- We have to put our hands back on the steering wheel, and we have to start getting serious about how do we put guardrails on the system.

Because what we know is that if you unleash Silicon Valley, they will burn the place to the ground to get to the finish line first, and we've got to start putting guardrails in place.

And the thing is, is that this is a borderless problem.

This can't be an only US or an only EU or an only- - That's a human problem.

- We have got to start thinking about this globally, and I don't think we are doing that very well.

(upbeat music) - Hello, everybody.

I sat down with Hany Farid, one of the leading voices in AI research.

He's a professor at UC Berkeley, where he works on digital forensics and misinformation, especially things like deepfakes, image analysis, and how we perceive fake content.

He's also the chief science officer at GetReal Labs, a company that focuses on the authentication of digital media, telling us what's real out there and what's fake.

So, yeah, he's busy, but he made time for us.

This isn't your standard AI interview because I know you all have heard too much about AI and you're tired of that same old conversation.

This time, we talk about how AI actually works and how he can detect the fake stuff, and, of course, what we all want to know, what he thinks the future holds.

If you enjoy the show, I'd love your help spreading the word.

So take a moment to rate, review, or leave a comment, and don't forget to subscribe so you don't miss anything.

Your support really helps us reach new audiences.

So, let's go.

Hany, welcome to "Particles of Thought."

- It's great to be here, Hakeem.

- All right, man.

So, look, this isn't going to be like your normal interview because this first question I got for you is something that the people need to know.

- Good.

- All right.

- I can tell you, I'm a little nervous already, but go ahead.

- All right, so listen.

You're an expert on AI and you specialize in identifying deepfakes.

So there's been three occurrences in recent history that has everybody in a tizzy.

The first one was the movie "The Matrix."

The second one is when physicists came up with their holographic theory, which seems to indicate that life could be a simulation.

All right?

And the third one now is AI.

And the question everyone has is, "Is reality a deepfake?"

And since you are an expert in uncovering deepfakes, are we code, man?

Is reality what we think it is?

And if it is a deepfake, would a deepfake expert have an idea of how to uncover that?

- Yeah.

I mean, if this was all a simulation, my job would be a lot easier, honestly, because then you sort of give up, right?

There is no more reality anymore.

I think that this is sort of where we are, Hakeem, is we are now questioning everything, not just what I see on social media, but our existence, we are starting to question, and our existence today and yesterday and in the future.

And it does feel, to me, I've been thinking about these problems for 25 years, that we went from...

If I was having a podcast with you 25 years ago, your questions would've been of the form, "Hey, let's talk about Photoshop and how people splice together two images, and maybe we would talk about the Lee Harvey Oswald photo or the moon landing photo."

And today, a perfectly reasonable question is, "Let's talk about the nature of reality," and that's happened in 25 years.

Imagine the next 25 years, the kind of conversations we're going to be having.

So to get back to your question, I don't know.

I honestly don't know, and I also don't know where this latest AI boom is taking us.

I don't think anybody knows, honestly.

But here's what I can tell you, if you look at the personal computer revolution, that took roughly 50 years to unfold, which at the time felt really fast- - Where do you start from like 1945?

- Let's start in 1950, 1945, and let's go to about 2000 when more than half of US homes had a personal computer.

Then you look at the internet revolution, right, from the beginning of the HTML protocol, Tim Berners-Lee, to about half the world's population being online.

That was 25 years.

The mobile revolution was less than 10 years.

Well, the AI revolution of course started in 1950, but the one we are experiencing now was about two to three years.

- Wow.

- We have gone from 0 to 100 miles an hour where we are now talking about existential threats of AI, we're talking about 50% of jobs being eliminated in the next five years, we are talking about general artificial intelligence, we're talking about "The Terminator," and we would not have had this conversation five years ago.

And, by the way, on top of all of that, we don't even know what's real anymore because we consume all of our content from online sources.

Online sources have been polluted for a while.

They're getting more polluted, thanks to AI.

And suddenly, our whole notion of reality is up in the air.

It's unsettling.

I think people feel unanchored.

I don't know how to help everybody with that.

- Yeah.

All right, well, listen man, let's step back because everybody listening may not know what's even meant by AI, what's meant by artificial general intelligence, what's meant by deepfakes.

So give us some bearing- - Let's define it.

- And foundation and define AI.

And I will tell you, I started in big data back in 2008.

And I was doing what was called machine learning, right, classified, I mean, supervised and unsupervised learning, and I was working on data sets and astronomy and astrophysics.

And it was for the Vera Rubin telescope, which has just debuted its images this year.

And I was being told, "Hey, you're building the software infrastructure for analyzing that data now."

No, I wasn't.

No, I wasn't because AI didn't exist.

- Good.

- I think of machine learning as just statistics.

Define AI for us, define deep learning, I mean, deepfakes, and define general intelligence.

- All right.

Let's start with AI because I think you raised a really good point, which is that everything we are talking about today is in fact not AI, it is machine learning, it is statistics.

Almost everything we see from the ChatGPTs of the world to deepfakes, which I'll define in a minute, to all of the things that we are seeing impacting our day-to-day lives is machine learning.

So let me rewind to 1950.

1950 was when the term, give or take, when Alan Turing and John McCarthy conceived of the term or the concept of AI, and the idea was then quite bold, given where computers were, it was, "Can we imbue intelligence the way we conceive of intelligence with humans into machines?"

And the idea was, "Yes, we should be able to do this," and I think it's because when we think about our own brain and how we operate, it seems like it should be straightforward, but of course it's not.

Talk to any neuroscientist, and they'll tell you that.

But for a long time, for about 30 years, the field of AI struggled to be relevant because there was no path to creating human intelligence in a machine.

And then, around the 1980s, the field splintered and the field of machine learning came about, which was what you were just referring to.

And here, the idea is, we are not going to imbue intelligence into machines.

We are going to learn it from data.

We are simply going to present to a machine a bunch of data and have it infer the rules and the logic that we think that we have.

And frankly, it failed for about 20 years.

So 1980 to about 2000, the field really struggled for relevance, and it struggled for relevance for two reasons.

One is there was not enough data because there was no internet, and there was not enough computing power because there was no Nvidia.

So what happened at the turn of the millennium?

Right?

The internet rose.

What did the internet give rise to?

Lots of things, but one of the things it gave rise to is a boatload of data from us.

Right, the ultimate irony, by the way, is if AI comes for us, we created it by giving it all of our data to learn from.

So around the turn of millennia, around 2000, huge rise in data being pushed online, and then, of course, a huge rise in computing power with things like Nvidia.

- So define Nvidia for the audience, for people who don't know.

- Good.

So Nvidia is the company, of course, that makes GPUs, graphical processing units.

This is the core computational power- - It's a type of computer chip.

- That's like a computer chip, yeah, that is particularly good at the types of computation you need to do machine learning.

And suddenly, we saw this explosion.

It really started around 2015, 2016, and, in the last decade, has really exploded with phenomenal amounts of data, phenomenal amounts of computing power.

And you got to give credit to the statistics and machine learning community, some real insights on how to do machine learning and how to do the types of things that you did.

You were way ahead of the curve in the early noughts.

And then we got better at doing that because we had some insights from statistics and physics and mathematics and computer science and engineering.

So everything you hear today, in my opinion, that we talk about AI is machine learning.

We've sort of conceded that we're just going to call it AI because that's just the term that has gotten adopted.

But underneath it, understand that everything that is happening is pattern matching, that you take a bunch of data and you learn the rules implicitly, not explicitly, from the data.

That's both very powerful and very dumb.

(laughs) It's powerful because you can learn complicated patterns, but it's dumb because you don't know what the rules are.

Like, it's not learning that gravity is minus 9.8 meters per second squared.

It's simply learning that if you film something falling, this is how it will fall, but it doesn't know what the physics of it are.

Right, it simply knows, this is where the ball will be at any given moment.

So that's AI/machine learning.

Artificial general intelligence is a term that I don't think anybody knows how to define, but I'm going to give you my definition of it, or at least we don't agree on a definition.

So typically, historically, when you've looked at machine learning and AI systems, they've been bespoke, so you'll have a system that does medical imaging, you'll have a system that does drug discovery, you'll have a system that drives self-driving cars.

The idea of AGI, artificial general intelligence, is that it does everything.

It's what humans do, right?

- Yeah, or stem cells.

- That's right, exactly.

That's a good example of it, right?

So, is that a year away, 5 years, 10 years?

I don't know.

I don't think anybody knows, but there's a lot of speculation about that.

I would argue, by the way, that we have some notion of AGI.

Because if you go over to your favorite large language model, like ChatGPT or Claude, you can ask it a lot of different questions about physics, about medicine, about computer science, about televisions, anything.

And so, is that AGI?

I don't know.

I suppose it depends on how you define it.

- That takes me to the Turing Test, right?

That was the big...

So what was the Turing Test?

Is it if you can't tell the difference between when you're talking to a machine or- - Yeah.

By the way, you got to have respect for Alan Turing.

- Oh, absolutely.

