America's first AI-fueled war is unfolding. How'd we get here?

This message comes from Total Line and More.

College Basketball's biggest tournament is here,

and watch parties are underway.

Shop Total Line and More in-store or online to get ready today.

Spirits not sold in Virginia and North Carolina.

Drink responsibly, be 21.

This message comes from Carvana,

who makes car selling easy.

Enter your license plate or VIN, get a real offer in minutes,

and have your car picked up from your door.

Sell your car the easy way with Carvana.

Pick up V may apply.

This is Fresh Air, I'm Tanya Mosley.

America's first AI field war is unfolding right now.

Over the last three weeks,

the US and Israel have launched strikes against Iran,

hitting 1,000 targets, and the first 24 hours alone.

Nearly double the scale of the 2003 shock

and all campaign in Iraq.

The system helping to enable much of this

is called the Maven Smart System,

and running inside of it is clawed

from the company Anthropic,

an AI model that millions of people interact with every single day.

On the very first day of the war,

a US Tomahawk missile struck a girls elementary school

in southern Iran,

killing more than 165 people,

most of them school girls.

A preliminary military investigation

found the strike likely resulted from outdated intelligence.

And while the role of AI has not been confirmed,

the Pentagon is still investigating

whether Maven played any part.

At the center of this story,

is a little no marine colonel named Drew Koukour,

who spent decades fighting to bring AI to the battlefield

and whose obsession has quietly changed the future of war.

My guest today has been reporting on Koukour for years

and how we got here.

Katrina Manson is an award-winning Bloomberg reporter

who covers cyber, emerging tech, and national security.

Her new book is Project Maven,

a marine colonel, his team, and the dawn of AI warfare.

Katrina Manson, welcome to Fresh Air.

Thanks for having me.

You have been reporting on this Maven smart system for

a couple of years now,

and now you're watching it used in a real-time war.

Take us a little bit into how the Maven smart system actually

works and specifically what Claude's role is inside of it.

How do those two things work together?

If you imagine looking at something like Google Earth,

you begin to have an idea of the display

that US military operators will be looking at.

Some people have described this to me as Windows for War

or an operating system for war.

It's essentially a digital map.

What makes that map special from the US military point of view

is the number of intelligence feeds that are coming into it.

At one public event, it was made clear that is more than 160

separate intelligence feeds.

Now, to crunch that data,

they're using digital data analytics,

but they are also using a few other tools that rely on AI.

There's computer vision to analyze some of the objects

that are showing up on the maps that could be potential targets,

also where US forces are.

And then Claude is doing something different,

that is not computer vision,

that is an AI tool based on a large language model

that can crunch data.

And what I've been told before is that Claude and LLMs

inside Maven smart system help speed processes.

So the sorts of processes you need to get sign off on a target,

everything short of sign off Claude can help with.

And it can also help plan courses of action,

help pair weapons to targets.

It can assist everything that the US military needs to do

when it comes to making a decision

short of actually making the decision.

On the very first day of the war,

this missile struck the girl school,

and there is some reporting about this case

that the United States was likely responsible.

There is no indication yet that AI had a role to play in here,

but the coordinates they used were more than a decade out of date.

What does that specific incident tell us about some of the lapses

in data keeping and potentially what could be

a challenge for AI models as they move through

when are used more often in war?

Adherence of AI warfare regularly emphasized to me

how important accountability is.

In every war, there are bad strikes.

Whether the US is prepared to investigate it

and make public what has gone wrong in this case

if the US is responsible will be a real test

for those claims of accountability.

AI is meant to make warfare more auditable.

Now, whether this is a case that the school was on a targeting list

that predates AI and wasn't updated

and whether AI drew from that targeting list,

all of that will be important to reveal.

Any system, particularly one that uses AI,

will only ever be as good as the data that feeds it.

And if they are drawing on a database that is old,

the AI, if it's set up that way, can't do anything about that.

And in numerous occasions, I've found examples

of poor, weak, or flagrantly erroneous data

that have fed systems.

If this is a US attack, it won't be the first one

against a mistaken civilian target.

In 1999, the US struck the Beijing Embassy in Belgrade.

And that case, the CIA came out in public

and said, we had the map labeled wrong.

And if a map is labeled wrong, which we don't yet know

is the final analysis of what happened here.

But if that girl's school was in a database,

no AI can beat that unless you start using AI

in other places.

If Google Maps, for example, showed that it was a girl's school,

it would be quite simple to draw from that information, potentially.

