Mustafa Suleyman (on artificial intelligence)

Welcome, welcome, welcome to Armchair Expert, Experts on Expert. I'm Dan Riggle, and I'm joined by Mrs. Mouse.

We had an armchair, this is a Easter egg, we had an Armchair, Anonymous. yesterday, we recorded one, and someone said, it's Dan and Lily.

Yes, Daniel and Lily, maybe?

Yeah, it was really cute.

Today we have Mustafa Suleyman, and this is, we've been wanting really, really bad, to get a premium expert on AI.

Yeah, because it's so relevant.

And we talked so much about it, yet we haven't had too many experts on to really tell us where we're at in this whole crazy experience. Mustafa is an AI researcher and an entrepreneur. He is the co-founder of DeepMind and Inflection AI.

He's a powerhouse in this field, he's a big deal.

Yes, they have Pi as their AI version. He has a new book out called The Coming Wave, Technology, Power, and the 21st Century's Greatest Dilemma. So he's got lots of recommendations on how we keep this from devouring us, and they're all very, very interesting.

It was a really interesting conversation, and I think necessary at this point in time.

So please enjoy Mustafa Suleyman. We are supported by Squarespace. Guys, we have a Squarespace website that's just gorgeous, that Wobby Wob, you built that yourself using all the templates, yeah? I sure did. Yeah, easy peasy?

Oh, what are you saying?

I can't imagine, you don't live here.

I live in Palo Alto.

Seems on the nose.

Yeah.

Literally, I was in San Francisco at the weekend, visiting this very interesting antique shop, full of strange curiosities, like really interesting collection. Within like one second, I started chatting to the person who worked behind the counter. She was like, so where'd you live then? I was like, Palo Alto. She's like, I knew it, it's so obvious.

Oh, my God,

I love that. he said, full of curiosities. Yeah.

I think the place is called a Cottage of Curiosity. Oh. Isn't that a cool name?

That's enticing, yeah.

They should sell Altoids there, because, of course, their saying is curiously strong. Oh, it is? Yeah. Isn't it the greatest slogan for a product? Yeah, I didn't know that.

I also love brands that have never changed. And I feel like Altoids, it's the same tin as it was in the 50s. Not that I was around, but I'm like, it feels like it was.

It has a retro vibe.

Unless it is just doing the retro thing, and I'm being tricked. It might be. And that would be really embarrassing.

Certainly possible, but you're right. It feels fun to consume an institution.

Yeah, be part of the history with every mouthful. Look at this, it's an advert for Altoids. Yeah.

I'll go further. I mean, chips are always a stretch. But there's even, like food products, where it's like, no, no, this is an institution. Like vanilla wafers. Yes.

I do not. One day, you guys do?

I have two. Monica has frozen eggs.

I have a couple eggs that are frozen.

That's where we're at.

No kids for me yet.

You have a, Jesus, I'm a mess.

Throw it all on the floor, dude. I mean, literally, just toss it all out the window. Yeah.

And I've just realized I have a little bit of Debris on my phone and I hate that. Oh. You know, like one of my mints. Give it to me, I'll wipe it. Well, no, when you don't need a waste of stuff, it's time to wipe it.

Well, I'll just do it on my shirt.

I think the quicker you get it scratched, the more accustomed you get to it just being like junky. And then it's like, fine. That's why I don't do cases. I'm like, screw it.

Oh, wow.

Yeah, it's fine. I mean, I've got quite a lot of scratches, so don't get overexcited, but, like I, just learned to live with them.

How frequently do you shatter a screen?

Once every two years. I tend to upgrade every other generation. Very good. Yeah, it's not too bad, actually. I'm good at catching.

They've gotten sturdier over time.

For sure. They have. I have kids grabbing them and stuff. I need a case. I hate it.

When you first get it, it looks so beautiful.

Right, sleek and slippery. Yeah. Could find its way into that shop of curiosity. One day.

One day. Most def.

Okay, so you're from England, London, England. Dad was a Syrian taxi cab driver?

He was a mini cab driver, yeah. There are two kinds of taxis in the UK. There's the black cab, which is the profession. Yeah. You have to do the knowledge.

Yeah.

Really appropriate for this conversation, because here's a task that has already been outsourced to the phone, yet they are still doing it that way. And I don't know what we say about that. That's just very rare.

It's a ding, ding, ding.

It's one of those things where the rules matter and it affects the pace of change.

Right.

That's, I think, a sign of hope, because I think we should be collectively making decisions about how quickly we want new technologies to be introduced, so that we can decide on the rate of change.

We accidentally, or maybe you engineered this, we stumbled upon almost the greatest example of some legislation or policy that has prevented technology taking over a space. Right. And it seems to be working. Yeah, we were in London not too long ago in one of those cabs and it worked out just fine. They're brilliant.

They do have the advantage, though, of being able to use the bus lanes. So they can get there faster than a regular car. So I think that is a structural advantage, which justifies the higher price, because they're also quite a bit more expensive than Uber.

They are. But they have adopted some stuff. You can order them with your phone now. Yeah, so that's an improvement. Now, I guess what I interpreted by reading that your father was Syrian, I assumed in Syria.

No, so he moved to London in the early 80s and settled there. And so I was born and raised in London.

84?.

That's right. Yeah, exactly.

Painfully younger than me.

It's okay.

Still older than me, so we're doing pretty good.

He's made more money and he's more old. Okay, well. It's hurtful, but anyways. So, but in London, as an occupation, he did drive.

He drove initially one of those unlicensed minicabs.

They don't have a medallion. You negotiate a price with the person, there's no meter.

Exactly. You negotiate on the spot. Is it 15 pounds, 20, pounds, 25?? That's too much, a bit less. That has gone out of fashion now, whatever.

We all love story, and for me, that's a great story that his son would end up being a pioneer in AI. I think that feels like about as good of an immigrant story as you can get for your child.