- In the 1950, this guy was thinking about things that were really outrageous to be thinking about at the time.

- Maybe he was a time traveler.

- That's right.

That's really funny.

(laughs) Don't give anybody any ideas, by the way.

- Right.

- Okay.

So what is the Turing Test?

So when Alan Turing first conceived of this notion of imbuing intelligence in machines, he came up with a mechanism to determine if you have solved that problem.

And the idea was that you would have two computer screens, and behind one computer screen was a human and behind another computer screen was a AI system.

You, of course, couldn't see what's behind it and you were allowed to interact with it with a keyboard.

And you can ask a question, you can have a conversation, the way you and I are having right now.

And if you could not tell the difference, then that machine has passed the so-called Turing Test and it has what probably today we would call AGI.

Okay.

So now we've done AI and we've done AGI.

So let's talk about deepfakes, which is this sort of sliver of all of this.

So deepfakes is an umbrella term for using machine learning, AI, to, whole cloth, create images, audio, and video of things that have never existed or happened.

So, for example, I can go to my favorite deepfake generator and say, "Give me an image of Hakeem in a studio doing a podcast with Professor Hany Farid."

And actually, it would do a pretty good job because you have a presence online, I have somewhat of a presence online, it knows what we look like, and it would generate an image that's not exactly this, but something like that, or I can say... By the way, I still say please, when I ask AI for things.

One of my students told me that this is a good idea because when the AI overlords come, they're going to remember you were polite to them.

I actually really liked this advice.

- Wait a minute.

So I read an article- - That it costs tens of millions of dollars.

- That's right.

(laughs) It is the energy.

- Yes.

Just saying please and thank you.

I still do it, by the way.

And even in my head right there, when I was asked, I still in my head say please.

- Well, listen, I have AI connected to my AI, and so my AI corrects my AI prompts to proper grammar and it's like, "Please."

It puts please in there.

- I know.

I know, and it does cost tens of millions of dollars for that extra token.

Okay.

So I will ask it for an image of a unicorn wearing a red clown hat walking down the street of Times Square, and it will generate that image.

I can ask, "Generate an audio of Professor Hany Farid saying the following."

I can generate a video of me saying and doing things I never did.

You can clearly see the power of that technology from a creative perspective.

If you and I are having a conversation and, in post, we said something we didn't mean to, we can just fill it in with AI now.

- Well, here's the thing that makes me... You just mentioned how we're only two, three years into this.

However good it is now, what is the- - This is the worst it will ever be.

So I can tell you, by the way, how good it is.

So in addition being trained as a computer scientist and applied mathematician, I've been somewhat trained as a cognitive neuroscientist, and we do perceptual studies.

So what we do is we recruit participants, we show them images, audio clips, and video, and we tell them, "Half of the things you're going to look at are real.

Half of the things are AI generated."

We explain to them what AI generated is.

We give them examples of that.

And for images, as of last year, people are roughly at chance at distinguishing a real photo from an AI-generated photo.

- So what you mean by that is, if you had a monkey behind the keyboard- - Yes, flipping a coin.

- Flipping a coin.

- Yeah, the monkey's probably better than you, by the way.

I'm going to go off and guess.

So with audio, so we play a clip of somebody speaking, like you, and then we play an AI-generated version.

They're slightly above chance, like 65%.

- On image, at chance, at audio, a little better than chance.

- In video, they're a little bit better.

But all of those trends are going towards chance.

So here's what we know.

Everything in the next 12 months, 18 months, 24 months, I don't know what the number is, it will be indistinguishable to the average person online, and that is... - Scary?

- That's a weird world we're living in because think about how much... First of all, the vast majority of Americans now get the majority of their information from online sources, and unfortunately, from social media too and because it is so easy to create this content.

Understand, all this is is a text prompt away.

I type, "Please, give me an image of this.

Generate this audio.

Generate this video."

There are dozens of services that will do this extremely inexpensive or for free, and you can carpet-bomb the internet with fake images of the conflict in Gaza, fake images of- - And I have seen these.

- I have seen them too.

Fake images of the flood in Texas, fake images and video of the fires in... Name it, across the boards.

Right?

Fake images of people stuffing ballot boxes, now, we have a threat to our democracy.

- Wow.

- So suddenly, our sense of reality, coming back to your first very good question, is up in the air because I can create whatever reality I want.

And understand that there's sort of three things happening here when we talk about deepfakes.

There's the creation of it, that's what we've been talking about.

There's the distribution, which we democratized 20 years ago.

So anybody can publish to the world, and that's very powerful and very terrifying because there's no editorial standards on social media.

And then there's the amplification that we have become so polarized as a society that when you see things that conform to your world view, you are more than happy to click like, reshare.

And now you have creation, distribution, amplification.

- Wow.

- That's the ballgame.

Right, that's the ballgame for spreading massive lies, conspiracies, and disinformation campaigns that affect our global health, our planet's health, our democracy, our economy, everything.

Everything.

- So let's get into how these fakes are generated.

- Good, yeah.

- So, start with images.

- Good.

So let's start with images because, in some ways, it's the easiest one, but all of these have a similar theme.

And one of my favorite techniques for generating images is called a generative adversarial network, or a GAN, and here's how it works.

- Wait a minute.

Wait a minute.

Adversarial?

- Adversarial, yeah.

- So that means that you're fighting your computer?

- Two computer systems are fighting each other, and this is sort of the genius of this technique.

So here's how it works.

You have two systems.

One system's job is to make an image of a person or a landscape or whatever you want.

And so what it does, it starts by...

This is literally true, it just splats down a bunch of random pixels.

And so I say, "Generate an image of a person," and it says, "Okay, here's a bunch of..." So think the monkey's at the keyboard typing randomly.

Let's see if this is Shakespeare.

And then it takes that image and it hands it to a second system and it says, "Is this a face?"

and that system has access to millions and millions of images that it's scraped from the internet that are faces.

- I see.

- And that system says, "That thing that you generated doesn't look like these things over here," and it gives the feedback to the generator and it says, "Nope, try again."

Modify some pixels, send it back to what's called the discriminator.

"Is it a face?"

"No, try again."

And they work in this adversarial loop, so it's like somebody's checking your homework.

- It seems like it could get stuck never getting to a face.

- You would think, and that's what's amazing about the GANs is that they converge.

They converge.

And part of that is the way they've been trained, but that's what's the genius of this is that the generator is not very smart because all it's doing is modifying pixels, and the discriminator is actually quite simple.

It's simply saying, "Does this thing look like these things?"

And because you pit them against each other in this adversarial game, this sort of amazing thing happens out the other side.

- So here's the question, on average, how many iterations does it take?

And then how much time does that translate to the real world?

- Yeah, that's a great question.

So typically, the time is in seconds.

So there's two phases.

You train the GANs, that's a really long process.

But then what we call inference, which is that, "Run this thing," it happens in seconds, and the reason it happens in seconds is... By the way, that is hundreds of thousands of iterations.

- Wow.

- But it's on a GPU, which is very powerful and very fast.

And then there's these tricks to make it even faster.

You start with small images and then you make them bigger over time, so there's these tricks to make, but it is literally seconds to make that image.

And what the brilliance of that is the two systems are competing with each other.

And then this thing that seems like intelligence come out, even though it's not.

If you think about those two individual components, they're pretty basic.

- They're pretty dumb.

- But then you have this like emergent behavior almost.

It's like, "You know how to generate images of people.

That's amazing."

- So let's have a little fun.

I understand that you brought me some fakes and some real images to put to the test, to see if I can discern the difference.

- Yeah.

I'm going to play for you a couple of audios.

Before I do this, let me say, I've been doing this for a long time and I'm pretty good at it.

I'm pretty good at what I do.

And I had created three audio samples.

I'm going to play them for you.

- Wait, are you allowed to say that, that you're good at what you do?

I'll say that.

Hany is really good at what he does.

- I said pretty good, by the way.

- He's amazing.

- This is a true story, by the way.

So I made three audio clips for you of me talking, and you and I have been talking for a little while, so you now know what my voice sounds like.

And I got off the plane and I was in the car coming over here.

And I wanted to make sure they worked and I played all three of them, and I couldn't tell which one of me was real or fake.

I wasn't 100% sure.

- Wow.

- And I do this for a living, and it's my voice.

- Right, yeah.

- Okay, so that is...

Okay.

- So, wait a minute, which AI did you use?

Was this something that you created or something that's generally available?

- So here's the thing you have to understand about AIs, this is so readily available.

So here's what I did.

I went to a service, it's a commercial service.

I uploaded, I think it was about three minutes of my voice.

I said, "Please, clone my voice," and it clones my voice.

And what I mean by that is that it learns the patterns of my voice, what I sound like, the intonation, my cadence, how fast I speak, where I put the pauses.

And then I can simply type and have it say anything I want to say.

I'm going to have you listen to three sentences.

I'm going to give you a hint.

One of them is fake and two are real.

- Okay.

- Let's see what we can do.

- All right.

- Okay, here we go.

