Okay.

This is talk about how powerful the models are, and they're trying to pump up multi-billion dollar valuations for an IPO, right? They have no business model.

These companies make no sense, but they wanna cash out before everything crashes.

So anything that says how powerful and scary and effective these models are is better for the company.

The employees, I mean, the best diagnosis, if I can use that word, I've heard, is that these people are incredibly overworked.

They've been running full-tilt for a year or more.

You know, and like the sailors on that aircraft carrier, they're tired.

Okay

Is that what we're talking about now?

You know, I really often dislike worst case examples because then we're just discussing extremes.

I think the worst, I guess the worst case is the fiery death of everybody on the planet plus a zombie apocalypse.

At the same time, we have some hot zone pandemic.

I don't know.

That sounds pretty bad.

It sounds pretty bad, right? It probably makes sense to back off that pretty quickly.

with

Yeah.

And the way to think of an agent is an AI that uses tools in a loop.

So what I mean by that is the AI has access to resources, memory, your files, the internet, other AIs, and it can do things.

send an email, control your thermostat, buy a book, book an airline ticket, drive a car from here to there.

And it does things.

It gets feedback and does another thing.

So you're not there talking to it.

10 million AI agents systematically searching the entire rule space is something else.

So when we talk about a decentralized web, what are we actually talking about? Like what becomes possible in a decentralized system that is difficult and impossible on the internet as we know it today? And any of you can start, or I'm also happy to call you out.

Well, the web is a decentralized system.

It's like email.

Anybody can put up a website.

You can put a web server on your computer, on my computer, on someone else's computer.

You can have Google hosted.

You don't like Google, you can move it somewhere else tomorrow.

All right,

A-and that would include, are you, um, averse to making, uh, your own projection for when the possibility of uncontrolled superintelligence, uh, emerges?

You know, it seems to be any time between next week and never, and pinning it down more than that, uh, is, is impossible.

A-and nobody can.

You know, e-even harder than predicting what will happen is when it will happen, and really, we don't know, right? The models are getting better.

They're not gonna get better forever.

They're gonna reach some asymptote, right? They're gonna level off.

There'll be new technologies.

Right.

Bruce, do you have any to add in that, in that tour, or is that a, is that a good coverage?

I mean, there, there's a lot of things in politics.

Uh, Germany, uh, where there's a, a lot of political parties, they regularly produce, uh, voter guides so voters can know what's going on.

Uh, recently, they added a chatbot functionality to that, where instead of reading a dry piece of paper, you can engage in a conversation, uh, to learn about different political parties.

That's another political one.

Uh, Chile.

Chile has a rule that when you introduce legislation, you need to explain how it affects existing legislation.

Can you say more about that, why you think that's so, and whether that's a good thing or a bad thing? I know you guys often say it's good and bad, but, um, I'm curious, um, what-- how is that gonna change the traits that make for effective political leaders?

Mm

...

in a sense it's self-fulfilling prophecy.

Uh, I mean, uh, on- only a slightly related example, uh, RSA Conference, right? The biggest conference in our, in our industry.

It's really expensive to exhibit there.

Why do you exhibit there? To prove you can exhibit there.

You know, and I, so I think Anthropic spends that kind of money to prove that they can spend that kind of money to justify their astronomical valuation so that the next, you know, person believes it.

Novel

to? Well, before we get to sort of the doom and gloom that we're going to talk about in the context of AI in society, what are some of the positive ways in which countries, governments have been using AI to deliver better services, to communicate better with constituents? What have you seen?

So there's a lot.

And, God, I mean, the doom and gloom bothers me just as much as the utopia people do.

AI is a tool.

It enhances your power.

If you want more democracy, AI will help you.

If you want less democracy, AI will help you as well.

What do you say? Do you agree with somebody who says that? Or do you push back and say this is anomalous?

From the world of Legally Blonde, watch Elle, a new original series only on Prime Video July 1st.

[instrumental music] [audience applauding] So security is two different things, right? It's a feeling and it's a reality.

A- And they're different.

