WEBVTT
00:00:00.080 --> 00:00:01.520
Hello and welcome to Tech Won's Save Us.
00:00:01.600 --> 00:00:02.799
I'm your host, Paris Marks.
00:00:02.879 --> 00:00:16.000
And this month we are continuing our re-airing of the Data Vampire series that I made about this build-out of hyperscale data centers, the demands that they have on communities, and why the tech industry is actually doing this in the first place.
00:00:16.160 --> 00:00:17.359
What is really driving them?
00:00:17.519 --> 00:00:28.960
And of course, we're doing that to mark the release of my new book, Hyperscale, which looks into those very same issues and comes out on October 20th in the US, Canada, the UK, and more broadly in Europe.
00:00:29.120 --> 00:00:32.640
If you like this series, I think you are really going to like the book as well.
00:00:32.799 --> 00:00:38.320
And of course, you can find out more information and pre-order a copy for yourself at hyperscalebook.com.
00:00:38.479 --> 00:00:54.320
Now I think it's actually really interesting that this is the episode of Data Vampires that happens to coincide with this week, because as I'm talking to you in 2026, for the past week or so, we've been hearing a lot about the threat of AI to humanity, basically.
00:00:54.479 --> 00:01:07.200
You know, a former worker at Anthropic came out and warned against the risks that generative AI and the AI models that Anthropic is creating, the risks that they pose for humanity itself.
00:01:07.359 --> 00:01:21.680
And continuing this myth that we've been hearing for the past few years that AI is getting closer and closer to being super intelligent and being able to take over the world and effectively control us humans, if not wipe us all out.
00:01:21.920 --> 00:01:27.200
Now, if you've been listening to the podcast for long enough, you will know that I think that is completely science fiction.
00:01:27.359 --> 00:01:59.120
But actually, this episode of the series digs into the wider costs of these generative AI technologies and of course the data centers that power them, and how the real threats are not, you know, this notion of AGI artificial intelligence or or even AI superintelligence, but the actual demands of the technology, what it means for the climate, what it means for communities, as you know, I was talking about last week on the show, and how these are the real problems, not these kind of fantastical science fictional, future-focused solutions.
00:01:59.359 --> 00:02:09.360
All that they do is distract us from the real harms that are being caused in the present that these companies really don't want us to be focused on, and that they certainly don't want regulators to be focused on.
00:02:09.520 --> 00:02:27.039
And of course, that's not to mention how I would argue there are just a lot of people in this industry who are kind of deluded and who have collectively deluded themselves into believing that they're building the science fictional threats that they've seen in movies and novels that they encountered as younger people.
00:02:27.199 --> 00:02:32.879
And that doesn't mean that we should believe everything that they tell us about the technologies that they're building.
00:02:33.120 --> 00:02:38.000
So that's all to say, I think that this episode of the series is very relevant to what we're seeing right now.
00:02:38.159 --> 00:02:47.280
And I think you're going to really enjoy it, whether you've never heard it before, or even if you heard it a couple of years ago, but maybe don't remember all the details that we included in this series.
00:02:47.360 --> 00:02:51.680
You know, the thing that it opens on is wild enough on its own, and you'll hear that in just a minute.
00:02:51.919 --> 00:02:59.439
But then to think about the other statements that these companies and CEOs and, you know, just workers are making about this technology, it's it's wild.
00:02:59.520 --> 00:03:02.560
And that that doesn't mean that we should believe exactly what they're saying.
00:03:02.719 --> 00:03:13.759
And of course, just a final note that if you do enjoy the series, I think you're really going to enjoy my book, Hyperscale: The Ambition and Excess of Big Tech's Data Empires, which, as I said, comes out on October 20th.
00:03:13.840 --> 00:03:16.800
And you can find more information at hyperscalebook.com.
00:03:16.960 --> 00:03:20.479
So with that said, please enjoy this week's episode of Data Vampires.
00:03:21.120 --> 00:03:26.319
We do need way more energy in the world than I think we thought we needed before.
00:03:26.400 --> 00:03:31.039
And I think we still don't appreciate the energy needs of this technology.
00:03:31.280 --> 00:03:37.759
That's Sam Alvin, the CEO of OpenAI, speaking to Bloomberg in January 2024 at the World Economic Forum.
00:03:37.919 --> 00:03:42.159
In that interview, he was lightly pressed on the climate cost of his generative AI vision.
00:03:42.319 --> 00:03:44.000
And he was remarkably honest.
00:03:44.159 --> 00:03:49.199
The future he wants to realize is one that will require an amount of energy that's hard to even fathom.
00:03:49.280 --> 00:04:03.439
And all that energy needs to come on stream in record time at the same moment we're supposed to be phasing out fossil energy in favor of less emitting alternatives like solar, wind, hydro, or in some people's minds, a ton of nuclear energy.
00:04:03.759 --> 00:04:08.479
The good news to the degree there's good news is there's no way to get there without a breakthrough.
00:04:08.639 --> 00:04:16.800
We need fusion or we need like radically cheaper solar plus storage or something at massive scale, like a scale that no one is really planning for.
00:04:17.680 --> 00:04:30.160
So we it's totally fair to say that AI is gonna need a lot of energy, but it will force us, I think, to invest more in the technologies that can deliver this, none of which are the ones that are burning the carbon.
00:04:30.319 --> 00:04:36.000
The way Altman talks about the massive energy demands his AI ambitions are creating is typical of tech billionaires.
00:04:36.240 --> 00:04:40.079
The climate crisis is not a political problem, but simply a technological one.
00:04:40.240 --> 00:04:51.360
And we need not worry because our technocratic overlords will deliver a breakthrough in energy technology so they can continue doing whatever they want, regardless of whether it makes any real sense to do so.
00:04:51.600 --> 00:04:55.759
Even though Altman refers to this as good news, it's hard to see it that way.
00:04:55.920 --> 00:05:09.360
He's basically acknowledging that warming far beyond the 1.5 or 2 degrees Celsius limit we're supposed to be trying to keep to is essentially locked in because of industries like his own, unless they come up with a technological breakthrough in time.
00:05:09.600 --> 00:05:13.600
There's no guarantee that will happen, and in fact, it's highly likely it won't.
00:05:13.759 --> 00:05:23.839
That's why so many of their scenarios assume we're going to overshoot on emissions, but hope we'll be able to use some future technology to pull all those greenhouse gases back out of the atmosphere.
00:05:24.000 --> 00:05:28.480
And again, another massive gamble with the planet and everything that lives on it.
00:05:28.639 --> 00:05:38.240
But in the interview, Altman wasn't just candid on how much energy the widespread rollout of generative AI will require, but also about that more grim scenario.
00:05:38.800 --> 00:05:49.279
I still expect, unfortunately, the world is on a path where we're gonna have to do something dramatic with climate like geoengineering as a as a as a band-aid, as a stopgap.
00:05:49.439 --> 00:05:52.160
But I think we do now see a path to the long-term solution.
00:05:52.480 --> 00:06:04.800
Altman and his fellow AI boosters want us to gamble with the climate to such a degree we have to try to play God with weather systems, all so they can have AI companions and imagine that one day they might be able to upload their brains onto computers.
00:06:04.959 --> 00:06:07.839
It's not only foolish, it verges on social suicide.
00:06:08.000 --> 00:06:12.399
And I don't think that's a trade-off that many people will openly accept.
00:06:34.399 --> 00:06:45.920
Over the course of this series, we'll learn more about data centers and the extreme vision of the future these powerful people in the tech industry are trying to foist on us, regardless of whether we want it or whether it will even make the lives of most people any better.
00:06:46.079 --> 00:06:59.360
In this week's episode, we'll be digging into how generative AI hype is accelerating the data center build out and presenting a series of threats, from worsening climate catastrophe to further social harms, that will only become more acute the longer this is allowed to continue.
00:06:59.600 --> 00:07:10.319
This series was made possible by our supporters over on Patreon, and if you learned something from it, I'd ask you to consider joining them at patreon.com slash techwon'save us so we can keep doing this important work.
00:07:10.480 --> 00:07:18.959
Plus, enjoy premium full-length interviews with the experts I spoke to for the series, and I've put them together in a special collection whose link you can find in the show notes.
00:07:19.120 --> 00:07:22.639
Become a supporter at patreon.com slash techwon'save us today.
00:07:22.800 --> 00:07:29.920
So, with that said, let's learn more about these data vampires, and by the end, maybe we'll be closer to driving a stake through their hearts.
00:07:32.879 --> 00:07:39.120
Since the release of ChatGPT in November of 2022, talk of artificial intelligence or AI has been everywhere.
00:07:39.279 --> 00:07:41.279
Let's be clear, AI is not a new thing.
00:07:41.439 --> 00:07:44.800
The term has been in use for decades and has referred to different things since then.
00:07:44.959 --> 00:07:55.199
When you type a message on your phone and the keyboard suggests the next word, or when you're putting together a document in Microsoft Word and a squiggly line appears beneath the word to tell you it's spelled wrong, that's AI too.
00:07:55.360 --> 00:08:00.800
It's just not the same kind of AI as what powers the chatbots and image generators that are all the rage today.
00:08:00.959 --> 00:08:02.319
That's generative AI.
00:08:02.399 --> 00:08:04.319
And it's what's fueling a lot of these problems.
