Original subtitles

Viewers like you make this program possible.

Support your local PBS station.

♪

MILES O'BRIEN: Machines that think like humans.

Our dream to create machines in our own image

that are smart and intelligent

goes back to antiquity.

Well, can it bring it to me?

O'BRIEN: Is it possible that the dream of artificial intelligence

has become reality?

They're able to do things

that we didn't think they could do.

MANOLIS KELLIS: Go was thought to be a game where machines would never win.

The number of choices for every move is enormous.

O'BRIEN: And now, the possibilities seem endless.

MUSTAFA SULEYMAN: And this is going to be

one of the greatest boosts

to productivity in the history of our species.

That looks like just a hint of some type of smoke.

O'BRIEN: Identifying problems before a human can...

LECIA SEQUIST: We taught the model to recognize

developing lung cancer.

O'BRIEN: ...and inventing new drugs.

PETRINA KAMYA: I never thought that we would be able

to be doing the things we're doing with A.I..

O'BRIEN: But along with the hope...

[imitating Obama]: This is a dangerous time.

O'BRIEN: ...comes deep concern.

One of the first drops in the feared flood

of A.I.-created disinformation.

We have lowered barriers to entry to manipulate reality.

We're going to live in a world where we don't know what's real.

The risks are uncertain and potentially enormous.

O'BRIEN: How powerful is A.I.? How does it work?

And how can we reap its extraordinary benefits...

Sybil looked here,

and anticipated that there would be a problem.

O'BRIEN: ...without jeopardizing our future?

"A.I. Revolution"

right now, on "NOVA!"

[whirring]

♪

Tell me the backstory on inflection A.I..

[voiceover]: Our story begins with the making of this story.

PI [on computer]: The story of Inflection A.I. is an exciting one.

O'BRIEN [voiceover]: I was researching

an interview subject.

Who is Mustafa Suleyman?

[voiceover]: Something I've done

a thousand times in my 40-year career.

PI [on computer]: Mustafa Suleyman is a true pioneer

in the field of artificial intelligence.

[voiceover]: But this time, it was different:

I wasn't typing out search terms.

What is machine learning?

O'BRIEN [voiceover]: I was having a conversation with a computer.

PI: Sounds like an exciting project, Miles.

[voiceover]: It felt like something big had changed.

PI: Machine learning, ML, is a type of artificial intelligence.

O'BRIEN [voiceover]: And as it happened,

I was focused on one of the innovators of this revolution.

Okay, so if I do this...

[voiceover]: Mustafa Suleyman is co-founder

of a startup called Inflection.

It makes an artificial intelligence assistant

called Pi.

So now you can speak...

O'BRIEN [voiceover]: I met them both in London.

It's fundamentally different, isn't it?

Yeah, it's a conversational style.

All of us humans learn through stories,

and through narrative, and through interactive dialogue.

And now, the machine can kind of come alive,

and talk to you about whatever it is

that's on top of your mind.

O'BRIEN: Tell me about the PBS program "NOVA."

[voiceover]: Chatbots can offer up quick answers, write poems,

finish essays, and translate languages

among many other things.

PI [over phone]: "NOVA" is a science documentary series...

O'BRIEN [voiceover]: They aren't perfect,

but they have put artificial intelligence in our hands,

and into the public consciousness.

And it seems we're equal parts leery

and intrigued.

SULEYMAN: A.I. is a tool

for helping us to understand the world around us,

predict what's likely to happen, and then invent

solutions that help improve the world around us.

My motivation was to try to use A.I. tools

to, uh, you know, invent the future.

The rise in artificial intelligence...

REPORTER: A.I. technology is developing...

O'BRIEN [voiceover]: Lately, it seems a dark future is already here...

The technology could replace millions of jobs...

O'BRIEN [voiceover]: ...if you listen to the news reporting.

The moment civilization was transformed.

O'BRIEN [voiceover]: So how can artificial intelligence help us,

and how might it hurt us?

At the center of the public handwringing:

how should we put guardrails around it?

We definitely need more regulations in place...

O'BRIEN [voiceover]: Artificial intelligence is moving fast

and changing the world.

Can we keep up?

Non-human minds smarter than our own.

O'BRIEN [voiceover]: The news coverage may make it seem like

artificial intelligence is something new.

At a moment of revolution...

O'BRIEN [voiceover]: But human beings have been thinking about this

for a very long time.

I have a very fine brain.

Our dream to create machines in our own image

that are smart and intelligent goes back to antiquity.

Uh, it's, it's something that has,

has permeated the evolution of society and of science.

[mortars firing]

O'BRIEN [voiceover]: The modern origins of artificial intelligence

can be traced back to World War II,

and the prodigious human brain of Alan Turing.

The legendary British mathematician

developed a machine

capable of deciphering coded messages from the Nazis.

After the war, he was among the first to predict computers

might one day match the human brain.

There are no surviving recordings of Turing's voice,

but in 1951, he gave a short lecture on BBC radio.

We asked an A.I.-generated voice to read a passage.

TURING A.I. VOICE: I think it is probable, for instance,

that at the end of the century,

it will be possible to program a machine

to answer questions in such a way

that it will be extremely difficult to guess

whether the answers are being given by a man

or by the machine.

O'BRIEN [voiceover]: And so, the Turing test was born.

Could anyone build a machine

that could converse with a human in a way

that is indistinguishable from another person?

