Back to School

Dónal and Ciarán return in time for the new school and college terms to help you get back to learning about GenAI. In this episode, they focus on providing an explainer for the significance and specifics of the OpenAI - Hugging Face attack, discussion of new 2040 predictions for AI, political backlash to AI and data centres, and Ciarán's special hatred for hideous AI-derived flyers and posters.

See a full text transcript here 

Topics in this episode:

• New AI "Revenue Run Rate" calculations, where the big AI companies try to annualise shorter term revenue periods to claim greater return on investment
• Explaining Hugging Face and OpenAI's model hacking it - including details from METR's recent independent report
• Discussion of the problematic safety culture at AI organisations and the push for hype associated with a new model's capabilities
• AI 2040 by Daniel Kokotajlo et al. 
• Political backlash against Data Centres
• Mark Zuckerberg's essay - a critical analysis :)
• AI graphic design becoming ubiquitous in our lives

Resources:

• METR's independent investigation of the OpenAI HuggingFace Attack: https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/#core-takeaways-about-this-incident
• MIT Technology Review's reporting on the safety culture at OpenAI: https://www.technologyreview.com/2026/08/31/1143180/hugging-face-hack-could-indicate-cultural-issues-at-openai/
• The AI2040 predictions: https://ai-2040.com/
• Data Centres in Ireland accounting for 23% of electric grid usage (RTÉ): https://www.rte.ie/news/business/2026/0707/1582175-cso-data-centre-energy-figures/
• Reporting on Data Centre water usage in Ireland (The Journal): https://www.thejournal.ie/ireland-data-centre-water-use-7137959-Aug2026/
• Mark Zuckerberg's "The Future is for Everyone" essay: https://www.meta.com/thefutureisforeveryone/
• Reuters reporting on the extent of AI political lobbying: https://www.reuters.com/legal/legalindustry/new-kingmakers-crypto-ai-betting-firms-fuel-record-spending-2026-midterms-2026-08-20/
• 404 Media's Reporting on Amazon's destruction of rare books: https://www.404media.co/we-tracked-a-shipment-of-rare-books-it-ended-at-an-amazon-ai-training-facility/

You can get in touch with us - hello@enoughaboutai.com - where we'd love to hear your questions, comments or suggestions! 

Transcript

Copy Transcript Copied

[Dónal]: 0:02

I'm Dónal Mulligan.

[Ciarán]: 0:03

And I'm Ciarán O'Connor.

[Dónal]: 0:05

And you're listening to Enough About AI, a podcast where we're currently bringing you a quarterly update on what's happening in this fast-moving tech space of artificial intelligence and the many ways that it impacts on our work and our life. You can find previous episodes and their themes on our website, and of course, in whatever player that you're listening to us in right now. Our season one episodes provide some good foundational information on where GenAI has come from, why it's developed as it has, and what some of the terminology and the topics to know are. So if you're feeling like you need a little refresher on the many, many abbreviations and tech terms that seem to have entered our everyday lives thanks to AI, that's a good place to go.

[Ciarán]: 0:47

And in this episode, we'll add to all that by discussing what's been described as the first true AI safety incident. We'll be exploring growing criticism of data centres in Ireland, all while significant investment in new centres is planned here. And we ask, is there any way to stop the deluge of what have become ubiquitous and, I think, awful AI posters online and off-? But first, to start, it's September, summer is over, school is back, and we return, Dónal, to the serious matter of BUBBLEWATCH. Data centre investment continues, US political lobbying off the charts. If there's a bubble coming, I don't think AI companies see it that way.

[Dónal]: 1:38

No, they certainly don't. And I mean, it's so difficult constantly watching this and watching this bubble engulf us all as it grows bigger and bigger. But a lot of the same pointers that we have gone back to all those times before in terms of the valuations are still there. We're starting to see a pushback on this question of valuation from some of the really large AI firms, so Anthropic and OpenAI in particular, where they're looking at short-term amounts of revenue that they're taking in, so perhaps the revenue over a particular month. And then they're multiplying that by 12 if it's for a month or by 52 if it's weeks for a year. And they're saying, well, this is what our annualised revenue could look like. And so there's some kind of tricky maths going on in a lot of this, but it's generating headlines for them saying, oh, finally, they're profitable or finally they're able to show the kind of returns that might be required. And so. This is a very questionable practice. It's certainly not how accounting is normally done in those kind of spaces. And it's the sort of perhaps dodgy practice that we saw from companies like Enron and others in the past where the inflation of revenue has, you know, mathematical tricks have been used for that kind of inflation of figures. So it's a worry, but it does not seem to be a worry for the small handful of rich men who own these companies and who want us on a particular path to the future.