And in fairness, this is not the best speaker, but okay... - Are there guardrails in our law?

- Good.

So first of all, when I went to do this service, I uploaded my voice and there's a button that says, "Do you have permission to use this person's voice?"

And I did because it was my voice, but I can upload anybody's voice and click a button.

The laws are very complicated and they actually vary state to state and, of course, internationally.

So there are almost no guardrails on grabbing people's likeness.

And even if there were, there's- - You could still do it anyway.

- There's no stopping this.

There's no stopping it.

Okay.

All right, number one... Oh, and, by the way, the three...

This is part of a talk I gave recently on deepfakes, so you'll hear a consecutive thing.

Okay, ready?

And if you invite me back next year, almost certainly everything will have changed, the nature of creation of deepfakes, the risk of deepfakes- - That's the deepfake right there, man.

- And the detection of deepfakes is changing.

Hold on.

Hold on.

That was good.

It is a fast moving field, and we have to start thinking seriously and carefully about the threat of misinformation.

Okay, good, and one more.

We are living through an unprecedented time where we are relying more and more on the internet for information, for information that affects our health, our societies, our democracies, and our economies.

- Can I hear number one again?

- Yeah.

You're a little less sure than you were a minute ago.

- Yeah.

- And if you invite me back next year, almost certainly everything will have changed.

The nature of creation of deepfakes, the risk of deepfakes, and the detection of deepfakes is changing.

- I think it's the first one still.

I got it right?

- Yeah.

I struggled with it, by the way.

Honestly, I couldn't remember.

- I'm from the future.

- (laughs) You're the time traveler, it turns out.

(both laughing) - Wow.

Well, you know what?

So I started my media work in audio being a voice actor and, very quickly, I was able to pick up on music and commercials and movies where they were dropping in, you know, pickups versus- - The reason I figured it out is there's a difference in the background noise.

Like, one had more reverb than the other, which is how I then remembered it.

But you got to admit, all three of them sound like me.

- Oh, they all do.

They all sound like you.

- Oh, and, by the way, so not only can- - Well, let me tell you what has gotten me recently is I'll get these social media announcements, "Oh, there's a new song by Tupac and Eminem," and I started listening to it and, halfway, I'm like, "No.

This is AI."

- I know.

- But at the beginning they gave- - It's coming from music.

It's coming from music as well, by the way.

So this is one of my favorite videos, by the way.

Let me just show this to you.

And if you invite me back next year, almost certainly everything will have changed.

The nature of the creation of deepfakes, the risk of deepfakes- - That's fake?

- That's real.

Wait for it.

(Hany speaking Spanish) I don't speak Spanish.

(Hany speaking Spanish) - And your mouth is doing it.

(Hany speaking Japanese) - I don't speak Japanese.

(Hany speaking Japanese) Doesn't it sound like me?

- Yes, it does.

- I know.

So, now, I can do full-blown video, any language.

- Wow.

- Any language.

By the way, here's what's really cool about this.

Here's a really cool application.

I like foreign films a lot, but I can't stand bad lip syncing.

- Yeah, I'm with you on that.

- It makes me crazy.

- Same.

- But you don't need it anymore.

- You don't need it.

- We're now going to make videos in any language you want, and it's going to be perfect.

- How did you do that?

- This is also a commercial software.

You upload a video, say that you have permission to do it, and you say, "Please, translate this into Japanese, Korean, Spanish, French, German," anything you want.

It's amazing.

- That is nuts.

The fact that the mouth changed to voice the words- - Yeah.

And by the way, the way this works, this is really amazing, is you upload a video of you talking.

And what it does is it takes the audio and transcribes it, so it goes from audio to words, and then it translates from English to Spanish, and then it synthesizes a new audio in Spanish, and then it puts that audio back into the video.

Every one of those is an AI system, by the way, and it does that in about three minutes.

- Wow.

- And it's amazing.

So if you wanted to take this podcast and distribute it in Spanish, French, German, just upload it and you're done.

- I'm just hitting India, China, Southeast Asia- - Yeah, that's right.

Two and a half billion people, done.

- That's right.

(laughs) - Done.

- Ten cents each- - That's right.

- We're good to go.

This podcast is from the producers of NOVA.

NOVA is supported by Carlisle Companies, a manufacturer of innovative building envelope systems with buildings responsible for over a third of total energy use and energy demand on the rise.

Carlisle's mission is to meet the challenge head on.

Carlisle's energy-efficient solutions are built to reduce strain on the grid.

For example, Carlisle's UltraTouch Denim Insulation made from sustainable recycled cotton fibers delivers high thermal performance for improved energy efficiency while being safe to handle and easy to install.

Made from 80% recycled denim, UltraTouch diverts nearly 20 million pounds of textile waste from landfills each year.

Operating nearly 100 manufacturing facilities across North America, Carlisle is working towards a more sustainable future.

Learn more at carlisle.com.

We have systems now that detect AI text, AI audio, AI images, AI video.

Give us the nuts and bolts of how you detect these fakes.

- My bread and butter as an academic and also now as a chief science officer of GetReal is to build technology to distinguish what is real from what is fake.

Okay?

And so, here, there is some AI there, there are some more classic techniques.

I want to talk, if I may, about my favorite one and I think this may resonate with you as a physicist.

What you have to understand about generative AI, deepfakes is that it is fundamentally learning how to generate images, audio, and video by looking at patterns in billions and billions of images, audio, and video, but it doesn't know what a lens is.

It doesn't know what the physics of the world is.

It doesn't know about geometry.

It doesn't know about the physical world.

It's not recreating this thing that you and I are in right now.

Take any image outdoors.

Sunny day here in Virginia, go outdoors, and, because the sun is shining, you will see shadows all over the place.

And those shadows have to follow a very specific law of physics, which is that there's a single dominant light source, sun, and it is giving rise to all those shadows.

So we have geometric techniques that can say, "Given a point on a shadow, and the part of the object that is casting it, tell me where the light is that's consistent with that," and we can do that not once, not twice, not three times, but as many times for shadows that we find.

- For every shadow in the image.

- Every shadow.

And if we find that they are not converging on a single light source, the sun, then we have a physically implausible scene.

- It seems like that will be easy for AI to figure out to do.

- You would think, but here's why it can't.

Because what I described to you is a three-dimensional process, that's happening in the three dimensional world, but the AI lives in 2D.

It lives in a two-dimensional world.

And reasoning about the three-dimensional world is not something it does.

Now, it can sort of fake it pretty well, the way artists fake it.

Lots of things in paintings are not physically plausible, but our visual system doesn't really care.

We're looking at a pretty picture.

So that's one of my favorite techniques.

Here's another one that I love.

Go outside and, well, you shouldn't actually do this, but stand on the railroad tracks.

I don't actually advise doing that.

I did this the other day with one of my students and like, "What are you doing standing on the railroad tracks?"

"I wanted to take a picture of railroad tracks."

And the reason I wanted to take a picture of railroad tracks is that, when you're standing on the railroad tracks, those railroad tracks of course are parallel in the physical world and they remain parallel as long as the track continues going.

But if you take a picture of it, those train tracks will converge to what's called the vanishing point.

This is a notion that Renaissance painters have understood for hundreds of years, and why is that?

It's because, when you photograph something, the size on the image sensor is inversely proportional to how far it is for me.

So as the train tracks recede, it looks like they're converging.

This is called projective geometry, a vanishing point.

It's a very specific geometry, and this is true of the parallel lines on the top and bottom of a window, on the sides of a building, on a sidewalk.

Anything that you have a flat surface, like this table that we're at.

Take a photo of this table and all these parallel lines will converge to a vanishing point.

So we can make those measurements in an image.

And when we find deviations of that, something is physically implausible.

The image is violating geometry.

Okay?

All right.

Let me move to a sort of different side of it.

This is actually one of my favorite techniques is when you go to your favorite AI system and you ask it to make an image, it will create all the pixels, but then it has to bundle it up into a JPEG image or a PNG image or some format.

And it actually does that in a very specific way, and so here's an analogy.

When I buy something from an online retailer, there's the product I get, but that product is also packaged in a box.

And different retailers have different ways of doing it.

Apple has a very specific way of doing beautiful packaging.

Other retailers just shove it in a box and send it off.

So the packaging, when I create an image on OpenAI or on Anthropic or on Midjourney, all these different generators, they package it up differently, and it's different than the way my phone packages up the pixels, and it's different than the way Photoshop packages up the pixels.

So when we get an image, or an audio or video for that matter, we can look at the underlying package and saying, "Is this a packaging that is consistent with OpenAI or Anthropic or a camera or whatever it is?"

- So it doesn't have package emulators?

- Yeah, it does not.

It doesn't know about it because it doesn't care.

Why would you care about it?

I'm the only person in the world who probably cares about this.

You certainly don't care how it's packaged because, what do you do?

You open the package, you throw it away, and you've got your product, the image.

So we can look at the packaging.

There's a whole nother set of techniques.

So everything I've described is sort of after the fact.

Right, you wait for the content to land on your desk and you start doing these analyses.