In 1956,

a group of pioneering scientists spent the summer

brainstorming at Dartmouth College.

And they told the world that they have coined

a new academic field of study.

They called it artificial intelligence

O'BRIEN [voiceover]: For decades, their aspirations remained

far ahead of the capabilities of computers.

In 1978,

"NOVA" released its first film on artificial intelligence.

We have seen the first crude beginnings

of artificial intelligence...

O'BRIEN [voiceover]: And the legendary science fiction writer,

Arthur C. Clark was, as always, prescient.

It doesn't really exist yet at any level,

because our most complex computers are still morons,

high-speed morons, but still morons.

Nevertheless, we have the possibility of machines

which can outpace their creators,

and therefore, become more intelligent than us.

At the time, researchers were developing "expert systems,"

purpose-built to perform specific tasks.

So the thing that we need to do

to make machine understand, um, you know, our world,

is to put all our knowledge into a machine

and then provide it with some rules.

♪

O'BRIEN [voiceover]: Classic A.I. reached a pivotal moment in 1997

when an artificial intelligence program devised by IBM,

called "Deep Blue" defeated world chess champion

and grandmaster Garry Kasparov.

It searched about 200 million positions a second,

navigating through a tree of possibilities

to determine the best move.

RUS: The program analyzed the board configuration,

could project forward millions of moves

to examine millions of possibilities,

and then picked the best path.

O'BRIEN [voiceover]: Effective, but brittle,

Deep Blue wasn't strategizing as a human does.

From the outset, artificial intelligence researchers

imagined making machines

that think like us.

The human brain, with more than 80 billion neurons,

learns not by following rules,

but rather by taking in a steady stream of data,

and looking for patterns.

KELLIS: The way that learning actually works

in the human brain is by updating the weights

of the synaptic connections

that are underlying this neural network.

O'BRIEN [voiceover]: Manolis Kellis is a Professor of Computer Science

at the Massachusetts Institute of Technology.

So we have trillions of parameters in our brain

that we can adjust based on experience.

I'm getting a reward.

I will update the strength of the connections

that led to this reward-- I'm getting punished,

I will diminish the strength of the connections

that led to the punishment.

So this is the original neural network.

We did not invent it, we, you know, we inherited it.

O'BRIEN [voiceover]: But could an artificial neural network

be made in our own image? Turing imagined it.

But computers were nowhere near

powerful enough to do it until recently.

It's only with the advent of extraordinary data sets

that we have, uh, since the early 2000s,

that we were able to build up enough images,

enough annotations,

enough text to be able to finally train

these sufficiently powerful models.

O'BRIEN [voiceover]: An artificial neural network is, in fact,

modeled on the human brain.

It uses interconnected nodes, or neurons,

that communicate with each other.

Each node receives inputs from other nodes

and processes those inputs to produce outputs,

which are then passed on to still other nodes.

It learns by adjusting the strength of the connections

between the nodes based on the data it is exposed to.

This process of adjusting the connections

is called training,

and it allows an artificial neural network

to recognize patterns and learn from its experiences

like humans do.

A child, how is it learning so fast?

It is learning so fast

because it's constantly predicting the future

and then seeing what happens

and updating their weights in their neural network

based on what just happened.

Now you can take this

self-supervised learning paradigm

and apply it to machines.

O'BRIEN [voiceover]: At first, some of these artificial neural networks

were trained on vintage Atari video games

like "Space Invaders"

and "Breakout."

Games reduce the complexity of the real world

to a very narrow set of actions that can be taken.

O'BRIEN [voiceover]: Before he started Inflection,

Mustafa Suleyman co-founded a company called

DeepMind in 2010.

It was acquired by Google four years later.

When an A.I. plays a game,

we show it frame-by-frame, every pixel

in the moving image.

And so the A.I. learns to associate pixels

with actions that it can take

moving left or right or pressing the fire button.

O'BRIEN [voiceover]: When it obliterates blocks or shoots aliens,

the connections between the nodes that enabled that success

are strengthened.

In other words, it is rewarded.

When it fails, no reward.

Eventually, all those reinforced connections

overrule the weaker ones.

The program has learned how to win.

This sort of repeated allocation of reward

for repetitive behavior is a great way to train a dog.

It's a great way to teach a kid.

It's a great way for us as adults to adapt our behavior.

And in fact, it's actually a good way

to train machine learning algorithms to get better.

O'BRIEN [voiceover]: In 2014, DeepMind began work on an artificial neural network

called "AlphaGo"

that could play the ancient,

and deceptively complex, board game of Go.

KELLIS: Go was thought to be a game where machines would never win.

The number of choices for every move is enormous.

O'BRIEN [voiceover]: But at DeepMind,

they were counting on

the astounding growth of compute power.

And I think that's the key concept to try to grasp,

is that we are massively, exponentially growing

the amount of computation used, and in some sense,

that computation is a proxy

for how intelligent the model is.

O'BRIEN [voiceover]: AlphaGo was trained two ways.

First, it was fed a large data set of expert Go games

so that it could learn how to play the game.

This is known as supervised learning.

Then the software played against itself many millions of times,

so-called reinforcement learning.

This gradually improved its skills and strategies.

In March 2016,

AlphaGo faced Lee Sedol,

one of the world's top-ranking players

in a five-game match in Seoul, South Korea.

AlphaGo not only won,

but also made a move so novel, the Go cognoscenti

thought it was a huge blunder.That's a very surprising move.