00:08:04.639 --> 00:08:09.439
Sasha Lucioni is the co-founder and chief scientific officer at Sustainable AI Group.
00:08:09.519 --> 00:08:13.600
And I asked her why this new generative AI is so much more computationally intensive.
00:08:13.759 --> 00:08:14.720
This is what she told me.
00:08:15.040 --> 00:08:22.079
If you compare a system that uses, I guess, extractive AI or good old-fashioned AI, to search the internet and find you an answer to your question.
00:08:22.240 --> 00:08:26.639
It's essentially converting all these documents, all these like web pages from words to numbers.
00:08:26.720 --> 00:08:33.519
And when you're searching for a query, like, I don't know, uh, what's the capital of Canada, it will also convert that query into numbers using the same system.
00:08:33.600 --> 00:08:35.679
And then matching numbers is like super efficient.
00:08:35.840 --> 00:08:36.960
This stuff goes really, really fast.
00:08:37.039 --> 00:08:38.080
It uses no compute at all.
00:08:38.159 --> 00:08:40.159
It's like it can run on your laptop, it can run anywhere.
00:08:40.240 --> 00:08:48.639
But if you're using generative AI for that same task, instead of finding existing text numbers, it's actually generating the text from scratch.
00:08:48.720 --> 00:08:55.759
And I guess the advantage, quote unquote, is that instead of just getting Ottawa, you'll get like maybe a full sentence, like the capital of Canada is Ottawa.
00:08:55.919 --> 00:09:00.159
But on the flip side, the AI model is generating each one of these words sequentially.
00:09:00.240 --> 00:09:04.320
And so like the longer the sentence, the output, the more compute it uses.
00:09:04.559 --> 00:09:12.080
And you know, when you think about it, for tasks, especially like question answering, like finding information on the internet, you don't need to make stuff up from scratch.
00:09:12.159 --> 00:09:13.360
You don't need to generate things.
00:09:13.519 --> 00:09:14.879
You need to extract things, right?
00:09:15.039 --> 00:09:24.480
So I think fundamentally speaking, what bothers me is that like we're switching from extractive to generative AI for tasks that are not meant for that.
00:09:24.960 --> 00:09:31.120
So basically, there's a lot more work that goes into generating text or images than simply trying to identify what you're looking for.
00:09:31.279 --> 00:09:39.679
These generative AI tools are built on general purpose models that were trained on almost any data these companies could get their hands on, often by taking it off the open web.
00:09:39.840 --> 00:09:50.799
That includes everything from Hollywood movies and published books to paintings and drawings made by all manner of artists, and even many of the things you or me have posted on social media and other parts of the web over the years.
00:09:50.960 --> 00:09:54.559
And the vast majority of that data was taken without anyone's permission.
00:09:54.720 --> 00:10:08.639
Now it forms the foundation of the AI tools and models that kicked off all this hype and that have companies of all sorts rushing to adopt generative AI and push it onto regular users, regardless of whether it's really necessary for the task they're trying to accomplish.
00:10:08.799 --> 00:10:10.320
And that all comes with a cost.
00:10:10.720 --> 00:10:18.240
When you're switching between a good old-fashioned extractive AI model to a generative one, like how many times more energy are you using?
00:10:18.480 --> 00:10:24.559
We found that, for example, for question answering, there's like 30 times more energy for the same task, for like answering a question.
00:10:24.639 --> 00:10:31.200
And so what I really think about is like the fact that so many tools are being switched out to generative AI, like what kind of cost does that have?
00:10:31.440 --> 00:10:34.159
Someone recently was like, Oh, I don't even use my calculator anymore.
00:10:34.240 --> 00:10:35.360
I just use Chad GPT.
00:10:35.440 --> 00:10:38.639
And I'm like, well, that's probably like 50,000 times more energy.
00:10:38.720 --> 00:10:43.440
Like I don't have the actual number, but you know, like a solar-powered calculator versus like this huge large language model.
00:10:43.519 --> 00:10:48.320
Nowadays, people are like, I'm not even gonna search the web, I'm gonna ask Chad GPT, I'm not gonna use a calculator, right?
00:10:48.480 --> 00:10:50.480
All of that, what the cost to the planet is.
00:10:50.799 --> 00:10:55.200
And for all that energy, there's no guarantee the outcome is even going to be better or more accurate.
00:10:55.360 --> 00:10:59.759
As Sasha explained to me, these tools operate not based on understanding, but probabilities.
00:10:59.919 --> 00:11:03.200
Again, think of when the keyboard on your phone is suggesting the next word.
00:11:03.360 --> 00:11:10.960
It doesn't know what you're doing, it's using probabilities based on the data it has to see what word has the highest likelihood of coming after what you've already written.
00:11:11.120 --> 00:11:17.120
That's why we so often see examples of ChatGPT and other chatbots generating completely incorrect outputs.
00:11:17.279 --> 00:11:25.279
There's no real understanding there, despite how often tech CEOs try to make us believe their large language models are on the cusp of sentience like a human being.
00:11:25.440 --> 00:11:35.759
But for those generative AI tools to work, they need a ton of computation, which is why Sam Altman says we either need a technological breakthrough in energy technology or to start geoengineering the planet.
00:11:35.919 --> 00:11:38.720
The notion of scaling the tech back is unacceptable.
00:11:38.879 --> 00:11:51.120
But there are only a small number of massive companies that have access to nearly the amount of computation to properly compete in the generative AI game, which is why the massive tech companies, especially Microsoft and Google, have become so involved.
00:11:51.279 --> 00:12:00.240
They're not only providing the cloud infrastructure to power the generative AI hype, they're also making sure they have a lot of influence over the startups finding success in this financial cycle.
00:12:00.480 --> 00:12:06.559
Here's Cecilia Rickapp, the University College London professor from the first episode in the series, explaining how that works.
00:12:07.039 --> 00:12:11.759
In 2019, Microsoft decided to invest one billion dollars in OpenAI.
00:12:11.919 --> 00:12:19.200
Of course, Microsoft, with all the profits it makes annually, has a lot of liquidity and can decide to invest in many different things.
00:12:19.519 --> 00:12:26.879
But Big Tech in particular have decided to pour a lot of money into the startup world as corporate venture capitalists.
00:12:27.120 --> 00:12:32.000
So Microsoft did this with OpenAI, but the main motive is not financial.
00:12:32.080 --> 00:12:36.480
It's not that they want to make more money just like by investing in the company.
00:12:36.559 --> 00:12:47.679
But the way to make more money, it's uh actually about how OpenAI is developing technology, what technology OpenAI was working on, and how Microsoft can steer that development.
00:12:47.840 --> 00:12:52.480
And by doing it, you can eventually get access to that technology earlier.
00:12:52.639 --> 00:13:03.039
So you can adopt it earlier, as Microsoft did with OpenAI, but you eventually may also be able to make extra profits if the company you invested in is successful and starts developing a business.
00:13:03.360 --> 00:13:18.960
In early 2023, Microsoft invested another 10 billion into OpenAI, but Semaphore reported months later that a significant portion of that investment wasn't in cash, but credits for Microsoft's Azure Cloud Computing Platform, what OpenAI needed to train its models and run its business.
00:13:19.200 --> 00:13:27.519
The company is reportedly losing$5 billion a year, but can continue to operate because of the support of powerful and deep pocket benefactors like Microsoft.
00:13:27.679 --> 00:13:38.720
On top of that, Microsoft, Amazon, and Google have effectively rated the talent at Inflection AI, Adept AI, and Character AI, respectively, to the degree that regulators are investigating them.
00:13:38.879 --> 00:13:45.440
Meanwhile, Amazon and Google have both put billions of dollars into Anthropic, and Microsoft has an investment in Mysterio AI.
00:13:45.600 --> 00:13:57.360
This ensures that on its face, the AI ecosystem looks like there are a bunch of new tech companies rising, but those companies are still completely dependent on the dominant players, not just for funding, but also for computation.
00:13:57.519 --> 00:14:00.320
There's one more angle of this, Cecilia pointed out to me though.
00:14:00.480 --> 00:14:11.440
Yes, generative AI is dependent on the centralized computation of major cloud providers by the hype around it and the perception that if companies adopt it, they'll see their share prices rise has accelerated its adoption.
00:14:11.519 --> 00:14:16.879
And by extension, the demand for computation and the energy and water needed to run all those data centers.
00:14:17.279 --> 00:14:21.840
Just because everyone is talking about AI these days, as a big company, you don't want to be left out.
00:14:22.080 --> 00:14:32.159
Basically, what has happened is a much faster adoption, not only of generative AI, but widely of the cloud and widely of all the different forms of AI.
00:14:32.320 --> 00:14:52.559
And because behind all this, we have the power of Amazon, Microsoft, and Google, not only because of the cloud, but also because they have been investing as venture capitalists in pretty much every single AI startup in the world, they keep on expanding not only their profits, but also their control over capitalism at large.
00:14:52.799 --> 00:15:02.080
So, in a way, it has its own specificities, but if we want to put it just in a nutshell, it has fast-forward something that was cooked from way before.
00:15:02.639 --> 00:15:11.519
So, in short, the AI boom isn't just creating this stock market bubble and allowing companies like OpenAI to rise up the ranks with the support of the existing dominant tech firms.
00:15:11.679 --> 00:15:16.080
The growth of generative AI isn't a challenge to companies like Amazon, Microsoft, and Google.