There's a whole nother set of techniques that are what are called active techniques.

So Google recently announced that every single piece of content that comes from their generators, image, audio, or video, will have what's called an imperceptible watermark.

We don't use currency that much anymore, but take your $20 bill out of your wallet and hold it up to the light and you'll see all kinds of watermarks that prevent or make it very difficult to counterfeit.

So what Google has done is they have inserted an invisible watermark into images, audio, and video at the point of creation that says, "We made this."

And then when I get that piece of content, I have a specialized piece of software because we, over at GetReal, have a relationship with Google that says, "Is there a watermark in there?"

It's a signal and you can't see it.

And the adversary can't see it, but I can see it.

So that's really cool.

And by the way, if this comes into the phones, so if Apple decides, "We're going to watermark every single piece of content that is natural."

I've got a signal that is built in.

So we've got lots of different techniques from things that we rely on third parties, like the Googles of the world, to measurements that we can make in an image, a video, or an audio.

I'll give you one of my favorite audio ones, by the way.

So if you're listening to this, you won't be able to see us.

But if you're watching this on YouTube, you will know we're in a really nice studio and there are soft walls around us, and we have really nice microphones, and so the amount of reverberation that you hear is quite minimal.

This audio is going to sound really good because you guys are pros here, but the amount of reverberation is dependent on the physical geometry around us, how hard those surfaces are, and that should be fairly consistent over an audio.

But what you see with AI generation is you see inconsistencies in the microphone and the reverberation because it's not physically recording these things, it's synthesizing.

- So even in a single recording- - You'll see modulation- - You get those modulations.

- That are, quote, unquote, "unnatural."

A lot of what we do is look for patterns you expect to see that mimic the physical world.

And then I talked about the active techniques of watermarking.

And then there's a whole nother set of techniques that I'm going to talk about a little, but not a lot, and you'll understand it in a minute why not.

So the other side of what we do is we try to understand the tools that our adversary uses.

So if you're using an OpenAI or Anthropic or some open source code, we actually go into the... Well, we can't do this for OpenAI, but for anything that's open source.

There are so-called face-swap deepfakes where you can take somebody's face, eyebrow to chin, cheek to cheek, and replace it with another face, and these are all open source libraries.

We can dig into the code and we can say, "Okay, what are they doing?"

"All right, the first thing they're doing is this, and then they do this, and then they do this, and then they do this."

And then we'll say, "Ah, that second step should introduce a very specific artifact."

So I'll give you one example, but not more than one.

One of the things that a lot of these swap faces do is they put a square-bounding box around the face, they pull the face off, they synthesize a new face, and then they put it back.

But when they put it back, it's with a bounding box and they do it very well.

You can't see it, but we know how to go into the video and discover that bounding box that was there.

- Wow.

- All right.

So that's an example where we reverse engineering because we understand how the adversaries made something.

Now, we have lots of other ones, which I don't want to tell you about.

- Yeah, I understand why.

- Yeah.

You could see now?

Because it's adversarial.

- Exactly.

Right.

Right, right.

Man, it sounds very systematic.

I have a decent understanding now if I want to make a lap to do this, some techniques to do it.

But, you know, the average person out there isn't a scientist.

So how can people, how can I, how can my mother identify real from the fake in the world of AI?

- Yeah, they can't, and this is the reality of where we are right now, and this is important to understand because I don't want you to walk away from this podcast thinking, "Okay, I understand a little bit now.

Now, when I'm scrolling through X or Bluesky or Facebook or Instagram, I'm going to be able to tell."

You won't.

You won't be able to tell.

And even if I could tell you something today that was reliable, six weeks from now, it'll not be reliable and you'll have a false sense of security.

So, I get this question a lot, and the thing you have to understand is this is a hard job.

It is really hard to do this, and it's constantly changing.

And the average person doom scrolling on social media cannot do this reliably.

You can't do it reliably.

I can barely do it reliably, and this is what I do for a living.

So here's some things.

Stop, for the love of God, getting your news and information from social media.

This is not what it was designed for and it's not good.

If you want to be on social media for entertainment, that's fine.

I don't care.

I don't think you should, by the way, but I don't care.

But this is not where you get news information.

You know where you get it?

Things like this.

You get it from news outlets that have standards, that have serious smart journalists who work hard to get you information.

And we have to come back to some sense of reality of where we get our news.

- Man, rigor around determining truth and measuring uncertainties is not something that we're generally taught.

And when you become a scientist, it is unnatural.

It's an unnatural way to think.

- That's right.

I agree, yeah.

And look, you know, you and I have both fallen.

You were telling me in the greenroom before this, right?

You heard this story.

You thought it was true.

You assumed it was true.

Somebody called you out on it.

You went and figured out like, "Oh, God, I was wrong."

Right?

- Yeah.

- And now, imagine that at scale of the thousands of posts you're seeing every day.

- Well, the reason why I thought it was true was because everybody else was saying it was true.

- That's right, exactly, and that's what social media is.

Everybody is saying the same thing.

Millions of views, "Oh, this must be true," but that's the way social media works.

We have to get back to getting reliable information from reliable sources.

That's number one.

Number two- - And I'll tell you the other thing about it though is that even though I assumed it was true because everyone else assumed it was true, and these others were scientists just like me, is I still knew that I didn't know, that I had not confirmed it for myself.

And I think that is where the average person can... You know, if we know the difference between knowing and not knowing- - Good.

- Then you can check yourself.

- I agree.

- Even if everybody's saying it, you don't know that that's the truth.

- I agree, and this gets to number two.

You said it in a nicer way than I was going to.

But understand that the business model of social media is to draw your time and attention by feeding you content that outrages you and engages you, to deliver ads, to get you to buy stuff you don't need.

And so understand that you're being manipulated and that you are being fed information that the social media companies believe you are going to engage with.

And first of all, that should make you angry that you are being manipulated.

But you're 100% right, is that, you know, we live in these distorted bubbles when it comes to social media and it's very easy to forget that you actually don't know what is happening in Gaza, in Ukraine, in Texas, in Los Angeles.

You just think you do, and that is incredibly dangerous.

It gives you this false sense of security, and so it comes back to... What I say is, "You got to have some humility."

You have to have humility that this is a complicated world, it is fast moving, and you have a responsibility to get reliable information because not only are you being lied to and being deceived in making decisions on bad information, you're also spreading that bad information.

So you are actually part of the problem now because when you like and share and send this to your friends, you are now a carrier of disinformation.

- So we have these AIs that do things for us, and we're sort of managing them.

- Yeah, that's right.

- You know, I might write something and I'm like, "Oh, give me an edit on this," and then it comes back and it will tighten it or whatever.

But you can move to a point where you're like, "Give me the first draft."

- Yeah, that's right.

Yeah, and it's coming.

- And you get to the point where it's like, "Okay, do it."

- Yeah, "Start responding to my emails."

I don't even have to read my emails anymore.

And by the way, there's a really weird future that you could imagine where emails are being sent by each of our agents and we're hanging out on the beach.

I mean, what are we doing?

- Exactly.

- I don't know.

- That's the point.

It's almost as if we are making ourselves obsolete in many, you know... You need a human being to build your house.

Right?

You need humans there swinging hammers.

AI can't do that.

- Yep.

Not yet.

(both laughing) - Well, AI robots, right?

- Yeah, that's right.

That's right.

But this is sort of the ultimate joke in some ways or the irony of all this is, I think if you go back 50 years, what were people worried about?

That we were going to take blue collared jobs away, manufacturing jobs, physical labor jobs.

That we were going to build robots to do those jobs.

What did we end up doing?

We took out the white collar jobs, we took out those high-paying computer science jobs, we took out those jobs that AI is now doing better than most humans, and that's a weird world.

I can tell you, I'm on the campus almost every day, and there is a lot of anxiety among students about what the future holds for them.

Are there going to be jobs because you're seeing unemployment go up in these historically high-paying jobs.

- Listen, in my own house, my son, he's 20, he's a senior in college this fall.

Guess what his major is, computer science.

- Yeah, and he's struggling.

Everybody is.

But I can tell you, so I'm at UC Berkeley, one of the top, let's say, five CS programs in the world.

And our students typically had five internship offers throughout their first four years of college.

They would graduate with exceedingly high salaries, multiple offers, they had to run the place.

That is not happening today.

They're happy to get one job offer, so something is happening in the industry.

I think it's a confluence of many things.

I think AI is part of it, I think there's a thinning of the ranks that's happening part of it, but something is brewing.

For people like your son, by the way, who four years ago were promised, "Go study computer science, it's going to be a great career.

It is future-proof," that changed in four years.

That is astonishing.

And I get this question, by the way, from students all the time is, "How do I prepare for this?"

- Yeah, exactly.

- And honestly, I'm sort of at a loss.

My best advice, and I don't know if it's good advice, by the way, is I used to tell people, this is what I used to tell people, is you want a broad education.

You should know about physics and language and history and philosophy, but then you have to go deep, deep, deep, deep into one thing, become really, really good at one thing.