There's no question to me that these A.I. models

are creative.

They're incredibly creative.

O'BRIEN [voiceover]: It turns out the move was a stroke of brilliance.

And this emergent creative behavior

was a hint of what was to come:

generative A.I.

Meanwhile,

a company called OpenA.I. was creating

a generative A.I. model

that would become ChatGPT.

It allows users to engage in a dialogue

with a machine that seems uncannily human.

It was first released in 2018,

but it was a subsequent version that became a global sensation

in late 2022.

This promises to be the viral sensation

that could completely reset how we do things.

Cranking out entire essays

in a matter of seconds.

O'BRIEN [voiceover]: Not only did it wow the public, it also caught

artificial intelligence innovators off guard.

YOSHUA BENGIO: It surprised me a lot

that they're able to do things that

we didn't think they could do simply by

learning to imitate how humans respond.

And I thought this kind of abilities would take

many more years or decades.

O'BRIEN [voiceover]: ChatGPT is a large language model.

LLMs start by consuming massive amounts of text:

books, articles and websites,

which are publicly available on the internet.

By recognizing patterns in billions of words,

they can make guesses at the next word in a sentence.

That's how ChatGPT generates unique answers

to your questions.

If I ask for a haiku about the blue sky

it writes something that seems completely original.

KELLIS: If you're good at predicting

this next word,

it means you're understanding something about the sentence.

What the style of the sentence is,

what the feeling of the sentence is.

And you can't tell whether this was a human or a machine.

That's basically the definition of the Turing test.

O'BRIEN [voiceover]: So, how is this changing our world?

Well, It might change my world-- as an arm amputee.

Ready for my casting call,right?

MONROE [chuckling]: Yes.

Let's do it.All right.

O'BRIEN [voiceover]: That's Brian Monroe of the Hanger Clinic.

He's been my prosthetist

since an injury took my arm above the elbow

ten years ago.

So what we're going to do today is take a mold of your arm.Uh-huh.

Kind of is like a cast for a broken bone.

O'BRIEN [voiceover]: Up until now, I have used a body-powered prosthetic.

Harness and a cable allow me to move it

by shrugging my shoulders.

The technology is more than a century old.

But artificial intelligence,

coupled with small electric motors,

is finally pushing prosthetics into the 21st century.

Which brings me to Chicago

and the offices of a small company called Coapt.

I met the C.E.O., Blair Locke,

a pioneer in the push to apply artificial intelligence

to artificial limbs.

So, what do we have here? What are we going to do?

This allows us to very easily test how your control would be

using a pretty simple cuff; this has electrodes in it,

and we'll let the power of the electronics

that are doing the machine learning

see what you're capable of.All right, let's give it a try.

[voiceover]: Like most amputees,

I feel my missing hand almost as if it was still there--

a phantom.

Everything will touch. Is that okay?

Yeah.Not too tight?

No. All good.Okay.

O'BRIEN [voiceover]: It's almost entirely immobile, stuck in molasses.

Make a fist, not too hard.

O'BRIEN [voiceover]: But I am able to imagine moving it ever so slightly.

And I'm gonna have you squeeze into that a little bit harder.

Very good, and I see the pattern on the screen

change a little bit.

O'BRIEN [voiceover]: And when I do,

I generate an array of faint electrical signals in my stump.

That's your muscle information.

It feels, it feels like I'm overcoming

something that's really stuck.

I don't know, is that enough signal?

Should be. Oh, okay.

We don't need a lot of signal,

we're going for information

in the signal, not how loud it is.

O'BRIEN [voiceover]: And this is where artificial intelligence comes in.

Using a virtual prosthetic depicted on a screen,

I trained a machine learning algorithm to become fluent

in the language of my nerves and muscles.

We see eight different signals on the screen.

All eight of those sensor sites

are going to feed in together

and let the algorithm sort out the data.

What you are experiencing

is your ability to teach the system

what is hand-closed to you.

And that's different than what it would be to me.

O'BRIEN [voiceover]: I told the software what motion I desired,

open, close, or rotate,

then imagined moving my phantom limb accordingly.

This generates an array of electromyographic,

or EMG, signals in my remaining muscles.

I was training the A.I. to connect the pattern

of these electrical signals with a specific movement.

LOCK: The system adapts,

and as you add more data and use it over time,

it becomes more robust,

and it learns to improve upon use.

O'BRIEN: Is it me that's learning, or the algorithm that's learning?

Or are we learning together?LOCK: You're learning together.

Okay.

O'BRIEN [voiceover]: So, how does the Coapt pattern recognition system work?

It's called a Bayesian classification model.

As I train the software,

it labels my various EMG patterns

into corresponding classes of movement--

hand open, hand closed, wrist rotation, for example.

As I use the arm,

it compares the electrical signals I'm transmitting

to the existing library of classifications I taught it.

It relies on statistical probability

to choose the best match.

And this is just one way machine learning

is quietly revolutionizing medicine.

Computer scientist Regina Barzilay

first started working on artificial intelligence

in the 1990s, just as rule-based A.I. like Deep Blue

was giving way to neural networks.

She used the techniques

to decipher dead languages.

You might call it a small language model.

Something that is fun and intellectually very challenging,

but it's not like it's going to change our life.

O'BRIEN [voiceover]: And then her life changed in an instant.

CONSTANCE LEHMAN: We see a spot there.