00:15:16.240 --> 00:15:21.279
It further cements their power, especially as other companies, non-tech companies, adopt it.
00:15:21.440 --> 00:15:32.960
Because every time they do so, they're becoming more dependent on the cloud businesses of those three dominant firms, further increasing their power, their scale, and driving a further build-out of major data centers across the world.
00:15:33.120 --> 00:15:39.039
And as we've touched on in the previous episode, all of that comes with a massive environmental impact.
00:15:41.919 --> 00:15:45.279
For quite some time, tech companies have wanted to be seen as green.
00:15:45.440 --> 00:15:52.480
In the picture they painted, digital technology was clean and green, the sustainable alternative to the dirty, polluting industrialism of the past.
00:15:52.720 --> 00:15:57.840
That was always more marketing campaign than reality, though, as the internet doesn't emerge out of nowhere.
00:15:58.000 --> 00:16:04.639
All the technologies that underpin it have serious material consequences that create plenty of emissions and environmental damage of their own.
00:16:04.799 --> 00:16:09.440
But as efforts were ramping up to tackle the climate crisis, they wanted to keep that image alive.
00:16:33.519 --> 00:16:34.480
That's Brad Smith.
00:16:34.639 --> 00:16:40.000
He's the president of Microsoft, and that clip is from an interview he gave to Bloomberg back in January of 2020.
00:16:40.159 --> 00:16:42.399
Microsoft was rolling out a new climate pledge.
00:16:42.559 --> 00:16:46.799
It would not just achieve net zero emissions, but become carbon negative within a decade.
00:16:46.960 --> 00:16:52.639
The company called this a carbon moonshot, indicating it was ambitious, but a goal they thought they could achieve.
00:16:52.799 --> 00:17:02.240
Well, that was before generative AI became the next big thing that virtually everyone in Silicon Valley felt they had to chase, and that Microsoft saw could significantly expand its cloud business.
00:17:02.399 --> 00:17:05.359
Here's Brad Smith again in May 2024 this time.
00:17:10.960 --> 00:17:14.079
Our goal of being carbon negative by 2030.
00:17:14.400 --> 00:17:18.160
That was before the explosion in artificial intelligence.
00:17:18.720 --> 00:17:23.680
So in many ways, as I say across Microsoft, the moon has moved.
00:17:24.160 --> 00:17:28.960
It's more than five times as far away as it was in 2020.
00:17:29.039 --> 00:17:35.680
If you just think about our own forecast for the expansion of AI and its electrical needs.
00:19:05.200 --> 00:19:08.799
But why can these companies claim to emit so much less than they really do?
00:19:08.960 --> 00:19:12.400
One expert The Guardian spoke to call it a form of creative accounting.
00:19:12.559 --> 00:19:18.240
Basically, they buy a bunch of offsets and act as though having done so means their emissions have been negated.
00:19:18.480 --> 00:19:27.599
Probably the most important of those tools are renewable energy certificates, which shows they've bought renewable energy that can be produced at another time of day or on the other side of the world.
00:19:27.839 --> 00:19:35.119
As long as it was generated somewhere, the companies use it to pretend they didn't actually generate the emissions, they very much did add to the atmosphere.
00:19:35.279 --> 00:19:43.119
And some tech companies are lobbying hard to ensure the rules on carbon accounting are rewritten to make it look like they're emitting way less than they really are.
00:19:43.359 --> 00:19:54.559
According to reporting by the Financial Times, Amazon and Meta are leading the charge to ensure their deceptive accounting mechanisms are legitimized by the Greenhouse Gas Protocol, which is an oversight body for carbon accounting.
00:19:54.720 --> 00:20:04.079
Google is pushing a competing proposal that would force companies to at least buy renewable certificates that are closer to where they're actually operating, but still relies on offsets at the end of the day.
00:20:04.240 --> 00:20:07.359
Companies like Amazon say even that would be too expensive.
00:20:07.519 --> 00:20:15.359
Matthew Brander, a professor at the University of Edinburgh, who spoke to the Financial Times, gave a pretty good example to show why this is all so ridiculous.
00:20:15.599 --> 00:20:24.559
He said allowing companies to buy renewable certificates is like if you paid a Fidder colleague of yours for the right to say you bite to work when you really drove your gas-powered car.
00:20:24.720 --> 00:20:32.000
It's foolishness, but this is how they're planning to keep expanding their data center networks while claiming they're reducing, if not eliminating, their emissions.
00:20:32.160 --> 00:20:35.359
It's a recipe for disaster on a global scale.
00:20:38.079 --> 00:20:46.000
We've talked a lot about why AI is using a ton of computation and further fueling the climate crisis, but what is all the compute we're putting into it really achieving?
00:20:46.160 --> 00:20:50.880
Maybe there's a world where all those resource demands are justified because the benefits are so great.
00:20:51.039 --> 00:20:56.480
And indeed, that's what tech CEOs like Sam Altman or supposed luminaries like Bill Gates would have us believe.
00:20:56.640 --> 00:21:01.200
But the truth is that the rollout of this technology only presents a further threat to much of the public.
00:21:01.440 --> 00:21:16.079
We're used to hearing about AI as forming the basis for a series of tools that can do all manner of tasks, but I was struck by how two of the people I spoke with described the broader project that AI seems to be part of when you consider who is developing it and how it's actually being deployed.
00:21:16.240 --> 00:21:18.079
Let's start with Ali Al-Khatib.
00:21:18.240 --> 00:21:22.559
He used to be the head of the Center for Applied Data Ethics at the University of San Francisco.
00:21:22.799 --> 00:21:32.319
When I asked him how he would describe AI, he began by noting how the term itself is decades old, but there was a troubling through line between its various permutations over the years.
00:21:32.799 --> 00:21:59.519
I think the thing that we would all recognize all the way through, continuously, like is the techno-political project of taking decisions away from people and putting consequential, life changing decisions into a locus of power that is silicon or that is automated or something along those lines, and redistributing or shifting and allocating power away from collective and social systems.
00:22:00.160 --> 00:22:02.720
And into technological or technocratic ones.
00:22:02.880 --> 00:22:08.480
And so this isn't really like a definition of AI that I think a lot of computer science people would appreciate or agree with.
00:22:08.640 --> 00:22:19.680
But I think it's the only one that, again, if you were a time traveler, kind of like going back 20 years and then 20 more years and then 20 more years, you would see totally different methods, but I think you would see basically the same goals, basically the same project.
00:22:20.079 --> 00:22:30.960
Ali's description is unlike anything you'll hear from industry boosters who want you to see AI as a way to improve many aspects of human life, or on the extreme end, thinking it could end humanity if not done right.
00:22:31.119 --> 00:22:42.319
They don't want to talk about that more political angle, the way it's used to cement their power in a way that can be harder to immediately identify than, say, the outwardly authoritarian actions of a politician or leader.
00:22:42.559 --> 00:22:52.240
AI much more quietly erodes the power of much of the public over their own lives, taking away their autonomy by shifting decisions to unaccountable technologies and the people who control them.
00:22:52.319 --> 00:22:58.160
This is something Dan McQuillan, a lecturer at Goldsmiths University and author of Resisting AI, identified too.
00:22:58.480 --> 00:23:07.279
AI is a specific in our faces example of a general technological phenomenon which claims to solve things technically.
00:23:07.519 --> 00:23:13.839
And you know, we see that across the board from tricky social issues all the way up to the climate crisis.
00:23:14.079 --> 00:23:35.839
But I think that that sort of diversion aspect is really an important aspect of contemporary AI, exactly because these issues are so urgent and other forms of collective, social, uh grounded community action and worker action are so urgently needed, that something that successfully, even semi-successfully diverts us from those things is extremely toxic.
00:23:36.000 --> 00:23:51.359
So I'm really talking about their AI as a narrative, AI as an idea, AI as a real technology that appears to do certain things, you know, that can emulate certain things or synthesize certain things in a way that provides people with a plausibility argument that maybe this could fill the hole in health services or education or whatever.
00:23:51.440 --> 00:23:52.240
So that's the technology.
00:23:52.480 --> 00:24:00.799
In Dan's telling, AI isn't just a digital technology made up of complex algorithms and underpinned by the material computational infrastructures that drive it.
00:24:00.960 --> 00:24:29.680
It's also a social technology, one that's deployed so the powerful can claim to be addressing what are very pressing problems in society, the lack of healthcare, inequitable access to education, growing poverty and inequality, not to mention the accelerating climate crisis, without having to actually take the extent of the difficult political measures that would really be necessary to tackle them, measures the elites in our society likely don't want to see taken in the first place, as it might erode their power and certainly require their wealth to be taxed at much higher rates.
00:24:29.920 --> 00:24:35.839
Instead, AI, like too many other digital technologies, can be presented as a seemingly apolitical solution.
00:24:36.079 --> 00:24:40.640
It doesn't require sacrifice and doesn't challenge the hierarchy of capitalist society.
00:24:40.799 --> 00:24:43.599
Indeed, if anything, it further solidifies it in place.
00:24:43.759 --> 00:24:57.680
And all we need to do as a public is have a little patience as our saviors in the tech industry perfect their technofixes so they can deliver us a digital utopia, which, it probably doesn't need to be said, never actually arrives as those deeper issues just keep getting worse.