Now, I think I'm telling people, "Be good at a lot of different things because we don't know what the future holds."

- And you need options.

- And you need options.

- Yeah.

- And so depth is, in some ways, less relevant, particularly if you know how to harness the power of AI to get the depth that you need in this particular area.

But if you have a broad knowledge, I think you're probably today more future-proof than if you're very narrow in one area.

And the best line I heard about this was, this is in the framing of the legal system, is that, "I don't think AI's going to put lawyers out of business, but I think lawyers who use AI will put those who don't use AI out of business," and I think you can say that about every profession.

So I think two things are going to happen is that you're going to have to learn, like every technology, how to harness the power of AI.

That was true of computers, that was true of the internet, that was true of everything.

But because of how powerful the AI systems are, it is absolutely going to reduce the workforce.

And I think the big question here, is the question always that happens with disruption of technology, is we are going to eliminate and reduce certain jobs and, the question is, do we create new jobs and what do they look like.

- Exactly.

- And I don't think anybody knows the answer to that right now.

- How is AI going to affect blue collar jobs?

Is AI going to affect blue collar jobs?

We have this vision of white collar jobs, but there could be an effect.

What do you think?

- Yeah, it's a great question.

So now let's come to blue collar.

So, where is it coming first?

So here's what I can tell you.

It's coming for the self-driving cars.

It's coming for drivers.

- Taxis.

- Yeah.

Go to San Francisco and you can get a car that is self-driving, it's weird, by the way.

You go to San Francisco and just stand on the street corner and look at how many cars drive by you without a driver.

- What about trucking?

- So I think it's coming for trucking.

I think it's coming for the lifts, the Ubers, the taxis, the limos, and I think that's in our lifetime.

- What about shipping?

- Yeah, everything.

Everything.

I think that long haul truckers that go a thousand miles, I think once that truck is on the highway, I think that's completely autonomous, within the next 10 years.

- Wow.

- So I think when it comes to moving people and objects, A to B, that is probably coming.

I don't think it's coming for plumbers, I don't think it's coming for electricians, I don't think it's coming for construction workers, because I don't think the robots are even close to that.

But driving is a relatively self-contained relative to what a plumber has to do, which is come into your house, find the bathroom, diagnose the problem, fix things, which is very bespoke.

When you drive on a highway from point A to point B, the rules are pretty clear, right, get from A to B and don't hit anything.

You can't say that about plumbing.

It's a much more complicated process.

I do think it's coming for some of those jobs, but I think other ones, frankly, are probably much more secure, which again is the ultimate irony is that the nerds like me are putting themselves out of business.

- Wow.

Yeah, yeah.

Trade school is going to make a comeback.

- I tell my students, and I'm only half joking, "You better learn how to fix a toilet."

- Right.

- I mean, no kidding.

I spent my first 20 years at Dartmouth College, and I very much believe in the liberal arts education.

I don't like when students go really deep, very quickly, all engineering all the time for four years, I don't think it's good intellectually, but I think we need to start rethinking higher education.

We have to rethink- - Absolutely.

- Everything.

Everything has to be thought.

- I completely agree.

- We are in a brave new world.

And I think if we don't... And by the way, I think we have to start thinking about how do we teach students to use AI to help them think.

Because, you know, what you're seeing at universities right now is one of two things.

Either bury your head in the sand and pretend ChatGPT and large language models don't exist, and that's a bad idea.

Ban it and saying you're not allowed to use it.

And neither of those are good solutions.

We have to think about how to incorporate this into the curriculum and, as an academic, you know that the university is not well-known for being very nimble and fast-moving.

And so we have to move a little bit faster.

- But you know what also comes to mind?

It's something I experienced myself, but then recently, I read an article that says that, "No, there's actually been a biological change."

And what is that?

I used to...

I lived my life driving cross country from throughout my childhood.

I never lived in the same state two years in a row.

When I became an adult, I kept it up, kept going.

And so I used to rely on paper maps, and I would get a map when I moved to a new town and I would study that map and I would have a bird's eye view of my town in my mind.

And I'm the type of dude I like to take the back streets.

I don't like to take the main streets.

GPS comes out.

I could not navigate from where we are now back home because I now rely on GPS.

And the article I read said that our brains, the part of it that has this memory map, has actually- - Shrunk.

- Shrunk.

- The hippocampus.

- It is modified.

- Yeah, it is.

- Because of this reliance on GPS.

So now, AI is going to come around and it's going to do so much more for us.

Right?

You don't have to calculate, you don't have to memorize as much anymore.

We're evolving ourselves.

- In a very short period of time.

I think this is 100% true.

By the way, there's no question that's changing the way we think.

And I think the question is, that's not fundamentally bad or it doesn't have to be bad because the question is...

So I get in a car now, and I'm the same way as you, I couldn't navigate my way half a mile from my home.

No matter where I am, I GPS it.

But what I do during that time is I think about other things, so the GPS enables me to think about the problems I want to think about.

And so, is that overall good or bad?

Okay, I'm not good at navigation.

So if my phone breaks down, I'm not going to be able to find my way out of paper bag, but I get a lot of my time back in the car to contemplate and to think.

And so that is really powerful.

And will AI do the same thing?

I'll give you an example.

I use AI every day in my work to help me write code, and so I'm probably becoming a worse coder.

I think that's probably true.

But I can now build prototypes and build systems at a scale that I couldn't do two years ago.

Okay, so it's a trade-off.

I'm less good at the nitty-gritty, but I'm good at building much more complicated systems because the AI can help.

And here's the thing about computer science, particularly, that you have to understand is that we've been moving in this direction for 80 years, because think about what computer science is all about.

All of computer science from the very early days has been about abstracting more and more detail out of the system.

So we used to start with punch cards.

We had to literally physically push little things into punch cards to program.

And then we programmed in assembly language, this very low-level language where we had to know about the hardware to program.

And then we went to high-level languages, and then higher-level languages.

And now, it's natural language.

And so it's part of an evolution.

But the thing is it's also...

Okay, so some people, purists can say, "Well, that's bad.

You should understand this," but I would say, "Hey, if I can empower somebody who is not a computer scientist to build systems, that's an incredible thing to do."

- It is.

It is.

It reminds me of the day I stepped foot in the graduate school.

Right, I was a rural dude in the deep woods, and I had a academic education that did not size up to my peers when I arrived, but I was like, "Can't nobody in here skin a squirrel better than me?"

(Hany laughing) Right?

I'm the only one who know how to grow some crops.

- That's right.

That's right.

- And so what I'm saying is, is that we are replacing one set of skills and knowledge with another set of skills and knowledge for the world we live in today.

So if we're not going back to an agrarian society, you know, nobody knows how to shoe a horse anymore.

- That's right.

Yeah.

- It was required.

No one knows how to take a piece of flint and nap it to make a stone arrowhead anymore, right?

- That's right.

- And we continue on.

- And I think the big question... Because you're 100% right, we have been doing this for our entire existence.

I think you do have to acknowledge, or we should acknowledge, AI is different.

It is different.

It's different than a flint versus a lighter because it is so fundamental to so much what we do.

And I think the question is, "Will the disruption be so great that unemployment is 30%?"

Because that is a problem.

- That's a problem.

- Let's agree on that because, now, the entire social contract is broken.

- You have instability in your society now.

- We have to start talking about universal basic income, if we're going to talk about this, right?

So, I don't know, and I don't think anybody knows, but here's the thing that I would argue is, for the first 20 years of the technology revolution, we largely took our hands off the steering wheel.

We largely said, "Hey, look, let's let the internet be the internet," and I don't think most people are looking at the internet today being like, "Let's do more of this."

It is not great.

There's lots of great things that came from it, don't get me wrong, but there's horrific things that have been happening.

- You ain't got to tell me, man.

- Horrific things.

And if you have kids, as you do, you know.

- Oh, myself, man.

I started graduate school in '91 and, at the time, it was newsgroups, right?

- Sure.

- Just that.

Some people on the planet thought it was really funny to label something as one thing when it was really something horrific that you didn't want to see.

- And then you opened it, yeah, and it's only gotten worse.

- Yeah.

- And it's only gotten worse.

We have to put our hands back on the steering wheel and we have to start getting serious about how do we put guardrails on the system.

Because what we know is that if you unleash Silicon Valley, they will burn the place to the ground to get to the finish line first, and we've got to start putting guardrails in place.

And the thing is, is that this is a borderless problem, right?

This can't be an only US or an only EU or an only- - That's a human problem.

- We have got to start thinking about this globally, and I don't think we are doing that very well, honestly.

The EU is probably the leaders on this because they have the AI safety bill.

The United Kingdom is doing fairly well, the Australians are doing well, we, I think are lost at sea right now.

That's not, by the way, a partisan issue.

I don't think either party has done particularly well on this regulation because, frankly, nobody on the hill, not too far from where we are right now, really understands this.

So I think we have to get smart very fast and we have to start thinking about how to put guardrails that allows for the innovation, but also protects individuals, societies, democracies, and economies, and I don't think we're doing that right now.

