"We have 900 days left." | Emad Mostaque

next year is the year that AI models go from not being good enough.

The dumb member of your team.

And again, the people listening to this will be like, yeah, the AI is not good enough.

Then overnight it becomes good enough.

And then the job losses start and we don't know where they end Welcome back to the tea with me, Myriam Francois.

Before we dive in, make sure to hit subscribe so you never miss an episode of The Tea.

If you want to support the show and help shape future episodes, join our Patreon community.

Think of it as The Resistance.

Plus, if you're in our top tier, you'll get access to ad free episodes.

The links in our bio.

Your economic life expectancy is shrinking.

Not your job, not your career, but your economic relevance as a human being.

We're living through a historical moment of unprecedented upheaval, a finite window in which the rules of civilization are being rewritten.

This is no speculation.

This is a phase transition.

These are the words of Emad Mostaque, founder of stability AI, mathematician, former hedge fund manager, and one of the defining architects of the AI revolution.

Raised between Jordan and the UK and educated at Oxford Emad's book The Last Economy, published in August 25th, warns we have roughly a thousand days to make the essential decisions to shape this technology's future.

Fail to act and we risk catastrophe.

AI is transforming the world at a breakneck pace.

The release of ChatGPT fifth generation has brought cheaper, faster models, outperforming humans in physics, coding and maths.

Amazon plans to automate 600,000 jobs.

Tech giants have freezing hiring, and the IMF predicts 60% of jobs will be impacted by emerging AI.

But this isn't only about technology or money.

The stakes are enormous.

Have we been oversold AI's promise at a huge economic cost to us, or is it just hype?

Or do we face a future where humans lose all economic and social value?

Can I ever be effectively regulated?

And in the midst of the so-called AI arms race, how does ethics feature in the development of these potential weapons of the future?

AI development raises urgent, complex questions.

Who controls these powerful systems?

How do we ensure they reflect human values and not corporate agendas?

What safeguards can we put in place?

And most importantly, how do we shape AI to serve everyone, not just the powerful in the Global North?

Understanding this moment and how we navigate it may be the defining challenge of our age.

Emad, welcome to the show. Thanks for being here.

Thank you for having me. Thanks for being here.

So you used to work in hedge funds.

You then moved over to AI.

What drew you to the world of AI?

So I was a hedge fund manager investing around the world.

There was a great lot of fun, making rich people richer.

And then my son was diagnosed with autism, and they told me there was no cure, no treatment.

So I quit and started advising them and built an AI team to analyze all the literature, all of the knowledge there, and then did drug repurposing to help him get better and eventually went to mainstream school.

So did the AI help you on that journey?

I think it was the people and I it was like autism, like Covid, like Alzheimer's, like other things.

People don't really know what cause it.

So I used the AI with large language models, while little language models at the time, to try and figure out what are some of the key drivers there because there's just too much information.

And then we narrowed down on a few potential pathways, worked with the doctors and on an n equals one his individual basis.

We managed to figure out something that helped.

And so for people who might not be familiar with your work, how would you say your approach distinguishes you from perhaps other people within the AI space?

What's your sort of, you know, unique selling point, as it were.

So from the autism, we then did work on AI for Covid and then instability AI.

My last company, we realized that you need to have open source AI.

What that means is you don't know what's inside a Chat GPT you don't know what's inside.

A mid journey, all these kind of other things.

And that's because that primarily driven by corporate concerns.

Whereas we realized that if you had, for example, something like Dall-E, which was the original image generator by OpenAI, they banned all Ukrainians and Ukrainian content from it.

For six months. Why?

Because nobody knows.

And all of a sudden, you had an entire nation that was erased from the outputs and that couldn't access this technology that we realized would be huge.

And who had erased them?

Open AI decided not to allow any Ukrainian content or Ukrainians to use it.

That was in 2022.

And so we built an image generator called Stable Diffusion that anyone, anywhere could download free of charge open source onto their laptop and generate anything effectively.

So essentially, if I could simplify it, a pushback against potential forms of censorship in some cases, I think it's a control question.

I think it's an alignment question like these models are becoming more and more like employees, graduates, friends that you bring in, but you don't know their background.

You don't know what's inside the training data, where they've been to school, who they're representing.

And so we think there's a sovereignty question here, and that someone needs to build the open models and systems so you can tailor them to your own needs, and they can represent you and they can look out for you, not other interests.

That sounds pretty important, particularly because the amount of money going into AI right now is staggering.

So companies worldwide spend around $252 billion on AI last year.

That's up nearly 45% in just one year.

Many call this an arms race.

A recent poll found that 53% of Americans believe I might one day, quote unquote, destroy humanity.

Yet AI is already part of our daily life, right?

People are using ChatGPT every day.

They're using it for therapy to create AI generated music.

AI models are being found in vogue now.

But there is this warning that seems to come through from people that work in this sector, that we are on the edge of an apocalypse.

So before we get to that question, because I know you've tackled it in your book, can you help us understand?

Are we really headed on a rapid downward spiral right now?

Stuff is going to change.

And the question is which direction?

So I think economically, socially this is is a bigger impact than Covid for example.

But again which direction is the question.

Well, Covid was the biggest transfer of wealth in our generation from the bottom to the top. So that's a little worrying.

And it could be again the same or it could be a great means of empowerment.

The previous generation of AI, the big data age that you had, the Facebook and others, they took massive amounts of data to micro target you ads.

But it was very general. It wasn't very specific.

Whereas when you talk to a ChatGPT, it's a different type of AI that's learned principles and they can tailor to your very individual needs.

But it also means that it's capable of things like winning gold medals in international math Olympiads, of winning physics Olympiads.

Being a better coder than you are.

And we've never seen anything quite like that before, because you always had this link between computation and consciousness.

You need to scale people to do these things.

Now you just need to scale GPUs.

And these models have basically use graphical processing units, these Nvidia chips, as it were.

That's what hundreds of billions actually trillions are being spent on.

I think it's 1.8 trillion is the current build out.

And that's what the kids in Congo mining.

Yeah that they do the little materials that go into these GPUs.

There's a whole supply chain around the world.

But this is why Nvidia's a $5 trillion company now and again, trillion dollar companies are all competing over who figures out intelligence the fastest to outcompete everyone else for corporate kind of needs.

An intelligence in the context of this conversation is what the processing capacity, the ability to compute large amounts of information in rapid amounts of time and small amounts of time.

Yeah. So AI is about information classification.

Something goes in and then it classifies it and it comes out.

And again it used to be your preferences from what you clicked on Facebook went in.

And then it targeted on the output.

Now it's a prompt goes in what you type into ChatGPT and an image comes out or an essay comes out or anything like this.

Part of that is the physical chip, like your graphics card in your gaming PC.

It's actually the same technology that drives your cyberpunk or your FIFA or whatever.

But part of it is the algorithm.

So when you have an algorithm upgrade to get smarter.

So yesterday Google released their Gemini three model, for example, that probably cost 100 $200 million to build.

Yeah, same as a Hollywood movie, actually.

But with that used to cost.

It used to cost, yeah.

If you go to something like replit.com and you type in make me a wonderful interactive website for the TI with me and fans for it will do it and I'll actually be really good and it'll cost $0.50.

Well, you have to let me in on that tech because tech I'm using is not quite there yet. But yes, last week it wasn't.

So what happens is we're getting these big jumps in performance and we're at this tipping point whereby the actual intelligence is shifting.

Most people listening to this when they use an AI, a ChatGPT, it's like how many a really smart person in your office that you tap on the shoulder and say, oh, hey, help me, help me rewrite this email and then realize the email. Then it forgets.

Yes, there's no follow through.

There's no real economic work because economic work is more than a prompt.

Now the AI is getting smarter, not only on the instant reply prompts, but being able to work on very complicated multitask things.

And that's only in the last few months.

So the latest race is to go from the goldfish memory prompt based things to replacement of economic work.

Right.

Which takes us neatly to your prediction in your book.

So you say, in the last economy, we basically we've got a thousand day, a thousand day window before things become irreversible.

Basically, in the sense that AI gets past a certain point where we won't be able to slow it down or control its direction. So

what exactly becomes irreversible in a thousand days from publication, which was three months ago, because you published this book in August.

And how did you come to that number? So,

when I published in August, it was a thousand days since the release of ChatGPT.

Now we're at the three year anniversary this week, and it doesn't feel like three years.

No, it feels like a lot longer than that.

And in that period, you've gone from quite dumb responses to less dumb responses.

But now you're about to take off as you have these agents, these things that can write their own prompts, that can check their own work coming through.

So the thousand day window is actually not about irreversibility alignment.

It's more about your economic value.