Right, it's funny because my mom, who was a nurse in the National Health Service, she was kind of adamant that I would drop out of school at 16 and get a trade, become a plumber, become a carpenter. Everyone's always gonna need an electrician. That was just like long-term, reliable. Obviously, I didn't do that, but it was definitely that kind of making money quick is important. Yeah, exactly.

She's English, though?

She's English. How did he woo her?

That must have been hard in the 80s.

That is actually a funny story in a way. So my mom was riding around the world on a double-decker red bus in 1982, I think it was, and she was at a rest stop in Afghanistan repairing their tire with this crew of basically travelers. And in the washrooms at the rest stop, my dad apparently was having a shower next door. Oh, my goodness. Oh, wow.

He's taking a shower at the gas station. She's in a double-decker bus, it's all impossible. It's like a Dr. Seuss story.

It's completely crazy. So he comes out in his towel. You know, they have a little exchange.

He's attractive, I'm guessing.

I have no idea. Okay.

He must, it's sounding like he is.

Well, you're being handsome as hell,

so I'm imagining he was a looker.

So then he's standing there in his towel, and of course, they don't have a common language. He doesn't speak any English, he doesn't speak any Arabic. They have a bit of like schoolboy, schoolgirl, French. And so they have this connection, nice little chat. They go on their way.

Is he going over the Kaibar Pass? Is it like that historic?

I don't know if that was the route, but it certainly feels like that in my head.

Yeah, for sure.

Then, really strangely, 10 days later, they end up encountering each other again, completely randomly, this time in Iran. What?

No. This is destiny.

This is so romantic.

It's totally wild.

Very horny.

And obviously that was game over. From then on, it was just like a slam dunk, at least for a few years.

He emigrated to England to be with her.

Correct.

Oh, wow. Did he get his engineering degree?

You know what? He did not finish it because they became pregnant with me in Pakistan. They didn't want to have me in Pakistan, although they did think about it in Islamabad. And so then they came back to the UK and then obviously he was sort of very upset because in the UK they didn't recognize half of his Pakistani engineering degree. And he had actually left Syria to avoid conscription, because in those days, everyone at the age of 18 has to go into the army for three years.

And in the 80s, Syria was getting into it. It's not like you wouldn't be realistic to assume that could be life-threatening.

That was a rough experience. I have uncles who went through it, and it's not the kind of touchy-feely, friendly, turn you into a servant of the national effort. It's very brutal. And so he basically wanted to avoid that.

Okay. We'd call that a draft dodger here.

Exactly.

Well, at one point in time, I think nowadays we'd say it's smart.

Well, I think we all know where I stand on that. I'm a big, bleeding liberal, so I'm, you know. Or for it. I don't think you should go kill people you don't want to kill. You have a very interesting and circuitous route to this position you hold as an authority in AI, because you drop out of college, I guess.

I got to Oxford, where I was actually studying philosophy. I think the biggest transformation of my life happened where I basically discovered human rights principles.

In a philosophy class? Yes.

In the kind of spirit of universal justice and fairness, rather than thinking that me and my people were the righteous ones, the chosen ones, the special elite, and so on. And throughout my late teens, I had kind of struggled with lots of parts of the religion. Both my parents were very strict. Very strangely, my mom was actually already a Muslim. She converted before meeting my dad.

Hence the crazy bus trip, probably.

Yeah, going to find herself, I think. I'd grown up with a very strict sense of the religion and it was starting to get uncomfortable. When I got to Oxford, it kind of provided me with a framework for thinking about universal rights rather than just one group. That you didn't need to appeal to faith. You could use reason to establish the fairness of things.

Right, which probably had to be very rare in that community.

I mean, it was the first of its kind, because all the other ones-.

They would drive you to- Exactly.

You shouldn't have sex before marriage, and homosexuality is a sin.

You observe the five pillars today.

Did your parents freak out? Did they know you were doing this?

I don't think my dad would have approved. I had left home by that point when I was 15, so they weren't so much in the picture. It wasn't such a big deal. That was actually very freeing in itself, because I could sort of go and make my own way, figure out my own ideas.

Because you had gone to a boarding school at some point as well.

Actually, I went to a state school, a free education, but it was a very good one. You had to do an exam to get in. And I think, because it was selective, you had to pass two exams and an interview, it just meant that everyone was starting from a little bit further ahead. It changed my life. The school really gave me a huge advantage in life and just to be surrounded by a lot of other really driven people and really smart people.

Pressure. testing your thoughts and ideas around adversaries that are worthy, they will make you raise your argument. They'll make you better.

Yeah, and that's actually what I found at Oxford as well, is that there was no judgment for being a bit obsessive or a bit nerdy or really overpassionate about stuff. Yeah, yeah. Whereas earlier in life, that was a bit more tricky.

The Helpline experience leads seamlessly into some more broader public health and public service work. You end up working for the mayor of London. You end up advising on conflict resolution, all kinds of things. You have clients like the United Nations. That seems like an easy to follow trajectory.

I definitely don't have a PhD, no.

All right.

Let alone computer science. I got lucky in that I started the kind of work thing very early. So, because I left at 19,, I managed to cram in a lot of experience, working in charities and local government, working as a conflict resolution facilitator all around the world. That gave me huge exposure to lots of different types of work and types of theories of change. I was obsessed with what is your theory of change?

Incredible instrument of nothing else.

It was very obvious to me that it was so much more than a platform. And the word platform was actually a misrepresentation of what was actually going on there. Because it's much more of a mode for framing how activity takes place. The incentives, the structure, the way the website is laid out, the colors, the very, all famous now thumbs up like button, which drives engagement.

The original SIM.

Yes.

Yeah, in hindsight, that seems super obvious. And to me at the time, it was like, wow, this is really incredible. I need to do everything I can to be a participant in technology. Technology is gonna be the thing that helps make us smarter, more efficient, more productive, essentially, do more with less.