- So let's jump off to a part of AI that is not typically discussed in these sort of forums, and that is AI infrastructure.

So when I look at it, there is like energy, there are data centers, right, then you have companies like Nvidia chips, then you have software.

And I imagine that all of that is going to evolve, higher efficiencies, new architectures.

So I could even imagine that there's going to be new computer hardware that's made specifically to serve the needs of AI.

So where do you see that entire infrastructure?

How do you see that evolve?

- Good, okay.

So let me preface what I'm about to say by saying nobody knows what the next five years are going to look like, and here's how I know that.

Because if you had asked any of us five years ago, "Would we be here?"

We would've said no.

So let's be honest about the future is that it's very hard to predict.

But here's what we are seeing.

I just saw an article the other day.

Mark Zuckerberg wants to build a data center the size of Manhattan, the size of Manhattan.

This guy over here is nodding his head.

He read the same article, by the way.

So we are talking about data centers that are massive and they're also gobbling up water, gobbling up energy, at a time, by the way, when we are in a crisis in our climate.

And I'm not sure that this is the right direction we want to be moving in with respect to the broader picture.

I'm not anti-technology, I'm not anti-innovation, but we have to think about the broader picture here.

So right now, what troubles me...

I think you asked the right question, which is how do we make these things more efficient, how do we make them require less power, less water, less data, and that is not what we are doing because there is a race to, quote, unquote, "win."

And you are seeing all the big tech companies just say, "Carpet-bomb the place, build data centers, gobble up all the water, gobble up all the energy, and go."

And by the way, what you should understand about energy, there's actually two places the energy is being consumed.

One is of course you have to power these machines, but what really the energy is being used is to cool them.

What you have to understand about GPUs and these specialized hardware is they pump out a phenomenal amount of heat.

So the amount of cooling that has to come into these systems is enormous.

And now, you are literally gobbling up city-wide, city-scale energy consumption.

Now- - Does it matter if it's, like, built in Alta, Norway?

- I mean, ideally, that's what you would do, right?

So 5, 10 years ago, there were discussions about building these data centers in the Pacific Ocean where you could use the inherent cooling of the ocean.

- I've seen some technology where the servers were actually inside of liquid.

- That's right.

Yeah, exactly.

- Is it non-conductive liquid?

Is that how that works?

- I don't know.

You're outside of my area of expertise, but now what they're doing is going into rural parts of the country and just taking over, sucking it all up.

- Well, where I live, in Northern Virginia, the data centers are popping up all over the place.

- Everywhere, yeah.

- That's right.

- There are some people who think that in order for this really to scale, we're going to have to figure out how to do this more efficiently.

This is outside of my area of expertise.

I'm not a hardware person.

- Well, listen.

So from the energy side, we're trying to, it's not coming about in the next 5, 10 years, fusion energy, right?

That's cheaper- - Or quantum computing.

I don't think that's going to happen in the next 5 to 10 years.

We are not going to have machines that suddenly become 100, 1,000 times more efficient.

The trend right now is simply to take the things that we have, GPUs, NVIDIA, and just scale them to massive scale.

And you know this better than anybody.

The advances in physics happen at a very different time scale.

This is not happening over a one or two years.

These are decade long.

And by the way, if I may, at a time when we are investing less and less and less in research, in academia- - Yeah, big mistake.

- Huge mistake, and we are going to pay the price for that for generations to come.

- Yeah.

So something you said earlier I found really interesting is that is you have some experience with neuroscience.

So now we see large language models are pretty much predict the next word, from my understanding, right?

- Correct.

- And we have these brains that we think, "Oh, we're so complex and high-powered and spiritual and magical."

Has the research in AI thinking illuminated in any way how our brains are working, and could our brains be as simple as binary code?

- All right.

So first, I have to apologize to my wife before I answer this question because my wife is a very serious computational neuroscientist and it makes her crazy when I talk about the brain.

I'm sorry, sweetie.

(both laughing) So, I don't know is the simple answer, but here's the more complicated question.

There is a story that, although we feel like we are very sophisticated, we are sophisticated pattern-matchers.

Right?

We see this, we do this.

I don't think that fundamentally we really understand how the human brain does what it does.

But I think there is a story there, that it's actually simpler than we think it is.

And again, it's an emergent behavior that a lot of very simple computations give rise to very complicated human beings.

I think, right now, in my view, AI has not illuminated a lot about the human brain.

I think the human brain has motivated a lot of how AI architectures are built using sort of similar types of neural architectures.

Whether it will illuminate more about the human brain, I don't know.

But the human brain also, there's a lot going on.

We are able to move, we're dexterous, I can pick up this cup without bumping into the microphone, I recognize you now that I met you just a few minutes ago, I now recognize you.

We are pretty complicated.

The short answer is I don't know.

I honestly don't know.

But I think more likely than not, we are simpler than we think we are.

I actually think, because the one thing you have seen about AI is that it can generate language, images, video, audio, fairly simply.

And the thing that's amazing about, you said it right, large language models, they are basically one-word predictors.

You start with a beginning of a sentence and you predict, "What's the next word?

What's the next word?

What's the next word?

What's the next word?"

Is that what I'm doing as I'm speaking right now?

No.

In my head, it seems more complicated, but that could be an illusion.

So I don't know.

- And what we see in nature is sometimes the same solution occurs by different mechanisms.

One that I was just studying had to do with the evolution of light skin, right, and how it happened in Europe and Asia, but via different genes.

So the same problem of, you know, getting lighter to get more vitamin D was solved in two different ways.

- Two different paths to the same thing.

- Two different paths to the same thing, so it could be that we are generating a new form of what we do ourselves.

Right?

- Yeah, I think that's right.

- Yeah.

All right.

So now, I'm going to put you to the test, my friend.

You put me to the test, I'm going to put you to the test.

- Good.

- So I'm going to give you some AI headlines and there's going to be two real and one fake.

- Okay.

- Here we go.

- I guess I had this coming.

- Man, listen, you brought it on yourself.

I didn't want to do this.

I did not want to do this.

All right, so here we go.

"Google makes fixes to AI-generated search summaries after outlandish answers went viral."

"Threaten an AI chatbot and it will lie, cheat, and, quote, unquote, 'let you die in an effort to stop you,' study warns."

- Good.

- And the third one is, "Delaware Court rules in AI can be named as co-inventor on a patent."

- Okay.

I know number two is real, which is I know...

I've read the Anthropic report, so I'm cheating here.

So first of all, this is a terrifying study.

And let me first give credit to some of the AI companies like OpenAI and Anthropic, which are two of the big AI companies that are doing safety studies, mostly internally, to understand how are these AI bots and these agentic AIs going to respond in real-world situations.

And what both of these companies, Anthropic and AI, found is that- - Wait, Anthropic and OpenAI?

- Yeah.

- So what is Anthropic's AI?

- Yeah.

Anthropic does Claude.

OpenAI does ChatGPT.

- Okay.

- Two are basically the same thing.

Large language models, you ask it questions, it gives you responses.

Don't forget to say please.

So what they found... We should understand this because some of the media reports about this were incorrect, this did not actually happen.

So what Anthropic will do is they will create simulations and see how their AI responds in a simulation and, to Anthropic's point, also push to an extreme what we would call an edge case, but nevertheless disconcerting that when the AI was presented with information that it was going to be shut down, it tried to blackmail an engineer that it believed was trying to shut it down.

So it had that notion somehow of self-preservation, which was worrisome to begin with, but it was nothing compared to the second study that found that, given a situation where it could lock a door automatically that was deprived of oxygen, knowing an engineer was in it, it would do that if it thought the engineer was going to shut the system down.

- Holy cow.

- Yeah.

So this is what the science fiction movies were made of, that we were warned about.

And the AI systems are doing these things in a way that we don't understand.

They haven't done them, please understand, but they have the ability to do it, which is disconcerting.

And I think what is particularly worrisome here is that it has internalized this notion of self-preservation without it being programmed to do that, and that is the very thing that people are concerned about is that, when you ingest these massive amounts of data to train these systems, you don't know what these systems are learning.

And this is a perfect example of that.

- So what really bothers me is the story of blackmail.

Does that mean that the AI found personal information about this person?

- Yeah.

So Anthropic did something very clever.

They put some information into Anthropic's knowledge base about an extramarital affair by one of the engineers.

So they gave it the ability to do it and they saw, "Will you do it?"

It actually did.

It actually did, and I think even the folks at Anthropic... And, by the way, we've been talking about- - Wait, so Claude is a snitch?

That's what we're getting at.

- Claude is a rat.

It's going to rat you out.

- Wow.

- Yeah.

It's funny to some degree, but it is terrifying because...

Here's a scenario which is not that far off.

We have self-driving cars coming out.

So what if those self-driving cars are being run by the very AIs that we were just talking about, and there's an engineer in a car that it thinks is writing a kill switch.

Is it going to drive that car into a wall at 100 miles an hour because it doesn't want that engineer to do it?

