can play games against yourself, to your parts content. And this is how you got to alphago.

O now, math, now there's a problem, which is you can generate themes, but what makes you a good theory? And I can prove themes all that long, just I put, I meet my machine right, just just by pushing the button. So these are different. I am generally extremely sceptical of a notion that in theory there, or simulation is gna get us that because we have fifty years of experiencing A F that not working.

And in fact, when you bring in the machine learning and actually gets worse, because the machine learning system uncandidness to keen on the stuff that you can use to as the task, but is completely revving to what your actual channel and to over feeling to the simulator is, is almost inevitable. And by the way, deep mind, right? They came out doing this because I think they didn't know the history, right? He worked, thought for them, but actually they haven't succeeded in that path. Protein folding is different because the same tent is still .

is a very simple world, of course. So for me, this is very interesting from from an economic investor standpoint. And like i'm wondering if we're gone to see like A A split or a schism where on one hand, the economics are phenomenon when the comes to generating creative content.

And I would say dollar waited. That's probably the number one uses of A I right now actually. And that that has no formal notion of correctness. Ss like listening itals or are probably a good thing at etra. And then there's, on the other hand, of these systems that are like very tuned to highly automatic things like you're saying, like protein folding does I mean that we can end up two types of?

No, I I don't think so. I mean that I think those things are just points on the spectrum. And most of the test that we have to solve in the real world have a combination of the two of example, if you just want to generate to cover for a book, all you have to do is please the author or the publishing, and ultimately know the readers.

The let me up with this way, the rewards function is very flat in areas where the word function is very flat. Generation does great, right? But the promise of, for example, one of the piece of the gas that that I often give to people, companies these days, this achieve that point.

I would use IT eternally before I use IT externally, because then there were function starts to get a little sharp right? You don't want to be in canada, and they have a return policy, refund d policy that was made up by your chat pot, and you have to enforce IT. So even like here's IT, a test that has been people have between to that.

Doing that for decades is like I am going to schedule your travel and manage your blah, blah, blah, right? Like we would all love that. But the problem is that is that the rewards function that is surprising with sharp because you should double book me or something, i'm going to be very annoyed, right? So for a lot. So I would say the kilo spot right now is where it's very flat and you can generate a lot of stuff like you thin molex of a you generate like these management plans. Well, yeah, I mean, like there's so much scope to say stuff there, right?

So maybe a very specific question on this. Like, let's say, one of these large model companies gets really, really good at math and you know solves all these outstanding themes and winds all the contest.

Does that work contribute meaningly to the use cases that are around creativity? Or or are they these little like separate research pass separate points in the design space? And like now we're going to actually start to see schism because what's been so phenomenal as one model is used for both of these things today. Just one of the research seeing that pulled apart just because .

like the data different well at s versus a one right L M S generate. But once you start to read and like they started this project couple years ago, because they saw that the lens weren't going to get you to the reliable, and for that you needed reasoning. Now the thing about all this that is slightly ironic, is that people in symbolic, I already know how to do that stuff. Well, you don't nearly lamps to prove IOS and and all of .

that like in the sixties yeah I mean, like and there .

are famous examples of firms being proved by symbol because if all you doing is manipulating symbols, really all this stuff is massive overall. If you want to pass a yet math exams, you need to understand, being said, but, but, but that for a mathematician.

that's not the important part. So I had, again, have lay intuition on this. My lay intuition is, is the reason alarms are good at the math is because they can read the english, and then, and then they do something pretty simple on the math for a computer. Is that like the way to think more plate?

This is a controversial question OK. Some people say yes, some say no. I think so. Certainly reading english helps. But, but, and, and a lot of the feeling would do the alams, like when IT comes time to do the real method, just fall ridiculously flat, right? So we have this machine that the buzz of floating point computations, and you ask you to add two numbers and to get them .

wrong here about the kind of like her full on this, where they found out at any time, like dot nine one one was involved, IT would get something wrong. And IT just turns out that because of nine, eleven year on themselves, like I so hard to test apart, like i'm using these numbers. For like addition verses.

all of the other good or another example is, you know, you ask you to generate, run the numbers.