O'BRIEN [voiceover]: In 2014, she was diagnosed with breast cancer.

BARZILAY [voiceover]: When you go through the treatment,

there are a lot of people who are suffering.

I was interested in

what I can do about it, and clearly it was not continuing

deciphering dead languages,

and it was quite a journey.

O'BRIEN [voiceover]: Not surprisingly, she began that journey with mammograms.

LEHMAN: It's a little bit more prominent.

O'BRIEN [voiceover]: She and Constance Lehman,

a radiologist at Massachusetts General Hospital,

realized the Achilles heel in the diagnostic system

is the human eye.

BARZILAY [voiceover]: So the question that we ask is,

what is the likelihood of these patients

to develop cancer within the next five years?

We, with our human eyes,

cannot really make these assertions

because the patterns are so subtle.

LEHMAN:Now, is that different from the surrounding tissue?

O'BRIEN [voiceover]: It's a perfect use case for pattern recognition

using what is known as a convolutional neural network.

♪

Here's an example of how CNNs get smart:

they comb through a picture with many virtual magnifying glasses.

Each one is looking for a specific kind of puzzle piece,

like an edge, a shape, or a texture.

Then it makes simplified versions,

repeating the process on larger and larger sections.

Eventually the puzzle can be assembled.

And it's time to make a guess.

Is it a cat? A dog? A tree?

Sometimes the guess is right, but sometimes it's wrong.

And here's the learning part:

with a process called backpropagation,

labeled images are sent back to correct the previous operation.

So the next time it plays the guessing game,

it will be even better.

To validate the model, Regina and her team gathered up

more than 128,000 mammograms

collected at seven sites in four countries.

More than 3,800 of them led to a cancer diagnosis

within five years.

You just give to it the image,

and then the five years of outcomes,

and it can learn the likelihood of getting a cancer diagnosis.

O'BRIEN [voiceover]: The software, called Mirai, was a success.

In fact, it is between 75% and 84% accurate

in predicting future cancer diagnoses.

Then, a friend of Regina's developed lung cancer.

SEQUIST: In lung cancer, it's actually

sort of mind boggling how much has changed.

O'BRIEN [voiceover]: Her friend saw oncologist Lecia Sequist.

She and Regina wondered

if artificial intelligence could be applied

to CAT scans of patients' lungs.

SEQUIST: We taught the model

to recognize the patterns of developing lung cancer

by using thousands of CAT scans

from patients who were participating

in a clinical trial.

From the new study? Oh, interesting.Correct.

SEQUIST [voiceover]: We had a lot of information about them.

We had demographic information,

we had health information,

and we had outcomes information.

O'BRIEN [voiceover]: They call the model Sibyl.

In the retrospective study, right,

so the retrospective data...

O'BRIEN [voiceover]: Radiologist Florian Fintelmann

showed me what it can do.

FINTELMANN: This is earlier, and this is later.

There is nothing

that I can perceive, pick up, or describe.

There's no, what we call, a precursor lesion

on this CT scan.

Sibyl looked here

and then anticipated that there would be a problem

based on the baseline scan.What is it seeing?

That's the million dollar question.

And, and maybe not the million dollar question.

Does it really matter? Does it?

O'BRIEN [voiceover]: When they compared the predictions

to actual outcomes from previous cases, Sybil fared well.

It correctly forecast cancer

between 80% and 95% of the time,

depending on the population it studied.

The technique is still in the trial phase.

But once it is deployed,

it could provide a potent tool for prevention.

The hope is that if you can predict very early on

that the patient is in the wrong way,

you can do clinical trials, you can develop the drugs

that are doing the prevention, rather than treatment

of very advanced disease that we are doing today.

O'BRIEN [voiceover]: Which takes us back to DeepMind and AlphaGo.

The fun and games were just the beginning,

a means to an end.

We have always set out at DeepMind

to, um, use our technologies to make the world a better place.

O'BRIEN [voiceover]: In 2021,

the company released AlphaFold.

It is pattern recognition software

designed to make it easier for researchers

to understand proteins,

long chains of amino acids

involved in nearly every function in our bodies.

How a protein folds

into a specific, three-dimensional shape

determines how it interacts with other molecules.

SULEYMAN: There's this correlation between

what the protein does and how it's structured.

So if we can predict how the protein folds,

then say something about their function.

O'BRIEN: If we know how a disease's protein is shaped, or folded,

we can sometimes create a drug to disable it.

But the shape of millions of proteins remained a mystery.

DeepMind trained AlphaFold

on thousands of known protein structures.

It leveraged this knowledge to predict

200 million protein structures,

nearly all the proteins known to science.

SULEYMAN: You take some high-quality known data,

and you use that to, you know,

make a prediction about how a similar piece of information

is likely to unfold over some time series,

and the structure of proteins is,

you know, in that sense, no different to

making a prediction in the game of Go or in Atari

or in a mammography scan,

or indeed, in a large language model.

KAMYA: These thin sticks here?

Yeah?They represent the amino acids

that make up a protein.

O'BRIEN [voiceover]: Theoretical chemist

Petrina Kamya works for a company called

Insilico Medicine.

It uses AlphaFold

and its own deep-learning models

to make accurate predictions about protein structures.

What we're doing in drug design is we're designing a molecule

that is analogous to the natural molecule

that binds to the protein,

but instead it will lock it, if this molecule

is involved in a disease where it's hyperactive.