If there were a way to analyze other location data

that might indicate there were children in the area,

and an additional factor will be where are the checks

and balances on an old database?

And what role could AI play in checking work

and in cross-referencing other data if, indeed,

the girl's school is labeled on something

as accessible as something like Google Maps?

I want to talk about some news this week that is coming

to bear because of a court case.

The Pentagon blacklisted anthropic for refusing

to allow claw to be used in autonomous weapons.

And within hours open AI stepped in,

they publicly then announced the same exact restrictions

anthropic was punished for holding.

Is that an accurate way to describe this?

It's one way it's been described, but not in my reporting.

The open AI deal I've reported is slightly different.

It's not clear if it maintains exactly the same safeguards

as anthropic.

And anthropic also, of course, it's really important

to frame, really lend in to working on classified cloud

for the Pentagon.

They were the first AI company to decide to offer AI

on a classified platform.

And from my reporting, it is not possible for them

to know every use case, every specific example

of the way their AI tool is used in classified operations.

And the classified level is where America fights its wars.

So that decision to lean in to what the American military calls

war fighting was already a very significant decision.

Open AI had not taken that decision.

It was not on classified cloud.

It now will be.

It does seem to have allowed more open acceptance

of how its tool could be used.

But I think we'll have to see because it's

a very politicized divide.

When you have the president calling anthropic, left wing,

not jobs, calling them a radical left company,

even though they were working on classified cloud,

clearly there's a technical debate, there's a policy debate,

but there is also a political flavor to this falling out.

Can you explain, maybe in layman's terms,

how a classified cloud actually works?

Almost.

If you imagine the cloud that we all use for, let's say,

our email or for documents that are loaded up into the cloud,

the same can be done for military data

and it can be accessed and shared.

Now, for a US military or for the intelligence services,

they don't want that information to get hacked.

And so there are a number of safeguards

that are introduced that can uphold a higher classification.

So information that the US system deems and secret

or top secret or compartmentalized information

that only a few people can access,

even at that top secret level.

And each has its own network that can, in theory,

secure that information so that it can't be hacked,

penetrated, ruined in some other way.

Of course, multiple times in history that's gone wrong.

All the time, those systems are under strain from hackers,

potentially also from insider threats.

So the US is constantly trying to safeguard its information.

I was reading about some researchers at King's College London

who recently put cloud and chat GPT and Gemini

into simulated nuclear crisis scenarios.

And 95% of the time, the AI reached for tactical nuclear weapons

as a strategic option.

You have spent years inside of this world

of these people who are building these systems for war.

And I just am curious, what do you think

when you hear a finding like that?

Well, I also reach for the word terrifying.

Clearly, that kind of tool is one that you really need

to put safeguards around.

So the US has said it doesn't want to put AI

into the nuclear controls.

So that's one step.

But there will be pressure on that system

and decision making is already speeding up.

But I've certainly spoken to US military advisors

who've brought me similar information.

They emphasize that AI can be escalatory as you just describe.

And also almost a more problematic issue, sycophantic.

There is a tendency to agree with the person

asking the question, so shall I go to war?

Would it be a good idea to launch this missile?

If the question is asked in that way,

assuming an intent or an action,

there is a tendency within AI also to buttress that opinion.

So as a check on opinion forming,

you need to consider AI in a really careful way.

Now, the US military knows this.

This was a very advanced computer scientist telling me this.

And he had been an advisor to US Central Command,

the very command that is now using these chatbots.

What he and others have told me

at the National Geospatial Intelligence Agency

is that they're aware of these risks

and they are trying to add in checks and safeguards

what they call underneath the hood.

So if a commander said,

shall I strike this now, is it a good idea?

Even if they were to prompt the chatbot in that way,

the claim was made to me

that the chatbot runs through a very fast series of checks.

It red teams the question, which is to say,

it pretends it is an attacker.

It checks for escalation bias.

It checks for a number of different things

and by the time it spits out the answer,

all of those potential problems have been factored in.

Now, I haven't seen that happen in real life

and I've certainly come across a lot of people

who are very frustrated by the answers that chatbots give

even within the military,

sometimes fabricating attacks that haven't even happened.

And if you can imagine the US needs to respond to attacks,

if they're responding to an attack that was fabricated,

there is constantly this risk for escalation.

And in that sense, it's always about that critical thinking,

that framing, what question are they asking of AI?

Can I win quickly if I start a war with Iran

or what are the risks that this could proliferate,

that US service members will be harmed,

that civilians will get hit?