And forty two comes up surprisingly well. So yes, so I I actually use I use another them for my programing, and I actually plotted the random a horrible, it's a horrible ble number. Then that's what you'd expect.

Now here's the thing is easy to look at that and laugh. But what I believe is the change. And not just believe this emergences, like the transformers, have the ability in principal to do this kind of compositionally that at in the days what math a theme, what is proving a theme right is like you have these acids, you combine them to produce another step, you bring in another action, like they do have this power yeah, right now.

Well, representations, ally. But in terms of of being able to learn to do this, that's where people often forget this, right? They confuse being able to represent something.

We'd being able to learn IT by backroads, and they clearly not learning IT. So clearly, there's a gap there. I do believe we need to find the mental and new organs for that.

I want to switch the topics to you personally and then to your book. Do you know how dave charging is? That is a professor stand like the thing with that, a charging is like polemic. But he was like often right.

like this very part, like him ah I think I have many a couple of times and .

like I was but he was like, right so off like my area, which is distributed systems and networking so of kind of view, just like the AI version of this, which is like you you often have kind of a non consensus view, but you're so often correct. And so the simple question is, are you intentionally like i'm a political or is IT just kind of naturally how you are?

No, no, no. I'm not.

How much energy goes into like being non consensus.

So i'm being intentional, but my intention is not to be politic. c. My goal in life is to have the positive impact that I can. Part of that is seeing things that I believe are different from what the mainstream believes, but need to be sad than seeing. So I believe a lot of things in which I agree with everybody, but what's the point of me saying them like there's demining returns to that, right?

When I see that people are very wrong about something where they say I A society, something like I I could be wrong, but then I think maybe i'm contributing something, but at least putting that out there, right? And I mean, but is also true that certain people's personnel is lend themselves to this more than others. Many people wouldn't able to do this.

I from when I was that he had no problem disrespecting authority. Just saying what I believe I think what happens and i've seen this happen over over there that like people come to appreciate that there's somebody there who they may be wrong or they may be right, but but they are telling the truth as they see IT. And it's a use I like I I have people in various fields that I like to listen to.

I mean, like i'll give you a small example. Seben husson fell red in physics, right? She's such a great anti that to at all the other physics are saying because it's actually telling a lot of truth. So I think you're created actually goes up over time also because the future answers the question. I say it's gonna be a yes, it's gonna be and then a few years will see.

okay. So your book is actually fantastic for those listening to recommend you read that. I think I actually a great kind of historical survey in a way to think about A I where we are now, maybe just for the the listeners the overview of the book and may be motivated you to to write, you know.

why did I write a satire about the tech in? This is a very strange thing for a computer scientists to do. But i've always felt for a long time that someone, for example, I have journalist friends that want to to write the great american novel.

And I was said to them like, that's too hard. Look at what you competing with, right? But, you know, silicon valley, and somebody should write the book about that. And my model was the bonfire of the vantage, one of the the great novel of the eighties.

Because, like, what I did was IT, saturius, ed all state at the time that I was the iconic industry, and he doesn't satirized IT wasn't just what I was. The whole world in new york, in the politicians, the activists, the journalists and the brilliant job. And like, somebody needs to do that for A I, but I couldn't convent anybody to do this as well.

I'll do IT myself also because back when, as a great student in my copy's three time, I went to clarian west, which is this famous thing where for a month and a half, you learn how to write science fiction. So at least I knew how to do IT right? Like was, I wasn't like starting from scratch, but then the, but I didn't have a concrete idea about how to do this.

What actually finally made me right there was two things. First was that, like, they need to do this, particularly with A I in the last few years, has just exploded, right? But the other one was, there was this journalist who I really blame her for, this SHE asked me, and this was doing the trump administration.

So would you vote for an eye for president? And I was like, well, no. But if the other can was trump, then maybe.

What would you say now? I had to publish .

this novel now because it's gonna be out there. Everything I put in to coming true. And again, the novel makes one of the tech industry of a ee, but also of politics.