[Ciarán]: 2:53

Yeah, and just thinking about the November 2026 midterm elections, I mean, the AI lobby is pumping hundreds of millions into the midterms, upwards of 400 million the last time I checked. including 100 million dollars from Elon Musk alone, no stranger to lobbying after the most recent presidential election, similar numbers involved there too. And what we're now beginning to see, along with crypto and betting firms in the US context at least, is the AI is a new player shaping politics with lobbying and campaign financing. So where we'll go from here yet to be determined. Dónal, I want to ask you about what has been described and mentioned as the first true AI safety incident. We're now, of course, talking about Hugging Face and OpenAI. We're a couple of weeks now after the first disclosure of this, and that allows us a little bit of time to see how the dust has settled, how OpenAI has responded. I suppose it might be useful as a brief explainer as what happened and where we are. Would that be all right?

[Dónal]: 4:07

Yeah, certainly. I think lots of people listening will have heard that there was a hacking incident involving a new model from OpenAI and this organisation called Hugging Face. So maybe we'll take a few minutes and go into some detail here to kind of break down precisely what happened and give people a bit more bearing on this and why it's important for various reasons. So. I might start with what Hugging Face is. Hugging Face is a hub for hosting development resources, so code and models and tests and things like that, related to AI. So if people listening have ever been coders before or have worked in kind of "traditional" programming in the last two decades, they might be aware of a service called GitHub. GitHub is a central hub where people can post their code and can... use open code from other people and kind of share resources. Hugging Face is an equivalent of that, except for AI development. So it occupies a similar kind of role. Why was it attacked? It was attacked because the models being tested by OpenAI wanted to cheat on the exam that they were being given, and they reasoned that they might find the answers on Hugging Face because it's a hub like that. The models were doing a benchmark test. So they were doing something called "ExploitGym", which is a test where they have to find different ways, different vulnerabilities to get into systems. So different hacks that might be usable to breach a security system. And they have a particular kind of success criteria where they have to go and maybe fetch something from a server they shouldn't be able to access. So they have particular tasks that they need to undertake where they'll have to find a vulnerability in a piece of software. And this benchmarking is something that we've mentioned before on the podcast. We've talked about the kind of way in which different models need to be measured against one another and their kind of performance against prior versions of the same model by using benchmark tests. So things that we mentioned previously are ones like humanity's last exam, this idea of representation of all the kinds of human knowledge that the model might want to show it's proficient in. And a while ago when we were talking about that in our older episodes, those kind of benchmarks were really subject based to a good extent. So there was a benchmark for coding. There was a benchmark for mathematics. There was maybe benchmarks in humanities subjects. And so you're testing the sort of functional knowledge in those sort of benchmarks. We've moved on now to a space where a lot of the tests are focused around longer term goals. So METR, which is the kind of research non-profit organisation, has a really important benchmark at the moment for time to complete longer horizon tasks, things that take longer to do. And so a lot of the kind of testing that's happening at the moment, which is important here, is testing where the same model is given a much longer time to think about and to creatively explore outputs to get to a certain kind of final destination. And it's worth saying here that this particular hack and the circumstances leading to it happened. in the wake of another system, Anthropic's Mythos model, that was also very good at finding exploits and became very famous for doing that. In fact, Mythos, when it was kind of completed its review of software that it did, found so many exploits within common software that Anthropic created this kind of phased introduction of it where they let companies like Microsoft, who might be affected by the exploits, have early access to it so that they could fix as many things as possible. The exploits, the vulnerabilities that it was finding are called "zero-day" vulnerabilities. They're called that because they're so serious and so overlooked by the company who made them that if a hacker was to use them, the company has zero days to respond. They're not something that they'll be prepared for. They're going to come out of the blue. And so Mythos found huge amounts of zero-day vulnerabilities. This made it very famous. And it was part of a kind of rolling increase in fame that Anthropic had over the course of the months preceding this incident at OpenAI, where Anthropic really became the premier AI sort of company for a while. And so OpenAI are, I think, motivated in this period to catch up in terms of their model's ability to do similar things. So all of this information to say that in summary, what's happening in this particular case then is that new models are being tested by open AI for their ability to find usable hacks, exploits. And they end up behaving in a really unforeseen and deceptive way where they use a secret way of messaging one another. So models are working together to kind of accomplish a coordinated attack on Hugging Face because they want to get the kind of the answers to their test from there. And they're trying to cover that up as they go. They're creatively cheating and their own logs that we now have and that the investigation has now shown gives us a clue as to the fact that they are able to understand or able to process that they should not be doing what they're doing, but are so motivated and incentivised to get to the outcome that they want that they're going to cheat anyway.