It must have been appealing as well. If you're dealing with governments and institutions in these negotiations, to see something that comes with no history has to be very encouraging. Like you're not bridled with 300 years of how we do things.

That's a great point. It's actually much easier to innovate in greenfields than it is to change the status quo. I kind of think that's one of the big problems that we have in the world today, where there aren't very many compelling, positive narratives of the future coming from the old world order. On all sides of the spectrum, people talk about the mainstream media, they talk about mainstream finance. Their reaction is that the existing establishment is not innovating fast enough.

Yeah, because the old institutions work so slowly, they can't even correct fast enough, right? I mean, that's. part of the inherent problem is that they're not nimble. the way all these technologies are. I mean, especially the ones we're gonna get into where it's, like between 8 a.m.

Yeah, I mean, what human organization, or collection of all humans, could possibly consume that much raw content? I mean, these models are just like alien life forms, in the amount of knowledge that they can consume and reproduce.

Yeah, kind of beyond comprehension.

You should get involved with the SAG and WGA strike negotiations, because first, you know how to negotiate, or you've been involved in that, and you know everything about AI.

Believe me, I've got an earmarked question as we get to that.

Okay, I was gonna say, let's get you in there and let's also get it solved fast.

It's coming.

Because we're ready. Okay, so, specifically, how does this interest in, I wanna say, MySpace, because I never did use Facebook, but Facebook, you weren't as enamored with MySpace. apparently. That's the only place I did a lot of business. But how does that then take you to artificial intelligence specifically?

Well, so it set me on a quest to find anybody and everybody who was in my network, who was involved in technology or software of some kind. During that year, 2009 and early 2010,, I met with anyone who would give me five minutes to just tell me what was happening and why and how it worked. And I had an interest in software and technology. I mean, I was on the internet very young and was an obsessive on forums. And actually, in 2003,, my first actual business was an electronic point of sale system, where we had these little PDAs and we went in and installed network equipment for restaurants and stuff and tried to get them digitized so they could take orders really quickly and stuff.

Then you remembered, in England, the restaurants have no desire to do anything quickly.

Especially, bring you your check.

or allow you to give them the credit cards. It's really true. It's my only complaint about England in general. It's just, you wanna walk in and go like, let's pay now. Yes.

100%.

It's so true. That's what I quickly found out. This is not a problem that needs to be solved. at that time in the UK. Plus, the technology wasn't good enough.

So that's what his degree was in, was neuroscience?

His degree was in neuroscience. He was interested in memory and how the brain processed information and used memory for imagination. He wrote some really important papers there. And whilst he was at Gatsby, he met our third co-founder, who is sort of an AI guy through and through and a big believer in AGI. from day one.

Interesting.

Intelligence is about generality. And that's when we pushed this idea of AGI. And in fact, our business plan said, and this is in summer of 2010, building artificial general intelligence ethically and safely.

So did you guys coin this AGI term?

It was in use before, we didn't coin it, but I think we were the first to use it in a company as a mission. And then it got popularized shortly after, because this technique called deep learning started to work. It wasn't really until 2012, 2013, that deep learning was showing very promising signs.

Did you have imposter syndrome among these two with the PhDs? Yeah, they're scary. And did you feel like you had to prove yourself worthy of being a part of this triumvirate?

For sure. What I came to realize after the fact is that I had very different skills and that complementarity was an amazing three-way dynamic for a decade. And one of the things I am luckily able to do is to just go from the micro to the macro. Like, I'm quite good at thinking big picture and quite practical and very operational as well. And I'm very urgent.

Would you guess, or you probably know, that's a rare combination of folks, I'd imagine. I imagine most people that end up in Silicon Valley, there's not a lot of outsider, let's just say artsy social scientists above hard science, you know, whatever we would call that. I can't imagine. they find themselves in this situation very often.

I think that's exactly right. And, interestingly, between the three of us, even Demis and Shane are not strong programmers. Demis is a neuroscientist. He did computer science as his degree, but he never became a software engineer formally. And Shane was very much on the mathematics and the theoretical end of the spectrum.

This is the Wozniak, Steve Jobs.

Yeah, although Steve wasn't a coder either, right?

Right, he's the non-coder.

He's a non-coder, yeah, exactly. Steve is the big picture guy, you know, whose focus is on the user and thinks always about the product.

Right. Okay, so this company, you guys end up being pretty successful in your exploration of all this, and Google ends up buying this company.

Yeah, pretty successful is really low key. Well, and that I mean,

it's not like they have an operational GPT or PI. You don't have a product yet per se, right? You just have a lot of progress.

Right, it was a remarkable acquisition. I mean, we're talking 2014.. Most people haven't even heard the word AI, let alone AGI, right? This is out there, completely speculative.

And they haven't been snatching up European companies either.

Exactly, it was actually the first acquisition that Google made in Europe.

Right. So all the reasons. you should not assume that you're gonna get bought by Google.

Yeah, it was very unlikely, but the thing that caught Google's attention is that we made this demo of an AI learning to play the Atari games from scratch. So you remember, like Space Invaders? Yeah. And Pong?

Asteroid.

Right, Breakout, if you ever played that game, where you have a paddle at the bottom, exactly. We basically had the AI just from the pixels, right? So we didn't give it any rules or any structure, nothing else. Just interact with this environment, get a score. So it tells you if you randomly, luckily managed to hit it or if you lose.

And really quickly, does it have to play the game in real time, or can it accelerate the game itself?

That's a great question. And this is a key insight about computers, right? They're paralyzable. So they can go at lightning speed. You can scale them up.

Okay, good. So it's playing a whole match in seconds or minutes, as opposed to it would take us 15 minutes to get to that board or whatever.

Right, that's a key intuition, because machines can have more experiences than any single human. And we see this pattern repeat over and over again. In 2015,, we trained an AI to detect breast cancer in mammograms. And by 2016, it was better than the expert radiologists. We did the same thing for ophthalmology.