What are the chances of achieving regime change

if I seek a quick war?

How many quick wars become medium term wars and long wars?

Is there still that human hubris where AI is put

will only ever be as good as the data and the question?

And there, there may still be a gap.

And all of this testing is happening

during an active war right now.

A lot of this testing that I've reported on happened before then,

but even at the time in February 2024,

I was able to report that the US did use this system

to narrow down some of the 85 targets

that the US military struck in Iraq and Syria.

This was in reprisal for the death of three US military personnel.

And that is the first large scale up until today's operations.

Case I know of of US central command using this system

to try and bring speed and scale to war.

It had been used before to assist others.

It had been used on a piecemeal scale

for US special operations command,

but they tend to be much smaller.

Getting into the big army, the big military formations,

this really is war at a very joined up

and connected scale involving every service.

And as we speak today,

Senkham has hit more than 9,000 targets.

And that certainly has relied on the system, Maven Smart System.

Katrina, there's a man at the center of your book

in this story that most people have never heard of,

a marine colonel named Drew Kukor.

Tell us who he is and why this moment basically exists

because of him.

Drew Kukor is this very absorbing retired marine

who I met, who was chief of this project called Project Maven.

He wasn't the director, he was the doer,

the leader of this effort to bring AI to the way

that America makes war.

And it started publicly at least as a very narrow effort.

The idea was to bring AI to rifleing through drone footage,

copious video that the US was taking

in various countries around the world

as part of what many military operators

called their G-Wat, the global war on terror.

Now, Drew Kukor had a long and frustrating career

inside the Marine Corps as an intelligence officer.

And he was repeatedly fed up with the tools

that he had to go into battle and to support

other military operators.

He was in Afghanistan in October 2001, after 9-11,

lugging around a large computer.

He felt that he couldn't support the US military operators

that intelligence was meant to keep safe.

And they simply weren't able to get frontline troops,

the kind of information they needed,

as these very rudimentary, unsophisticated,

improvised explosive devices started to name

and kill American service members.

And so there was a constant frustration

that the US could bring to bear enormous firepower,

precision firepower, but couldn't put it in the right place.

And you see, as you see in all wars,

what's known as friendly fire, allied fire,

so the US mistakenly harming their own

harming partners and allies,

and also harming and killing civilians by mistake.

And this number of problems he began to feel

could be solved with better intelligence.

And if there was a way to reduce that loss

when he was in Afghanistan, there were Marines dying

the whole time when he was in Iraq,

there were hundreds of Marines dying.

And he simply felt not that AI so much was the solution,

but better information.

And in the modern world, better information

has come to mean AI.

And in 2011, he worked on an effort to bring

technology from a company named Palantir Technologies

to Afghanistan to start to track

where these improvised explosive devices had been before.

So we're 10 years into this 20-year project

that Kukur envisions.

He has always said that he feels the Department of War,

which during the time of your talking

was the Defense Department,

needed to function more like a software company

than a weapons factory.

But looking at Iran right now,

the scale and the speed is this war he envisioned.

There's no doubt that this is an AI-infused war.

And the other element of safety, accuracy, scope, scale

is people are claiming that AI makes war more efficient.

Often, what happens when things are more efficient

is you can simply do more of it.

And to hit a thousand targets in the first day,

now 9,000 targets, and not yet have finished the war,

the Iranians are still continuing.

The Strait of Hormuz is closed.

There is a question about overconfidence,

about how much you can rely on these systems,

and if expanding the pace of war gets you there.

And this is a long-term debate.

If you go back to 1899,

there was a Polish banker, Ivan Blok,

who brought out a paper called Is War Now Impossible,

because he looked at these claims for mass-produced rifles,

that now the ways of killing was so industrialized

at such scale no one would dare declare war

against someone else.

And instead, he argued long before World War I started,

that actually the mass production of weaponry

would lead to stalemate, human harm,

long wars.

And it raises this idea of, is there ever a way

to deliver palatable killing?

Our guest today is Bloomberg journalist Katrina Manson.

We'll be right back after a short break.

I'm Tanya Mosley, and this is Fresh Air.

This message comes from Total Wine and More.

College Basketball's biggest tournament is here,

and watch parties are underway.

Total Wine and More has options that appeal to everyone,

plus a few choices for cocktails,

with low prices and a wide selection,

everything can be found in one place,

whether hosting or bringing something to share,

shop Total Wine and More in-store or online

to get ready for the big tournament today.

Spirits are not sold in Virginia and North Carolina,

drink responsibly, be 21.