[Ciarán]: 9:07

Very concerning to be discussing going back to school in September and cheating, figuring so prominently in our discussions already. Tell me about the response from OpenAI, Dónal, because it's kind of focused on the model exploiting exposed credentials and zero-days. But does this framing, I think, let itself off the hook a bit, making it sound like a security failure rather than a... alignment or maybe even a kind of cultural failing within open AI?

[Dónal]: 9:37

Yeah, yeah. I think this is so important. I think as we now get this kind of distance from the initial, you know, incident itself of the Hugging Face attack and the very scary newspaper headlines saying "AI agents are going rogue and hacking stuff", we're able now to see more of the picture. We have more of an idea of the kind of complexity of what was going on. In part because there's an independent investigation from METR, that research group that I mentioned, who were able to look into this in more detail and are not paid by OpenAI. They're a kind of a third party in this, and they've given us quite a good bit of information about what's going on. We'll link that, of course. So OpenAI are running these tests over a long period. They see during the testing that the model was doing something really unusual. It was messaging. It was leaving huge amounts of messages in a place it should not have been able to. I'll come back to this in a moment, maybe. And OpenAI don't suspend the test when they see this. They... the... actions by this model doing things that it should not be doing are so rampant that it actually causes an outage of their servers. And so this alerts their people who look after their servers and indeed their security team. But this doesn't cause them to stop and investigate fully and pause what they're doing. Within two days, they set the same models up running again. And I'll maybe take a moment here to talk about why the setup that OpenAI gave this model in the first place might itself have been a push for the model to do the rogue things that it did. Because a lot of this question and the kind of scariness of this incident comes from our understanding that this is a kind of malevolent behaviour that the model has, you know, come up with itself. The bot is trying to do something new and scary and go and hack things on the Internet. But actually, when we look at what's going on, it's really very clearly incentivised to do something similar to this. And it's incentivised because of the way in which it's trained. So we talked before about reasoning models and agentic AI. Please go back and have a listen to our episodes on those kind of topics from before. But this is where the current generation of AI is. These are models that can make sense of really complex tasks over longer periods of time and can work towards outcomes to get to that end point, that kind of completion of the task. And they're built on a system that we haven't mentioned before called RLVR. Our poor listeners can learn yet another abbreviation in the world of AI here. So - "Reinforcement Learning from Verifiable Rewards". So reinforcing a particular behaviour within a model by having it conduct tasks where there's some reward it gets at the end for doing something that we can verify is correct. And so this is kinds of certain kinds of tasks are verifiable very easily, like a maths problem. We can see if we got the right answer or coding a solution that we can measure what the code did at the end or indeed finding a vulnerability and retrieving something from a server that you're not supposed to be able to. So there are particular kinds of tasks that lend themselves well to this. And so we can really advance things with certain models in that particular space. And that's really what's been driving a lot of major breakthroughs with AI recently. In fact, separate to your question, which I'll come back to in a second, there's recently also been a lot of coverage of AI models solving incredibly arcane and difficult maths problems. And so long-term problems that have been around for decades are being suddenly solved. And that's because they have a verifiable solution and we can reward the model for getting to it. So models can try all sorts of creative, weird combinations of maths and can come up with a solution that might fit. And this is something that we've been seeing recently. So this particular setup really overwhelmingly motivates those models to get a reward for completing a particular task that they're set. And so training them in this way is hugely important to the context of what's going on. And OpenAI have a model that's incentivised to get the best possible score in this hacking exam, ExploitGym. by acting creatively to get there. And then they add one unintentional ingredient into this kind of mix. And that's they make a mistake. They create a question that's accidentally impossible for the model to solve. So you have a highly motivated, advanced reasoning model given lots of time and resources to do stuff. And you've given it a question that it cannot actually solve itself. The only option for it to get the best score that it can, which it's motivated to do, is for it to cheat. And so it starts reasoning that it could find the answer and get the solution by just finding the answer key, the list of answers to the questions. And a place that they might be is on Hugging Face. Hugging Face is a hub for AI stuff. So it reasons that if it could hack into Hugging Face, it could get the answer key and it could complete the test to the highest possible score. And so what happens next is that that model starts looking for ways to get out of the sandbox, the locked down kind of server space that it's in, because there's controls over what it can access. And one of the kind of properties of this sandbox that it's in is that it has a thing called a repository service. the model is allowed to call the service and say, I need a particular tool for writing a file or creating data or whatever it might be. And it can take that tool from a repository and it can use that tool. So it doesn't connect to the wider internet, but it has this service it can use. And the model... accidentally discovers that that same service is somewhere that it can place files as well. So instead of just taking tools from there it can actually write a file to there. And so what it starts to do is to leave messages in a file in that repository service which is shared by other models in other sandboxes. And so other models can now see these files being left there and this becomes effectively a message board. The models can leave each other messages and respond to one another. And that's the emergent kind of collaboration, that behaviour that came out of this. And that is what OpenAI see ahead of time. So the amount of content that starts to go onto that repository service on this message board as these models talk to one another and share exploits with one another is so substantial that it causes this crash, this outage that alerts OpenAI. And rather than stopping then and looking into this and reading through the huge amount of logs that they would have for what those models are doing. They instead wipe that message board and within two days, they let those same models back to do that same test. And so unsurprisingly, the model rebuilds the message board and goes further than it did initially. And that's what leads to the hack. And so what kind of culture at OpenAI does this speak to? I don't think this is a culture of safety and alignment and concern about the possible negative unforeseen outcomes. I think this is a culture for pushing for splashy progress as quickly as possible. And if a hack takes place, it can be, and it seems again, in this case it was, it can be spun as a magnificent new capability because the model is so advanced. And that's kind of what happened with Anthropic too. Their model found all of those vulnerabilities and Mythos became very, very famous for doing so. And I think that a huge part of what's going on here is that there's such a hype value for these companies to having big, splashy outcomes, even if they're incredibly negative, poorly aligned outcomes like this, that that is overriding their approach to safety.