So the blood of the novel ChatGPT runs for president, yet the end of the day, the book, you know, without any spoilers. A lot of big disaster has happened in the book, but IT has an optimistic message at the end. And like there are two versions of press bot, press open with a chatbot.

The president to pino is where I think I should go and how I think I can make society Better. So as much as I make fun of all these things to kind of like illuminate them, right? A good satire based on observation.

It's like a cricketer. You can often see a people straight Better from a character than from a photograph. So no animal farm is not about animals .

are farming .

twenty four, not about twenty forty. It's it's a mirror being held out to twenty twenty, forcing like if we stay on this path. And of course, this is all part of IT happens in central isco that is now like, you know, organized but IT into the day, like as much as I try to show these things, I also try to show the way. So like I hope for twenty four is that IT does not come true. I want that novel to be the opposite of actually happens.

So I need to hear this. I think maybe people listening to hear this. Are you at least guardedly optimistic? We can figure that out.

Are you camp that we're fucked? Because reading the book of my, oh my goodness, there is no right path here. And so, like, I would love to, at least here.

you think there is a path. No, there is a thing. Again, prey about two point of this crowd. Source the ei instead.

Like what I was talking about that you know the I A couple days ago, I do think we can use the I to build a Better democracy than we have today. And this is not right wing or left doing, is just using art in like the founding father didn't have a eye. So I do think we can do way Better than we are doing today.

And we must, we really must now. And then the book illustrates is that the following is that, like I am in the short term pessimistic. I think the dynamic that will rain right now, fortunately, I can stop, you can stop.

The things are gonna get worse. And you think about like the cultural and the left.

there's multiple asic dispute. Of course, the most one is the polarization, right? Like the polarization that I make is really, really bad, right? And like, I don't know how to stop IT, but I do think in the long term I am optimistic.

But you know, I am maybe a good of those. Like i'm a worried optimist like this also a true of the eye in general and all this words about a eye. If we worry about a eye, it'll turn out well.

And if we don't want, which is what has happened to prove from the technology. My big problem with a lot of the AI words in particular, that people are worrying about the wrong things that discharges from the real problems. And I tried to redirect attention to what .

I see as the real problems just to find up I would love your thoughts because I know that you're working on now on A N A S I of like roughly just to make a prediction when you think we're going na have glimmers of real A G I.

I think there's a very good chance that the next few years of A I will see progress that makes the progress that we think so far look tiny really? Yes, absolutely. So you know, there are excuse, but they ask for A S so tall and we are showing the beginning that is entirely possible that like how you know twenty twenty that that was just the Price on the other hand, IT is also possible that things were entirely possible in a favorite question of journalists is like, you know, when and you think we'll get to ig and and my stop cancer is like when in one hundred years give our take an order of my ude because, I mean, I really could .

be answer yeah .

and but but the deeper answers is alan k, right? The best way to put the futures to invented this is not some exigence thing. It's going to happen if we, the AI research is able to solve the problems, we will get that exponential. And if for not, we want. In fact, my hope right now is that because the hype has got no way from reality, that we will make enough progress in the next two years to justify the hype that is there that is not, justify that the technology is, because if not, there could be a big dowsers ing in the crash.

and not which I would like to avoid. Something like this epoch versus the internet is in the internet, you know, a lot of the build up was done on debt. And the internet was an entirely new thing where a lot of this wave is actually funded by large companies that have hundreds of billion of dollars on the baLance sheet. And so it's quite unlikely we will have the same type of crash that we did back then, but we certainly could have another winter.

This is the thought that recurs, in my mind, is like, if there was a god, right? I personally am a native like you, like i'm going to do the internet boom so these companies can have hundreds of billion dollars in the bank.

I was great place of petra. Thank you so much for joining.

Thank you. This is fun.

There you have IT, I told you was insight ful, and I think I delivered on my promise. If you enjoy the episode or enjoy the podcast all, please do rate us and recommend IT far and wide.