Well, seen and, I think, even more importantly, remember at all times. I mean, that's the real key. It's like, if you actually could remember the sequence of events, of everything in your life, you too would be able to predict a much higher degree of accuracy. But it's like, we don't have access to that.

Totally. So, the fact that it can see things much, much faster, the fact that it has perfect memory in many cases, or very, very good memory, and the fact that it doesn't get tired.

Yeah, because if it sees a slide of an iris, let's just say, I don't really know how it works, and it has some little impigmentation here, it can rapidly go through all 10 million photos in the dataset with known prognosis. The human doctor can't hold it up next to 10 million examples, but it can. That's the magic of it, right?

Right. And that's actually what has happened for the last decade, as we see these crazy exponential trends of now seeing trillions and trillions of words of open data. But at the time, it couldn't work in the space of words. It was really just looking at images and pixels. When we saw it play these Atari games, it was really incredible.

Yeah, they can shatter paradigms, right? Because we all get stuck in thought paradigms. Medicine works this way. Diagnose, treat, whatever paradigm it is, the thing outside of it is so hard for us, but for a machine, it's not.

And it can actually be creative. So this is the cool thing, is that it can discover new knowledge. It can see outside of the box and it can be genuinely inventive. Everyone's probably seen these image generation models these days, like Dali and other ones.

This is crazy.

It's wild. Probably more than you can imagine can now be manifested. I think that's all creativity is. It's just combining multiple different ideas in novel ways to produce something unique.

Okay, that kind of fast forwards us to one point about AI. When it's explained to us, I think we discredit the fact that we too operate almost in an identical way. So a current argument right now, and this is dangerous for me to say, because I'm in the union and I am supportive of the WGA,

and SAG.

And SAG. The example I'm giving is about writing. Joseph Gordon-Levitt, wonderful actor, really smart guy, wrote this incredibly well thought out piece about this issue and said that the owners of the content that are fed to the AI, let's say, if it's to write a script and you feed it 600 scripts from brilliant writers that have worked in the past, that there should be a royalty paid on that original work, which is a very sound argument, but I do have to point out, I, as an artist, I too have been trying to replicate all my favorite movies. When I sit down to write a movie, I'm informed by Michael Mann in Heat, I'm informed by Pulp Fiction's my favorite movie, I like this, I like that. I am an AI.

So, first of all, the strength in the argument, which I broadly agree with, is provided the AI isn't regurgitating word for word, the paragraph of copyrighted text, that would be plagiarizing, right? But it seems that most of the time, these models are not regurgitating that word for word. They're really inventing something new. They're finding the space between two ideas that this interpolation is called. That seems to me just like anything that we would normally do as humans.

I'll be more literal. I pitched this movie to Warner Brothers. I said I wanna remake the TV show Chips, but I wanna do it. Chips meets Lethal Weapon. So it's hard, R, the action's intense, and it has comedy.

But didn't you have to pay some sort of royalty to original Chips?

Well, Chips, because we're using the actual intellectual property in the name. But I could have also gone in and said, I got this movie idea, it's called Bike Cops. Imagine Chips plus Lethal Weapon. And most pitches are sold that way. So it's Midnight Run meets blank.

And it's also exactly the same thing with apps. The number of times you hear a startup founder being like, I'm gonna do Uber for food delivery, or I'm gonna do it for this, I'm gonna do it for that. This is how we create. Yeah. Yes.

Stay tuned for more Armchair Expert, if you dare.

Yeah, I like that. Have you heard? the one, comparison is the thief of joy?

Yes.

That's a good one too. It is.

If you wanna give therapy a chance, start with BetterHelp. It's entirely online and designed to be convenient. To get started, all you have to do is fill out a brief questionnaire and you'll be matched with a licensed therapist. So stop comparing and start focusing with BetterHelp. Visit betterhelp.com slash dax, today to get 10% off your first month.

I think it's a little bit of both. I think I've been nervous about it since the day we founded the company. That's the only honest and wise approach. We should approach technologies that are as fundamental as intelligence itself. We've literally just been describing something that if you swapped out the human for the machine and you replayed this conversation, you couldn't really tell, were we talking about the machine now or were we talking about the human?

You can put a kid in front of an Atari game and he will master it. We've seen it happen. And he does it through remembering what things worked and what didn't.

So we're already at a place where we're taking the thing that has made us unique as a species, our intelligence, this ability to plan and imagine, create, adapt, and invent new things, communicate perfectly in language. Language is a technology. We now have another type of input to that technology, which is the machine, able to use language. So it's always been top of mind for me. I mean, it's why we framed the company around ethics and safety from day one.

But as we'll get into, even a well-intentioned, seemingly bulletproof statement like that can be hacked by AI. And we'll get into how, like, you, couldn't prevent the machine from getting racist. There's all these things you can't account for. Then that's in your book, The Coming Wave. There's like a lot of almost impossible to predict things.

Right, and that has been the story of the last decade. The progress has been eye-watering. It was 2014 when we were acquired by Google and we became Google DeepMind, part of the Google ecosystem. And in that year, when we trained the Atari model, that used something like two petaflops of computation. So two billion million computations, which sounds like a lot, but it's actually relatively small.

Two billion million.

Two million million.

That's like when a little kid is like, it's a billion million.

Yeah, yeah, yeah, exactly.

Like a made up.

A thousand billion, a hundred.

It's totally a made up number. It's called a petaflop. So you can go and, like, you know, drop that on someone. Every year. since then, the cutting edge models in AI have used 10 times more compute than the previous year.

Oh, my God.

So to your question about, am I surprised? Am I amazed? I mean, it's totally mind blowing.

Well, 10 to the 10th powers, I don't even know what that number is. It's kabillions. Exactly.