This message comes from Comcast.

Nothing brings people together quite like Team USA

at the Olympic Winter Games,

from NBC Universal's iconic storytelling,

to the innovative technology across Xfinity and Peacock,

Comcast brings the Olympic Games home to America,

sharing every moment with millions.

When Team USA steps on to the world stage,

people are not just watching, they're cheering together.

This winter, everyone's on the same team.

Comcast, proud partner of Team USA.

This message comes from Jerry.

Are you tired of your current Sharon's rate going up,

even with a clean driving record?

That's why there's Jerry,

your proactive insurance assistant.

Jerry compares rates side by side from over 50 top insurers

and helps you switch with ease.

Jerry even tracks market rates

and alerts you when it's best to shop.

No spam calls, no hidden fees.

Drivers who save with Jerry could save over $1,300 a year.

Switch with confidence.

Download the Jerry app or visit jerry.ai slash NPR today.

This message comes from BetterHelp.

As a dad, BetterHelp president Fernando Madera

relates to needing flexibility

when it comes to scheduling therapy.

I have kids under 18, so time is very limited.

That's why at BetterHelp or Therapist try to have sessions.

Sometimes at night, depending on the therapist

or during the weekend.

So I think that's what we need to tell the parents,

you are not alone, we can help you out.

If a flexible schedule would help you,

visit betterhelp.com slash NPR

for 10% off your first month of online therapy.

This is fresh air, I'm Tanya Mosley,

and my guest today is Bloomberg journalist Katrina Manson.

She's written a new book titled,

Project Maven, a marine colonel, his team,

and the dawn of AI warfare.

The book traces how marine colonel Drew Kukor

became instrumental in the decade-long creation

of America's AI warfare capabilities,

which are now being used in the act of war in Iran.

I wanna talk to you a little bit about Kukor's relationship

with Palantir.

It seems to be one of the most complicated threads

in your book, and Palantir for those who aren't familiar

is a data analytics company.

It helps organizations make sense of massive amounts

of information, and Kukor became one of the most powerful

internal advocates for Palantir.

How did that relationship begin?

And why was it so controversial?

Kukor learned about Palantir in the late 2000s

when it was really quite a young company,

and he was looking for this data analytics solution

that could bring data together

and deliver him a picture of war.

As he said to me, it's just a very hard question

to know where the enemy is and where your own people are.

And this for him became a tool that he really believed in.

And others in the defense tech world

who in their military service relied on Palantir

have spoken favorably of it as a tool to me.

He continues this relationship, and he flies over to see them

and he explains his entire vision

for what becomes Maven Smart System,

a digital map and operating system with white dots,

with coordinates that ultimately can pair a target

to a weapon and shoot it.

And at the time Palantir doesn't really want to do this

because he's asking them to do two things

they don't see themselves as.

One, AI and two, to create a user face

and they didn't see themselves as creating pretty user interfaces.

They saw themselves as the data analytics,

the crunching of that aspect.

But they went along with it and a very senior person

at Palantir, Aki Jane, told me that it really is Kukor himself

who convinced Aki Jane to what he said is revisit his priors.

He had a bias against AI, so did all of Palantir.

And they begin to listen to Kukor to understand

how AI might support their data analytics.

In addition to that, Palantir was already controversial

within the Pentagon.

They had actually sued the army in 2016

to gain access to a contract.

This is a time where you really have young, hungry companies

beginning to say give us a contract.

There's this sense that contract awards in the Pentagon

are very old fashioned, it's in function too slowly.

So Palantir has succeeded in getting a foothold in the Pentagon

but was seen as very arrogant by many

because it had sued and it continued to claim

its tech was the best.

Whether that was true or not,

the manner in which they said this

irked several people.

And Kukor himself guided them not only on AI

and what he wanted, but also on the manner

in which they should conduct themselves.

He said, we think you're great, but you need to turn it down.

How would you characterize Palantir in this story?

Is it an honest actor in it?

I think it's really fair to see it as a very divisive company.

You have got people who cheerlead for it with great passion

who feel that Palantir's tech saved their lives.

You also have people who think they are arrogant,

risk being monopolistic, charging too much

and simply make tech that is good,

but not as good as everyone makes out.

Even as late as 2023, a senior commander

who was using Maven Smart System

awards at a grade C+.

So right the way through, you have problems of Palantir

and multiple members of the military lined up to tell me,

okay, we're using Palantir,

but if something else better comes along, we'll switch.

I want to talk a little bit about some other active wars,

particularly the war in Ukraine.

