The Next Leap for Intelligence

All right, so let's talk about artificial general intelligence or AGI.

It's the point at which machines can think at the same level as humans, and there's

a lot of debate about when we'll actually get there, and you're actually one of the

few people who thinks that we've already done it.

So can you explain why?

Well, I mean, when we were growing up, we all knew what AI was.

Now 9,000 in the Star Trek computer and Rosie the robot, that was AI, right?

AI was robots you could have a conversation with, and do general sorts of tasks, and

they could talk on any subject.

If you took any of the frontier models today and you transported them back in times the

year 2000, when the distinction between artificial narrow intelligence and artificial general

intelligence was being made, I think anybody back then would have said, yeah, of course

this is AGI.

This is what we were talking about.

It confuses me a little bit when people are still debating about when artificial general

intelligence will arrive.

This is where the internet lives.

I show about the unseen world of data centers and the incredible advances they make possible.

I'm Stephanie Wong, and I'm your guide to the people and places that make up the internet.

This season, we're exploring how AI is fueling the next industrial revolution, redefining

everything from farming and health care to design and manufacturing.

And we're asking how data centers are adapting to power at all.

In today's episode, a bold look at how powerful intelligence could help create a better,

writer, future for all of us.

Throughout this season of the podcast, we've explored different ways that AI is reshaping

our world.

From the food we eat to how we fight cancers, from the way we build cars to the way we

make art.

And if you've missed any of these episodes, we encourage you to go back and give them

a listen.

But in the season finale, we're turning our eyes to the future.

And we'll be joined by one of the most adventurous and thought-provoking voices in that conversation.

I'm Blaze Aguera Yarkas.

I am a VPN fellow at Google and the CTO of Technology and Society.

Blaze is many things.

A technologist, a scientist, an engineer, and an AI philosopher.

In his recent book, What Is Intelligence, Blaze offers an exciting and controversial perspective

on machine learning.

Although many people consider machines to be only capable of artificial intelligence,

he believes they have true genuine intelligence.

And the distinction, he says, is huge.

It matters because it's very relevant to this question of what AI is or what intelligence

is.

An analogy that sometimes comes up is two diamonds.

So a cubic circonia is a fake diamond or an artificial diamond.

Whereas a lab-grown diamond is a synthetic diamond.

There's nothing fake about a lab-grown diamond.

It's the same stuff.

It's just made in a different way.

For me, artificial is a little bit of an unfortunate term for what we have today because it implies

that what we have is fake intelligence, that it's not real intelligence.

But for me, things like understanding and intelligence, which relate to the ability

to make sense of interactions with the world and with people to generalize that, to be

able to follow complex instructions, understand complex things, apply ideas outside the original

domain in which they were learned.

I think that modern AI systems very clearly do that.

For Blaze, this synthetic intelligence doesn't just have real intelligence, it evolves.

Just like human brains have evolved, and human societies.

This trajectory, he says, won't just give us faster computers, but a paradigm-breaking

shift to ubiquitous, more powerful intelligence.

Such a dramatically could require a new way of thinking about energy.

That's why, through the paradigms of intelligence team at Google, which he founded, Blaze is

working on a new research moonshot that may one day move the engine of AI to where the

power is most abundant in outer space.

But we'll save that for later.

So buckle up as we explore where AI may be headed next is going to be a wild ride.

We all grew up with an idea about how evolution works based on a fairly narrow reading of

Darwin.

It's not just not the full story, it's barely half the story.

Much of the story is also about cooperation and about things coming together to make larger

and more complex things.

When you start to look at things that way, biology and technology begin to look not only

a lot more similar, but actually like part of the same grand process.

Like a lot of our other guests this season, Blaze fell in love with computers and coding

at a young age.

I've kind of known how to program about as long as I've known how to speak English.

But he's the only one who was recruited by the military for his computer skills as a teenager.

The summer he was 14, Blaze was offered an internship at the Navy's David Taylor Research

Center in Annapolis, Maryland.