00:24:57.920 --> 00:25:02.160
There are many harms we can talk about with generative AI, and some of the more common forms of it too.
00:25:02.319 --> 00:25:11.279
We could talk about how companies are stealing all this data and using it to harm the prospects of workers in different industries, like in visual media, writing, journalism, and more.
00:25:11.440 --> 00:25:25.200
Or we could talk about the waves of AI generated bullshit flooding onto the web, some with malicious intent like non-consensual, deep fake, and AI nudes, but much more of it being made just to try to make a buck through social media engagement or tricking people into scams.
00:25:25.519 --> 00:25:37.519
Those things are important, but the deeper issue to me seems to be those that Ali and Dan are describing, and which Alex Hannah, the director of research at the Distributed AI Research Institute, outlined in a bit more detail when I spoke with her.
00:26:16.319 --> 00:26:24.240
We're seeing more and more at the border intense amounts of AI and automated decision making with biometrics.
00:26:24.480 --> 00:26:32.720
That is not necessarily generative AI, but there are other kinds of things that could be looped in with generative AI, which are used at the border.
00:26:32.960 --> 00:26:37.680
Healthcare, education, legal access, virtually anything that happens on the border.
00:26:37.839 --> 00:26:42.079
And the list of all the places they're trying to falsely present AI as a solution goes on.
00:26:42.319 --> 00:26:53.599
Ultimately, generative AI is another one of the tech industry's financial bubbles, where its leading figures hype up the next big thing to drive investment and boost share prices until reality starts to creep in and the crash begins.
00:26:53.839 --> 00:27:10.880
We saw it most recently with cryptocurrencies and NFTs, but there are already questions about how long the generative AI bubble is going to last, with everyone from Goldman Sachs to Sequoia Capital starting to join the existing chorus of critics in calling out the aspects of generative AI that are clearly inflated and poised to crash.
00:27:11.119 --> 00:27:16.559
Even after that crash, generative AI won't fully go away, just as other forms of AI have stuck around as well.
00:27:16.720 --> 00:27:20.720
It won't be everywhere, or have the widespread implementations the companies promised.
00:27:20.880 --> 00:27:24.400
But that doesn't mean there still won't be threats that emerge from its ongoing presence.
00:27:24.559 --> 00:27:29.920
As Ali explained to me, we'd be foolish to think it can be seized and redirected to mostly positive ends.
00:27:30.160 --> 00:27:37.440
If people are designing these systems to cause harm, fundamentally, then there kind of is no way to make a human-centered version of that sort of system.
00:27:37.599 --> 00:27:52.559
In the same way, legislation that makes it slightly more costly to do something harmful doesn't necessarily fix or uh even really discourage tech companies that find ways to amortize those costs or kind of absorb those costs into their business model.
00:27:52.799 --> 00:28:03.920
One example that I think I've given recently in like conversation was that there are all sorts of reasons or all sorts of powers that cause us to behave differently when we're driving on the streets.
00:28:04.079 --> 00:28:11.839
Because as individual people, the costs of crashing into another car or of hitting a pedestrian or something like that are quite substantial for us as individuals.
00:28:12.079 --> 00:28:24.000
But if a tech company that's developing autonomous cars is going to put 100,000 or a million cars out onto the streets, it really behooves them to find a way to legislatively make it not their fault to hit a pedestrian, for instance.
00:28:24.079 --> 00:28:34.400
And so they find ways to sort of defer the responsibility for who ultimately like caused that harm or who takes the responsibility for whatever kind of incident or whatever.
00:28:34.559 --> 00:28:43.200
And so that creates like these really wild perverse incentives to find ways to sort of consolidate and then offload responsibilities and consequences for violence.
00:28:43.440 --> 00:28:50.240
And I just don't see a good way with design out of that, or even with a lot of legislative solutions and everything else like that.
00:28:50.480 --> 00:28:54.559
When the harms are acknowledged, the discussion around AI is about how to properly regulate it.
00:28:54.720 --> 00:29:08.960
But even then, all too often the conversations about those regulations are dominated by industry figures who shape the process and sometimes even present outlandish scenarios like AI presenting a threat to the human race itself to completely sidetrack the discussions.
00:29:09.200 --> 00:29:21.839
The idea that maybe some of these technologies shouldn't be rolled out at all, or that some use cases become off-limits, become harder to contemplate because the narrative we have about digital technology is that once the tech is out in the world, it can never be reined in again.
00:29:22.000 --> 00:29:32.559
A perspective that not only feels defeatist, but is clearly proliferated by the industry to serve its own interests and prevent any public discussion or democratic say over our collective technological future.
00:29:32.720 --> 00:29:41.440
In my view, that can't stand, either when it comes to AI or to data centers, because that's the other piece of this discussion about AI and the bubble currently fueling it.
00:29:41.599 --> 00:29:47.599
Once the crash comes, the generative AI might not fully go away, but neither will the infrastructure that's been built to support it.
00:29:47.759 --> 00:29:50.640
Namely, all those massive hyperscale data centers.
00:29:50.960 --> 00:29:57.279
The data centers are not going to be decommissioned, there's this huge capital expenditure, it's a fixed asset.
00:29:57.440 --> 00:29:59.119
They're going to try to do something with them.
00:29:59.359 --> 00:30:04.400
Data centers are not going to go the way of malls, which like malls are now just skeletons of their former selves.
00:30:04.720 --> 00:30:11.200
There's going to be a demand for computation, but maybe it's not AI, and that's going to have lasting environmental impacts.
00:30:11.519 --> 00:30:14.160
What uses will all that additional computation be put to?
00:30:14.240 --> 00:30:17.759
It's hard to say for now, but we can be pretty certain it won't be for the social good.
00:30:17.920 --> 00:30:24.240
But rather will expand corporate power and further increase the profits of Amazon, Microsoft, and Google.
00:30:28.559 --> 00:30:46.319
We started this episode with an honest but troubling statement from Sam Altman that the future he imagines, where generative AI is integrated through society, regardless of whether it truly has a beneficial impact, will require an unimaginable amount of energy, and that means we either find a breakthrough in energy generation or we begin geoengineering the planet.
00:30:46.559 --> 00:30:53.039
The notion that maybe his vision for the future isn't the ideal one, or the one the rest of the public might not agree to, cannot be fathomed.
00:30:53.200 --> 00:30:59.359
This is the path that he and many of his powerful buddies in the tech industry want to put us on, and thus it must be pursued.
00:30:59.519 --> 00:31:01.279
The rest of us do not have a say.
00:31:01.519 --> 00:31:04.559
We must simply accept it and hope it works for us.
00:31:04.799 --> 00:31:06.079
But that's not good enough.
00:31:06.319 --> 00:31:11.519
Dan argues this isn't just about AI or data centers, but something greater, and I tend to agree with him.
00:31:11.680 --> 00:31:22.079
It's a remaking of society by the new dominant group, who not only want it to function in a certain way, but also want to protect their privileges and are willing to deploy their vast power and wealth to realize it.
00:31:59.200 --> 00:32:01.200
Neoliberalism has kind of run out of steam.
00:32:01.359 --> 00:32:02.240
It's fracturing.
00:32:02.400 --> 00:32:04.160
There's a need for a restructuring.
00:32:04.480 --> 00:32:09.519
There's no desire to involve any kind of social justice or redistribution in that restructuring.
00:32:09.680 --> 00:32:14.720
So we've got to find uh both a mechanism and a legitimation of what we're gonna do instead.
00:32:14.960 --> 00:32:17.519
And AI is one of the candidates for that.
00:32:17.680 --> 00:32:25.440
And I think despite the fact that generative AI is demonstrably bullshit, it's still gonna serve some kind of function in that, whether we like it or not.
00:32:25.759 --> 00:32:27.599
A restructuring sounds about right.
00:32:27.759 --> 00:32:31.200
And it's not just one that doesn't consider the broader concerns of the public.
00:32:31.279 --> 00:32:32.799
It's one we have little say in.
00:32:32.960 --> 00:32:48.720
The effort to roll out these AI technologies, ensure digital technology is at the core of everything we do, and increase the amount of data collected and computation required is wrapped up in all of this, as are the social harms that are already emerging from it, and the broader climate threat presented by the intense energy demands.
00:32:48.880 --> 00:32:56.960
It's no wonder communities are pushing back locally, but stopping these data centers and the new world they're designed to fuel will require even more.
00:32:57.200 --> 00:33:02.160
Next week, we'll explore this ideology more deeply and what another path might look like.
00:33:02.480 --> 00:33:07.200
Data Vampires is a special four-part series from TechWon't Save Us, hosted by me, Paris Marks.
00:33:07.440 --> 00:33:12.079
This original series was produced by Eric Wickham, and updates were made by our producer, Kyla Hewson.
00:33:12.240 --> 00:33:17.119
This series was made possible through the support from our listeners at patreon.com slash techwon'save us.
00:33:17.359 --> 00:33:23.759
We've already uploaded the uncut interviews of some of the guests I spoke to for this series exclusively for Patreon supporters.
00:33:23.839 --> 00:33:32.559
So make sure to go to patreon.com slash techwon'save us to support the show, and make sure to order your copy of Hyperscale, the ambition and excess of big tech's data empires.
00:33:32.720 --> 00:33:36.079
You can find more information about that at hyperscalebook.com.