O'BRIEN [voiceover]: If the molecule fits well,

it can inhibit the disease-causing proteins.

So you're filtering it down

like you're choosing an Airbnb or something to,

you know, number of bedrooms, whatever.To suit your needs.

[laughs]Exactly, right.

Right, yeah.That's a very good analogy.

It's sort of like Airbnb.

So you are putting in your criteria,

and then Airbnb will filter out

all the different properties

based on your criteria.

So you can be very, very restrictive

or you can be very, very free...Right.

In terms of guiding the generative algorithms

and telling them what types of molecules

you want them to generate.

O'BRIEN [voiceover]: It will take 48 to 72 hours of computing time

to identify the best candidates ranked in order.

How long would it have taken you

to figure that out as a computational chemist?

I would have thought of some of these,

but not all of them.Okay.

O'BRIEN [voiceover]: While there are no shortcuts for human trials,

nor should we hope for that,

this could greatly speed up the drug development pipeline.

There will not be the need to invest so heavily

in preclinical discovery,

and so, drugs can therefore be cheaper.

And you can go after those diseases

that are otherwise neglected,

because you don't have to invest so heavily

in order for you to come up with a drug,

a viable drug.

O'BRIEN [voiceover]: But medicine isn't the only place

where A.I. is breaking new frontiers.

It's conducting financial analysis,

helps with fraud detection.

[mechanical whirring]

It's now being deployed to discover novel materials

and could help us build clean energy technology.

And It is even helping to save lives

as the climate crisis boils over.

[indistinct radio chatter]

In St. Helena, California,

dispatchers at the

CAL FIRE Sonoma-Lake-Napa Command Center

caught a break in 2023.

Wildfires blackened nearly 700 acres of their territory.

We were at 400,000 acres in 2020.

Something like that would generate a response from us...

O'BRIEN [voiceover]: Chief Mike Marcucci has been fighting fires

for more than 30 years.

MARCUCCI [voiceover]: Once we started having these devastating fires,

we needed more intel.

The need for intelligence

is just overwhelming in today's fire service.

O'BRIEN [voiceover]: Over the past 20 years,

California has installed a network

of more than 1,000 remotely operated

pan, tilt, zoom surveillance cameras on mountaintops.

PETE AVANSINO: Vegetation fire, Highway 29 at Doton Road.

O'BRIEN [voiceover]: All those cameras generate petabytes of video.

CAL FIRE partnered with scientists at U.C. San Diego

to train a neural network

to spot the early signs of trouble.

It's called ALERT California.

SeLEGUE: So here's one that just popped up.

Here's an anomaly.

O'BRIEN [voiceover]: CAL FIRE Staff Chief of Fire and Intelligence Philip SeLegue

showed me how it works while it was in action,

detecting nascent fires,

micro fires.

That looks like just a little hint

of some type of smoke that was there...

O'BRIEN [voiceover]: Based on this, dispatchers can orchestrate

a fast response.

A.I. has given us the ability to detect and to see

where those fires are starting.

AVANSINO: Transport 1447 responding via MDC.

O'BRIEN [voiceover]: For all they know,

they have nipped some megafires in the bud.

The success are the fires

that you don't hear about in the news.

O'BRIEN [voiceover]: Artificial intelligence

can't put out wildfires just yet.

Human firefighters still need to do that job.

But researchers are pushing hard

to combine neural networks

with mobility and dexterity.

This is where people get nervous.

Will they take our jobs?

Or could they turn against us?

But at M.I.T.,

they're exploring ideas to make robots

good human partners.

We are interested in making machines

that help people with physical and cognitive tasks.

So this is really great,

it has the stiffness that we wanted...

O'BRIEN [voiceover]: Daniela Rus is director of M.I.T.'s Computer Science

and Artificial Intelligence Lab.

Oh, can you bring it to me?

O'BRIEN [voiceover]: CSAIL.

They are different, like, kind of like muscles

or actuators.

RUS [voiceover]: We can do so much more

when we get people and machines working together.

We can get better reach.

We can get lift,

precision, strength, vision.

All of these are physical superpowers

we can get through machines.

O'BRIEN [voiceover]: So, they're focusing

on making it safe for humans

to work in close proximity to machines.

They're using some of the technology that's inside

my prosthetic arm.

Electrodes that can read

the faint EMG signals generated

as our nerves command

our muscles to move.

They have the capability to interact with a human,

to understand the human,

to step in and help the human as needed.

I am at your disposal with 187 other languages,

along with their various

dialects and sub tongues.

O'BRIEN [voiceover]: But making robots as useful

as they are in the movies is a big challenge.

♪

Most neural networks run on powerful supercomputers--

thousands of processors occupying entire buildings.

RUS: We have brains that require

massive computation, which you cannot include

on a self-contained body.

We address the size challenge by

making liquid networks.

O'BRIEN [voiceover]: Liquid networks.

So it looks like an autonomous vehicle

like I've seen before,

but it is a little different, right?

ALEXANDER AMINI: Very different.

This is an autonomous vehicle

that can drive in brand-new environments

that it has never seen before for the first time.

O'BRIEN [voiceover]: Most self-driving cars today rely,

to some extent, on detailed databases

that help them recognize their immediate environment.

Those robot cars get lost in unfamiliar terrain.

O'BRIEN: In this case, you're not relying on

a huge, expansive neural network.

You're running on 19 neurons, right?

Correct.