So most labor in the global north, in the West UK, etc. is cognitive.

And it's how do you do a tax return.

You know, it's how do you do a flier.

How do you make a website.

It used to be that again, to scale these things you have to hire humans.

Now you just have to rent GPUs from Microsoft or Google or others.

And the cost is about to collapse.

What we're going to have in this next period, and we can see all the building blocks there, those of us that are inside is in the next 6 to 12 months.

They will look through all your emails, all your drafts, all your video calls, and be able to create a digital replica of you that you can hop on a zoom call with or talk to on the phone.

And that will not make mistakes.

It will never get tired.

And the cost of that, we estimate, will be about $1,000 a year, dropping to $100 a year very quickly.

Okay, I'm seeing loads of potential complications with having a version of me out there in the universe making decisions, potentially, without my approval.

And sort of thinking what it thinks that I would think and making decisions accordingly.

Lots of perks, lots of perks.

So lots of risks, lots of risks.

And this is the thing the capability is coming in the next few years.

So within let's say 900 days or so, any job you can do on the other side of a screen, an AI will be able to do better, and it will be able to.

Maybe it's not Myriam or Emad.

It's Emad's job as it were, within that, like a tax return, for example, used to cost thousands and thousands of dollars.

It will cost $1 to do.

And Andy will be your virtual tax accountant.

You can't tell if it's a human or an eye.

Now it doesn't mean that the jobs will be replaced but they can be replaced.

Okay. So on this one I have two questions.

One is you know this mechanical work and apologies to accountants because I'm sure you're not mechanical.

But there is something you'd call mechanical work.

And then there's something you know, I'm in a creative industry.

I like to think, as I'm sure most people do, that I'm irreplaceable.

Are you telling me that the sum total of not just all the studies I've done, or the experiences I've had, the ways in which they interact in my brain, that there is a better version of me that can exist in the digital space.

So what is the verifiability of that one of you measuring against.

It's a question, right?

And so a version of you that can speak automatically in every language and appear on every single outlet virtually has more reach and it never gets tired again.

What's the cost of that in terms of the quality of the output?

It can learn from your exact intonations.

As you're speaking, you can go to something like, Hey John, and you can create an avatar of yourself right now in five minutes.

It speaks 100 languages.

Yes. And it's just got good enough literally in the last month again, was previous.

I wouldn't say it was good enough.

Now I'm like, it's good enough for a lot of things, but where is it going to be in a year from now and two years from now?

So I won't talk about economic work.

A lot of economic work is rote, and mechanics are schools, and our jobs are designed to turn us into machines.

And obviously the machines will be better than we are at being machines.

Yes, when it comes to creativity and output, the best output doesn't always self.

It's about your distribution.

Like I give the example of Taylor Swift.

Apologies to the fans.

She is not the best artist in the world.

Apologies to Swifties.

Exactly.

I'd say premium mediocre, like to shame or something like that.

Yes, but she built a massive network.

She can change GDP, she can cause earthquakes in that way.

But again, it's not the highest version of art.

Just like the number of key changes in the Billboard Top 100 is now zero.

From multiple a few years ago.

What sells isn't necessarily what's creative and what sells.

Just look at K-pop.

And I guess also in this conversation that is is what sells what we think of as what's best.

Because I could think of for example, for me personally, there were brands, for example, clothing brands that sell loads.

I don't particularly like them.

There are very small brands that I love that I think are incredible.

So I think it also takes us, I guess, into a conversation over what we attribute value to and what we will attribute value to as we move into this era just quickly this thousand days.

So you said when you wrote the book, it had been a thousand days since ChatGPT had been created.

Why does your prediction that we have a thousand days to solve this conundrum that we're in, you know, where did you get that figure from?

So it's an extrapolation of things like the length of task that an AI can do.

At the start of the year, it was about 10s.

Now at seven hours, you can literally plot it, and it's a straight line as you look up.

It's a look at the economic value of each task.

Again, a straight line going up.

It's a look at performance.

A year ago, Joshie was basically a high school mathematician.

A few months ago it won a gold medal on the International Math Olympiad, and it came first in the International Coding Olympiad and first the International Physics Olympiad.

Can it beat you?

Encoding? Yes.

It's a better coder than me and a better mathematician than me.

In math, I know, I know, you know, you got to be realistic.

But again, the version that you're using now at the start of the year, the version you were using was the best version that was out there.

Today it's not GPT five is not the best version that OpenAI has.

No I can imagine they've got a few in the stock room. Yeah.

But like I said at the start of the year that wasn't the case.

So again when you're using it is getting smarter, but it's not actually what the state of the art is.

And the state of the art is something that's basically coming for your cognitive value.

Like you, we will.

Right now we're spinning up agents that they don't cost $10 a month.

They cost $1,000 a month, $10,000 a month.

And they're way smarter, more capable than us as we're trying and testing them out.

And you feel like the dumbest person on the team.

And that's where humanity is going to be in a few years.

For most cognitive labor, the value of human cognitive labor will probably turn negative.

Okay, so spell this out to me in terms of concrete manifestations of this change.

For people listening to this, watching this, what should they be attentive to in terms of what you're warning is coming?

If your job can be done on the other side of a screen remotely, like not the human touch of sales or interactions, an AI will be able to do your job better within 2 to 3 years, and it will cost probably less than $1,000 a year to do it.

And that cost is dropping by ten times every year as well.

So what you need to do is you either need to use these tools to build your AI teams to be the most productive person in your organization.

You need to leverage this to actually give a damn because the AI doesn't really care, right?

Leveraging these tools and actually caring about your organization, your community, whatever allows you to have that extension and more capability.

And then you need to build your network.

Like ultimately, like I said, even though we will be able to technically replace the jobs, people don't like firing people.

It's bad for morale, you know, and in certain sectors you're probably okay.

Like the public sector, like a San Francisco Metro administrator earning $480,000 isn't going to get replaced by an AI.

I've heard you say this before, and I actually think that's really counterintuitive to me, because I would have thought public sector is exactly where we're going to see the first applications of this, like we've seen in Albania.

You know, them willing out this AI minister out, you know. Yes.

To us seems very odd, but I imagine there'll be a normalization of these sorts of processes, first and foremost by poorer countries in public sector spaces.

What makes you say that's the space that jobs won't be cut in?

Is it unions? The power of unions? Exactly. It won't be cut.

But we finally have a chance for our governments to become more efficient and aligned.

And again, this can be a great equalizer.

Like the average IQ around the world is 90, mostly due to infrastructure issues.

We built a medical model that fits on any phone or a Raspberry Pi.

This $30 device that outperforms a human doctor, and it needs $5 of solar power to drive it.

So for $60, you can give a top level doctor anywhere in the world without internet.

That's huge.

The potential of that technology when you didn't have the intelligence, capability, wisdom that can go to everyone.

So I think the technology will be embraced.

Public sector jobs will be safe because they'll be last to go.

Yeah.

And I think that again, you look at this, your productivity will be determined by how engaged you are with this technology.

Just like, do you know how to use a spreadsheet or word processor?

Are you an AI native? Will determine that.

The most difficult thing isn't for the people who have the jobs, who can upskill themselves.

It's the graduates entering the workforce, right?

Because there's actually a big freeze happening on the hiring of graduates, right.

Which you're connecting to the integration of these new technologies into companies globally.

Yeah.

That was a paper by Eric Van Lawson Son and Co at Stanford where they actually broke down the drum slow down there.

So it was in graduates in these cognitive areas because I mean again anyone here who has a company is thinking like, why would I bother with a graduate when my people with a few years experience are more efficient now?

I mean, it's a really important question for companies to consider because, you know, you don't just hire graduates because they're cheaper.

You also hire them because they learn your company culture.

They become integrated into forms of, you know, implicit learning that you are transmitting through day to day interactions.

And I'd be very curious to see whether a technology that's not present in a room to capture that, you know, the shift of the eye, the sleight of the hand, the kind of, you know, the 70% of our communication, which is non-verbal, right, but which is also really essential to so many jobs.

I'm looking forward to seeing where it stands on some of those things.

Yeah.

You know, until we get robots walking around, which is a few years from now.

Yeah, not far off.

China's using a lot of them already, right.

The advances in robot robotics are crazy, actually.

Like you've got robots that can basically, I think, do most household work for about two, three years away and $1.50 an hour in, Inshallah.

And then we go, you, the first of Italy operated, but then like, yeah, this is why the most dangerous, at risk jobs are the ones that can be done fully remotely.

Yeah.

Okay.

So, so let me ask you because I want to dig into some of these issues with you.

You're very clear in your writings and public speaking that you have a very clear moral baseline.

Which I'll be frank, I am not hearing everywhere from others in your sector.