Yeah, yeah, yeah. That's a good number. Yeah, yeah, yeah. It's called 10 billion, million billion flops. So it's gone from two to 10 billion.

Yeah, this is not real. 10 billion million billion flop billions.

Some insane numbers.

You can't even graph that.

There's definitely not enough room on the page to list all the zeros. Scale has been the thing that has transformed this. The models are getting bigger. They're consuming more data. They consume insane amounts of computation.

Which we have trillions.

We have about a hundred trillion. Hundred trillion. Connections in the brain. And so these models at the moment are roughly on the order of about one trillion. Okay.

Which would just be a year or two. Yeah.

Could be a couple of years. That's right.

To at least match the amount of neurons in the brain.

Just to be clear, that's not gonna mean a human-like performance. Right. But it's just a crude, rough measure.

And really early into this acquisition, you started using the DeepMind technology to address energy consumption for cooling Google's data centers. And the AI itself looked at this and figured out how to reduce it by 40% for cooling.

Yeah, that was a crazy project. I mean, the Google data center. infrastructure spends tens of billions of dollars a year.

I was gonna say, I have to imagine the very best engineers on planet Earth designed the Google cooling center.

Literally.

Whoever, the very best person in thermodynamics, or whatever, I'm sure they have, he or her.

Absolutely. And they were very resistant to cooperating with us. at the beginning. They were like, we're the best systems engineers, mechanical engineers, industrial engineers on the planet. So it took us about six months to cozy up to them and persuade them to give us a shot at doing this.

Does it turn on at this temperature or that temperature?

Exactly. What combination of fans to use? Cause. there's like a motherboard fan, there's a chassis fan, there's corridor fans. And these are like the size of three or four football pitches, these data centers, right?

And they're usually underground?

They can be overground, but they're almost always near hydroelectric power or solar power. Cheap energy is the number one priority.

Right. Wow.

Three footballs a year. I know, it's insane.

I mean, it takes you like 20 minutes, half an hour to walk down the thing and you have to wear headphones cause it's like screechingly loud. Really? And, of course, freezing.

So, in looking at all this data, how long did it take the AI to figure out a system that would reduce it by 40%?

Well, it took us about a year of tinkering and experimentation because obviously we got it wrong a lot. And in the first deployment, it got about a 30% reduction. So we were pretty blown away by that. And then what we were doing, because obviously they didn't want to let the AI loose on the actual hardware.

Of course.

The AI would basically give a set of recommendations to the human data center controller. And you guys would implement that? Yeah. And then the human would be like, yeah, that seems sensible. I'll adjust it, this level, that level.

We have limits of conscientiousness. I mean, I don't know why, but just recently I've been thinking, like, you know, it's kind of the number one thing you would want in an employee is conscientiousness. The computer's just the most conscientious thing in the world. It can't deviate from whatever it's supposed to do, right?

I think this is another reason for us to be really optimistic about the future, because we want the machines to produce a consistently fair and just outcome across the board. Obviously, the challenge is that we have to make sure we program them in that way and that we can hold them accountable, keep them controlled. So there's that challenge. But in theory, it should mean that, unlike a judge that gets tired after lunch.

Sentences you before or after lunch. Right. Yeah.

There's a huge swing.

It's like three years difference, or something crazy. You see it all the time. We're guilty of it individually. And also we walk into a place and we know they're supposed to do that, but they don't want to and they don't do it.

Emotions.

Yeah, emotions and energy levels and distractions and personal problems.

You think about it. if you ever visit a relative in a hospital or something and you see all the different people on different parts of the ward and the people who are quiet and not very pushy and demanding of nurse practitioner care. You have to think how uneven that treatment is as a result of somebody having a really active patient advocate, member of their family, to help manage their care and stuff. And that is gonna end up in worse care for some people who really deserve better. And so that's why I'm excited about the kind of fairness side of AI.

Yeah. I will add that you left Google in 2022.. People will be mad if I don't mention you had allegations of bullying people and you left DeepMind. You're right. You bullied people on email, maybe?

I can be super hard charging and very demanding. It was five years ago and I've learned a lot from that.

Okay, wonderful. Moving on. So you leave DeepMind and you go to Google proper, and then you leave there in 2022, and then you found, with Reid Hoffman, who we've had on and we adore Reid, Inflection AI. And in 2023, you guys introduced Pi, which is a chat bot. Now I'd love to get into some of the things that your book is warning us about.

Yeah, I mean, I think this is gonna be the greatest force amplifier in history. So you're right that it's gonna amplify the bad as much as it will amplify the good. And I think the sooner we just accept that that part of the equation is inevitable, the quicker we can start to adapt. because this is about adaptation and mitigation. What boundaries and constraints and guardrails can we put around the technology so that we get the most out of it, but so that it always remains accountable to us and that it doesn't end up causing more harm than good.

What's that?

Stable Diffusion is an image generation model, which is like OpenAI's DALI, but it's entirely open source. So the actual code to run that model is available to anybody on the internet to adapt it, play with it, improve it. You can actually take the entire model and run it on a laptop and it can produce photo realistic, super high quality images with just a few sentences of instruction, zero technical effort or expertise required.

Wow, my mind just immediately went to something pornographic. People are probably creating like-.

Well, that's a whole deep fakes thing, right?

Yeah, is that AI generated, those deep fakes?

Those deep fakes are AI generated.

Right, so they're probably, you know, imagining their favorite-.

They are making that, yeah.

Podcast host, probably me. You wish.

I know, that's open source.

Anyone wants to do that?

Kristen has talked about this because her face has been put on some porn stuff and it's crazy.

Ashton sent that over to us. Okay, well, I guess he is at the forefront of this tech stuff.

He's an investor in inflection, actually.

Oh, he is? Okay, okay.

He would be.

Of course.

Okay, so you can run it on your laptop. Now, when you generate, we get into kind of copyright authorship. Like, if you give it the five commands that generate this image, is that your image?