00:00:00.080 --> 00:00:01.520
Hello and welcome to Tech Won's Save Us.
00:00:01.600 --> 00:00:02.799
I'm your host, Paris Marks.
00:00:02.879 --> 00:00:16.000
And this month we are continuing our re-airing of the Data Vampire series that I made about this build-out of hyperscale data centers, the demands that they have on communities, and why the tech industry is actually doing this in the first place.
00:00:16.160 --> 00:00:17.359
What is really driving them?
00:00:17.519 --> 00:00:28.960
And of course, we're doing that to mark the release of my new book, Hyperscale, which looks into those very same issues and comes out on October 20th in the US, Canada, the UK, and more broadly in Europe.
00:00:29.120 --> 00:00:32.640
If you like this series, I think you are really going to like the book as well.
00:00:32.799 --> 00:00:38.320
And of course, you can find out more information and pre-order a copy for yourself at hyperscalebook.com.
00:00:38.479 --> 00:00:54.320
Now I think it's actually really interesting that this is the episode of Data Vampires that happens to coincide with this week, because as I'm talking to you in 2026, for the past week or so, we've been hearing a lot about the threat of AI to humanity, basically.
00:00:54.479 --> 00:01:07.200
You know, a former worker at Anthropic came out and warned against the risks that generative AI and the AI models that Anthropic is creating, the risks that they pose for humanity itself.
00:01:07.359 --> 00:01:21.680
And continuing this myth that we've been hearing for the past few years that AI is getting closer and closer to being super intelligent and being able to take over the world and effectively control us humans, if not wipe us all out.
00:01:21.920 --> 00:01:27.200
Now, if you've been listening to the podcast for long enough, you will know that I think that is completely science fiction.
00:01:27.359 --> 00:01:59.120
But actually, this episode of the series digs into the wider costs of these generative AI technologies and of course the data centers that power them, and how the real threats are not, you know, this notion of AGI artificial intelligence or or even AI superintelligence, but the actual demands of the technology, what it means for the climate, what it means for communities, as you know, I was talking about last week on the show, and how these are the real problems, not these kind of fantastical science fictional, future-focused solutions.
00:01:59.359 --> 00:02:09.360
All that they do is distract us from the real harms that are being caused in the present that these companies really don't want us to be focused on, and that they certainly don't want regulators to be focused on.
00:02:09.520 --> 00:02:27.039
And of course, that's not to mention how I would argue there are just a lot of people in this industry who are kind of deluded and who have collectively deluded themselves into believing that they're building the science fictional threats that they've seen in movies and novels that they encountered as younger people.
00:02:27.199 --> 00:02:32.879
And that doesn't mean that we should believe everything that they tell us about the technologies that they're building.
00:02:33.120 --> 00:02:38.000
So that's all to say, I think that this episode of the series is very relevant to what we're seeing right now.
00:02:38.159 --> 00:02:47.280
And I think you're going to really enjoy it, whether you've never heard it before, or even if you heard it a couple of years ago, but maybe don't remember all the details that we included in this series.
00:02:47.360 --> 00:02:51.680
You know, the thing that it opens on is wild enough on its own, and you'll hear that in just a minute.
00:02:51.919 --> 00:02:59.439
But then to think about the other statements that these companies and CEOs and, you know, just workers are making about this technology, it's it's wild.
00:02:59.520 --> 00:03:02.560
And that that doesn't mean that we should believe exactly what they're saying.
00:03:02.719 --> 00:03:13.759
And of course, just a final note that if you do enjoy the series, I think you're really going to enjoy my book, Hyperscale: The Ambition and Excess of Big Tech's Data Empires, which, as I said, comes out on October 20th.
00:03:13.840 --> 00:03:16.800
And you can find more information at hyperscalebook.com.
00:03:16.960 --> 00:03:20.479
So with that said, please enjoy this week's episode of Data Vampires.
00:03:21.120 --> 00:03:26.319
We do need way more energy in the world than I think we thought we needed before.
00:03:26.400 --> 00:03:31.039
And I think we still don't appreciate the energy needs of this technology.
00:03:31.280 --> 00:03:37.759
That's Sam Alvin, the CEO of OpenAI, speaking to Bloomberg in January 2024 at the World Economic Forum.
00:03:37.919 --> 00:03:42.159
In that interview, he was lightly pressed on the climate cost of his generative AI vision.
00:03:42.319 --> 00:03:44.000
And he was remarkably honest.
00:03:44.159 --> 00:03:49.199
The future he wants to realize is one that will require an amount of energy that's hard to even fathom.
00:03:49.280 --> 00:04:03.439
And all that energy needs to come on stream in record time at the same moment we're supposed to be phasing out fossil energy in favor of less emitting alternatives like solar, wind, hydro, or in some people's minds, a ton of nuclear energy.
00:04:03.759 --> 00:04:08.479
The good news to the degree there's good news is there's no way to get there without a breakthrough.
00:04:08.639 --> 00:04:16.800
We need fusion or we need like radically cheaper solar plus storage or something at massive scale, like a scale that no one is really planning for.
00:04:17.680 --> 00:04:30.160
So we it's totally fair to say that AI is gonna need a lot of energy, but it will force us, I think, to invest more in the technologies that can deliver this, none of which are the ones that are burning the carbon.
00:04:30.319 --> 00:04:36.000
The way Altman talks about the massive energy demands his AI ambitions are creating is typical of tech billionaires.
00:04:36.240 --> 00:04:40.079
The climate crisis is not a political problem, but simply a technological one.
00:04:40.240 --> 00:04:51.360
And we need not worry because our technocratic overlords will deliver a breakthrough in energy technology so they can continue doing whatever they want, regardless of whether it makes any real sense to do so.
00:04:51.600 --> 00:04:55.759
Even though Altman refers to this as good news, it's hard to see it that way.
00:04:55.920 --> 00:05:09.360
He's basically acknowledging that warming far beyond the 1.5 or 2 degrees Celsius limit we're supposed to be trying to keep to is essentially locked in because of industries like his own, unless they come up with a technological breakthrough in time.
00:05:09.600 --> 00:05:13.600
There's no guarantee that will happen, and in fact, it's highly likely it won't.
00:05:13.759 --> 00:05:23.839
That's why so many of their scenarios assume we're going to overshoot on emissions, but hope we'll be able to use some future technology to pull all those greenhouse gases back out of the atmosphere.
00:05:24.000 --> 00:05:28.480
And again, another massive gamble with the planet and everything that lives on it.
00:05:28.639 --> 00:05:38.240
But in the interview, Altman wasn't just candid on how much energy the widespread rollout of generative AI will require, but also about that more grim scenario.
00:05:38.800 --> 00:05:49.279
I still expect, unfortunately, the world is on a path where we're gonna have to do something dramatic with climate like geoengineering as a as a as a band-aid, as a stopgap.
00:05:49.439 --> 00:05:52.160
But I think we do now see a path to the long-term solution.
00:05:52.480 --> 00:06:04.800
Altman and his fellow AI boosters want us to gamble with the climate to such a degree we have to try to play God with weather systems, all so they can have AI companions and imagine that one day they might be able to upload their brains onto computers.
00:06:04.959 --> 00:06:07.839
It's not only foolish, it verges on social suicide.
00:06:08.000 --> 00:06:12.399
And I don't think that's a trade-off that many people will openly accept.
00:06:34.399 --> 00:06:45.920
Over the course of this series, we'll learn more about data centers and the extreme vision of the future these powerful people in the tech industry are trying to foist on us, regardless of whether we want it or whether it will even make the lives of most people any better.
00:06:46.079 --> 00:06:59.360
In this week's episode, we'll be digging into how generative AI hype is accelerating the data center build out and presenting a series of threats, from worsening climate catastrophe to further social harms, that will only become more acute the longer this is allowed to continue.
00:06:59.600 --> 00:07:10.319
This series was made possible by our supporters over on Patreon, and if you learned something from it, I'd ask you to consider joining them at patreon.com slash techwon'save us so we can keep doing this important work.
00:07:10.480 --> 00:07:18.959
Plus, enjoy premium full-length interviews with the experts I spoke to for the series, and I've put them together in a special collection whose link you can find in the show notes.
00:07:19.120 --> 00:07:22.639
Become a supporter at patreon.com slash techwon'save us today.
00:07:22.800 --> 00:07:29.920
So, with that said, let's learn more about these data vampires, and by the end, maybe we'll be closer to driving a stake through their hearts.
00:07:32.879 --> 00:07:39.120
Since the release of ChatGPT in November of 2022, talk of artificial intelligence or AI has been everywhere.
00:07:39.279 --> 00:07:41.279
Let's be clear, AI is not a new thing.
00:07:41.439 --> 00:07:44.800
The term has been in use for decades and has referred to different things since then.
00:07:44.959 --> 00:07:55.199
When you type a message on your phone and the keyboard suggests the next word, or when you're putting together a document in Microsoft Word and a squiggly line appears beneath the word to tell you it's spelled wrong, that's AI too.
00:07:55.360 --> 00:08:00.800
It's just not the same kind of AI as what powers the chatbots and image generators that are all the rage today.
00:08:00.959 --> 00:08:02.319
That's generative AI.
00:08:02.399 --> 00:08:04.319
And it's what's fueling a lot of these problems.