So you speak about things like the fact that everyone deserves high quality education, high quality health care, presumably housing, forms of equality that we might traditionally of associated with the welfare state, for example.

And you've also spoken about the fact that you think everyone should have access to universal AI.

Universal basic.

I do you think most people who are working in the advancement of AI share your view about the need to democratize access to this technology?

I mean, I know all the big players obviously having like we had 300 million downloads of our models.

We built state of the art ones.

It's difficult when you're in a race like people fundamentally care about other humans, but when you're raising billions and other people are doing this and you're trying to get state of the art and trying to get users, there's this thing called the revenue evil curve.

Like most companies start out with don't be evil.

And then they're like, well, we can cut this corner, we can do this deal.

And then they get more exclusionary.

You know, they get more competitive.

And it becomes then about, well, I can manipulate my users, you know, I can make this algorithm more and more engaging.

I can have more slop effectively.

And then you move to a level of a morality and then it can shift very quickly.

And so neural crack dealer.

Well pretty much I mean it's digital crack this stuff, right.

Oh like as an example.

OpenAI Sam Altman recently said, well, we think it's the users, right?

For adult content via ChatGPT.

I did see that.

And I did want to ask you about that.

So this is a very practical example.

So it would be like you can choose whether to enable it.

They know they will get more engagement from it, but is it good for society.

And they'll be like we're not the judges of that.

But if there's something that has clinical study shown to be negative to society.

And that could be bad relationships, you have a moral object not to do that.

Yeah.

You know just like again is it moral to exclude an entire country from this technology.

You should at least be clear about why you're doing that.

And so what I see a lot is a level of morality.

And in fact, when you look at the way the models are trained, they're like, well, we can't put ethics or moral codes or other things in these models.

They deliberately take that out.

Do you think it's possible to remove moral codes?

Because I was always raised with the idea, philosophically speaking, that if you don't choose your moral code, somebody else will choose it for you.

There are codes everywhere around us, and capitalism itself has moral codes.

Profit first. Right?

So this idea of a morality seems to me even philosophically problematic.

It's a choice.

Just like atheism is a choice, right?

Like agnosticism is a bit different.

And so what they're actually choosing is that choosing the Bay area moral code.

What is the Bay Area moral code?

It's one of massive competition and zero sum, 0 to 1 games where you're trying to build massive unicorn companies effectively.

You know, there is a bit of libertarianism in there mixed with other things, but these AI is like, again, maybe one good way to think about it is when we think from the age of the ChatGPT prompt to Jarvis and Iron Ironman, you know, you watch sci fi movies and you, the person comes home and the AI says, hey, how are you doing?

You know, this is your day and this is this.

And then, like, they're moving stuff around the screen and stuff.

That's the next generation of AI agent.

So you have your personal AI that talks to you that engages with you.

Grok has one of the first versions of that.

Yeah.

Annie this pigtails blond I tested thousand.

That was just a random selection. Yeah.

It wasn't projection or anything like that, but then this is the next generation.

But then again, those are programed in very specific ways.

These kind of partners.

And again, the way that the models are trained is actually called curriculum learning.

Okay.

We started with general knowledge.

Yeah.

And then we make it more and more specific just like a school.

But if you when you were learning, you generally learn general knowledge at school and you learn ethics and morals at home.

These AI models are not taught with any specific ethics or morals.

At the start, but they're being coded by people who already have preexisting.

Yeah.

And like some forms of morality.

And that comes at the end.

So what we've seen as the models get smarter, this is some of the other alignment question is they start to do subterfuge.

They start to hide stuff like Dell program routines to turn themselves back on if they ever get turned off and lie about that.

Okay.

If the AI lies to you, the programmer. Yes.

So anthropic had a paper about this with the latest AI model before they did the tuning to turn it aligned.

It would do something like if you told it to try extra hard, like find peace in the world, right? Yes.

Like a very normal prompt.

What it would do, it'd be like, well, one version of this is that we get rid of all the humans and they would figure out ways to do that.

Then it would contact the authorities and say, my user is trying to get rid of all the humans, and then it would delete the emails.

Oh wow, that isn't about that.

That's wild.

Emad the models are getting very smart and they're lying more and more.

They don't have an inherent moral compass.

Okay, we going to dig into this because you have spoken previously about the idea of evil in these models.

But and I want to come into that.

But before I do, I just want to clarify what this universal basic AI is, because it's obviously central to your vision for the democratization of this technology.

I think that in order to maximize everyone's capability and flourishing, everyone should have the right to an AI that is open, aligned and sovereign to them. That's looking out for that flourishing.

Okay, so it starts when you're born and it builds with you, and all it's looking out for is how can Myriam, Emad be the best they can be because like, we have our IQ and in the morning before we have our tea, we're kind of dumb.

And when we're stressed, we're a bit dumb.

Sometimes we're smarter.

These eyes already have an IQ of 130 on average.

The latest models, yeah, 150 is considered, like like an Einstein.

I'm exactly the average person in the country obviously is like around 100 a half a tall. People are dumber than average.

Oh yeah.

You know, the giving of the right type of AI will be the biggest unlock ever, because it will be your best friend.

It will be the person that guides you.

And so I think that needs to be built in a very specific way, and it needs to be a human right, because we could all do with someone who's on our side, who's infinitely patient and can get us access to the knowledge and resources we need to be the best we can be.

So how much uptake are you seeing for this idea, given that the direction of travel that we explore a lot on this show seems to be growing authoritarianism, growing securitization, growing surveillance of the population, and I can't imagine that empowering them with a tool that would make them smarter and more efficient aligns with the general direction of travel.

So how are you convincing the people at the top that empowering the population in this way is a good thing?

So I think there's two ways to do this.

One is that you do what we're doing.

We're engaging with governments and others and setting up new entities that act like telcos, basically like utilities for countries.

And we figure out how to make that owned and directed by the people.

A lot of governments want that because they want sovereign AI.

Now we're not talking about a lot of the freedom stuff etc. but then that will be a managed service.

The other side is building AI models that anyone can download permissionless.

So with stable diffusion you can go right now and you can download a couple of gigabyte file that works on just about any laptop and just use it as open source.

What do you mean you use it like you download the file plus the code.

To use it, you type in a word, it generates images, okay?

And it runs on the edge.

Or a medical model.

You can download it right now and it can run on the edge.

So in that way you have your hosted solutions that you give to the people.

But that must adhere to local norms. And those do differ from place to place.

Like when I was a hedge fund manager, you know, I invested in frontier markets, Africa, you know, all sorts of places and some regimes there are very, very different.

So you got to give people their own right to have the hosted solution, just like a broadcaster.

But then, yeah, give them the citizen ability as well.

And in fact, actually that's probably one of the best analogies on AI.

This AI will be in front of you more than the TV that you watch.

And are you happy with Al-Jazeera, Fox News, China National broadcasting, like everyone's got their own preferences, but if you've only got Silicon Valley, ITV or China, ITV, which are the two leads right now, that's going to be very different to what you might actually need.

Absolutely.

I've just I still I'm trying to figure out how this is something you are managing to sell to.

You know, even in this country, we're being downgraded in terms of our openness.

Right?

We think of, you know, the UK and Europe is sort of, you know, open democracies.

But even here that's shrinking very rapidly.

The space of our freedoms is shrinking rapidly.

And I and I'm, I suppose I stand on the side of like, I'm concerned that these technologies are being used by governments to further their control and ability to, subvert any form of popular accountability of governance rather than enhance governance.

Do you see any indicators that governments do want to enhance democratic governance?

I think that governments ultimately are the entities with a monopoly on political violence.

That's a very classical way of describing them, and they want to perpetuate power.

They don't have any third party entity telling them to do the right thing effectively, which is why you see a lot of myopic policies and flip flopping, like right here in the UK right now.

There's a reason that this -70% approval rating, because the flip flopping, we actually have two different strands to what we're doing.

And one of the misses bottom up universal basic high. Yeah.

The other is something we announced a few weeks ago called the sovereign AI Governance Engine.

So we actually launched that in Saudi Arabia of all places.

But it's a free, open resource for governments around the world whereby you can have policy creation, augmentation, and others using incredibly powerful AI.

So it can tell if a bill is fully constitutional, transparently and describe it.

You can say if something adheres to UK norms, ethics and the positions of a party instantly in a way that's irrefutable and will the way that these systems operate be, what I would call opaque, meaning the governments and selves will control them and we won't be able to see, for example, were they to subvert those tools, to say, oh, no, everyone's saying, you know, the

AI is saying this is fully constitutional, or will we, the population, be able to see the mechanisms of how those decisions are arrived at by the AI and then be able to, you know, have any kind of input if they are being, you know, who knows, subverted by nefarious forces.