This is not far-fetched, and I'm not saying the probability is high.

But if the probability is greater than zero, we have to have a very serious conversation.

And what we have learned about these safety studies is that the probability is non-zero, and that, I think, is worrisome.

- Well, this raises the obvious question of how do we control potential bad outcomes from AI?

- Okay.

So here's the reality of the systems today is I don't think we know how to do it.

Given the way the systems have been trained and the way they're being deployed, we don't know technically how to control them.

So here's what I think has to happen.

This is an imperfect solution, please understand.

But what we have learned from the physical world is that when you create liability for companies, they build safer products.

The reason why we have safer products that we bring into our home, that we ingest, medicines, et cetera, is because we created liability.

We told companies that, "You create a product that you knew or should have known as dangerous and it creates harm, we're going to sue you back to the dark ages."

We've not done that with the technology sector.

But if we tell these AI companies, "Your AI systems start creating damage, we're going to sue you back to the dark ages," they will then have to internalize that liability and they will start to build safer products.

We know this.

So I think some of this has to come from liability.

I think some of it has to come from regulation.

I think technology can get you there to a certain degree, but none of these are perfect systems.

- So it's almost like you're saying that there is an engineering workflow fix that has to be done.

It is like, "Okay, you need to install scrubbers in your smokestack," but now you're saying, "In your engineering, you need to behave certain ways with your training sets.

You need to-" - Or, not release products until you know they're safe.

When that plane taxis down the runway for the first time, it has gone through an awful lot of safety testing.

Cars are safety-tested to oblivion.

- But if the AI determines...

So essentially, if the AI determines a theory of mind and an idea of deception, then it could hide.

- It can, and that is terrifying.

So what if it knows you're putting it through a safety test and it's being deceptive.

What if it starts learning deception?

This is exactly what science fiction writers have been warning us for 50 years, and we have an inkling that this is starting to happen.

So does it mean you have to have a kill switch that can't be manipulated by AI.

Is that possible?

I don't have great answers here and, frankly, I don't think anybody has great answers, but we now have an inkling of the art of the possible.

So I think number two is real.

- Wait a minute now, wait a minute.

All right.

So you know the probability test, where if I give you three doors, and I reveal that one is not the answer, you're supposed to change your answer.

- I know.

I know, I'm not going to.

I think the first one is real and the third one is fake.

- You nailed it.

- Yeah.

- Not only did you nail which headline was the fake headline.

That headline was AI-generated.

It was not created by the NOVA production team.

- What was the prompt, by the way, to get it?

Do you know?

- I do not know.

- All right.

(both laughing) - That wasn't my job.

- I didn't mean to put you on the spot, man.

- I know.

It's fine.

The spot is fine.

It's okay by me.

I love when I say, "I don't know."

- Good.

- When I was in graduate school, I felt like the greatest thing that I learned was the difference between knowing and not knowing.

- I agree.

- Right?

It had never struck me before.

And you'd be surprised how many times I've said to people this thing about the difference between knowing and not knowing and they immediately go like, "Oh, yeah, I know the difference," and very rapidly- - I know.

- Illustrate that they don't.

- This is what I tell all my grad students.

The paradox of a PhD is you will feel stupider graduating than when you started- - Absolutely.

- Because you realize how much you don't know, and that's a gift.

It's a gift.

- It's a gift.

It's a gift to know that you don't know.

- And by the way, I'm not sure how long that third one is going to be true for.

You know, will AI be able to be a patent?

Will you be able to use AI as a lawyer in the court of law instead of hiring a lawyer?

I don't know for how long we're going to be able to hold off.

- Or, for example, you know, as a person who writes books, instead of having my blurb of what Bill Nye thought of my book, maybe we could say, "Oh, here's what Claude thinks."

- That is probably within the next six months, you're going to start seeing that.

- Wow.

- I don't know what the next five years are going to look like, but I think something is brewing and I think it's going to make the last 20 years of the computer, internet, and mobile revolution look like a day in the park.

I think something is happening.

I don't think anybody really fundamentally understands it, but there's hardly a day goes by where I don't have a conversation about this.

- Wait a minute, you are a co-founder of an AI startup.

So that means that you do have to have some sense of what's coming next, and I think you're holding back.

- I think I can see 6 to 12 months out, but even that is... We are more often wrong than we're right.

- So in the semiconductor field, we had Moore's Law.

Right?

- Yeah, that's right.

- Is there a similar metric in AI and what is its basis?

- Yeah, that's a great question.

- So in Moore's law, it was a number of transistors on a chip would increase, double every 18 months.

- Double every 12 to 18 months.

Okay, so Moore's Law, and, by the way, phenomenal that continues to be true.

By the way, how many times have people said Moore's law is over and they were wrong?

They were wrong, they're wrong, they're wrong, keeps going, keeps going.

So what's amazing about Moore's Law, and I'm glad you brought this up, is that we would measure the doubling of transistors on a chip in a 12 to 18-month period, and that was really sort of the way we tended to think about changes in technology.

The cadence was about every 12 to 18 months.

It is now every 12 to 18 days.

- No way.

- I mean, no kidding.

Think about how quickly the next version of ChatGPT, the next version of Claude, the next version of...

I mean, it is happening at a pace of weeks, weeks.

I'll give you an example of this, by the way.

Just before the Christmas break last year, so this is December of 2024, three students are in my office saying, "We want to work on detecting full-blown AI-generated video," what's called text-to-video, where you type a prompt and, instead of giving you an image, it gives you a full-blown video with audio, like, essentially making short clips.

I'm like, "This is a dumb problem.

There's no way this is going to be out in the next 18 months."

About a month ago, Google's Veo 3 emerged and they did it, and I was wrong.

I was wrong.

There it was a couple months later.

The students were right.

I got it wrong.

But here's what I know is I don't know what the equivalent of Moore's law is.

I don't know if it's 12 days or 24 days or 48 days, but we are seeing a rapid development.

Now, there are some people out there who are saying this is going to plateau, that you can only get so far with these mechanisms.

And there are other people who are saying it is endless, we're going to keep going.

But here's the thing you have to understand about AI, I mean, you know this as a former academic, as an academic, is that this all started in the academy.

All of these technologies started in the academy, but now where is it?

It's in Silicon Valley with hundreds and hundreds of billions of dollars being poured into it.

And that train doesn't stop.

When you start pouring that kind of resources and that kind of talent and that kind of money, this is not stopping.

- So who's doing it now?

Because if there are AIs that have not been released to us, that are being kept behind by these Silicon Valley companies, is it now AI that's building the AI?

Have we gotten to that point?

- Good.

First of all, yes.

So if you talk to leaders in Silicon Valley, they will tell you somewhere between 20 and 50% of code being written for the next generation is now being written by the AI systems.

Now, there are still humans in the loop, there are still humans that are being part of the designing, but a lot of what is happening and part of the reason for acceleration is that, when you have AI that helps you write code or writes it entirely, the systems just move faster.

- But what if they're keeping a secret from you?

Can it do that?

Can it make a leap?

- Yeah.

Yeah.

And, by the way, the thing you will hear from AI engineers is, "We don't understand our systems anymore.

They have gotten so complicated, they behave in ways," and this is not what you want to hear from engineers.

- No.

- That we don't understand why it does this.

We would not tolerate that, by the way, with aero engineers who are building airplanes, "Yeah, we don't really understand why it did that."

Well, we're going to start grounding planes until you figure that out- - That's true.

- But we don't do that with AI.

We're like, "Well, okay, we'll figure out as the plane is taxiing down the runway," and I say that as a person who's about to get on an airplane in a few hours.

I think you're absolutely right, and let's come back to that conversation, is that if the AI starts to have a sense of self-preservation and you unleash it to write the next generation of code, how is it going to incorporate that?

Is it going to start putting its own guardrails in place, so you can't shut it down?

So I come back to something I said earlier.

We have to start thinking about how to put reasonable guardrails on these technologies.

- What about a filter?

You can take your old AI that's had all this data and say, "We're going to pass that data through a filter in order to protect humanity."

And related to that is, I feel like there are clearly geopolitical implications here because it is a competitive world that we live in and we see now, with the wars in Ukraine and the wars in the Middle East, that there is this rapid innovation that's taking place.

Do you see what the future of geopolitical...

There is that document, AI 2027, that takes that stage of geopolitical competition, adds AI, and it ends with human extinction.

- Yeah, that's right.

So you can't look at what China's doing right now and the investment that the Chinese government is putting into AI and not take that seriously.

If you look in the academic literature now, you see a domination of published papers coming out of China.

They are investing phenomenal amount of resources and talent into this at a time, by the way, when we are doing almost exactly the opposite here in the United States.

We are cutting funding, we are cutting research, we are attacking universities that are the very place where this innovation happens.

We're moving in exactly the wrong direction.

I think there are serious geopolitical implications here because these AIs will do more than large language models and creating images.

They will start flying our jets, driving our tanks.

There will be the incorporation of AI into warfare, and it's already started.

It started here in the US, it's in China, it's in Russia.