O'BRIEN [voiceover]: Computer scientist Alexander Amini

took me on a ride in an autonomous vehicle

with a liquid neural network brain.

AMINI: We've become very accustomed to relying on

big, giant data centers and cloud compute.

But in an autonomous vehicle,

you cannot make such assumptions, right?

You need to be able to operate,

even if you lose internet connectivity

and you cannot talk to the cloud anymore,

your entire neural network,

the brain of the car, needs to live on the car,

and that imposes a lot of interesting constraints.

O'BRIEN [voiceover]: To build a brain smart enough

and small enough to do this job,

they took some inspiration from nature,

a lowly worm called C. elegans.

Its brain contains all of 300 neurons,

but it's a very different kind of neuron.

It can capture more complex behaviors

in every single piece of that puzzle.

And also the wiring,

how a neuron talks to another neuron

is completely different than what we see

in today's neural networks.

O'BRIEN [voiceover]: Autonomous cars that tap into today's neural networks

require huge amounts of compute power in the cloud.

But this car is using just 19 liquid neurons.

A worm at the wheel... sort of.

AMINI [voiceover]: Today's A.I. models

are really pushing the boundaries

of the scale of compute that we have.

They're also pushing the boundaries

of the data sets that we have.

And that's not sustainable,

because ultimately, we need to deploy A.I.

onto the device itself, right?

Onto the cars, onto the surgical robots.

All of these edge devices

that actually makes the decisions.

O'BRIEN [voiceover]: The A.I. worm may, in fact,

turn.

The portability of artificial intelligence

was on my mind when it came time

to pick up my new myoelectric arm...

equipped with Coapt A.I. pattern recognition.

All right, let's just check this

real quick...

O'BRIEN [voiceover]: A few weeks after

my trip to Chicago,

I met Brian Monroe

at his home office outside Washington, D.C.

Are you happy with the way it came out?Yeah.

Would you tell me otherwise?

[laughing]: Yeah, I would, yeah...

O'BRIEN [voiceover]: As usual,

he did a great job making a tight socket.

How's the socket feel? Does it feel like

it's sliding down or

falling out...No, it fits like a glove.

O'BRIEN [voiceover]: It's really important in this case,

because the electrodes designed to read the signals

from my muscles...

...have to stay in place snugly

in order to generate accurate, reliable commands

to the actuators in my new hand.

Wait, is that you?That's me.

[voiceover]: He also provided me with

a human-like bionic hand.

But getting it to work just right

took some time.

That's open and it's closing.

It's backwards?

Yeah.Now try.

If it's reversed,

I can swap the electrodes.There we go.

That's got it.Is it the right direction?

Yeah.Uh-huh. Okay.

O'BRIEN [voiceover]: It's a long way from the movies,

and I'm no Luke Skywalker.

But my new arm and I are now together.

And I'm heartened to know

that I have the freedom and independence

to teach and tweak it

on my own.

That's kind of cool.Yeah.

[voiceover]: Hopefully we will listen to each other.

It's pretty awesome.

O'BRIEN [voiceover]: But we might want to listen

with a skeptical ear.

JORDAN PEELE [imitating Obama]: You see, I would never say these things,

at least not in a public address,

but someone else would.

Someone like Jordan Peele.

This is a dangerous time.

O'BRIEN [voiceover]: It's even more dangerous now than it was in 2018

when comedian Jordan Peele

combined his pitch-perfect Obama impression

with A.I. software to make this convincing fake video.

...or whether we become some kind of [bleep] up dystopia.

♪

O'BRIEN [voiceover]: Fakes are about as old as

photography itself.

Mussolini, Hitler, and Stalin

all ordered that pictures be doctored or redacted,

erasing those who fell out of favor,

consolidating power,

manipulating their followers through images.

HANY FARID: They've always been manipulated,

throughout history, but--

there was literally, you can count on one hand,

the number of people in the world who could do this.

But now, you need almost no skill.

And we said, "Give us an image

"of a middle-aged woman, newscaster,

sitting at her desk, reading the news."

O'BRIEN [voiceover]: Hany Farid is a professor of computer science

at U.C. Berkeley.

[on computer]: And this is your daily dose of future flash.

O'BRIEN [voiceover]: He and his team

are trying to navigate the house of mirrors

that is the world of A.I.-enabled deepfake imagery.

Not perfect.

She's not blinking, but it's pretty good.

And by the way, he did this in a day and a half.

FARID [voiceover]: It's the classic automation story.

We have lowered barriers to entry

to manipulate reality.

And when you do that,

more and more people will do it.

Some good people will do it,

but lots of bad people will do it.

There'll be some interesting use cases,

and there'll be a lot of nefarious use cases.

Okay, so, um...

Glasses off. How's the framing?

Everything okay?

[voiceover]: About a week before

I got on a plane to see him...Hold on.

O'BRIEN [voiceover]: He asked me to meet him on Zoom

so he could get a good recording

of my voice and mannerisms.

And I assume you're recording, Miles.

O'BRIEN [voiceover]: And he turned the table on me a little bit,

asking me a lot of questions

to get a good sampling.

FARID [on computer]: How are you feeling about

the role of A.I. as it enters into our world

on a daily basis?

I think it's very important, first of all,

to calibrate the concern level.

Let's take it away from the "Terminator" scenario...

[voiceover]: The "Terminator" scenario.

Come with me if you want to live.