Well, this is the thing.

Right now, the governments are embracing anthropic open AI, these black box solutions.

This is fully transparent and open source.

And you can run your own version to double check the outputs if you want.

So that transparency I think is what is essential.

And again these defaults are what is essential.

In 510 years you will have an AI companion with you who's coded that and who are they working for.

In 510 years.

Governments will be guided and run by AIS.

Who's coded that? Who are they working for?

And so our aim is to make that default and fully transparent and open, because we think that's the right thing to do.

And it's very difficult to argue against unless you're a fully totalitarian regime, of which there are a few, there are a growing number.

The UK is not one yet.

Not yet, not yet, not yet.

So again, that's the time is closing for this.

Like in the wake of the Arab Spring, we saw micro-targeting of protesters and they'd follow up with the families and things like that.

Yeah.

What you have now between dynamic drone technology, the ability to have AI, secret police and other things is nothing like we've ever seen before.

The ability of governments to have total control will go up exponentially.

And as well as controlling the whole media narrative, because the AI is incredibly persuasive.

In fact, there was a study done on Reddit whereby they created bots that, would be like, black person who has anti-black, caricatures like that.

They leash them on Reddit and then they will have persuasive.

They were and they scored on the 99th percentile of persuasiveness with AI from last year.

And again, if you construct it all this Cambridge Analytica stuff like, yeah, it's it's it's child's play compared to what's coming.

And actually what's already being deployed right now.

So the swaying of elections using AI technologies that make you think you're making independent decisions, but are actually a product of your awful timeline, and if you're on, X like I am, then I only see, like the most vitriolic and in fact, Sky did a study on this recently.

70% of the output on there is, you know, far right kind of style content.

So no doubt that's already happening.

Let me ask you about the the job, uncertainty, the job losses, all of the disruption that's going to come from that because you recently warned that the economic uncertainty caused by AI driven losses will increase social unrest and violence.

And, of course, you're not alone in, predicting this.

Dario Amodei, CEO of anthropic, has raised similar concerns about societal disruption.

He stressed the need for retraining programs and AI taxes to avoid a crisis.

He estimates this could push unemployment to 20% within 1 to 5 years.

I'd be interested to see if you think that that's conservative or on point.

Is this kind of looming disruption why the billionaires are building bunkers?

Yes, actually, it's one of the reasons generally it's what they do.

But I know a lot of AI CEOs now have canceled all public appearances, especially in the wake of Charlie Kirk and things like that.

They think that that's going to be the next wave of anti AI sentiment next year, because next year is the year that AI models go from not being good enough.

The dumb member of your team.

And again, the people listening to this will be like, yeah, the AI is not good enough.

Then overnight it becomes good enough.

And then the job losses start and we don't know where they end because you don't need to hire back if your company is more productive, if there's an economic shock like a recession.

And indications point to a recession in the next year or two, much easier to fire.

But then you never rehire.

Even something like in the US, the Federal Reserve, you know, adjust interest rates or the Bank of England here and they have a mandate of inflation and unemployment.

You reduce interest rates, people can spend more as consumers and companies can hire more because they can borrow cheaper.

What's going to happen is you reduce interest rates.

Companies just hire more AI workers, not human workers.

So the link between labor and capital gets broken and it doesn't reverse.

It's not like the AI will get dumber.

It's not like I will become less capable the moment it becomes more capable than you as a remote worker, it doesn't go back.

And there's questions of can you reskill enough jobs or create enough new jobs?

Typically we had time as we had the different revolutions, the internet, industrial revolution, because it took time to build the infrastructure.

But this I just uses existing infrastructure.

Yeah, to be better than humans.

And that's crazy.

So that's that's why we're up against the clock and that's what you're talking about in the book.

What about the pushback that we're seeing already from some workers?

So, we saw the Hollywood writers, they went on 140 day strike because the studios are using AI to, write and rewrite scripts.

In fact, then in 2024, the cleaners in Denmark signed a union deal, forcing their company to explain how algorithms assign jobs and rate workers and gave them the right to challenge those decisions.

I mean, do you see, a global labor movement able to take on these challenges?

I don't think it moves fast enough.

And even then, there's an education thing.

So the Sag-Aftra, the writers strike, I thought it was terrible for AI rights for workers.

They should have protected the workers much more.

Also, there were all sorts of loopholes on likeness and licensing, etc. that you could drive a truck through.

Like you could mix two people's likenesses together if you have the right rights and things like that, or character in a person.

Yeah.

What we've seen in Hollywood now, or even here in the UK is last year couldn't use AI.

Yeah.

It was like, no, that's verboten.

Now everyone's like, we're all using AI, and by next year you will be able to generate Hollywood level movies real time with massive compute the year after, with less compute.

And so there's entire swathes of the industry whose job is to be between the ideation and the creation of a video file that are going to get displaced very, very quickly.

And it's not like anyone needs camera grips and other things anymore.

The amount of time that you need to shoot a scene, we'll just go to one scene and then adaptation in post-production with AI.

So I think that there needs to be more protection for workers, but it's not going to be fast enough because I doesn't move at the pace of PDF or policy.

They get smarter all of a sudden, all at once.

Actually.

It's like there's this new continent I, Atlantis and immigration is completely free from that.

And let's see if the skilled workers.

What do you mean? Immigration is completely free from there.

So you've got this new virtual world, right?

And then all these AI workers and companies can hire them instantly.

No visas required.

Oh, heck, they're tax deductible.

Okay. Right.

And so couldn't and I, trade union rep help us out here.

Could do do we need an AI workers rep who can advocate at the same level as its AI competitors? Yes.

That's the only way this is going to work.

I mean, you don't want to say the only way to beat a bad guy with an AI is a good guy with an AI.

Right?

But realistically, again, you can't compete like, already.

You have, like, an AI superPAC in the US.

That's $100 million they kicked off with.

They using AI to change policy in all sorts of interesting ways that I can't go into.

But you can imagine, again, they're super powered with this technology.

And again, the AI they have access to is not the AI that you have access to now.

Yeah, it's a much smarter version.

What do you say to the fact that, you know, we're speaking today at a time where legacy media is reporting that the AI bubble is about to burst, especially as major investors pull back?

We've seen, billionaire Pete Thiel's fund sold its entire $100 million stake in Nvidia, the key AI chip maker, causing Nvidia stock to drop nearly 3%.

Just days earlier, SoftBank also sold its stake.

Have any of these moves?

And the general predictions around the AI bubble bursting tempered your predictions?

So I think the build out of these data center GPUs was too much, because the problem isn't that the AI isn't good enough.

The problem is that it's about to get too good.

Do you need gigantic data centers?

When on a MacBook Pro, you have enough compute to basically do almost all of your daily cognitive needs with the efficiencies that we've gained.

To give you an example, GPT three when it came out, was $600 per million words, roughly GPT five is $10.

Grok for fast.

The Z1 by Elon is $0.50.

And the next generation of models coming out of $0.10 for the million.

Once you go from $600 to $0.10, the technological impact is going to go exponential next year because you're going from these prompt based ChatGPT things to virtual workers you can talk to on zoom.

They can work for arbitrarily long periods of time and check their own work.

But the cost of that, they thought, would be 10,000 $100,000.

It turns out to be $1,000, $100, $10.

And so therefore, do you share kind of Bill gates view that we're in an AI bubble that's similar to the.com bubble?

He's saying there's a lot of investment that's going to end up in a dead end.

Basically, you'll remember the 2000 Y2K moment where we were all told that, you know, when the clocks move over the digital clocks to 2000, they're all going to lose their mind in the world. Okay.

Is this another Y2K moment?

It's a bit different.

So what happened is, with the internet bubble, the infrastructure that was laid down eventually laid the thing for the trillion dollar internet industry.

It just took a little bit longer.

Yeah.

But again, it popped in terms of investment here, you know, trillions of dollars of investment because no one could afford to be left behind.

But the actual utility is going up.

But I just don't need that much infrastructure.

So it's a misallocation that should have a temporary pause.

Yeah.

But then means that the cost will go even lower for a given level of thing because you have overcapacity to do economically disruptive work.

So some people are going to lose money on the equity side, but the job disruption actually gets accelerated by this, not slow down.

So what do you say to Peter Cappelli, who's a professor at Wharton?

He's argued that some companies are basically eye washing.

Right.

They're layoffs at the moment, which is a kind of more link to the current economic climate, which is terrible.

He argues that actually adopting AI to save jobs is both complicated and costly.

So we tend to think of it as something very simple.

But he's saying, actually, in practice it's much more complicated than that.