You could feel secure even if you're not, and you can be secure even if you don't feel it.

I mean, really, we have two separate concepts mapped onto the same word, and what I wanna do in this talk is to split them apart, figuring out when they diverge and how they converge.

And language is actually a problem here.

There aren't a lot of good words for the concepts, uh, we're gonna talk about.

Actual mileage and range may vary.

Could this be one thing to... balance the costs and access? Could AI give those people in working class communities the same medical care of the so-called rich and famous?

It's not the AI, it's the medical system, right? US has this horrible, expensive system ponderous medical system where people are making all sorts of money on the backs of us.

So yes, you can imagine a world or a country where that happens, where AI gives benefits to people.

You can imagine a country where all events of AI are captured by whatever private equity firm owns the hospital down the street.

It's not the tech.

It's the economic system the tech is in.

I think a lot about AI being used to augment doctors or in places in the planet where there are no doctors, where the choice isn't AI or doctor, but it's AI or somebody who got a week's of training in health somewhere.

So do you believe, Bruce Schneier, that we have to beat China?

This is my moment to remind you that the internet brings you apparently water and power and your paycheck because Bruce says this comic is accurate.

I mean, the internet, it's cobbled together.

It's perpetually in beta.

We are relying on an infrastructure that is put together by companies with their own agendas, by hobbyists who are trying to do a nice thing, and there isn't a whole lot of deliberate, "Let's pay attention to this and make it reliable."

And yet even the guy who has articulated this to us, the guy who wrote Click Here to Kill Everybody, justDoesn't seem that concerned

I think most of us are rightly not worried about that.

It's like, do you worry about electricity failing? Kind of.

You have a flashlight.

Uh, is that warning just more Anthropic hype?

Yeah.

It's hard to know because you, again, you don't get the details.

So you, uh, gave the impressive numbers.

75 of these were patched.

What? Like, how is that...

I mean, uh, it, it...

Yeah.

We need states to play their role." Talk a little bit to us about where states should have the ability to put their own, uh, thumbprint on regulation.

I, I think this is a time when we need experimentation.

That's what you get from the states.

I think the states are closer to citizens.

In the same way states regulate, you know, so many aspects o- of people's lives, uh, AI is touching those, and we need regulation there.

I mean, the argument that, you know, too many regulations are hard to follow I think is just whining on the part of these multi-hundred-billion-dollar tech monopolies.

God, if you can't figure out 50 regulations, what are you doing in business? I have no sympathy for you.

What's your response to that?

What do you think of that sort of argument?

I mean, that's kind of a bullshit argument.

These are not tools like a hammer.

These are not passive tools.

These are models that are designed very carefully to do certain things and not do others.

It's more like an automobile.

I mean, we could probably use it as a handgun as a metaphor here.

If we imagine that the opposite kind of bill was passed, which does, I mean, we've kind of gone into it a bit already, but one that says you are specifically liable in the event of any deaths, of anything like that, what sort of effect would that have on the open source models? And would it maybe create an inadvertent monopoly of AI companies that can afford legal cases and to pay for overstepping on that line?

And there's a complete lack of comparison in anthropics announcement to traditional tools, like you mentioned earlier as well.

Right.

And I'm not sure how much that matters.

So starting in like November, December, these tools have gotten much better at software.

We sort of hit a plateau in many chat applications, but the writing of code, of sophisticated code, of finding vulnerabilities, of exploiting them and patching them, really changed in the past few months.

So let's take Anthropic out of this for a second.

So we're now moving into a world where these AIs can find vulnerabilities in software, whether it's old things like BSD or OpenSSL, which is another commonly used piece of software that a month ago, someone found a whole bunch of old vulnerabilities in using a different model or new software being written.

With that, I suppose, aside for a second, where do you see, what's your prediction on who this will favor in the long run?

What are the ones that kind of left the either filled you with the most awe or left the most sort of meaningful impression that you're most excited about?

Oh, wow.

So I can give I can give a few different stories.

Let me start with Germany.

This is an easy one.

So Germany has a complex political system, lots of parties, and they have long had a nonpartisan voter guide.