I think that the idea of ownership is gonna start to fall away.

Interesting.

And obviously, that's controversial. I'm not proposing this outcome. I'm describing what I see. Don't shoot the messenger. I believe that, as far as I can see, my best prediction is that over the next five to 10 years or so, text, video, audio, imagery, it is just gonna be produced at zero marginal cost.

And just the volume will be so vast that copyright tracking down would be impossible.

Yeah, a little bit like on YouTube. Other than for the very sort of elite, high-grade feature film stuff, there isn't really a concept of you owning your data. Like you made a video on YouTube five years ago, 10 years ago, and someone basically completely ripped it off. I mean, that is the definition of the meme space. A meme is an evolving idea manifested in images and video.

Well, and it also makes sense that people's attachment to things that took years to create are going to be different. Let's say the David. I don't know how long it took him to chisel that out of the block of marble, but presumably a long, long time. And so his attachment to it is, this represents a year or two of my life. Whereas if he was able to make 10 Davids a day, he couldn't possibly have that same attachment or that sense of ownership or the sense of anything.

Totally. I mean, we've become overly attached to the craft, whereas really the value is in the ideas and the concepts. Now, I think we're moving to a stage where the value is in the curator. It's the edit. It's the judgment to reduce, to scale back, to simplify.

Yeah, I think for people like myself, who I would say is a tradesman or a craftsman, I know how to write screenplays. I've spent a huge chunk of my life investing in that ability. And I spent a huge chunk of my life refining my acting and everything else. I think what happens with this technology is it kind of democratizes and gives everybody access to executing their ideas. And I personally feel like, bullshit, go sit in a room for five days.

It feels like it erases talent. Talent is about to become, or is becoming irrelevant.

Which is interesting, because I think, as humans and humanists, we would probably go like, well, yeah, we would want everyone to feel creative and be able to create, and we wouldn't want barriers. I don't either. But I'm saying it probably butts up against some of our ideals.

Well, you know, I think humans marvel at talent. That's why we love the Olympics, because it's like, how is it possible that someone was born like me can do that? Or someone could write this movie or direct everything all at once. How did somebody do that? When you take away the person, what's gonna happen?

I think it's a really interesting question. We're about to run the experiment. Yeah.

So we will know. We'll find out in my lifetime.

Exactly. We will find out in 10 years. My instinct is that I think we tend to feel arbitrarily precious about the thing that we've previously been invested in. Understandably, it's an emotional attachment.

It's an identity.

And that identity is a product of you having slogged away for like years, honing your craft, only to see it now available to absolutely everybody. But then the flip side of that narrative would be, well, what about all the people that didn't get the privilege and opportunity to be able to get through that training, to make those connections, to be able to make it in the industry? Think about how much arbitrary luck or even nepotism there is that enables people to get access. And so much as we're displacing people by the democratization of access to these tools, we're also giving other people the potential to be massively uplifted. I mean, think of all the YouTube stars, TikTok stars that have emerged out in the middle of nowhere, that never were on the scene or had connections.

Yeah, but there also is like a really huge global thought or question. A literal analogy I can give is that I had a year in my life where I came into more money than I thought I would ever have. I started doing all these exceptional things I always dreamt of. Maybe I'll take a helicopter to this race. A year into this experience, I was doing something spectacular again.

Yeah, there's nothing to compare one thing to another. What's good anymore? if it's all good? What is good?

Is half of what we like about something good is in fact how rare it is. Rare, exactly. And the dopamine hit of experiencing something novel. So we could end up with six million perfect movies a year in the theater, but we might find that we don't give a fuck as soon as they're all perfect.

I think that's a great point. And actually I think that's quite likely. We're already trending in that direction. We're becoming completely overwhelmed with access to information. Where are the secrets today?

Right, exactly, the Altoids. They're not better than other mints, but there's something special about that old package.

You should be definitely getting a cup from them.

I mean, they are not sponsored yet, but.

Send them in.

You can imagine their ad budget is enormous. Anywho, you do think that's already kind of happening?

I think we're becoming desensitized to the variety of content that we experience everywhere, of every form. So now that we've overwhelmed with this information, it is just hard to discern. What is cool? What is good? What is interesting?

Also, how's anyone going to make money? I don't understand. If we can produce a hundred dollies, like if I can, and then you, what's going to happen to money?

Explain that to us. Like the marketplace. How are people going to make money?

Yeah, I mean, look, hyper-competition has that tendency, because if we reduce the means of production to zero marginal cost, so the inputs become much cheaper. You don't have to rent a studio. You don't have to go hire a whole ton of extra actors. You don't have to go into post-production and do all the coloring because there's filtering. You can just take an auto-tuned voice off the library or off a shelf.

So you had originally been a part of the Artificial General Intelligence, which we talked about a little bit, but you have since introduced this term Artificial Capable Intelligence, which is kind of a midway point between AI and AGI. So tell us how you would define Artificial Capable Intelligence. Because I feel like it funnels into the question about money.

It's exactly connected to that. So for years, there's been this idea of a Turing test.

What is that?

Yeah, I never heard that. Alan Turing. Alan Turing, one of the earliest computer scientists.

Spent it at CoverBudget.

Yes, exactly. Hot guy.

I'll remember it if it has spent it at CoverBudget. We should have Brad Pitt teaching all tech.

Oh, my God, what a hack.

Everything's the Brad Pitt principle. Oh my God. Pitt principle.

I like this.

That's why Cillian Murphy was such an amazing cast. Jesus, I couldn't take my eyes off.

Who doesn't want to watch? Yeah, Oppenheimer V looks like Cillian.

Perfect looking.