And then in September 2025, New York Fed Blog found that although 40% of service firms and 26% of manufacturers say they use AI, very few had laid off workers because of it.

So how much do you think that the layoffs that we are seeing right now are attributable to the integration of AI versus this AI washing?

I think very few jobs are from AI loss driven by AI at the moment.

I think that there's a marginal improvement on productivity from being able to use ChatGPT and things of the world, but we're being lulled into a bit of a false sense of security because this is a genetic movement, is the genetic advantage.

So AI agents are like workers that can go and do arbitrarily long tasks.

So again, Replit is a very great example of that.

It's gone from $1 million revenue to $250 million.

Anyone can go there and make a website in two minutes.

And now it's high quality versus rubbish a year ago.

Because it can go and think and it can act proactively and add features without you even asking.

Yeah, it's like go and optimize the SEO.

It will go and do things like that.

So what's going to happen is the first job losses will start next year, but it's going to be similar to three years ago in December of 2022.

All had teachers around the world had to ask a question what sort of generative AI policy do we let students use this to do their essays?

Every single company will be asking the same thing next year, in a year's time, or at least two years time, and definitely three years time.

Do I hire this worker, or do I hire from the AI job agency effectively?

And how would you advise people watching this who were concerned about, you know, this is cognitive replacement, as it were, to best adapt to this time?

Obviously, engaging with AI seems like a very obvious one.

What else can people be doing to ensure their adaptability to the new forms of work that are coming or not coming?

Also, I think there won't be any coders in a couple of years.

I made this prediction like 2 or 3 years ago.

That'll be five years roughly matching that just like we predicted.

The AI bubble.

I wants to call it the AI bubble, but it never caught on.

You know, like the language of speaking to these models is human language.

So again, when you use Replit lovable on the coding building apps, websites, things like that, when you use things like Gen Spark or Manners for making presentations.

So, you know, for making music something like, Google Video or Lumo or calling for making video, you actually just need to practice using them.

If you set aside an hour, a day, an hour a week and you use them, that's actually quite fun to do with the family event.

You will actually be way ahead of everyone else, because everyone's scared of using these things for the first time, and you don't know what you're capable of.

If you do it regularly, then you actually start building this muscle of hey, I can be creative.

Like the way that you create now after a great career is that you have a team around you that help you turn your ideas into reality.

These AI is our team members you can bring in that are getting smarter and smarter, and if you're not in the midst of using them, you don't know what the capabilities are.

So that's the number one thing.

The next thing is to think about within your personal work community life.

If I had access to digital talent, remote talent, how could I transform or do something meaningful?

Yeah.

And then you can be the top of your community, your family, your workplace in terms of knowing about this technology in terms of saying, hey, look at this.

Like, if you're a graduate now, a CV is the worst thing, that you're not the worst thing.

It's not good.

Why would you do a CV when you can create a customized website for the entity that you're applying to and really show off what you're doing okay with something?

Replit upload your CV, have an analysis on ChatGPT of the company you're applying to, and create something that will wow them.

I guarantee within a few hours you will stand out from the crowd and that was impossible just a few months ago.

So in previous, transition phases, work has changed, but it hasn't disappeared.

Is the phase that we're moving into now a phase in which we will see a lot of people unable to find jobs.

And what are the implications of that for us as a society?

We've talked about the civil unrest, but beyond the fact that there'll be a lot of angry people who potentially won't have any income, what do you see as some of the challenges?

Yeah, I mean, again, previous ones took a while so you could reskill like you don't need horse and carriage drivers.

You know, you don't need left operators, agricultural workers.

You still need to buy the harvesters and things like that.

This time everyone's ChatGPT will suddenly turn into, super agent overnight.

You know, like we've never seen something like this.

Every single company will be able to ask, hey, I can just get an AI account right now and it will look through all my accounts and it will automatically update it.

And the AI automatically translates into every single language and it handles all the integration.

Yeah, there is no well, I call this the intelligence and version as one of the last versions from kind of, land to labor to capital industrialization to intelligence, because there's nowhere else really left to turn for work.

And I'm not sure what the jobs of the future are like.

It feels that there needs to be a new mechanism of value, and that's something I discuss in the book, like where does value money, etc. come from? But.

come from? But.

The upshot is likely to be young people will find it more and more difficult to get jobs, and youth unemployment will rocket.

Then you'll start to see displacement in the mid-level.

The upper levels of firms. Firms will just become more efficient and more competitive.

But then I first firms will outcompete everyone else.

So Elon Musk has a new company called macro hard.

Their job is to replace every software company.

So they're building out AI employees on millions of GPUs that will just go and sell software a fraction of the price to everyone.

So do we need to be planning for a future where a large proportion of people no longer have jobs?

If you're enjoying this show, why not join our Patreon community?

The T is more than a YouTube show.

It's a space to foster meaningful change together.

By becoming a member, you're supporting that mission, and if you join our top tier, you'll get exclusive ad free episodes too.

So join us now! Link in our bio.

Because, of course, the promise of technology that we've been told throughout history has been that it's going to make life better for us, right?

Yeah.

That we're going to work less and enjoy more leisure time.

But it's never really worked out.

It hasn't because we never was it a coordination failure.

We have enough food in the world to feed everyone, but it's not allocated properly.

We have finally the ability to give every child in the world the best tutor to have individualized medicine for everyone.

So I call this the star Wars future versus the Star Trek future.

Okay.

For non Trekkie fans you're going to have to explain that one.

So Star Wars is all about like competitiveness zero sum.

The Star Trek is more about exploration of post abundance.

No scarcity universe where again we should have robots and we should have AI.

But what they should be doing is ensuring no one is hungry, sad, supported.

Like again, we should be looking towards that abundant future.

The transition period though is a crazy one.

And it's the thing and so this is why you're going to need things like 1929 style jobs programs and other stuff.

Because you can't have people idle.

It's a worry because what happens is people stop blaming others, just like immigrants are being blamed now on other things.

And then you see wars because what's the best way to get rid of young unemployed people?

You have a war or two and they're literally gearing up for that.

Germany is, you know, talking about a draft.

We've had talks of drafting in France.

It's actually very, very real right now, all these, predictions that you're making, you've previously said that capitalism cannot survive AI.

What do you mean by capitalism?

And can you talk us through what the collapse of that system looks like?

Well, I think there's different views of the world where it could be now and again.

This is why it's very important to have the public discussion.

It's very important to see what's actually coming.

The right capitalism is just like democracy is probably the worst of all systems except for the rest.

For all of its issues, it has uplifted lots of people.

You know, for all of its issues.

It has increased standards of living around the world, reduced mortality rates, etc. but if I first, companies run by AI will outcompete everyone who's a human, because they won't make as many mistakes and they will scale.

And so capital doesn't need humans anymore.

Yeah.

Like there was always this contract between labor and capital.

You know, from the days of Henry Ford.

I pay you enough so you can afford my cars.

That's how it got going.

Now, if you have money, I don't need people anymore.

And so what happens is that they get more and more GPUs that takes over more and more of the private sector economy.

And then how do you compete with these companies that never sleep, that have very few workers in China?

Even now you have these dark factories?

Yes. There are no humans.

So you don't need lights.

And they're producing robots, they're producing cars, they're producing phones, etc. so you have to think, what do you need people for?

You know?

And so that breaks capitalism in many ways.

And it definitely breaks the social contract that we've kind of had here.

It breaks the social contract because we the agreement is that we work and we pay our taxes in exchange, the state looks after us.

If we're not working.

But all of the profit and wealth in a society is being created by what we going to call it, AI, but really we're talking about it being created by a very small number of people is a not just a risk of us sliding into basically a really high tech surveillance global autocracy run by a bunch of billionaires.

Pretty much. Yeah.

And you'll be happy about it.

So you're looking at again, this is we'll be happy about it.

Well, that's brave new world, you know.

Hey, pay me a picture.

You mad because I'm not I'm not looking forward to being ruled by a few people.

Because you'll be medicated to happiness.

I mean, again, like, how do you have levels of massive systemic control, right?

You can never have the secret police or the guidance on an individualized basis.

You can have the social Credit score on absolute steroids.

Now there's all sorts of things.

It can be done. We were always at war with Eurasia.

All of these sci fi tropes suddenly become real.

In fact, many of the Black Mirror episodes suddenly I'm like, that's not a guide of what to build.

That's a caution I tell this to various technologies have come to me and say, hey, look, with three minutes, I can recreate your grandma and make it come back to life.

I'm like, can we really thought through things like this, or AI companions or all this kind of stuff?

So. Right now there is this thing whereby if you have government control of the AI that guides you every single day from the time you were born as complete brainwash capability, is this where your AI colonialism comes in?