It seems the way that you've been writing about this,

that that's where AI warfare kind of became real at scale.

When Russia invaded back in 2022,

the US deployed Maven in support of Ukrainian forces,

but it almost immediately fell apart.

What happened and how did they fix it?

The computer vision had been trained on the Middle East,

think hot, think of sand,

and suddenly it was being asked to identify Russian tanks

in the snow in Ukraine.

So it wasn't delivering the detections

that the US wanted to rely on.

Secondly, the system wasn't loading.

I found out there were often eight second delays,

which in a war is a lifetime.

And that was because after a lot of investigation,

it turned out that the networking just wasn't up to it.

It was in fact criss-crossing the Atlantic,

sometimes as much as four times.

So that created delays

and sometimes even packets of data

could fall off the network

and you might miss crucial information.

So they really needed to work on the networking,

the sort of arteries of information.

And they also needed to very quickly gather up imagery

of Russian equipment and retrain the algorithms.

And that was going on at a very fast pace.

People complained about getting phone calls

two in the morning, others welcomed them

in order to be part of this effort to support Ukraine.

You know, in reading about that from you,

one of the maybe illegal lines in warfare

is kind of this difference between supporting an ally

and fighting their war for them.

And you report that the US, as you said,

was passing targeting coordinates directly to Ukraine,

sometimes through signal,

sometimes literally printed paper and walking it across.

By that measure, how close was the United States

to actually being in that war?

I suppose that becomes a diplomatic question.

And certainly the US wanted to frame itself

as being a supporter, but not a direct participant.

And that knife edge is really in the eye of the beholder.

Does Russia choose to see it that way?

Or does Russia say you've gone too far?

And so the US was very, very, very sensitive to that.

And the actual project may even operators

and those in the army who were using this system

were even more sensitive because some people

among their groups said we are going too far

and others said we have to help Ukraine

with everything we have.

And at the time, that debate was not public.

There's also some elegant language,

which is this term point of interest.

So rather than saying we're sharing targets,

we're passing targets to Ukraine,

they settled on this language of we're passing

points of interest to Ukraine.

Everything short of the decision to target,

which was a Ukrainian own decision.

But as even some of the people I spoke to for the book framed it,

it was almost a sort of Pinocchio like relationship.

The Americans potentially pulling the strings

on Ukrainian decisions.

And it got tighter and tighter and tighter.

One reason the Pinocchio metaphor isn't fully fair

is because also both sides have emphasized

to me in interviews that they really developed trust.

And so the Americans ultimately were finding

pieces of military equipment

that on Ukrainian information just looked like a truck.

But on US information, they were able to say, trust us.

Hit it.

And it was, in fact, a transporter erector launcher,

essentially a mobile missile launcher.

And that relationship got faster and faster and faster

until at one point, the US identified a target

in one example I'm told about.

And 18 minutes later, the Ukrainians were able to hit it.

Let's take a short break.

If you're just joining us, I'm talking with Bloomberg

journalist Katrina Manson,

whose new book, Project Maven,

Traces How the United States built its AI warfare capabilities.

And how those capabilities are being used right now

in an act of war in Iran.

We'll be back after a break.

This is Fresh Air.

This message comes from an PR sponsor, total line and more,

with so many great bottles to choose from.

It's easy to find your favorite Cabernet

or a new single barrel bourbon to try

with some help from one of their friendly guides.

And with every bottle comes the confidence

of knowing you just found something amazing.

Find what you love and love what you find,

only at Total Wine and More.

Curbside pickup and delivery available in most areas

visit TotalWine.com to learn more.

Spirits not sold in Virginia and North Carolina,

drink responsibly, be 21.

This message comes from Jerry.

Are you tired of your car insurance rate going up,

even with a clean driving record?

That's why there's Jerry,

your proactive insurance assistant.

Jerry compares rates side by side from over 50 top insurers

and helps you switch with ease.

Jerry even tracks market rates

and alerts you when it's best to shop.

No spam calls, no hidden fees.

Drivers who save with Jerry could save over $1,300 a year.

Switch with confidence.

Download the Jerry app or visit jerry.ai slash NPR today.

This message comes from Bombas.

When you're playing sports, you're focused.

Your socks should be too.

Bombas engineer socks to fight sweat

and cushion impact for every sport.

Visit bombas.com slash NPR and use code NPR

for 20% off your first purchase.

This message comes from Carvana

who makes car selling easy.

Enter your license plate or VIN,

get a real offer in minutes

and have your car picked up from your door.