O'BRIEN [voiceover]: You know, a malevolent neural network

hellbent on exterminating humanity.

You're really real.

O'BRIEN [voiceover]: In the film series,

the cyborg assassin

is memorably played by Arnold Schwarzenegger.

Hany thought it would be fun

to use A.I. to turn Arnold into me.

Okay.

O'BRIEN [voiceover]: A week later, I showed up at

Berkeley's School of Information,

ironically located in the oldest building on campus.

So you had me do this strange thing on Zoom.

Here I am. What did you do with me?

Yeah, well, it's gonna teach you

to let me record your Zoom call, isn't it?

I did this with some trepidation.

[voiceover]: I was excited to see what tricks

were up his sleeve.

FARID [voiceover]: I uploaded 90 seconds of audio,

and I clicked a box saying

"Miles has given me permission to use his voice,"

which I don't actually

think you did. [chuckles]

Um, and, I waited about, eh, maybe 20 seconds,

and it said, "Okay, what would you like for Miles to say?"

And I started typing,

and I generated an audio of you saying

whatever I wanted you to say.

We are synthesizing,

at much, much lower resolution.

O'BRIEN [voiceover]: You could have knocked me over

with a feather when I watched this.

A.I. O'BRIEN: Terminators were science fiction back then,

but if you follow the recent A.I. media coverage,

you might think that Terminators are just around the corner.

The reality is...

O'BRIEN [voiceover]: The eyes and the mouth need some work,

but it sure does sound like me.

And consider what happened in May of 2023.

Someone posted this A.I.-generated image

of what appeared to be a terrorist bombing

at the Pentagon.

NEWS ANCHOR: Today we may have witnessed

one of the first drops in the feared flood

of A.I.-created disinformation.

O'BRIEN [voiceover]: It was shared on Twitter

via what seemed to be

a verified account from Bloomberg News.

NEWS ANCHOR: It only took seconds to spread fast.

The Dow now down about 200 points...

Two minutes later, the stock market dropped

a half a trillion dollars

from a single fake image.

Anybody could've made that image,

whether it was intentionally manipulating the market

or unintentionally,

in some ways, it doesn't really matter.

O'BRIEN [voiceover]: So what are the technological

innovations that make this tool widely available?

One technique is called

the generative adversarial network,

or GAN.

Two algorithms

in a dizzying student-teacher back and forth.

Let's say it's learning how to generate a cat.

FARID: And it starts by just splatting down

a bunch of pixels onto a canvas.

And it sends it over to a discriminator.

And the discriminator has access

to millions and millions of images

of the category that you want.

And it says,

"Nope, that doesn't look like all these other things."

So it goes back to the generator and says, "Try again."

Modifies some pixels,

sends it back to the discriminator,

and they do this in what's called

an adversarial loop.

O'BRIEN [voiceover]: And eventually,

after many thousands of volleys,

the generator finally serves up a cat.

And the discriminator says,

"Do more like that."

Today, we have a whole new way of doing these things.

They're called diffusion-based.

What diffusion does

is it has vacuumed up billions of images

with captions that are descriptive.

O'BRIEN [voiceover]: It starts by making those labeled images

visually noisy on purpose.

FARID: And then it corrupts it more, and it goes backwards

and corrupts it more, and goes backwards

and corrupts it more and goes backwards--

and it does that six billion times.

O'BRIEN [voiceover]: Eventually it corrupts it

so it's unrecognizable from the original image.

Now that it knows how to turn an image into nothing,

it can reverse the process,

turning seemingly nothing, into a beautiful image.

FARID: What it's learned is how to take

a completely indescript image, just pure noise,

and go back to a coherent image, conditioned on a text prompt.

You're basically reverse engineering an image

down to the pixel.

Yeah, exactly, yeah.

And it's-- and by the way-- if you had asked me,

"Will this work?" I would have said,

"No, there's no way this system works."

It just, it just doesn't seem like it should work.

And that's sort of the magic

of when you get this much data

and very powerful algorithms and very powerful computing

to be able to crunch these massive data sets.

I mean, we're not going to contain it.

That's done.

[voiceover]: I sat down with Hany

and two of his grad students:

Justin Norman and Sarah Barrington.

We looked at some the A.I. trickery

they have seen and made.

Somebody else wrote some base code

and they got grew on to

and grow on to and grow on to and eventually...

O'BRIEN [voiceover]: In a world where anything

can be manipulated with such ease

and seeming authenticity,

how are we to know what's real anymore?

How you look at the world,

how you interact with people in it,

and where you look for your threats of that change.

O'BRIEN [voiceover]: Generative A.I. is now part of a larger ecosystem

that is built on mistrust.

We're going to live in a world where

we don't know what's real.

FARID [voiceover]: There is distrust of governments,

there is distrust of media,

there is distrust of academics.

And now throw on top of that video evidence.

So-called video evidence.

I think this is the very definition

of throwing jet fuel onto a dumpster fire.

And it's already happening,

and I imagine we will see more of it.

[Arnold's voice]: Come with me if you want to live.

O'BRIEN [voiceover]: But it also can be

kind of fun.

As Hany promised,

here's my face

on the Terminator's body.

[gunfire blasting]

Long before A.I. might take

an existential turn against humanity,

we will need to reckon with the likes...

Go! Now!O'BRIEN [voiceover]: Of the Milesinator.

TRAILER NARRATOR: This time, he's back.

[booming]

O'BRIEN [voiceover]: Who will no doubt, be back.