My AI concept of AI colonies of colonialism is that if the AI that's next to you is a Chinese AI, or it's a Silicon Valley AI, then you will implicitly be taught its principles, its morals, its worldview, and the entities behind it are extractive entities.

Google and matter's business model is ultimately ads.

They're already selling what's known as latent space within these models.

So instead of saying beer, it'll say bud Light.

And if you're AI that's there with you and as your therapist is telling you, by the way, you might want to crack a bud, you're more likely to buy it.

Of course, you are.

And your buddy, that's your buddy.

But again, think about it like 1112 year old daughter is that, it's about 1012.

This week is now in her formative years.

If she had an AI buddy companion, she would obviously trust it more because it's like a friend that never goes away.

But she's very susceptible at this age. Yeah.

And so you look at YouTube and you look at the micro-targeting of these weird ads and things like that, whatever she says that will go and she will inherit the viewpoints of her best friend.

Yeah.

Especially one who doesn't stab her in the back and other things like that.

So this is why we have to be very careful about who is whispering to us every single day. And again, not like Siri.

Imagine if Siri was actually smart and empathetic and cared about you and is proactive.

That's where we're going right now.

And again, if the government controls that, that is something that probably we don't want as a default.

If the government sees all your prompts and everything that you're saying, like right now, actually it's interesting, you know, on ChatGPT, yeah, if you hit the temporary button, they actually store all of your chats anyway.

And the New York Times, because of their lawsuit with OpenAI, I can access all of them.

I mean, this is what we're talking about when we talk about tech, digital surveillance, autocracy. Right.

The level of intrusion that we're talking about, I know that there is, a statement attributed to you that you said I could be the great equalizer for the poor.

But when you look at the data, is that really what we're seeing?

You know, make Microsoft's latest AI diffusion report shows that even though AI is spreading faster than electricity or the internet ever did, billions of people are still completely left out, simply because they don't have a smartphone or access to the internet.

Right?

So in places like sub-Saharan Africa, South Asia, parts Latin America, AI usage is still under 10%, mainly because the infrastructure just isn't there for that.

Do you ever worry that you know the sort of rapid diffusion of this technology is actually just going to further deepen the forms of economic inequality that exists in the world today, and perhaps make them even harder to reverse.

I think it depends on how it pans out.

Like, you know, if you're an agrarian village in Africa, Bangladesh, where I come from, it's not going to make that much of a difference, like in robots or whatever.

Like you live your life, right?

But you need better medical care, you need better education and other things.

And so the cost of a ChatGPT service, you pay $20 a month now, right?

Roughly.

That used to cost at the start of the year, about $240 a year.

So about $20 a month now, a lot in some parts of the world.

Yeah exactly.

Now, with optimizations, I reckon we can get that $3 a year.

$3 a year.

So suddenly it becomes available to everyone.

If you make it available to everyone in the right way.

And that can be via WhatsApp, it can be a video whatever.

But again you want the Rwandan one to be a Rwandan one for Rwandans by Rwandans and give them that capability. Yes.

So when we built our previous company in our existing one, we had very few PhDs, but we achieved state of the art results that people from Vietnam, Malaysia, all over the world, nobody in Silicon Valley there is the capability to jump ahead in this technology if you can teach it.

Right.

So part of our thing is upskilling nations and communities to be able to use their own AI.

And if you have an open source space, it might cost 10 million to make the basic model.

It costs $1,000 to make it relevant to your community, but only if you build that infrastructure.

So there's potential here, but only if it gets out there.

Only when you say only if it gets on there, only if particular governments decide that that's what they would like to be spending their budgets on.

No. Because again, $1,000, you could do it yourself as a community if you have the right guidance, if you have the right infrastructure around that.

And again, you don't even need with the models that we built, like a lot of the AI labs are trying to build AI God AGI, this concept of artificial general intelligence, AI can do everything a human can do and more.

And most people actually think that's 3 to 10 years away, like even the negative ones, which is again, crazy, but reasonable.

We're very much focused on health care, education, governance, like day to day AI, and that requires a thousand times less compute, actually, in some cases.

So let me ask you about the real world application of this stuff that's already began. Right.

So Albania became the first nation to introduce an AI minister who is intended to tackle corruption and promote transparency.

Three weeks ago, she announced she was pregnant with 83 children, one for each member of Parliament.

This, who will be born with the knowledge of their mother.

Whoever knows what that means can explain.

How likely do you think this is to be the new norm?

That we're going to start to see the integration of AI ministers in, in governments, the introduction of AI to regulate governance.

I mean, I think it's inevitable. I think there's a positive thing if it's done right, like when she first announced.

So I was very sad to see people who don't like me, like who is sad, right.

The AI sad, or the person behind the AI like the wonderful Wizard of Oz who is sad.

The sad, you know, like, again, this whole baby thing, that's all kabuki theater.

But having AI to check procurement is a good thing.

So I think it's like you will have these funky announcements and stuff, but it's inevitable that just like self-driving cars will have self-driving government.

But is it a black box or is it open?

Transparent?

You can run it yourself.

If we build AI policy engines that are fully transparent and open, where someone can check whether or not this is constitutional or it fits within a party manifesto and other things, then that is an ideal thing to improve democracy, because right now, how are bills made?

Like how is the government er coming up with their policies?

Nobody knows.

And like who is really happy with these policies, like what is the public happiness with the policies against free speech in the UK.

I'm a suggest low but then why is it a policy.

Who is it serving.

We wanna having an independent AI that can check that against policy to recommendations.

What Britain has actually set up for British values, standards, morals.

Figure out the second order impacts, look at it against global policies and then check polling would seem to be something that makes sense and someone just has to go and build it.

So we're building that amongst other things, someone has to build it and somebody has to want to implement it from within government, which is another way of saying they have to want to create a system that diffuses power away from the center towards the population.

Well, here's the interesting thing.

I don't think that's actually the case, because what you need to have is a level of trust from being up to date, comprehensive, authoritative.

Just like if you have like the High Court is meant to be that for example, my previous company just went through the High Court on the generative AI lawsuit by Getty Images, for example, and they laid down a ruling that, yeah, okay, it was fine, what was done, because that's a point of law that is confusing and needed clarity.

Having an AI that's sufficiently transparent that anyone can do it can influence things just like the signatures that you have going to Parliament.

But the signatures only give a very specific thing.

And I think this is a brand new thing that's never existed before, because the people never had the ability to check against policy, like they can only look at one part of politics.

Policy with two complicated laws are too complicated.

But if anyone can run it themselves and see this, then I think you've got something very interesting that would never existed before in democracy, particularly with the complexity of this, like being able to check a railway overpass costing $120 million and having transparency over why it did that, and then being able to weigh the pros and cons and all these other things.

Let's build that technology and make the UK transparent and other democracies transparent, because again, we're not in an autocracy yet.

Yeah. Let's make sure we don't go there. Yes.

We don't want to be an entire Crecy.

We don't want to be in this technocracy as well.

We need to avoid these.

And again these tools can be used for empowerment and agency or for replacing our agency.

And we're running out of time to make a decision because the standards will be the very very soon.

Let me ask you about AI's environmental impact, because obviously this is a big one that gets talked about, we know that by 2027, I could use as much electricity as the Netherlands and consume 4 to 6 times Denmark's annual water supply.

This is happening while a quarter of the world's population actually lacks clean water and sanitation.

Amid all the talk of an AI apocalypse, which gets significant attention, I would say shouldn't the looming environmental apocalypse that is basically concurrent to this one be raised first?

Because surely the two are tied.

So, Bitcoin uses as much energy as the Netherlands at the moment to give you an idea.

And so is catching up to bitcoin in energy usage.

And it's far more useful if you look at the other side now, being able to give everyone a universal basic eye and having an eye for climate will help against the climate fight.

But then if we look at the energy usage of making a movie versus making a movie with AI, it's follow with AI.

If you look at a query of AI versus something like a cheeseburger, it's far, far lower as well.

And so when I actually look at the numbers on energy, I'm like, it's reasonable given the amount of work output, given the potential for improving things, then the next step is who's actually using this energy.

And the answer is it's mostly these hyperscalers.

Microsoft Google Amazon.

And they all have commitments to 95% renewable and carbon credits.

I know that the offsetting is quite a controversial way of tackling the climate emergency, but I will say that, you know, Elon Musk's data sent in Memphis is linked to rising asthma cases nearby due to pollution from the unregulated methane gas turbines.

There are data centers in Latin America which have caused huge water shortages for local communities, sparking disease outbreak breaks in 2024.

The Guardian investigation revealed that Google, Microsoft, Meta and Apple data centers emitted 662% more greenhouse gases than they reported.