00:08:04.639 --> 00:08:09.439
Sasha Lucioni is the co-founder and chief scientific officer at Sustainable AI Group.
00:08:09.519 --> 00:08:13.600
And I asked her why this new generative AI is so much more computationally intensive.
00:08:13.759 --> 00:08:14.720
This is what she told me.
00:08:15.040 --> 00:08:22.079
If you compare a system that uses, I guess, extractive AI or good old-fashioned AI, to search the internet and find you an answer to your question.
00:08:22.240 --> 00:08:26.639
It's essentially converting all these documents, all these like web pages from words to numbers.
00:08:26.720 --> 00:08:33.519
And when you're searching for a query, like, I don't know, uh, what's the capital of Canada, it will also convert that query into numbers using the same system.
00:08:33.600 --> 00:08:35.679
And then matching numbers is like super efficient.
00:08:35.840 --> 00:08:36.960
This stuff goes really, really fast.
00:08:37.039 --> 00:08:38.080
It uses no compute at all.
00:08:38.159 --> 00:08:40.159
It's like it can run on your laptop, it can run anywhere.
00:08:40.240 --> 00:08:48.639
But if you're using generative AI for that same task, instead of finding existing text numbers, it's actually generating the text from scratch.
00:08:48.720 --> 00:08:55.759
And I guess the advantage, quote unquote, is that instead of just getting Ottawa, you'll get like maybe a full sentence, like the capital of Canada is Ottawa.
00:08:55.919 --> 00:09:00.159
But on the flip side, the AI model is generating each one of these words sequentially.
00:09:00.240 --> 00:09:04.320
And so like the longer the sentence, the output, the more compute it uses.
00:09:04.559 --> 00:09:12.080
And you know, when you think about it, for tasks, especially like question answering, like finding information on the internet, you don't need to make stuff up from scratch.
00:09:12.159 --> 00:09:13.360
You don't need to generate things.
00:09:13.519 --> 00:09:14.879
You need to extract things, right?
00:09:15.039 --> 00:09:24.480
So I think fundamentally speaking, what bothers me is that like we're switching from extractive to generative AI for tasks that are not meant for that.
00:09:24.960 --> 00:09:31.120
So basically, there's a lot more work that goes into generating text or images than simply trying to identify what you're looking for.
00:09:31.279 --> 00:09:39.679
These generative AI tools are built on general purpose models that were trained on almost any data these companies could get their hands on, often by taking it off the open web.
00:09:39.840 --> 00:09:50.799
That includes everything from Hollywood movies and published books to paintings and drawings made by all manner of artists, and even many of the things you or me have posted on social media and other parts of the web over the years.
00:09:50.960 --> 00:09:54.559
And the vast majority of that data was taken without anyone's permission.
00:09:54.720 --> 00:10:08.639
Now it forms the foundation of the AI tools and models that kicked off all this hype and that have companies of all sorts rushing to adopt generative AI and push it onto regular users, regardless of whether it's really necessary for the task they're trying to accomplish.
00:10:08.799 --> 00:10:10.320
And that all comes with a cost.
00:10:10.720 --> 00:10:18.240
When you're switching between a good old-fashioned extractive AI model to a generative one, like how many times more energy are you using?
00:10:18.480 --> 00:10:24.559
We found that, for example, for question answering, there's like 30 times more energy for the same task, for like answering a question.
00:10:24.639 --> 00:10:31.200
And so what I really think about is like the fact that so many tools are being switched out to generative AI, like what kind of cost does that have?
00:10:31.440 --> 00:10:34.159
Someone recently was like, Oh, I don't even use my calculator anymore.
00:10:34.240 --> 00:10:35.360
I just use Chad GPT.
00:10:35.440 --> 00:10:38.639
And I'm like, well, that's probably like 50,000 times more energy.
00:10:38.720 --> 00:10:43.440
Like I don't have the actual number, but you know, like a solar-powered calculator versus like this huge large language model.
00:10:43.519 --> 00:10:48.320
Nowadays, people are like, I'm not even gonna search the web, I'm gonna ask Chad GPT, I'm not gonna use a calculator, right?
00:10:48.480 --> 00:10:50.480
All of that, what the cost to the planet is.
00:10:50.799 --> 00:10:55.200
And for all that energy, there's no guarantee the outcome is even going to be better or more accurate.
00:10:55.360 --> 00:10:59.759
As Sasha explained to me, these tools operate not based on understanding, but probabilities.
00:10:59.919 --> 00:11:03.200
Again, think of when the keyboard on your phone is suggesting the next word.
00:11:03.360 --> 00:11:10.960
It doesn't know what you're doing, it's using probabilities based on the data it has to see what word has the highest likelihood of coming after what you've already written.
00:11:11.120 --> 00:11:17.120
That's why we so often see examples of ChatGPT and other chatbots generating completely incorrect outputs.
00:11:17.279 --> 00:11:25.279
There's no real understanding there, despite how often tech CEOs try to make us believe their large language models are on the cusp of sentience like a human being.
00:11:25.440 --> 00:11:35.759
But for those generative AI tools to work, they need a ton of computation, which is why Sam Altman says we either need a technological breakthrough in energy technology or to start geoengineering the planet.
00:11:35.919 --> 00:11:38.720
The notion of scaling the tech back is unacceptable.
00:11:38.879 --> 00:11:51.120
But there are only a small number of massive companies that have access to nearly the amount of computation to properly compete in the generative AI game, which is why the massive tech companies, especially Microsoft and Google, have become so involved.
00:11:51.279 --> 00:12:00.240
They're not only providing the cloud infrastructure to power the generative AI hype, they're also making sure they have a lot of influence over the startups finding success in this financial cycle.
00:12:00.480 --> 00:12:06.559
Here's Cecilia Rickapp, the University College London professor from the first episode in the series, explaining how that works.
00:12:07.039 --> 00:12:11.759
In 2019, Microsoft decided to invest one billion dollars in OpenAI.
00:12:11.919 --> 00:12:19.200
Of course, Microsoft, with all the profits it makes annually, has a lot of liquidity and can decide to invest in many different things.
00:12:19.519 --> 00:12:26.879
But Big Tech in particular have decided to pour a lot of money into the startup world as corporate venture capitalists.
00:12:27.120 --> 00:12:32.000
So Microsoft did this with OpenAI, but the main motive is not financial.
00:12:32.080 --> 00:12:36.480
It's not that they want to make more money just like by investing in the company.
00:12:36.559 --> 00:12:47.679
But the way to make more money, it's uh actually about how OpenAI is developing technology, what technology OpenAI was working on, and how Microsoft can steer that development.
00:12:47.840 --> 00:12:52.480
And by doing it, you can eventually get access to that technology earlier.
00:12:52.639 --> 00:13:03.039
So you can adopt it earlier, as Microsoft did with OpenAI, but you eventually may also be able to make extra profits if the company you invested in is successful and starts developing a business.
00:13:03.360 --> 00:13:18.960
In early 2023, Microsoft invested another 10 billion into OpenAI, but Semaphore reported months later that a significant portion of that investment wasn't in cash, but credits for Microsoft's Azure Cloud Computing Platform, what OpenAI needed to train its models and run its business.
00:13:19.200 --> 00:13:27.519
The company is reportedly losing$5 billion a year, but can continue to operate because of the support of powerful and deep pocket benefactors like Microsoft.
00:13:27.679 --> 00:13:38.720
On top of that, Microsoft, Amazon, and Google have effectively rated the talent at Inflection AI, Adept AI, and Character AI, respectively, to the degree that regulators are investigating them.
00:13:38.879 --> 00:13:45.440
Meanwhile, Amazon and Google have both put billions of dollars into Anthropic, and Microsoft has an investment in Mysterio AI.
00:13:45.600 --> 00:13:57.360
This ensures that on its face, the AI ecosystem looks like there are a bunch of new tech companies rising, but those companies are still completely dependent on the dominant players, not just for funding, but also for computation.
00:13:57.519 --> 00:14:00.320
There's one more angle of this, Cecilia pointed out to me though.
00:14:00.480 --> 00:14:11.440
Yes, generative AI is dependent on the centralized computation of major cloud providers by the hype around it and the perception that if companies adopt it, they'll see their share prices rise has accelerated its adoption.
00:14:11.519 --> 00:14:16.879
And by extension, the demand for computation and the energy and water needed to run all those data centers.
00:14:17.279 --> 00:14:21.840
Just because everyone is talking about AI these days, as a big company, you don't want to be left out.
00:14:22.080 --> 00:14:32.159
Basically, what has happened is a much faster adoption, not only of generative AI, but widely of the cloud and widely of all the different forms of AI.
00:14:32.320 --> 00:14:52.559
And because behind all this, we have the power of Amazon, Microsoft, and Google, not only because of the cloud, but also because they have been investing as venture capitalists in pretty much every single AI startup in the world, they keep on expanding not only their profits, but also their control over capitalism at large.
00:14:52.799 --> 00:15:02.080
So, in a way, it has its own specificities, but if we want to put it just in a nutshell, it has fast-forward something that was cooked from way before.
00:15:02.639 --> 00:15:11.519
So, in short, the AI boom isn't just creating this stock market bubble and allowing companies like OpenAI to rise up the ranks with the support of the existing dominant tech firms.
00:15:11.679 --> 00:15:16.080
The growth of generative AI isn't a challenge to companies like Amazon, Microsoft, and Google.