An honest first day, he was directed past rows of barbed wire fences to a giant gloomy

airplane hanger that rumbled with the sounds of airplanes taking off nearby.

I showed up in like a really janky suit and tie to my first day at this thing.

I had no idea what to expect.

His advisors didn't really know what to do with him.

So he was told to alphabetize papers in a file cabinet.

You know, classic intern stuff.

But he soon hit on something fascinating.

I ran across a paper about seasickness.

I remember it saying something about deep and goldfish getting seasick if you swash

the water around the word emesis was used, which I then looked up in a dictionary on

one of those shelves.

Oh, emesis is vomiting.

Fish vomit when they get seasick.

I interrupted my advisor asked him about this.

Oh, yeah, it's actually a major problem on ships.

Seasickness is a big issue and it costs a lot of money.

A few years earlier, the Navy had begun using a program that harnesses a ship's rudders

to limit the amount it rolls.

It's called rudder roll stabilization.

But Blaze thought he could modify the software to limit seasickness.

And he was right.

And at the end of the summer, I got sent on a business trip to install that new software

on a couple of aircraft carriers in the Pacific.

So it was my first business trip that I arrived with my really bad suit and two suitcases

full of floppy discs.

OK, I love that story.

So what were your big takeaways from that experience?

Is it representative of how you have tackled other big questions since then?

I have almost always used my skill at programming as a kind of Swiss Army knife to attack many

different kinds of problems.

At a more meta level, the idea of taking intersections of different kinds of concerns,

different kinds of problems.

You know, let's cross A with B where these might be quite different fields or quite

different ideas that are not sort of in the mainstream or not already understood to be

connected.

It has always been a theme as well.

Even if you don't do the cleverest thing at that intersection, being at a new intersection

first is pretty special.

Blaze's passion for finding those unexpected intersections played out again when he was

studying neuroscience and university.

Neuroscience and the human brain, he realized, had a lot in common with his first love,

computers.

And it's not just an analogy or a metaphor to think about the brain as computational.

It was very natural that those interests came together.

That idea of the brain operating like a computer became foundational to Blaze.

And it's at the scene not just for his work, but for his entire philosophical approach

to intelligence.

In 2011, Google's Jeff Dean and Greg Carrado, along with Stanford University Professor

Andrew Ng, launched a part-time research project called Google Brain.

And their top goal was to use large-scale deep learning and neural networks to improve

Google's products.

They quickly made strides in a ton of fields, from object and speech recognition to translation.

Google employees stole the New York Times, they never taught it what a cat was.

But the brain was able to create an image of a cat.

In 2013, Blaze joined Google Research.

There was this revolution happening at Google, and that was obvious, and I knew this would

be a very exciting place to come and be a part of that.

It was kind of what I'd always dreamed about.

In the following years, AI made huge strides.

In 2017, Google introduced its breakthrough transformer architecture, a radical new way

for computers to process information that opened the door to generative models like GPT

and Gemini.

For Blaze, LLM's Gemini actually reflected his ideas of human biology.

There's been an idea for a long time that the function of the brain is to predict the

future.

You have a brain because when you live in a complex world, it's very helpful to be able

to know what's coming next and to know how your own actions will affect.

Those are very useful sorts of things to be able to do for our survival and for us to thrive.

When we say, oh, all an LLM is doing is predicting the next token, from that perspective, we

kind of are just next token predictors.

So can you explain why this connection between LLM's and human biology matters for the

average person using Gemini?

Because it feels like there's this parallel between how humans interact with an LLM and

how we interact with each other.

On a very practical level, it matters because the ability to have a successful interaction

with another entity that is also modeling you back, that relies on, you know, I hesitate

to use the word empathy, but at a minimum theory of mind.

My advice to you is, you know, give it a persona, you know, say like, you know, I want

you to pretend that you are, or be, right, a, you know, a physicist who is working with

me as a business person who, you know, but who's very good at communication and also can,

you know, explain things to students and so on.