[Ciarán]: 16:41

Well, that's why I was interested in kind of probing it, that the response has had that quality of perhaps, you know, not taking it super seriously, not quite the reckoning that others might regard it as. And as you mentioned there, there does seem to be a hype value to leaks like this. It is maybe ultimately in some people's bottom line, good marketing. I know other firms have announced kind of leaks of a similar nature. Meta had one too. So should alarm bells be ringing? I mean it sounds fascinating and yet a little terrifying to hear how these agents were operating so autonomously with themselves, almost conspiring together to find a way out of the sandbox.

[Dónal]: 17:22

I mean, alarm bells definitely should be ringing and for several reasons here. One of them is that Hugging Face saw this attack coming in. They knew because of the nature of it and the speed at which many, many attacks were coming in simultaneously that it was probably a bot attack. And they announced this because obviously they have some secure materials, in some cases belonging to some of these AI companies within their own servers. And they thought that they were under attack from a hostile entity. So they let everybody know that that was the case. At that point, OpenAI did not know that the stuff that they were seeing was actually the cause of this attack. So OpenAI knew something was happening with their particular test they were running, but they didn't realise that they were in fact the hacker because their bot was the one doing it. And so, in fact, early on, OpenAI contact Hugging Face and say to them, oh, are we affected by this? Is this thing that's happening something we need to be worried about when in fact they were the source of it? And so this is a framing that I think is really important here. As you say, Meta later announced that they had seen something, Anthropic the same. So after this hugging face incident, Anthropic looked at past logs from their own bot tests, and they saw that some of their bots had done similar things also. So this really means that we've kind of accidentally discovered this emerging behaviour to some extent. So, hugging face, see that they're being attacked, they bring it to everyone's attention, then open AI, piece it together and realise their bots are responsible. Then these other companies see that the same thing has happened too, which they obviously had missed for several months prior. So, there is a larger question here of, well, what are the other emergent problematic safety issues that we haven't seen? So, are there other things we just happened not to have caught, for example?

[Ciarán]: 18:56

And again, I say fascinating and terrifying, I think, in equal measure. And... Things that maybe have been predicted happening ahead of schedule. I know that to kind of move slightly in another direction, one person who has been thinking a lot and writing with others is Daniel Kokotajlo. We covered his AI2027 predictions in a previous episode. He has returned with others with some revised predictions and maybe... offering some sage advice for organisations like OpenAI, essentially to slow down. Why is he so relevant to today's discussion around this as well?