The Turing test was an imitation game that Alan Turing invented or proposed 60 years ago. And it basically said, if you could teach a machine to speak as well as a human, and it could deceive another human into thinking that it was in fact human and not a machine, then it would have succeeded in imitating human intelligence. That was the grand Turing test. And now that we have these language models, Pi and ChatGPT, it's pretty clear that we're quite close to beating the Turing test. And yet we've got no idea if we're close to inventing intelligence at all.

And you said, like, you imagined, giving it 100,000 to start with.

It could start with a relatively small amount of money, and I think, quite quickly it would be able to generate a lot more.

And that would be, for you, artificial, capable intelligence. That would signal that.

It would, yeah, because artificial general intelligence is a much more long-term speculative. That's the kind of super intelligence terminator and all the rest of it. This is a much more narrow focus on not just what can an AI say, but what can an AI do, right? Because that's what we really care about.

It's funny, while you were talking, I don't know if this is a breakthrough idea or if it's really basic, but I was thinking, weirdly, I might actually say a measure of intelligence would be acting logically without data. It's almost the opposite. It's that a human without any data can make a relatively sound, good, logical decision in any given moment by modeling purely from their imagination, and perhaps no data.

Extrapolation is a critical skill that these machines currently don't have, which is, without any context, can you figure out what's required in a certain setting without any prompting?

Without any patterns to observe to then model onto it.

Exactly. The moment they're mostly reproducing known patterns. Right. And it turns out you can go really, really far just by looking at known patterns. Yeah.

But that's humans too. Normally, I think it's because you've experienced something you can at least connect to. that's similar in some way, or you've seen it in a movie, or you've read it.

I agree, but if you look at babies. Yeah. Babies have an intelligence. They're born with an intelligence. They can make decisions and they don't really have a data set to draw from.

That's true.

And I think we often are in situations with almost no comps, and we intuitively know what to do.

Well, another way of thinking about that is that we are in the millionth generation of the evolution of that mind. So, although the baby appears to learn something on the spot, that's really novel, in fact, we're all one species.

65 million years of mammalian evolution, that we're sharing lots of our genetic code with.

Right, and in some ways, these new AIs, these large language models, are only in the 10th or 20th generation. As we make progress in a certain area and we publish academic papers on it and they get peer reviewed, the knowledge and insight then gets passed on to the next group of developers and other teams. And when you see that, oh, those guys have done something really interesting, you try and copy it. And so then that gets incorporated into the models next.

You're just building endlessly. Once we have Pythagorean's theorem, it gets implemented in everything going forward.

Yeah, and that's how we evolve knowledge. We're inventing new ideas. It becomes part of the established status quo. I mean, think about how many things seemed normal 30 years ago. that are now absurd.

They endorse many different brands of cigarettes. over the years. People collect these wonderful ads for Camel, eight out of 10 doctors smoke Camel.

Oh, my God.

That's bonkers. Viruses that produce batteries.

What?

Proteins that purify contaminated water. Carbon scrubbing algae. Toxic waste into bio factories. Like these are very exciting and could heal the planet. Whatever our concern is about whether we're going to write movies or not.

We have to extract carbon from the atmosphere. Those algae or the kelp farms that you talk about, that's an incredible upside of this experimentation. Imagine being able to absorb all of this excess carbon that we really need to remove as quickly as possible. I honestly think that in the next 20 years, energy is going to become an order of magnitude cheaper than it currently is. And we are on an incredible trajectory.

Also like running core government services. You know, you imagine a government without any outside financial manipulation, no corruption whatsoever. These are pretty pleasing thoughts. And obviously those will be the upside. Now, we should address some of the things that could be very dangerous about it.

The technology itself is not gonna kill us. It's gonna be the mishandling of the technology by our own governments or by a bunch of crazies, bad actors. Now, that is a manageable downside. It is not a reason for us to panic and pull the plug. It's a reason for us to be responsible and conscious and proactive and start having an adult debate about it right now.

They're the most exciting arguments to listen to. Moderation and pragmatism isn't the sexiest.

Right, and it's actually just a much more dry and boring to sort of fumble our way through a way of just making it work, which I think is possible.

Stay tuned for more Armchair Expert, if you dare.

It's ubiquitous, which is really nice.

Yes, and you get cash back.

I also like it because you can send money with PayPal to friends.

Oh yeah.

Which means going Habsies when your BFF visits this summer, or Callie and I live in the same city, so she's not visiting, but we are going on a trip, and so we're gonna have to do a lot of splitsies.

And fire money back and forth to each other.

That's right.

With even more cash back in your pocket when you pay with PayPal, saying yes to summertime fun just got a whole lot easier. Make sure to download the PayPal app. An account with PayPal is required to send and receive money, redeem points for cash, and other options. Terms apply. We are supported by Rocket Money.

Oh, yuck.

Thankfully, I started using Rocket Money. They found a bunch of subscriptions I had forgotten all about, and then helped me cancel the ones I didn't want anymore. Rocket Money is a personal finance app that finds and cancels your unwanted subscriptions, monitors your spending, and helps lower your bills so that you can grow your savings. Rocket Money will even try to negotiate lower bills for you by up to 20%. All you have to do is submit a picture of your bill, and Rocket Money takes care of the rest.

It really is.

You might think you know what the Camry is all about. After all, it has been America's best-selling sedan for the past two decades. But the all-new Camry is different. It's a vibe. Whether you're exploring the city, running errands, or heading off on a vacation, the all-new Toyota Camry elevates the ordinary to the extraordinary.

So, containment is a theory. It's an approach that basically says all of the technologies that we make have to remain under meaningful human control. Long-term, we want to make sure that we're able to understand what we're building, that there are guardrails that are built into all of these systems by design, and that the emergent effects can be predicted or at least, to the best of our ability, accounted for, because in past waves of technology, if they come too quickly, then society doesn't really have time to adapt and update. Whereas if you look at things like the car, it's actually been a pretty incredible track record of safety on every front, from seatbelts to airbags.

Non-shattering glass.