This is coming.

And so, we have to start thinking very carefully... And by the way, if you go look at the EU AI safety bill, they did something that I really liked.

They have what's called a risk-based intervention.

So if you are using AI to predict what movie will Hany like to watch next, okay, fine, we don't need a lot of guardrails on that.

But if you are using AI to make critical life-saving decisions, hiring and firing decisions, criminal justice decisions, we need a much higher bar, and they have banned the use of AI in the military.

They have just said, "This is a line too far right now.

We are not willing to go there until we understand more."

So I think this is the way we want to think about this.

It's risk-based, right?

You put the AI here.

What is the implication?

Let me give you some examples.

People are using AI right now to make hiring and firing decisions.

Who gets an interview?

Do we fundamentally understand how these AI systems... Are they biased against certain individuals, certain categories?

We are using AI to make criminal justice decisions or sentencing decisions and bail decisions are being turned over to AI.

- And that's the thing that bugs me is, you know, it's almost like stereotypes.

They're based on averages, and averages don't tell you anything about an individual event.

- Yeah, 100% correct.

And let's go back to where we started this conversation, which is what these systems do is pattern matching.

They look at history and they repeat it.

So if you have had historical problems in a system, in the criminal justice system, in Silicon Valley with women making up less than 20% of the technical workforce, your AI is simply going to reproduce that because it's not thinking about any of these issues.

It's simply saying, "What is the pattern?

Reproduce the pattern."

- That's right.

- So when it comes to loan decisions, financial decisions, hiring and firing, university admissions, criminal justice, we, in my opinion, should not be unleashing AI systems without understanding them better than we do right now.

If it comes to help somebody write code, great.

If it comes to making predictions on what you're going to listen on Spotify next, fine.

I don't care.

- Right.

Let's go even bigger with the greatness.

- Good.

- So it's not all bad.

- No.

- AI has amazing...

When I look at society, I recognized, as a young man in my 20s, I was like, "There are certain things that we create: militaries, police forces, that by their very nature, religions, they have great good built into them and they have the ability for great bad built into them."

And our job is to harvest the good while keeping the bad, but recognizing that both are going to occur.

- 100%.

It's a mitigation strategy.

- It's a mitigation strategy.

So what does the positive future of AI look like and how can we bring that about?

- Good.

Okay.

So first of all, I 100% agree with you that everything we should be doing is how do we harness the power and mitigate the harm.

I would argue that for 20 years in the technology sector, we have harnessed the power, but we have not done a good job of mitigating harm.

I think the harm has been on balance with the good things that has happened when you look broadly at the issues that have come out of technology.

So let's go forward.

I have a prediction.

And you think I would learn not to make predictions, but here it is.

I think the next blockbuster movie is going to come from some kid in their bedroom making movies with this AI technology, because think about what you now have access to as somebody who has a creative mind.

You can now make full-blown videos with any character you want, any storyline, any narrative, any actor, any actress, and what else can you do?

You can distribute it, you can put it on TikTok, you can put it on YouTube.

So I think that the ability to create things without a multi, multi-million dollar budget, whether it's a book, music, a podcast, a full-feature film is going to be fully democratized.

- So essentially, as the human now, if you wanted to create something, you had to take your idea and then it had to be backed by a bunch of money and a bunch of skill and experts to bring it to fruition.

- 100% correct.

- So what you're saying now is that your mind and your AI business partner is all you got to need.

- That's all you need, right.

- So human creativity is going to be unleashed?

- I think it's going to be unleashed, and you're already seeing it.

And that's amazing.

I think if you wanted... We were talking recently about the company I co-founded.

You know, I had to go to somebody who had a lot of money and convince them to give me some of that money to start a company, so I could hire lots and lots of people to build this company over a multi-year process.

I think you're going to be able to start a company with one person.

And again, there are serious employment questions.

There's lots of...

But that's incredibly empowering.

And now, imagine that this can be somebody who is in Sub-Saharan Africa.

They don't need to have access to venture capitalists.

They don't need to have access to Hollywood studios.

They can do amazing things.

And so I think the creative energy now has zero barriers to entry very soon, and that's exciting.

- How do we put that into our education system broadly?

So one of the criticisms of our current education system that you see all over the internet is, "Hey, you know what they created?

Drones for working factories.

That's what our educational system is designed to do."

I don't know if that's true, but that's what's said all over.

But now, you know, if there's anything that we humans excel at, it is imagination.

Right?

And so now, we have this tool to bring our imagination to fruition.

So instead of having an education system that people say sort of, you know, tampers down our imagination- - Creativity, yeah.

- And creativity, how do we use AI?

Because another way that we could use AI is as an educator.

And here's another thing that's being talked about in the academic spaces, kids and their iPads, take away the kid's iPad and they're just all over the place.

How could we use AI, for example, to keep a brain from becoming addicted in that way?

- Yeah.

Good.

I'm going to back through those questions.

Let me talk about the second one first.

I'm glad you reminded me of this.

One of the hardest things about teaching is that, in a class of 50 students, you have a huge range of, let's call them, the top performing and the bottom performing students.

And it's not because they're smarter or dumber, it's because they come with different experiences.

- That's right, different training when you step at the door- - Different experiences.

And some of it, they naturally come to the material.

Some of them, it's less natural.

And one of the hardest things is how do you teach to all those students and keep them all coming, that the highest performing students, you know they're not bored, but the lowest performing students don't fall off the cliff.

Now, imagine that students can... All my lectures will be recorded and they will have an AI that summarizes it and that AI will understand what the student understands well and doesn't understand well, and it will be their essentially personalized tutor.

So if you're high performing, you move fast.

If you're lower performing, you move a little bit slower and we find ways to...

I love this idea as an educator.

I think it's incredible.

- So it's bespoke education.

- Bespoke education.

And to do the things that, even in a class of 50 students, I can't do.

I can't talk to every student three hours a week.

It's physically impossible.

- I know.

- Let alone a class, by the way, at UC Berkeley, with 2,000 students.

- Oh geez.

- You can't do this.

And so I think this is an enabling technology.

Now, of course, we have to be careful about it.

What information is it collecting about the kids?

Is it privacy-respecting?

And so you have to think about all these things.

- What about medicine?

Because one of the things that really is mind boggling to me is that I can't go to a doctor every year and say, "Hey, give me an ultrasound of all of my organs."

So, is it possible for us to have AI as a- - Yes.

Personalized medicine, yeah.

- Yeah, personalized medicine.

- And the answer is yes, and here's why, because a lot of that is pattern matching, right?

So, what is your age?

What is your ethnicity?

What is your- - What's your DNA?

- DNA?

What is your family history?

Do you smoke?

Do you drink?

What do you eat?

What do you not eat?

A lot of that are very good predictors.

And if the AI understands you and it is watching what you do, "How much do you exercise?

Are you getting enough of this?

Are you getting too much of that?"

it can start to say, "All right, it's time for an ultrasound right now."

Will it be perfect?

Of course not.

Will it be better than the system we have now?

Almost certainly, because it's highly personalized.

Now, again, we have to think, how do you do that in a way that's privacy-preserving?

Are you giving all that information up to an AI company that is going to feed that information to an insurance company that's going to affect your insurance rates because you're at high risk because you had a cigar that evening?

I say that as somebody who just had a cigar last night.

So, you know, we have to think about that.

So we have to think about the guardrails like, "How do we protect individuals while helping them?"

Because a lot of what happened in the last 20 years is we enabled things, we enabled phenomenal technologies with this, but we also created problems.

And so I think we just have to think about how to do that in a way that is respectful and careful.

- Man, this conversation is one that, for me, it is almost like when I watched the movie "The Matrix."

I stepped out of that movie thinking, "Am I living in reality or am I living in the Matrix?"

- I know.

- This is a conversation that just really left me with so much more that I want to explore and understand and keep track of.

And also, I feel like I need to go make a movie.

- Yeah, that's right.

That's what I'd say, and invite me back because I like talking to you.

No kidding, in a year from now, I think the future, five years from now, is going to look unbelievably different than today.

I think we are going through something.

And every once in a while and when I'm in San Francisco and I see these self-driving cars, I think, "Oh, my God, we're living in the future."

- And there's a generation being born today- - Yeah, into this.

- This is their normal.

- Yeah.

They're normal is, this is it.

I mean, think about, you and I grew up without these devices in our hands.

- That's right.

Yeah.

- Think about how different that world, it was just not that long ago.

- Well, listen, you probably confused the audience members because they look at me and they're like, "How has this teenager established himself?"

You know, like, this youthful beauty I possess.

Right?

But no, I'm a Gen X-er.

- That's the best generation, by the way, just saying.

- Gen X, I'm telling you, man, we did it without parents.

- That's right.

That's right.

(Hakeem laughing) You know what we did?

We had NOVA and PBS to raise us.

- We had Nova and PBS, absolutely.

Hey, man, I really appreciate you coming out.

This has been amazing.

- It's been great.

- Thank you so much.

- Thanks, Hakeem.

(upbeat music)