Now, a document that voters can read to figure out what the parties stand for and who to vote for.

Is it is that coming is keeping humans in the loop like is that a technological decision? Is that a human and societal decision? How is this going to unfold in the next couple of years?

H- how do you envision that coming about? I think it's a wonderful dream, but I'm, I am- I will confess, I'm making a struggle on seeing how do we get from, from here to there.

Yeah, the US doesn't do this, of course not.

I mean, we can't even fund, uh, you know, government weather service anymore.

I mean, the US is done.

Just assume that US government will just decay and not do any of this.

It probably couldn't before, just because we believe corporations should do all the things.

Other countries don't have that bias.

Maybe, maybe just a couple of standouts from your perspective.

It's important to make the point that if you no longer have any, um, solid confidence that most, like, 99% of your source of information is actually true beyond an AI video generator, how on earth can you make the calculations necessary in order to form political, um, judgments?

Yeah.

I mean, this is a real interesting question.

It- it's not new, right? You know, uh, doctored pho- doctored photographs is old as photography.

Photoshop's been around since the 1980s.

What's different is, it's now easier to fabricate images a- and voices and, and videos.

I think about it if, like, if we are ever gonna argue about what to do, right? I mean, if the, the actual business of politics, we have to agree on what the world is.

Could you just explain a little bit how that might work in practice?

I was wondering if you could give me an idea of who's doing what in this business.

You know, it's really interesting that all of the top-tier models are about the same.

You'd imagine with these different skunk works doing secret stuff, they'd be different.

But when you look at performance, they're all about the same.

My guess is that these large models will never be profitable.

That, you know, we don't want a chatbot that does everything.

We want a chatbot that is a good travel agent, or a good investment counselor, or a good talk show host, right? You know, we want to do a specialized thing.

Does it amaze you?

I was wondering if you could give me an idea of who's doing what in this business.

You know, it's really interesting that all of the top-tier models are about the same.

You'd imagine with these different skunk works doing secret stuff, they'd be different.

But when you look at performance, they're all about the same.

My guess is that these large models will never be profitable.

That, you know, we don't want a chatbot that does everything.

We want a chatbot that is a good travel agent or a good investment counselor or a good talk show host, right? You know, we wanna do a specialized thing, and we're seeing these smaller, nimbler, cheaper models.

Does it amaze you?

What does that mean to you?

So it beats a couple of things.

Power is the most important one.

If the AI system distributes power, it's generally good.

If it further concentrates power, it's probably bad.

So there are AI systems right now that help people run for office.

There's a lot of paperwork, there are things you have to do.

How can regulation ensure AI systems used in democracy are secure, unbiased, and transparent?

Correct?

You know, I don't buy the arms race argument.

I mean, this research happens in the open.

I mean, like, I mean, the DeepSeek model, which is a Chinese model that, that did so well because they couldn't get the, uh, the good American chips.

All that research is public.

We learn from that.

They learn from us.

Do people, should people be worried, or should there be somebody cognizant of not, uh, not AIing ourselves out of, uh, workers?

Do people, should people be worried, or should there be somebody cognizant of not, uh, not AIing ourselves out of, uh, workers?

I think the market will AI workers away, right? The market seeks efficiencies.

It doesn't seek, you know, the benefit to society.

I think it's gonna be a real problem.

I mean, think about apprentice-based professions, right? It's not plumber.

It's lawyer, it's doctor, it's accountant, it's architect, right? Junior level positions will...

There'll be fewer of them, at best.

But do you think they could help expedite some, with the backup and the courts?

So we're already seeing it and

You know, so these are attacks and it could be adversaries or our people, right? The NSA will do this too.

And they're sensibly there on our side.

So again, we're seeing AI for vulnerability finding.

So finding vulnerabilities in code, useful both attack and defense.

AIs for spam detection.

Also, that means AIs for spam detection evasion, right? So those two arms races, finding and hiding vulnerabilities, finding and hiding spam.

the horizon for our existing public key methods? Should we start to migrate now or is this going to need to happen in five to 10 years?