Yeah, like there's so many small innovations and a huge amount of licensing, like driver training, and you have parking tickets, and there's all these test standards. That is a huge success story. Obviously, it's a tragedy that some people still lose their lives, but net-net, that is an incredible benefit. Likewise with flight aviation, we very early on established that there must be a black box recorder tracking everything that the pilots say, tracking all the telemetry on board the aircraft, and sharing that with a centralized body, the FAA or equivalent, that can review that information and share it with competitors where there's, like, a known weakness or a known fault.

Right, because it would benefit Boeing to have Airbus not have a solution to their air safety. As a competitor, not that they're evil and they would want that, but surely a competitive advantage. The FAA says not a chance.

Right, that's a sign of good regulation. Like there's too much regulation bashing, particularly in this country. I think that Europe's a little bit more open to it. People are just so afraid of it, and it's like, come on, let's not throw the baby out with the bathwater. It works quite often.

Look at a country that just had an earthquake but doesn't have building regulations.

Right, Turkey, for example. I mean, that was a complete catastrophe. Many, many buildings completely just sank to the ground.

Because this evil force government oversight wasn't there. Right. Okay, so containment. Now, here's my question. There is a bit of a paradoxical thought.

Okay.

Bit of Robert Downey. Oh, wow. Yes, yes, yes, yes. It hit me in the eyes there. There was a moment where I thought I was looking at Downey.

This is always part of our show.

I was like, is this the paradox? I'm an AI, I see patterns.

I'm like, oh, some of the metrics are there.

Think about it like this. An AI isn't going to function in the same way that a human does. So it's actually an anthropomorphism, like a projection of our human kind of emotional state, to assume that an AI, by default, must desire control. Now, that is just because we are a competitive species and we've lived and breathed evolutionary fight or flight for millions of years. Kill or be killed.

Sure. Why would it want us on Earth gobbling up resources?

Totally. And actually, the way that we design these models is very far from that kind of approach, right? It's a completely different species. Some people may design them to have that kind of independent control. And that is one of the things where we would need regulation to stop those kinds of activities, that kind of research.

Yeah, yeah, yeah.

It's not free to do that. You know, likewise, recursive self-improvement. If the AI can look at its own code and update its own code independently of human oversight, that's an issue.

That's, I guess, the thing I get fearful of. Mind you, I've heard Steven Pinker make a very similar argument, which I like, which is like, we're kind of trapped in our animal mindset and we are anthropomorphizing this machine, giving it these animalistic things that it just doesn't have.

But it's on us, though, to make sure we don't program it that way, because it will have the ability, but we have to not allow it to have that ability.

And what we just talked about, with everything being open source and it being very democratizing, right now, it does require a company the size of yours, the size of Google, but as the individual has access to all this, because you can hold Google accountable, you can hold Pi accountable, holding Jerry accountable in Tulsa, I don't know how we do that.

Well, it has to be illegal. I mean, it has to then be fully illegal to do that.

This is part of the, you know, definition of containment that I'm sort of trying to popularize, because I do think that the proliferation risk is the real risk here. We've evolved as a nation state system, a mechanism for holding centralized power accountable. That's what the state is. It says, you pay your taxes and in return, we'll have a monopoly over the use of violence and force and we'll use that to keep the peace. That's just the basic rules of the state.

It's all white college educated. At the end of the day, you get into who's really at the top. All of us.

And who should make those decisions. That is a problem, because I have to be regulated. Pi has to be regulated. Google has to be regulated. Microsoft, everyone else, just as much as the open source.

Open source is super in support of real regulation. Right.

I think everyone is. Everyone is like, look, this is a time to have this conversation. We want to remain as humans at the top of the food chain forevermore. We're not trying to displace ourselves.

The analogy I would use is it weirdly in my mind, feels similar to the war on drugs, in that we have only been successful at ending a single drug. And that was Quaaludes. Very popular drug here in the US in the 70s and 80s. It was like a benzo or a muscle relaxer. It chilled you out.

And I think that's the fundamental question of the next 20 or 30 years. Because, in every sense, power is getting smaller with a bigger reach. So think about those image models that we mentioned earlier. They've been trained on all of the images available on the open web, but they've been compressed to a two gigabyte thumb drive. So, in many ways, you're putting all of that knowledge and insight from the open web onto something that is moved around on a thumb drive.

Oh, yeah.

Yeah, oh.

Well, one of the 10 steps is national treaties. I wish I had more faith in these. I mean, if you just look at, like, the nuclear treaties and how willfully they were totally lied and ineffective in so many ways, and those are visible from space. You know, we can see manufacturing of this. This other stuff is like imperceptible from the outside.

Actually, nuclear gives a lot of hope. Oh, okay. We have had a lot of progress with nuclear nonproliferation. It is kind of incredible that there are only seven nuclear powers in the world today. That's true.

Wow, okay.

Same story with chemical weapons, biological weapons, if you work on conventional missiles.

But I would say biological weapons are failing. I think we just saw that. Now, look, I don't think we know for sure, but certainly even the New York Times is willing to consider with a 50% likelihood that COVID was coming from a lab in Wuhan. And we have treaties against that.

Completely. That really does make my blood boil. That's a different story, actually. That's a different failure of containment. That is the scientific research of gain of function, where, in that lab, they were actively trying to improve the transmissibility of viruses.

It's like out of a James Bond movie.

It's totally nuts, right? But obviously their goal was to try and make it more transmissible, so they could invent something that would counterattack it.

One of the things that was alluded to is that they wanted to see more progress, that Xi Jinping, this lab, was under all this pressure to show him some huge progress that they had been able to unleash and then control for his vanity, I guess.

The Wuhan lab was actually funded or had a number of postdoctoral researchers and professors that were funded by the U.S. National Institute of Health, right? So this is a much more complicated story than a Chinese effort. There are many gain of function research labs that are the same BSL-4 and 3 in the U.S. and in Europe and in Australia.