Sell your car the easy way with Carvana.

Pick up V may apply.

This is Fresh Air, I'm Tonya Mosley

and my guest is Katrina Manson,

a Bloomberg journalist and author of Project Maven.

We've been talking about how the United States

built its AI warfare system.

Let's talk about Gaza for a moment.

Israel reportedly used AI targeting systems,

gospel and lavender in its campaign there.

What does Gaza tell us about where the guard rails

on AI warfare actually are?

Some defend the AI saying the way it is used

is down solely to policy.

And others have suggested that the way the IDF

was prepared to potentially accept collateral damage,

meaning civilian harm,

and that speed would not be palatable to the US.

To the US, it just isn't the way that the US

currently operates.

And I should say the IDF defends its action

saying they have not broken the law of war.

They have been proportionate and discriminate.

That's their position.

There are also these very stark numbers of 70,000 dead.

For me, a key question was to understand,

was this defense of AI?

Was it fair to try and separate AI from policy?

So for those who've expressed concern

at the way the IDF pursued targets and civilian harm,

they've blamed policy rather than AI.

So for several of the experts I've spoken to

who make the distinction that they're totally separate,

the tech and the policy.

Many others are there arguing that the more you have

an AI-infused killing machine, the more you can use it.

Which brings up something else for me.

You report that the US has already built weapons

that can fly and select their own targets and kill

without a human making the final call.

So autonomous weapons.

And you name these two classified programs in the book,

goalkeeper and whiplash.

Can you tell me briefly what they are

and what does it mean that they already exist?

These are efforts to bring drones in the air

and on the sea into life.

And this is for a very different conflict scenario.

This is the US thinking about the defense of Taiwan.

So if China were ever to attempt an invasion of Taiwan

and if another big if the US were to decide

to help defend Taiwan, there could be a very different

scenario from the one that Ukraine is facing in Russia

because of jamming.

So the fear is that China would disrupt US satellite

communications such that it couldn't control its own drones

and the drones that would protect and defend Taiwan

against a maritime onslaught would need to be able

to function autonomously without any internet connection.

And so the US has been developing these drones

in a pursuit of autonomy for several years.

Whiplash is an effort to put weapons on a jet ski

that can move autonomously.

And goalkeeper is an effort to weaponize drones

and have them fly about and be able to select

and hit a target under its own steam.

It exactly what campaigners from human rights watch

argued against at the dawn of Project Maven

and what UN Secretary General has called a pursuit

of something morally repugnant and politically unacceptable.

That is the pursuit of lethal autonomous weapon systems.

Well, I mean, what is standing in the way

then of any meaningful international regulation?

Because what does it actually mean

that we're already at war while these particular conversations

are still happening?

That's such a fascinating tension.

There have been discussions at the UN, a UN body

for more than 10 years now.

And they are still trying to define

what is an autonomous weapon system.

And the US position has been let's make it first

and then let's work out what we need to regulate.

That of course speaks to a fear that China might get there first.

The US has wanted to dominate this technology

and to be the ones who could deliver it

in a way that they felt they could use it and win.

But there is a push now to turn some of that work into a treaty.

And a treaty would, by all accounts, not include

the likes of the US or China or Israel or Russia.

Katrina, tell me if I'm right in this.

I mean, everything that we've been discussing may then

the autonomous weapons, this arms race between tech companies

to supply the Pentagon.

I mean, all of this exists in large part

because the US is preparing for a potential conflict

with China over Taiwan.

So what does this moment tell us about

whether we're actually ready for that?

The US has assessed that China wants to be capable

of taking Taiwan by 2027.

So next year, so this date has become this sort of drum beat

for the US to make sure that if it wanted to,

it could fend off a Chinese invasion of Taiwan

as soon as next year, but any time after that.

And there's been an increasing focus since 2018

on the prospect of China being a potential adversary,

not just a competitor on the global stage,

but also a military adversary.

And you see now senior US military commanders

saying quite clearly, China is rehearsing

for an invasion of Taiwan.

And how the US could prevent that

or help partners and allies prevent that

is a subject of some anguish within those

quite tight military circles that look at this.

There's a group that has really pushed for autonomy

to say there's no way we can defend Taiwan

without it, we need to do much more.

And I was told that often pentagon officials reassure allies

and say, look, there is nothing inevitable

or imminent about a Chinese invasion of Taiwan.

And if there is, we'll make sure we're ready.

But then they drop their voice in the corridors

of the Pentagon and whisper, we're not ready.