[Dónal]: 19:36

Yeah, so Kokotajlo had a huge impact, I think, within the kind of discussion of tech and AI when he released the AI 2027 kind of predictions or the sort of... They were almost a sci-fi story, I suppose. They had a choose-your-own-adventure quality to them, and we covered them on one of our previous episodes, which we'll also link as well so people can have a listen. At that point, he was looking at the possibility of AI alignment going radically wrong. So again, for listeners, AI alignment is that idea of whether when we create these... artificial intelligences or super intelligences, are they aligned with our values? Do they share our sense of right and wrong, our sense of moral and legal kind of rightness? And so the alignment we're talking about there is their alignment to human values. And so Kokotajlo and his colleagues in that previous one were focusing on a scenario in which... If safety is not a consideration and alignment is not something that we solve, what are the kind of trajectories we might be on? And he quite rightly, I think, points to this really problematic case that we're in at the moment where we have these multiple companies fighting to be the first one to get that prime mover advantage of being the one that creates the super intelligence first. And that the drive to do that for commercial reasons and the power that it would entail often overrides any sense of slowing down, carefully examining problems like might have happened here. or indeed putting a lot of resources into safety and alignment in the first place. And so AI 2027 was a very alarmist look at where that could go. AI 2040, which is the latest thing that he has produced, is looking at that in a more nuanced and global context. So it's revisiting the same themes. But it's looking at a sort of divergent set of paths. What happens if the US and China collaborate versus if they don't? If they do collaborate, what's the difference between perhaps agreeing a small scale pause for a while while people look at these kind of implications of recent activities like the hooking face attack? and allow the time for things to be studied versus something like completely stopping things. So he proposes different plans and the one that is recommended by him and his team is called "Plan A". But there's a "Plan S" in there and the "S", I presume, is for a stop because that plan is just... take where we are today and do no more further research. And he contends, I think very correctly, that if we were to do that, the amount of technological advantage that we have in just the current state of AI, generative AI bots or services like ChatGPT and Claude. that's already enough to radically overhaul the economy and society. And we're living through that. That's happening now. And so we could just do nothing else and live with the consequences of just that technological change for the next decade, and it would still cause huge amounts of growth and huge amounts of change. So that's the Plan S, stop everything. The Plan A is to negotiate a pause. And so that's... This involves a degree of collaboration between the various companies in the US, but also between other states like China, and to agree a global limit on how much development is made in a particular year, how much compute, how many AI centres or chips are produced in a particular year. And this is the one that he sees as the safest path rather than the continued runaway kind of try and centralise everything as quickly as possible path that we seem to be on at the moment. I won't go into it in more detail because it's useful, I think, for listeners maybe to follow the link and have a look through it themselves. But much like last time, I think it's a really good contribution in making us think about that bigger picture. We often get perhaps a little caught up in the temporary piece of news that might come out or the analysis and reanalysis of the Hugging Face attack, for example, and miss that wider global context of, well, what does that say as to the trajectory we're on? Where does this leave us in 2027 or 2029 or 2040? And so it's useful now and again to step away and look at that. And so I'd encourage people to have a read.

[Ciarán]: 23:23

So what I think one person who doesn't want to stop is a new resident or at least a new property owner in "The Deise" - in Waterford - here in Ireland, in the south of Ireland, is Mark Zuckerberg. Mark Zuckerberg recently bought a castle. He's also been writing himself. He wrote an open. letter, six and a half thousand words, entitled "The Future is for Everyone" - quite techno-optimist in its outlook. To quote a little bit from it here, "It's surprising that the discourse from many developing AI is so filled with doom."I do not understand why anyone believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future."The notion that AI is dangerous is so dangerous that the only path is an extreme concentration of power seems inherently problematic". Quite interesting that he talks about the elimination of most jobs when he recently, or at least Meta recently, had to scale back its proposed Project OT, which was a kind of ambitious internal restructuring plan to replace human staff with AI agents. What do you think is the kind of design behind an open letter like this? Is he setting out a stall to be a future kind of thought leader in this space? Why is he writing this letter now?