00:15:16.240 --> 00:15:21.279
It further cements their power, especially as other companies, non-tech companies, adopt it.
00:15:21.440 --> 00:15:32.960
Because every time they do so, they're becoming more dependent on the cloud businesses of those three dominant firms, further increasing their power, their scale, and driving a further build-out of major data centers across the world.
00:15:33.120 --> 00:15:39.039
And as we've touched on in the previous episode, all of that comes with a massive environmental impact.
00:15:41.919 --> 00:15:45.279
For quite some time, tech companies have wanted to be seen as green.
00:15:45.440 --> 00:15:52.480
In the picture they painted, digital technology was clean and green, the sustainable alternative to the dirty, polluting industrialism of the past.
00:15:52.720 --> 00:15:57.840
That was always more marketing campaign than reality, though, as the internet doesn't emerge out of nowhere.
00:15:58.000 --> 00:16:04.639
All the technologies that underpin it have serious material consequences that create plenty of emissions and environmental damage of their own.
00:16:04.799 --> 00:16:09.440
But as efforts were ramping up to tackle the climate crisis, they wanted to keep that image alive.
00:16:33.519 --> 00:16:34.480
That's Brad Smith.
00:16:34.639 --> 00:16:40.000
He's the president of Microsoft, and that clip is from an interview he gave to Bloomberg back in January of 2020.
00:16:40.159 --> 00:16:42.399
Microsoft was rolling out a new climate pledge.
00:16:42.559 --> 00:16:46.799
It would not just achieve net zero emissions, but become carbon negative within a decade.
00:16:46.960 --> 00:16:52.639
The company called this a carbon moonshot, indicating it was ambitious, but a goal they thought they could achieve.
00:16:52.799 --> 00:17:02.240
Well, that was before generative AI became the next big thing that virtually everyone in Silicon Valley felt they had to chase, and that Microsoft saw could significantly expand its cloud business.
00:17:02.399 --> 00:17:05.359
Here's Brad Smith again in May 2024 this time.
00:17:10.960 --> 00:17:14.079
Our goal of being carbon negative by 2030.
00:17:14.400 --> 00:17:18.160
That was before the explosion in artificial intelligence.
00:17:18.720 --> 00:17:23.680
So in many ways, as I say across Microsoft, the moon has moved.
00:17:24.160 --> 00:17:28.960
It's more than five times as far away as it was in 2020.
00:17:29.039 --> 00:17:35.680
If you just think about our own forecast for the expansion of AI and its electrical needs.
00:19:05.200 --> 00:19:08.799
But why can these companies claim to emit so much less than they really do?
00:19:08.960 --> 00:19:12.400
One expert The Guardian spoke to call it a form of creative accounting.
00:19:12.559 --> 00:19:18.240
Basically, they buy a bunch of offsets and act as though having done so means their emissions have been negated.
00:19:18.480 --> 00:19:27.599
Probably the most important of those tools are renewable energy certificates, which shows they've bought renewable energy that can be produced at another time of day or on the other side of the world.
00:19:27.839 --> 00:19:35.119
As long as it was generated somewhere, the companies use it to pretend they didn't actually generate the emissions, they very much did add to the atmosphere.
00:19:35.279 --> 00:19:43.119
And some tech companies are lobbying hard to ensure the rules on carbon accounting are rewritten to make it look like they're emitting way less than they really are.
00:19:43.359 --> 00:19:54.559
According to reporting by the Financial Times, Amazon and Meta are leading the charge to ensure their deceptive accounting mechanisms are legitimized by the Greenhouse Gas Protocol, which is an oversight body for carbon accounting.
00:19:54.720 --> 00:20:04.079
Google is pushing a competing proposal that would force companies to at least buy renewable certificates that are closer to where they're actually operating, but still relies on offsets at the end of the day.
00:20:04.240 --> 00:20:07.359
Companies like Amazon say even that would be too expensive.
00:20:07.519 --> 00:20:15.359
Matthew Brander, a professor at the University of Edinburgh, who spoke to the Financial Times, gave a pretty good example to show why this is all so ridiculous.
00:20:15.599 --> 00:20:24.559
He said allowing companies to buy renewable certificates is like if you paid a Fidder colleague of yours for the right to say you bite to work when you really drove your gas-powered car.
00:20:24.720 --> 00:20:32.000
It's foolishness, but this is how they're planning to keep expanding their data center networks while claiming they're reducing, if not eliminating, their emissions.
00:20:32.160 --> 00:20:35.359
It's a recipe for disaster on a global scale.
00:20:38.079 --> 00:20:46.000
We've talked a lot about why AI is using a ton of computation and further fueling the climate crisis, but what is all the compute we're putting into it really achieving?
00:20:46.160 --> 00:20:50.880
Maybe there's a world where all those resource demands are justified because the benefits are so great.
00:20:51.039 --> 00:20:56.480
And indeed, that's what tech CEOs like Sam Altman or supposed luminaries like Bill Gates would have us believe.
00:20:56.640 --> 00:21:01.200
But the truth is that the rollout of this technology only presents a further threat to much of the public.
00:21:01.440 --> 00:21:16.079
We're used to hearing about AI as forming the basis for a series of tools that can do all manner of tasks, but I was struck by how two of the people I spoke with described the broader project that AI seems to be part of when you consider who is developing it and how it's actually being deployed.
00:21:16.240 --> 00:21:18.079
Let's start with Ali Al-Khatib.
00:21:18.240 --> 00:21:22.559
He used to be the head of the Center for Applied Data Ethics at the University of San Francisco.
00:21:22.799 --> 00:21:32.319
When I asked him how he would describe AI, he began by noting how the term itself is decades old, but there was a troubling through line between its various permutations over the years.
00:21:32.799 --> 00:21:59.519
I think the thing that we would all recognize all the way through, continuously, like is the techno-political project of taking decisions away from people and putting consequential, life changing decisions into a locus of power that is silicon or that is automated or something along those lines, and redistributing or shifting and allocating power away from collective and social systems.
00:22:00.160 --> 00:22:02.720
And into technological or technocratic ones.
00:22:02.880 --> 00:22:08.480
And so this isn't really like a definition of AI that I think a lot of computer science people would appreciate or agree with.
00:22:08.640 --> 00:22:19.680
But I think it's the only one that, again, if you were a time traveler, kind of like going back 20 years and then 20 more years and then 20 more years, you would see totally different methods, but I think you would see basically the same goals, basically the same project.
00:22:20.079 --> 00:22:30.960
Ali's description is unlike anything you'll hear from industry boosters who want you to see AI as a way to improve many aspects of human life, or on the extreme end, thinking it could end humanity if not done right.
00:22:31.119 --> 00:22:42.319
They don't want to talk about that more political angle, the way it's used to cement their power in a way that can be harder to immediately identify than, say, the outwardly authoritarian actions of a politician or leader.
00:22:42.559 --> 00:22:52.240
AI much more quietly erodes the power of much of the public over their own lives, taking away their autonomy by shifting decisions to unaccountable technologies and the people who control them.
00:22:52.319 --> 00:22:58.160
This is something Dan McQuillan, a lecturer at Goldsmiths University and author of Resisting AI, identified too.
00:22:58.480 --> 00:23:07.279
AI is a specific in our faces example of a general technological phenomenon which claims to solve things technically.
00:23:07.519 --> 00:23:13.839
And you know, we see that across the board from tricky social issues all the way up to the climate crisis.
00:23:14.079 --> 00:23:35.839
But I think that that sort of diversion aspect is really an important aspect of contemporary AI, exactly because these issues are so urgent and other forms of collective, social, uh grounded community action and worker action are so urgently needed, that something that successfully, even semi-successfully diverts us from those things is extremely toxic.
00:23:36.000 --> 00:23:51.359
So I'm really talking about their AI as a narrative, AI as an idea, AI as a real technology that appears to do certain things, you know, that can emulate certain things or synthesize certain things in a way that provides people with a plausibility argument that maybe this could fill the hole in health services or education or whatever.
00:23:51.440 --> 00:23:52.240
So that's the technology.
00:23:52.480 --> 00:24:00.799
In Dan's telling, AI isn't just a digital technology made up of complex algorithms and underpinned by the material computational infrastructures that drive it.
00:24:00.960 --> 00:24:29.680
It's also a social technology, one that's deployed so the powerful can claim to be addressing what are very pressing problems in society, the lack of healthcare, inequitable access to education, growing poverty and inequality, not to mention the accelerating climate crisis, without having to actually take the extent of the difficult political measures that would really be necessary to tackle them, measures the elites in our society likely don't want to see taken in the first place, as it might erode their power and certainly require their wealth to be taxed at much higher rates.
00:24:29.920 --> 00:24:35.839
Instead, AI, like too many other digital technologies, can be presented as a seemingly apolitical solution.
00:24:36.079 --> 00:24:40.640
It doesn't require sacrifice and doesn't challenge the hierarchy of capitalist society.
00:24:40.799 --> 00:24:43.599
Indeed, if anything, it further solidifies it in place.
00:24:43.759 --> 00:24:57.680
And all we need to do as a public is have a little patience as our saviors in the tech industry perfect their technofixes so they can deliver us a digital utopia, which, it probably doesn't need to be said, never actually arrives as those deeper issues just keep getting worse.