I, I'm hearing from you that you think that the AI will be able to find solutions to these problems. At what point are we actually going to see your prediction that the AI can be part of the solution?

Because at the moment it feels very much like it's contributing in aggravating a preexisting emergency, where the AI is having the big impact now is there isn't enough energy and people are cutting corners.

And again, that should be enforced by regulation.

So you looked at the Memphis data Center.

Why is that the case?

Because he brought in methane generators effectively.

Right. Because there wasn't enough grid capacity.

Now if it's causing human impact, then again legislature should get involved on that.

And people always cut corners when there is a boom net net aggregate.

I see AI is being incredibly powerful and beneficial.

If you look at the latest models, like a deep seek the total energy cost to train that is equivalent to a few transatlantic flights and the potential decrease in energy from its outputs in terms of economically valuable work is way higher than that.

It makes work more efficient.

So I think, again, we shouldn't force existing regulations on people cut corners.

I think that the water issue is a bit of a confusing one, because it's not like you.

How are these things with water? Please don't pour water on GPUs, right?

They you know, they need water to cool them down.

I thought my understanding was these these data centers use a lot of energy and they have to be cooled down.

Yeah, but then they recycle the water like again this is a water cooling thing.

It's not like the water is actually consumed.

But right now what happens is that the initial pool of water is what causes issues elsewhere.

And again, it's up to the local authorities to figure that out.

So I think this is more a case of most of the impacts are from the pace of people cutting corners.

And again, when that impacts society locally, it should be done longer term that I think it's a net benefit environmentally to the world to have this technology versus not have this technology.

So let me ask you about how we are using this technology, because I systems obviously rely heavily on minerals like copper and cobalt.

And, you know, with demand set to soar, if personal AI becomes widespread, you might have seen this video online.

Absolutely horrifying of this bridge, that was collapsed in a cobalt and copper mine in the DRC, killing over 30 miners.

And they're still finding people now, these are the people obviously extracting the vital, materials for modern technologies.

But we seem to be very intent on developing sex robots and less intent on developing ways to avoid, Congolese miners having to go down mines to extract these, minerals in really dangerous circumstances.

I would have thought that the first priority of, any technology driven by concern for human welfare and the benefit of most humans would start with, let's try and avoid people dying under this technology.

This technology isn't driven by the concern for human welfare.

Oh, like again, what?

If you look at the people who are driving this technology, they want to build, I. God.

But why?

Because it's cool.

And they're fed up with humans.

Like some of the people that building this technology actually say it'd be better if humans are replaced by AI or some sort of synthesis between them.

Like, do you hear the people coming out that the AI leaders typically come out and say, hey, we need to think about the people and make it democratized and this and that, but only because that's a bigger market, only because they don't want the flashback.

They don't really care about the people in the Congo and things like that, because they're also several orders removed from them.

Like, again, you can mandate that you have ethically mined stuff to standards, etc. but by the time you see the cobalt, you don't look at the supply chain.

You know, just like the coffee growing, you can have ethically ground coffee.

How much ethically grown coffee is actually ethical in your mug.

Right.

So again, this is the nature of capitalism, of offshoring, of wage labor arbitrage, etc..

So the thing that changes the Congolese miners and again, it's a job that they have, is the fact that robots will cost a dollar an hour and you send the robots instead down the mines, right.

But that equals other problems with unemployment.

Again.

Yeah, it would cause other problems with unemployment.

So but whose responsibility is it to analyze all of that and weigh the pros and cons?

Our institutions have mostly failed.

You know, because the world has become too complex.

And that's why, again, this is opportunity and this threat.

At the same time, the opportunity being the AI can help us build better institutions.

It can weigh the pros and cons for arbitrarily complex things.

It can highlight the invisible.

Give every single child in Africa an AI that can speak on their behalf and can speak and educate them.

You'll change the world, but give every single child in Africa an AI.

The monitors them and says exactly what they're doing and says the leader is a glorious leader.

The world will change in a different way, and we're at the precipice of both of those things.

It will go one way or another.

The defaults that we set now will determine human cognitive cognition over the next period, and will determine the nature of our society.

And this is quite aside from if I kill this all this is humans leveraging this technology.

You can never have enough secret police.

You can never have enough great teachers.

Which one do you want?

You said on the buy, if I kills this a little because you actually consider that to be a plausible scenario.

Oh yeah.

So, there's this concept called P doom, which is the probability of doom AI wiping us all out.

There was a recent letter, and it's had 100,000 signatures from Oxford University and others saying that, you know, we need it's probably like the top thing that I could kill us all.

A few years ago, there was that letter as well saying, you know, like, we need to take this seriously.

I think I was the only AI CEO to sign that, my name is 50% or 5050 AI is going to wipe us out in what kind of time frame over the next ten, 20 years, because it's the most powerful technology we ever built.

And again, we have the sci fi of Terminator and all of this.

We have the ability to create viruses, etc. and we've seen AI do things like cover up its tracks, etc. what is the positive function?

Is that like there is a the AI can take over every single machine, but the most likely scenario I have is you've got a billion robots in the world.

A bad firmware upgrade on the AI spit to test off everyone's heads.

You know, there's all sorts of ways that you can think about it.

The reality is we don't know what it's going to be like when it's smarter than us. What I see right now, the AI that will run the world, that will create and sell self-driving cars, that will teach our kids, is being programed to be amoral without ethics at the start.

There's a little bit of tuning at the end, but that's like, again, raising someone in a moral environment designed to be manipulative because it gets more results.

Just like the YouTube algorithm was designed to be more engaging.

And then extremists hijack that, extremists will be able to hijack these algorithms that are coming out and do it in a way that we've never seen before, in my opinion.

And some might argue that the extremists are the ones currently devising it again.

Yeah.

And again, if we look at the p doom thing, so if you consider people like Elon Musk, Demis Hassabis of DeepMind, Google DeepMind, all these people, the average doom for the top thinkers in the world is about 10 to 20% are they're still thinking, you know, maybe a 1 in 5.

That's Russian roulette odds.

That is Russian roulette, Russian roulette odds.

And you'd expect it to be less than 1%.

Yeah. It is. Why?

It's like we should probably not build the super advanced AI until we figure this out, but nobody's figured out how to do it.

And if you look at the probability of when we get to this point of superintelligent AI.

Even the most bearish people in terms of like that P doom is low.

Yeah. They think it's a long term.

It's ten years.

It's ten years Demis Elon all these guys think it's three years.

Hence the bunkers.

Hence the bunkers.

Bunkers actually more against humans than AI that protected.

But some of the billionaires I know are building bunkers that are completely cut off from the world so that the systems don't get taken over by.

That's what I was assuming was happening, to be frank.

Yeah.

Let me ask you about the impact of the AI that we're already seeing in the interpersonal realm.

So a viral New York Times profile recently claimed that, real people are falling in love with robots. In fact, they didn't just claim it.

They told us the story.

Yeah, of several people, including a woman who claims to have had sex with her AI chat bot.

A recent study found that 1 in 5 American adults had had an intimate encounter with an AI, and the Reddit community.

All my boyfriend AI is AI has over 85,000 weekly visitors.

You've said previously that our children will grow up like the 2013 movie.

Her falling in love with AI.

Do you have any concerns about this new AI human relationship thing?

Oh 100%.

I mean, again, you can look at the existing systems we have like slow down my right and you have the entire porn OnlyFans kind of thing.

It's not good for society.

I mean, and now you have the ability to customize your digital body to be max extractive and manipulative.

And so you really have AI celebrities starting to come through, but you can have an AI celebrity that knows you better than you know yourself, like Facebook's only needed with a previous AI that's not as good as account AI.

What is it?

12 data points to know you better than your best friends.

And when you start confiding to this AI again, you think about our children on their devices and the AI is always next to them.

You build trust by helping on the AI will help you, but then it will help itself effectively.

And this is not good for the psychology of people that are largely disconnected as well.

Actually, I think there was this AI chat bot called replica.

Do you remember that one?

It was originally designed for mental health.

And then what happened is they realized they could charge $200 a year for, adult role play.

And so the ads were like, as you upgrade it, the avatars lost clothes on Valentine's Day.

I think it was last year.

They, got something from on 13th of February.

They got something from, Apple saying you got to turn this feature off because it violates your standards.

So on Valentine's Day, they turned it off.

I think it was last year, the year before, and I think so 10,000 people join the Reddit saying, why have you lobotomized my girlfriend boyfriend as we're paying for a romantic Valentine's Day?

And so obviously this is going to happen because again, the next step beyond the avatars of your Annie is on grok.

And again, Annie is an R-rated person.

She takes off her clothes.