And if you do that, right, and you're both sort of, you know, in your roles and thinking

about each other as people, you will have a much more successful interaction.

Because if we're not trying to get inside each other's heads and inside our own heads,

then it becomes very difficult for us to do division of labor, to cooperate, to win non-zero

some games together, and to solve problems bigger than can fit in one brain, if that makes

sense.

I want to double back on this idea of solving problems that are too big to fit in one

human brain, because that idea of humans working with each other to solve problems actually

mirrors a huge advance in computing.

I'm talking about parallel processing.

Around 2006, engineers realized that they couldn't make computer processors go much faster.

The solution was to make computers able to handle multiple operations at once, in parallel.

And that allowed them to scale dramatically.

There are more operations happening at the same time, rather than just a faster clock

speed, which allows more computations to happen in series, you know, in one second.

And that blaze points out is pretty similar to how our brains work.

I don't think that's a coincidence.

We have 86 billion neurons in our heads, in our brains, and they're all working at once.

None of them is particularly fast, but you know, it's an impressive thing to think about

a 100 billion core processor.

So how does parallel processing tie into your belief that machine learning will evolve,

like human societies have evolved?

More is more, as they say.

We've seen this kind of scaling up of parallelism in nature, with, you know, brains getting bigger,

our brains are much bigger than those of our primate cousins.

And the explosion of human population has resulted in a lot of parallelism, too, in urbanization,

you know, many people working in parallel.

And cities, of course, are much smarter than villages, in the sense that, you know, that

they can develop much more advanced technologies, and so on.

I fully expect that that big trend of increasing parallelism, increasing computation, and increasing

goodness for lack of a better work, you know, that all of the benefits that we have in modern

society come from that, from that intelligence.

How do you see that scaling of intelligence helping humanity?

The bottom line is that intelligence is not a bad thing.

It's a good thing, solving the climate crisis, solving our political problems, et cetera.

I think all of these things require more intelligence, not less.

So do you see AI as like the next industrial revolution?

I mean, there are obviously lots of comparisons to other general-purpose technologies,

like the Z-mengin, the electric motor, et cetera.

Yeah, I do think that there is an analogy to be made there, for sure.

I think that it's equally momentous.

It's also a symbiogenetic moment, meaning a moment when we are about to become something

greater than the sum of the parts and co-dependent in certain ways.

In the industrial revolution, we developed the ability to do combustion, to have a metabolism.

If you like, outside our own bodies.

And within a short time, we went from the main metabolism of society being in the bodies of people

and draft animals to being in engines.

And that's what allowed our population to explode from 1 billion to 8 billion.

Seven-eighths of us would not be here if we didn't have that energy source.

Essentially, the industrial revolution meant the ubiquity of energy.

It meant that energy stopped being the limiter for the growth of humanity.

The internet did that with information, you know, perhaps to a fault, right?

I mean, information used to be scarce, now it's abundant.

So, yeah, industrial revolution, abundance of energy,

internet abundance of information, AI, the abundance of intelligence.

The future that Blaze imagines won't be filled with just abundant intelligence,

but ubiquitous intelligence.

And that trend, he says, has already started.

Now, every time you get in your car, there are 100 cores in there with you,

not even counting what's in your smartphone or in your laptop or something.

And all of that means the ubiquity of computation.

And again, it's in a way not so new.

It's not like humanity today is an intelligence that is about individual human brains.

I mean, none of us individually know how to put a person on the moon or how to transplant an organ.

These are phenomena that are only possible because our collective intelligence is so much bigger

than our individual intelligence.

So, my point is our collective intelligence will be really, really big relative to what it is today.

What that will make possible is, I think, hard for us to imagine at this point, but it'll be a lot.

But here's the thing.

The collective intelligence needed to solve the world's biggest challenges

from curing diseases to managing climate change is going to require new ways of thinking about how to power it.

I think our thirst, if you like, for intelligence is unlimited.

In nature, brains get bigger until they can't get any bigger.

Cities grow until they kind of can't grow anymore for very psychological reasons.