And so there is this constant concern

that the US needs to go faster in developing autonomy

that could withstand the sort of onslaught

that might be involved in an attempt to take Taiwan.

One of the other things we're all kind of asking

is whether we are the best custodians of this technology.

And after everything that you've reported,

what is your feeling?

What do you come down to?

I know you're a journalist, but you're also greatly informed

and you have all of these facts in front of you.

When you meet people whose business is the business of war,

your perspective changes

because there is so much risk

and there is such a long tail of experience

of these forever wars.

Many of the people involved in project Maven

were involved in the forever wars in Afghanistan, Iraq,

and saw their friends die.

And they put this trust and belief in AI

that that could save their friends,

that could save them, that could save America,

and it could prevent if AI were big enough and bad enough,

China from ever daring go to war with America.

So there's this deep belief in AI

as some kind of panacea.

I think for me, it raises the question of,

what is this idea of a costless war?

If you can make killing more remote,

is that more palatable?

We know that drone operators

and drone screeners, drone analysts

also experience post-traumatic stress.

And AI won't have those same reactions to watching the go.

So there is that argument that you can protect operators.

I question whether you also can protect civilians

by pursuing that notion of remote war.

And the bigger question I have,

is does remote war make war more possible, more likely?

Does it mean that war option

will, someone will press play on it, not understanding

the long, deep impacts?

So for me, there is a lot more to be done

by the people who advocate for AI

to use it in this way they claim it can be used

to deliver a better outcome.

Katrina Manson, thank you so much for your reporting

and thank you for this book.

Thank you.

Katrina Manson is a reporter for Bloomberg.

Her new book is Project Maven,

a marine colonel, his team, and the dawn of AI warfare.

This is fresh air.

This message comes from an PR sponsor,

total line and more, with so many great bottles to choose from.

It's easy to find your favorite Cabernet

or a new single barrel bourbon to try,

with some help from one of their friendly guides.

And with every bottle comes the confidence

of knowing you just found something amazing.

Find what you love and love what you find,

only at Total Wine and More.

Curbside pickup and delivery available in most areas

visit TotalWine.com to learn more.

Spirits not sold in Virginia and North Carolina,

drink responsibly, be 21.

This message comes from Jerry.

Many people are overpaying on car insurance.

Why?

Switching providers can be a pain.

Jerry helps make the process painless.

Jerry is the only app that compares rates

from over 50 insurers in minutes

and helps you switch fast with no spam calls or hidden fees.

Drivers who save with Jerry could save over $1,300 a year.

Before you renew your car insurance policy,

download the Jerry app or head to jerry.ai slash NPR.

This message comes from Bombas.

You need better socks and slippers

and underwear because you should love what you wear every day.

One purchased equals one donated.

Go to bombas.com slash NPR and use code NPR for 20% off.

This is fresh air.

The rise of AI has had seismic implications for Hollywood.

Movie scripts can be written by bots

and one AI company has even created

a computer-generated actor.

But amid this transformation,

one director has created an art installation

that harkens to the old days of cinema.

In 2000, an unknown Mexican filmmaker made waves at can

with a film about a car crash titled Amores Peros.

The director Alejandro Inyadi too

has now turned the film's extra footage

into an art installation.

Contributor Carolina Miranda reviews the show.

Walk into the first floor gallery

at the Los Angeles County Museum of Art

and you'll be forgiven for thinking

that you've wandered into the building's machine room.

The clatter of industrial appliances

makes normal conversation a challenge

and the room is hot, even a bit steamy.

But move deeper into the space

and you'll find that you're actually

in the middle of a movie.

Large projectors display looped scenes

on six screens staged around the room,

all featuring snippets from director Alejandro Inyadi's

first film, Amores Peros,

which debuted to much acclaim in 2000.

On one screen, you catch a piece

of one of the movie's brutal dog fights.

On another, a hand reaches up a woman's skirt,

a car chase ensues, and a brutal crash.

Then that same crash plays again from another angle.

This is Sueño Perro, devised by Iñarito

with the help of a robust production team.

The installation takes the unused scraps

of his groundbreaking film

and transforms them into an environment

that not only plunges the viewer right into the movie,

but into the act of filmmaking.

You see slates marking the beginning of the action.

You see takes and retakes.

Occasionally, the strips of colored film

at the end of a real come into view,

casting an orange light on the room.

Sueño Perro, in Spanish, translates to dog dream.

And Iñarito's installation certainly feels

like a dream of the original movie.

Fragmented, chaotic, out of order.