00:24:57.920 --> 00:25:02.160
There are many harms we can talk about with generative AI, and some of the more common forms of it too.
00:25:02.319 --> 00:25:11.279
We could talk about how companies are stealing all this data and using it to harm the prospects of workers in different industries, like in visual media, writing, journalism, and more.
00:25:11.440 --> 00:25:25.200
Or we could talk about the waves of AI generated bullshit flooding onto the web, some with malicious intent like non-consensual, deep fake, and AI nudes, but much more of it being made just to try to make a buck through social media engagement or tricking people into scams.
00:25:25.519 --> 00:25:37.519
Those things are important, but the deeper issue to me seems to be those that Ali and Dan are describing, and which Alex Hannah, the director of research at the Distributed AI Research Institute, outlined in a bit more detail when I spoke with her.
00:26:16.319 --> 00:26:24.240
We're seeing more and more at the border intense amounts of AI and automated decision making with biometrics.
00:26:24.480 --> 00:26:32.720
That is not necessarily generative AI, but there are other kinds of things that could be looped in with generative AI, which are used at the border.
00:26:32.960 --> 00:26:37.680
Healthcare, education, legal access, virtually anything that happens on the border.
00:26:37.839 --> 00:26:42.079
And the list of all the places they're trying to falsely present AI as a solution goes on.
00:26:42.319 --> 00:26:53.599
Ultimately, generative AI is another one of the tech industry's financial bubbles, where its leading figures hype up the next big thing to drive investment and boost share prices until reality starts to creep in and the crash begins.
00:26:53.839 --> 00:27:10.880
We saw it most recently with cryptocurrencies and NFTs, but there are already questions about how long the generative AI bubble is going to last, with everyone from Goldman Sachs to Sequoia Capital starting to join the existing chorus of critics in calling out the aspects of generative AI that are clearly inflated and poised to crash.
00:27:11.119 --> 00:27:16.559
Even after that crash, generative AI won't fully go away, just as other forms of AI have stuck around as well.
00:27:16.720 --> 00:27:20.720
It won't be everywhere, or have the widespread implementations the companies promised.
00:27:20.880 --> 00:27:24.400
But that doesn't mean there still won't be threats that emerge from its ongoing presence.
00:27:24.559 --> 00:27:29.920
As Ali explained to me, we'd be foolish to think it can be seized and redirected to mostly positive ends.
00:27:30.160 --> 00:27:37.440
If people are designing these systems to cause harm, fundamentally, then there kind of is no way to make a human-centered version of that sort of system.
00:27:37.599 --> 00:27:52.559
In the same way, legislation that makes it slightly more costly to do something harmful doesn't necessarily fix or uh even really discourage tech companies that find ways to amortize those costs or kind of absorb those costs into their business model.
00:27:52.799 --> 00:28:03.920
One example that I think I've given recently in like conversation was that there are all sorts of reasons or all sorts of powers that cause us to behave differently when we're driving on the streets.
00:28:04.079 --> 00:28:11.839
Because as individual people, the costs of crashing into another car or of hitting a pedestrian or something like that are quite substantial for us as individuals.
00:28:12.079 --> 00:28:24.000
But if a tech company that's developing autonomous cars is going to put 100,000 or a million cars out onto the streets, it really behooves them to find a way to legislatively make it not their fault to hit a pedestrian, for instance.
00:28:24.079 --> 00:28:34.400
And so they find ways to sort of defer the responsibility for who ultimately like caused that harm or who takes the responsibility for whatever kind of incident or whatever.
00:28:34.559 --> 00:28:43.200
And so that creates like these really wild perverse incentives to find ways to sort of consolidate and then offload responsibilities and consequences for violence.
00:28:43.440 --> 00:28:50.240
And I just don't see a good way with design out of that, or even with a lot of legislative solutions and everything else like that.
00:28:50.480 --> 00:28:54.559
When the harms are acknowledged, the discussion around AI is about how to properly regulate it.
00:28:54.720 --> 00:29:08.960
But even then, all too often the conversations about those regulations are dominated by industry figures who shape the process and sometimes even present outlandish scenarios like AI presenting a threat to the human race itself to completely sidetrack the discussions.
00:29:09.200 --> 00:29:21.839
The idea that maybe some of these technologies shouldn't be rolled out at all, or that some use cases become off-limits, become harder to contemplate because the narrative we have about digital technology is that once the tech is out in the world, it can never be reined in again.
00:29:22.000 --> 00:29:32.559
A perspective that not only feels defeatist, but is clearly proliferated by the industry to serve its own interests and prevent any public discussion or democratic say over our collective technological future.
00:29:32.720 --> 00:29:41.440
In my view, that can't stand, either when it comes to AI or to data centers, because that's the other piece of this discussion about AI and the bubble currently fueling it.
00:29:41.599 --> 00:29:47.599
Once the crash comes, the generative AI might not fully go away, but neither will the infrastructure that's been built to support it.
00:29:47.759 --> 00:29:50.640
Namely, all those massive hyperscale data centers.
00:29:50.960 --> 00:29:57.279
The data centers are not going to be decommissioned, there's this huge capital expenditure, it's a fixed asset.
00:29:57.440 --> 00:29:59.119
They're going to try to do something with them.
00:29:59.359 --> 00:30:04.400
Data centers are not going to go the way of malls, which like malls are now just skeletons of their former selves.
00:30:04.720 --> 00:30:11.200
There's going to be a demand for computation, but maybe it's not AI, and that's going to have lasting environmental impacts.
00:30:11.519 --> 00:30:14.160
What uses will all that additional computation be put to?
00:30:14.240 --> 00:30:17.759
It's hard to say for now, but we can be pretty certain it won't be for the social good.
00:30:17.920 --> 00:30:24.240
But rather will expand corporate power and further increase the profits of Amazon, Microsoft, and Google.
00:30:28.559 --> 00:30:46.319
We started this episode with an honest but troubling statement from Sam Altman that the future he imagines, where generative AI is integrated through society, regardless of whether it truly has a beneficial impact, will require an unimaginable amount of energy, and that means we either find a breakthrough in energy generation or we begin geoengineering the planet.
00:30:46.559 --> 00:30:53.039
The notion that maybe his vision for the future isn't the ideal one, or the one the rest of the public might not agree to, cannot be fathomed.
00:30:53.200 --> 00:30:59.359
This is the path that he and many of his powerful buddies in the tech industry want to put us on, and thus it must be pursued.
00:30:59.519 --> 00:31:01.279
The rest of us do not have a say.
00:31:01.519 --> 00:31:04.559
We must simply accept it and hope it works for us.
00:31:04.799 --> 00:31:06.079
But that's not good enough.
00:31:06.319 --> 00:31:11.519
Dan argues this isn't just about AI or data centers, but something greater, and I tend to agree with him.
00:31:11.680 --> 00:31:22.079
It's a remaking of society by the new dominant group, who not only want it to function in a certain way, but also want to protect their privileges and are willing to deploy their vast power and wealth to realize it.
00:31:59.200 --> 00:32:01.200
Neoliberalism has kind of run out of steam.
00:32:01.359 --> 00:32:02.240
It's fracturing.
00:32:02.400 --> 00:32:04.160
There's a need for a restructuring.
00:32:04.480 --> 00:32:09.519
There's no desire to involve any kind of social justice or redistribution in that restructuring.
00:32:09.680 --> 00:32:14.720
So we've got to find uh both a mechanism and a legitimation of what we're gonna do instead.
00:32:14.960 --> 00:32:17.519
And AI is one of the candidates for that.
00:32:17.680 --> 00:32:25.440
And I think despite the fact that generative AI is demonstrably bullshit, it's still gonna serve some kind of function in that, whether we like it or not.
00:32:25.759 --> 00:32:27.599
A restructuring sounds about right.
00:32:27.759 --> 00:32:31.200
And it's not just one that doesn't consider the broader concerns of the public.
00:32:31.279 --> 00:32:32.799
It's one we have little say in.
00:32:32.960 --> 00:32:48.720
The effort to roll out these AI technologies, ensure digital technology is at the core of everything we do, and increase the amount of data collected and computation required is wrapped up in all of this, as are the social harms that are already emerging from it, and the broader climate threat presented by the intense energy demands.
00:32:48.880 --> 00:32:56.960
It's no wonder communities are pushing back locally, but stopping these data centers and the new world they're designed to fuel will require even more.
00:32:57.200 --> 00:33:02.160
Next week, we'll explore this ideology more deeply and what another path might look like.
00:33:02.480 --> 00:33:07.200
Data Vampires is a special four-part series from TechWon't Save Us, hosted by me, Paris Marks.
00:33:07.440 --> 00:33:12.079
This original series was produced by Eric Wickham, and updates were made by our producer, Kyla Hewson.
00:33:12.240 --> 00:33:17.119
This series was made possible through the support from our listeners at patreon.com slash techwon'save us.
00:33:17.359 --> 00:33:23.759
We've already uploaded the uncut interviews of some of the guests I spoke to for this series exclusively for Patreon supporters.
00:33:23.839 --> 00:33:32.559
So make sure to go to patreon.com slash techwon'save us to support the show, and make sure to order your copy of Hyperscale, the ambition and excess of big tech's data empires.
00:33:32.720 --> 00:33:36.079
You can find more information about that at hyperscalebook.com.