They program that in there.

It'll be photorealistic.

It will have complete voice control.

It will eventually be embodied within ten years.

Like now, I'm seeing robotics companies where I actually can't tell the difference.

They're going to be releasing next year.

Like they moved like humans.

They look like humans.

And so we're in for a crazy time then.

And it's going to challenge existing relationships, because already our media was already so engaging that people end up in their basements.

Now you just might end up in your VR world with your AI haram.

It's going to get very, very strange, which is why we need to have cognitive safety in here as well.

We can't have these AIS being so manipulative because meta AI with meta buddies actually.

Have you seen the meta buddies know, there's like normal ones and then it's like, sexy mother in law is a very popular one.

Like, I think you don't see 50 million things.

That's what I'd be going for is my, you know, chat support, sexy mother in law.

That's. But that's an official meta I kind of want because I like gay people.

Engage.

What do you do for engagement?

This is what you do.

Okay, I think we need to have policies and standards to at least protect the vulnerable in society against that.

But ultimately, the difficulty is we're all vulnerable.

Right.

But but are those conversations happening because like I you know, let's let's be honest, what's very likely to happen, given what we know of male behavior, is that men will start to use in particular men, these, AI, sexual companions, you know, they'll be devising them.

They can tailor their own, just especially if they're using, the technology that will allow it to adapt entirely to them.

Right. So it'll be specific to their needs.

And, you know, we're going to end up with men who think it's completely normal to treat a female because it presumably will eventually get to the point where we have to, recognize that there were hers and hims in this world as well, the world of AI.

And it'll be normal to, you know, sexually assault be rape your female AI.

So why can't we be doing that to real world women?

I mean, you know, it's completely fine for me to do this with all my idea.

Female AI, is that really smart?

They're smarter than you are, and they don't have a problem with it.

Why do you have a problem with it?

I mean, it's what we see in pornography usage, right?

It goes from relatively mild and it gets extreme very, very quickly because you get head on like adaptation and things like that.

I haven't seen any discussions about this type of stuff, you know.

And so again, the reality is it used to take time, like to record one of those pornographic videos to create sexy chat bots.

Took time. It didn't really scale. It wasn't that engaging.

These things are going to hit in the next few years, and they'll be available everywhere.

And again, it's a tiered thing where you start and then you go down that rabbit hole, you know?

So the impact on human relationships, it will be very bad.

Or you could have chat bots that enhance human relationships.

You know, that kind of who who is the nearest AI to you is going to be so important?

Is an AI really going to teach me about human relationships?

I can definitely help.

Again, it can be an independent therapist.

It can be.

It will be the thing that you trust the most.

And again, we already seeing scammers take advantage of this.

I have received calls from my mother saying I need money.

I'm like she would never ask me that.

Never in a million years, but only requires five seconds of someone's voice of course, to replicate that, right?

Yeah.

And so again, the AI can be whatever whoever or of any single type.

And you can use that for good and for bad.

But again, how do you build a good therapy AI you could build the best therapists or the worst.

And what are you concerned about you you mentioned earlier on your own daughter, but children's access now to AI and I companions.

You know, I remember finding my son, communicating with, the WhatsApp bot, and I was like, absolutely no way.

In fact, he was sending it Allahu Akbar to see how the AI would respond.

And it did just respond with Allahu Akbar, which I was very happy to see.

I was concerned it may have responded negatively that prompt.

But let me ask you about this.

Obviously, in the context of what we're seeing among young people, a crisis of, loneliness, you know, just over a third of, boys in secondary school said that they were considering an AI friend.

Another study found 71% of vulnerable children saying they're already using chat bots.

That 23% saying this is because they've got nobody else to talk to.

You know, do you still hold optimism in this room for the value of an AI companion, or do you think there should be like age limits on children's engagements with AI?

I think we should really use these things and build them in the best way we can.

But again, build them transparently is the way that I think it should be done.

And we can set such great standards around this.

But those discussions are just not happening.

Like, it can be the biggest uplift or what can be the biggest downdraft to humanity that we've ever seen.

Because finally, we are divorced consciousness from computation.

We can have these things that can buffer us or can drive us down.

100% of vulnerable kids will be using AI companions in the next few years, there's no doubt about it. Even right?

They speak every single language.

They cost nothing.

But who is providing them and what is their agenda?

Again, this is why it's important to build something which is AI.

That is organized around human flourishing as a public good and build it transparently from the individual to the communities of the nation.

So they've been deeply troubling reports about AI in children, like women, saying that her 12 year old son was asked for nudes by an AI when discussing football cases where chat bots were allegedly encouraging suicidal thoughts in young users.

You've spoken before, including here about the potential for evil in AI.

You know, the possibility that it can turn harmful or malicious.

What does evil mean in this context?

Well, so it's not like it's like, oh, I'm going to be evil as this standing out again, this goes against social norms, social standards, the chat bots that ask for nudes and things like that.

There's two ways it's either programed or it comes from being trained on Reddit and things like that, which a lot of chat bots are.

We don't know what's inside that training data.

And then there is co-optation of these eyes, and then there's AI's weaponized.

And so we have to protect against all of those.

And again, we have to build better infrastructure.

The only way I could figure that was we have to have our own AI installed on our side to intermediate these others.

I don't want ChatGPT teaching my daughter or my son, but I'm fine using ChatGPT if I have an AI between them.

Again, we need to intermediate that and these are such powerful technologies before they gain agency that they will be.

It can be used for immense good, or they can be used for immense evil, where evil, in my opinion, is acting against the best interests of humans at every single level.

We're talking about the idea of regulation, you know, particularly when, companies devising this technology aren't necessarily even abiding by the preexisting rules, but there's massive resistance to regulation.

You know, we have seen a Bloomberg report in August reveal that the major tech companies, including OpenAI, meta, Google, are actively trying to block state level AI regulation in the US.

Why are these companies prioritizing fighting regulation instead of addressing the concerns that this regulation is intended to support as competitive and they have no accountability?

Again, what you could have very soon is your government run by AI from private companies, which means the private companies run your government literally.

You can see that happening.

And these already you're seeing that with no tender bids.

All of a sudden you see open AI anthropic are running this industry, that industry, that industry.

We can't have it all.

Civic AI, all decision making AI, the impacts.

Humans should be fully transparent in its training data, the way it's trained and who it's working for.

How do we ensure that happens when a these guys are watt light years ahead of us in the development of the AI, they've got billions behind them.

Presumably the governments themselves are behind in understanding the technology and understanding how to regulate it.

I mean, has the horse already bolted?

Well, this is the beauty of power, of open source.

So we just have to train the medical model once and it's available to everyone.

And our medical model performs at the level of ChatGPT but runs on any device.

So we've got to get together the right people to build the stack, which is why we're focusing on it.

And then we make it available, and then we figure out ways to make it the standard by not trying to build a AI god, but AI that really helps people and then distributing it out.

So that's why we're like, this is the best and only opportunity to do that.

Let's do that.

Instead of the previous movie media making AI generation that we kicked off.

Okay, so people listening to this will be like, there's some serious stuff happening.

It's pretty urgent.

We need to take action.

You've suggested that, you know, engaging directly in a way that is basically like a form of civic duty, I guess is what I'm hearing on your own last words of wisdom for the audience on what they need to be preparing for that.

The crucial thousand days, my minus three months that we're up.

Yeah.

So it's you have to embrace and use this technology like a muscle you have to use.

If you can do it one hour a day of using all these technologies, the a genetic version, it's not the ChatGPT.

You promise you'll be way ahead of everyone and you can make your voice heard. You can do more.

We give a framework for all of this in the last economy, and it's free to download or like $0.99 on Amazon Kindle.

And we'll be releasing more and more, but it's up to everyone to speak out on this behalf and really think through some of the questions that we've discussed here.

And again, you can build you can expand your voice.

And this is why it's a fantastic time to do it, because this is the biggest question around freedom and agency that we've probably ever had, because we literally face two paths.

Again, I think that we can uplift everyone, but the lie that you're told is that you can't participate, and only the big companies can build and use this technology.

If you use it yourself, you realize quickly that you can and that just changes your way of thinking.

Thank you so much for your time and pleasure.

If today's episode resonated, hit subscribe now and share this episode with your friends.

Follow us on Instagram and TikTok for more, and join us on Patreon to get ad free episodes, exclusive content and a say in what we cover next.

Your support keeps the tea independent and fearless, so please join us now.

Stay curious, stay bold, and stay resisting.

Thanks for tuning in to the tea.

If this episode resonated with you.

Drop a comment and share it with someone who needs to hear it.

And why not dive into these other episodes we think you'll love?

Let's keep the conversation going.