Trust me.

O'BRIEN [voiceover]: Trust,

but always verify.

So, what kind of A.I. magic

is readily available online?

It's pretty simple to make it look

like you're fluent in another language.

[speaking Mandarin]:

It was pretty easy to do,

I just had to upload a video and wait.

[speaking German]:

And, suddenly, I look pretty darn smart.

[speaking Greek]:

Sure, it's fun, but I think you can see

where it leads to mischief

and possibly even mayhem.

[voiceover]: Yoshua Bengio is an artificial intelligence pioneer.

He says he didn't spend much time

thinking about science fiction dystopia

as he was creating the technology.

But as his brilliant ideas became reality,

reality set in.

BENGIO: And the more I read,

the more I thought about it...

the more concerned I got.

If we are not honest with ourselves,

we're gonna fool ourselves.

We're gonna... lose.

O'BRIEN [voiceover]: Avoiding that outcome

is now his main priority.

He has signed several public warnings

issued by A.I. thought leaders,

including this stark single-sentence statement

in May of 2023.

"Mitigating the risk of extinction from A.I.

"should be a global priority

"alongside other societal scale risks,

"such as pandemics

and nuclear war."

As we approach more and more capable A.I. systems

that might even become stronger than humans in many areas,

they become more and more dangerous.

Can't we just pull the plug on the thing?

Oh, that's the safest thing to do,

pull the plug.

Before it gets so powerful that

it prevents us from pulling the plug.

DAVE: Open the pod bay doors, Hal.

HAL: I'm sorry, Dave,

I'm afraid I can't do that.

O'BRIEN [voiceover]: It may be some time

before computers are able

to act like movie supervillains...

HAL: Goodbye.

O'BRIEN [voiceover]: But there are near-term dangers already emerging.

Besides deepfakes and misinformation,

A.I. can also supercharge bias and hate content,

replace human jobs...

This is why we're striking, everybody.[crowd exclaiming]

O'BRIEN [voiceover]: And make it easier

for terrorists to create bioweapons.

And A.I. systems are so complex

that they are difficult to comprehend,

all but impossible to audit.

RUS [voiceover]: Nobody really understands

how those systems reach their decisions.

So we have to be much more thoughtful

about how we test and evaluate them

before releasing them.

They're concerned whether machine will be able

to begin to think for itself.

O'BRIEN [voiceover]: The U.S. and Europe have begun charting a strategy

to try to ensure safe, secure,

and trustworthy artificial intelligence.

RISHI SUNAK: ...in a way that will be safe for our communities...

O'BRIEN [voiceover]: But how to do that

in the midst of a frenetic race

to dominate a technology

with a predicted economic impact

of 13 trillion dollars by 2030.

There is such a strong commercial incentive

to develop this and win the competition

against the other companies,

not to mention the other countries,

that it's hard to stop that train.

But that's what governments should be doing.

NEWS ANCHOR: The titans of social media

didn't want to come to Capitol Hill.

O'BRIEN [voiceover]: Historically, the tech industry

has bridled against regulation.

You have an army of lawyers and lobbyists

that have fought us on this...

SULEYMAN [voiceover]: There's no question that

guardrails will slow things down,

But, the risks are uncertain

and potentially enormous.

So, it makes sense for us

to start having the conversation right now.

O'BRIEN [voiceover]: For me, the conversation

about A.I. is personal.

Okay, no network detected.

Okay, um...

Oh, here we go.Okay.

And now I'm going to open, open, open, open, open...

[voiceover]: I used the Coapt app

to train the A.I. inside my new prosthetic.

♪

It says all of my training data is good,

it's four of five stars.

And now let's try to close.

[whirring]

All right.

Seems to be doing what it was told.

[voiceover]: Was my new arm listening?

Maybe.

I decided to make things simpler.

I took off the hand and attached a myoelectric hook.

[quietly]: All right.

[voiceover]: Function over form.

Not a conversation piece necessarily at a cocktail party

like this thing is.

This looks more like Luke Skywalker, I suppose.

But this thing has a tremendous amount of function to it.

Although, right now, it wants to stay open.

[voiceover]: And that problem persisted.

Find a tripod plate...

[voiceover]: When I tried using it

to set up my basement studio

for a live broadcast.

Come on, close.

[voiceover]: I was quickly frustrated.

[item drops, audio beep]

Really annoying.

Not useful.

[voiceover]: The hook continuously opened on its own.

[clattering]Damn it!

[voiceover]: So I completely reset

and retrained the arm.

And... reset, there we go.

Add data...

[voiceover]: But the software was

artificially unhappy.

"Electrodes are not making good skin contact."

Maybe that is my problem, ultimately.

[voiceover]: My problem really is

I haven't given this enough time.

Amputees tell me it can take

many months to really learn

how to use an arm like this one.

The choke point isn't artificial intelligence.

Dead as a doornail.

[voiceover]: But rather, what is the best way

to communicate my intentions to it?

Little reboot there, I guess.

All right.

Close.

Open, close.

[voiceover]: It turns out machine learning

isn't smart enough to give me a replacement arm

like Luke Skywalker got.

Nor is it capable of creating the Terminator.

Right now, it seems many hopes and fears

for artificial intelligence...

Oh!

[voiceover]: ...are rooted in science fiction.

But we are walking down a road to the unknown.

The door is opening to a revolution.

[door closes]

♪