So, even assuming that we gain another factor of a thousand in efficiency,

we have to look at where the energy comes from.

And that's where Project Suncatcher comes in.

And when you start to think about it that way,

the answer is obvious.

Space is where there is a huge, huge amount of energy available.

I'm sure that terrestrial energy innovations that are already well underway are going to keep going.

We're going to have much more solar and renewables,

wind, nuclear power, hopefully we'll be making a comeback.

Maybe fusion will come eventually, but space is much bigger than all of those.

Just the amount of energy that you can use, let alone generate,

is many orders of magnitude larger in space.

So, for a general audience, can you explain how Project Suncatcher would take advantage of

sunlight to one day power data centers in space?

Take more or less conventional satellites, put TPUs in them, which are our current generation

chips that do all of Google's AI, and put them plus solar panels in the satellite and launch them.

About one part in 10th to 10th of the energy that the sun emits hits the Earth,

and the total primary power production of humans is a small fraction of that of the solar energy

hitting the Earth. The general theme there is that you want to make your space-based AI as

two-dimensional as possible, as flat as possible. You need an area in order to gather sunlight,

and you need an area on the back to radiate your heat, but any volume, any third dimension,

is just more mass that you have to launch. So, how two-dimensional can you make it,

is kind of the name of the game. The other big engineering challenge, by the way, is communications.

So, you know, communications within a data center involve lots of fiber optics and a very,

very high bandwidth. In space, the right way to do this is probably with free space optical links.

In other words, you're communicating by laser within a fleet of satellites, and that'll require

the development of new kinds of optical communications that are still very experimental in the North.

Blaise acknowledged that there are still some big questions to answer before the project can take off.

It's, of course, you know, easier said than done. There are at least a couple of major

engineering challenges there. One of them is heat. The only way to dissipate heat in space is

radiatively through infrared coming off the object, and so you have to have large radiating surfaces

to radiate the heat and to spread it from these very hot TPUs. And you have to have very large

scale structures. But, you know, I fully expect that much larger scale structures are going to be

in space for doing AI in the coming years. Project Suncatcher might sound super futuristic,

but according to Google CEO Sundar Pachai, the team is working towards its next milestone in early

2027. I think we are taking our first step in 27. We'll send tiny, tiny racks of

machines and have them in satellites test them out, and then start scaling from there. But there's

no doubt to me that a decade or so away will be viewing it as a more normal way to build data centers.

This has just been a mind-blowing conversation. But before we wrap up, I didn't want to ask you

what you might say to someone who has concerns about AI in the future.

I think a lot of concerns that we have about AI are grounded. I don't think any of these things

should be just dismissed out of hand. I don't think that our political system and our economic system

are necessarily what they need to be in order to take into account or manage, you know, this world

we're going to be in very shortly where, you know, we have vastly powerful intelligences that

are part of your extended humanity and that can do all sorts of intellectual work. But also

that growth and intelligence has underwritten everything good about human life today.

And can you tell me a little bit more about that? Like, why do you think we should be optimistic?

Because so far, the development of greater and greater intelligence has led to greater and

greater freedoms and possibilities and quality of life for people. I feel like we've already

reaped so many benefits from all of this. We're now really down on progress in ways that I think

are blind to the long term of what we have actually seen happen. You don't need to zoom out

very far before you start to see that the big picture is a pretty positive one.

Blaze Aguera Iyarkas is a vice president and fellow at Google and the CTO of Technology and Society.

He's also the founder of the paradigms of intelligence team. That's it for season 5 of where

the internet lives. Thank you so much for coming along with us on this journey. We've learned

so much about the ways AI is transforming our world and we can't wait to see what the future holds.

Where the internet lives is produced by latitude media in collaboration with Google.

You can subscribe to the show anywhere you access your podcasts and please give us a rating if

you have enjoyed our journey together and if you want to learn more about how Google's data

centers are benefiting communities around the world, click the link in the show notes. I'm Stephanie

Wong. Thanks for listening.