At times you hear the convulsive explosion

of the film's climactic car wreck,

sometimes that same crash occurs in eerie silence.

Like an actual dream, it's then up to the viewer

to make sense of what the bits might mean.

Like any movie, the image is also function

as a timestamp of the past.

The old sedans look dated.

One of the film's stars, Guy Elgarcia Bernal, is still a teenager.

And the Mexico City of the film is one that

has not yet been gentrified by the digital nomads

of the 21st century.

As Iñarito writes in a book about the project,

a film is made of time and light.

But what makes Sueño Perro truly remarkable

is its analog nature.

Amores Perros was made before digital cameras

had completely transformed moviemaking.

Iñarito, a storyteller who embraces excess,

shot a million feet of film to make the movie.

But the final cut, which clocks in at about two hours

and 30 minutes, used only about 13,000 feet of that footage.

That left about 187 miles of film on the cutting-room floor.

In an era when the word movie has come to mean a video

you can shoot and edit on your phone,

Sueño Perro is a reminder that films once carried physical weight,

a 35-millimeter reel weighs about five pounds,

and the average film was about two reels long.

The use of celluloid film also involves photochemical processing

and displaying the work requires large projectors

that generate heat and noise.

Making a movie is a creative process.

It used to be an industrial one, too.

Sueño Perro makes this industrial nature visible and visceral.

In the gallery, massive reels rotate on the large format projectors

typically used in old movie houses.

Long strips of 35-millimeter film travel through elaborate looping systems

that reach a height of more than six feet.

In addition, the designers have pumped a small amount of fog into the gallery,

making visible the beams of light projected onto each screen.

To enter the space isn't simply to be surrounded by the images of Iñarito's movie,

but the mechanics that make it possible.

It's a reminder of all the physical things that have been lost to the immaterial pixel.

Vinyl records have given way to streaming, newspapers to websites and apps.

Directors used to haul around heavy reels to display at film festivals.

Now at most, they carry a small hard drive.

And as acts of creation have been turned over to artificial intelligence,

Sueño Perro stands as a reminder of what could go missing when you take out the human touch.

The physical world, full of love and pain, can be a really enthralling place.

Carolina Miranda reviewed Sueño Perro on view at the Los Angeles County Museum of Art through July 26th.

If you'd like to catch up on interviews you've missed,

like our conversation with Riz Ahmed on starring in the new series Bate,

as a British Pakistani actor whose audition to play James Bond,

sends his life into a spiral or with human rights lawyer Brian Stevenson

about reflecting on the harsh truths of our nation's history.

Check out our podcast.

You'll find lots of fresh air interviews and to find out what's happening behind the scenes of our show

and get our producers recommendations on what to watch, read and listen to.

Subscribe to our free newsletter at w-h-y-y.org slash fresh air.

Fresh air's executive producer is Sam Briger.

Our technical director and engineer is Audrey Bentham.

Our engineer today is Adam Stanishevsky.

Our interviews and reviews are produced and edited by Phyllis Myers,

Roberta Shorot, Anne-Marie Baldenado, Lauren Crenzel, Theresa Madden, Monique Nazarib,

Susan Yacundi, Anna Balman, and Nico Gonzales Whistler.

Our digital media producer is Molly CV Nesper.

They are a challenger directed today's show.

With Terry Gross, I'm Tonya Mosley.

Support for this podcast and the following message come from Green Chef.

Tired of locking in for another New Year's trend?

Take control of your health without the stress.

Green Chef's meal planning is low prep, low mess meals for every lifestyle.

Right now, go to greenchef.com slash fresh grasa and use code fresh grasa to get started with

50% off green chef and free grasa olive oil set in your second and third boxes.

This 50% off offer is only available for a limited time.

Greenchef.com slash fresh grasa.

This message comes from an PR sponsor, Charles Schwab.

When is the right time to sell a stock?

How do you protect against inflation?

Financial decisions can be tricky, and often your own cognitive and emotional biases

can lead you astray.

Financial decoder and original podcast from Charles Schwab can help.

Join host Mark Reepy as he offers practical solutions to help overcome the

cognitive and emotional biases that may affect your investing decisions.

Download the latest episode and subscribe at Schwab.com slash financial decoder.

This message comes from Alexa Plus.

The all-new Alexa Plus will do more so you don't have to.

Chat naturally while Alexa learns your style,

anticipates what's next, and makes it happen.

Free with prime on your Amazon devices.

Learn more at amazon.com slash Alexa Plus.