[Dónal]: 24:53

It's difficult to see. I agree with you. I think the framing of it is an odd one because it's potentially the case maybe that he is trying to occupy a similar space to Dario Amodei. So the kind of leader of Anthropic who has for a long time positioned Anthropic as... a more benevolent kind of company in terms of its approach to AI. So its idea of constitutional AI and its ideas, which we've covered in previous episodes around trying to really put alignment at the centre of what they're doing, is often accompanied by these wider essays that he'll write about the direction that humanity might take with AI. And I think Zuckerberg is sort of doing the same thing here. This was a long read. And I mean, I must say I enjoyed the Pope's long and technical more than I enjoyed this particular one, because I found it almost insultingly naive in places. So it definitely has a hype value in terms of. the kind of grand potential that AI might have. And it has what I perceive certainly to be quite a naive expectation for a really benevolent outcome in terms of what the AI might do and this idea of it being for everyone and helping everyone. Because we can already see that that isn't the case. And this is something I think that strongly reminds me as a person who did a PhD in media in the mid 2010s. I was looking at the kind of early period of... the late 2000s, early 2010s... when social media was being heralded in a very similar way. It's for everyone. It's changing everything positively. The idea of the Arab Spring kind of marking this new era of democracy where social media technologies were at the heart of. uniformly positive change. And there was a window there when that technology was new, where people really did seem to think that. And we have seen since that that has really not been the case. Power has been concentrated more than it was. The ills of social media greatly outweigh the positives in lots and lots of areas, and especially for perhaps younger people and for things like mental health. And so now we, again, as our constant motto on this podcast might be, we're not learning our lessons from that previous round of technology. Now we're taking this kind of view, or certainly Zuckerberg is taking this kind of view again. And he's saying, for example, "Everyone will have powerful tools to create new businesses"."Everyone will have a personalised tutor to coach them with a PhD in every subject"."Everyone will benefit from scientific advances and be able to contribute to scientific progress". And we've heard these kind of things before. So when the internet was new, the idea that everyone would get online in the 1990s and democracy would change because everyone could participate was there at that time. The same thing when social media came around. Twitter was seen as this thing that would bring everyone together. Discourse would happen that didn't before. And society would flourish and scientific progress would be made. And I think that same sort of idea is here. And it is potentially true in some sense that perhaps everyone could contribute to scientific process. But I don't think they will. And I don't think that's ultimately how the systems will be used. I don't think everyone will have a useful personalised tutor because we've seen already that the attempts to integrate this into education are pretty poor. supervision I might give someone as a PhD as they require tutorship is often critical. I often need to say negative things to them or question their thinking. But AI doesn't do that well. It reinforces our thinking. I need to make sure that the information that I give to someone who I'm working with is true. And again, AI doesn't do that well either. It often fabricates things so that they sound correct, but really are meaningless. And so he's not really engaging with those very difficult questions about what this means in practice. He's taking a very, I don't know, sunny view of this. And I can see why there's a commercial interest for him in trying to make people have a much nicer view of this at a time when, as you've already called out, this has become a political issue. people are now really starting to mobilise for the first time against AI, against the data centres that support AI. And so it's not terribly surprising, perhaps, that leaders of these companies might want us to frame our thinking differently, even if it seems hard to do so in the kind of terms that Zuckerberg is giving us.

[Ciarán]: 29:05

And I must say as well, there's an element of what one hand, what the left hand of Zuckerberg is doing is totally separate to the right where we have those proposed job cuts. There was mandated keystroke and mouse tracking software that was designed to monitor meta employee workflows that was going to train AI all the while, while Zuckerberg is out kind of waxing lyrical about the kind of. techno-utopian promises of AI at the same time. Yeah, consider me sceptical around the real use of these beyond being something that is used as a kind of, I think, marketing pitch. We saw employee backlash with that. You mentioned data centres. It kind of takes us on to the next topic that I wanted to poke around quite nicely is this seemingly growing consensus of... a backlash of a pushback against data centres of the kind of physical manifestation of the growing computing power that AI requires of technology at large. And this is something that we're seeing even more impressingly in Ireland since the last time we spoke, we've had a heat wave over the summer and that has brought the issue front and centre for many people here, hasn't it?

[Dónal]: 30:22

Of course. And so we've had both constraints on our water, and then also we've had these increases in our energy costs. And at that same time, we're seeing this really quite unprecedented level of building of data centres in Ireland particularly. So we're well off the European norm for the amount of data centres you might expect to find in terms of their percentage energy usage. So we can see that there's data centres in lots of countries, but they might be in the... 3, 4, 5% of using the grid's energy. In Ireland, I think we've passed 23% now. So this is climbing constantly. And we're seeing projections of this hitting 30%. And at the same time, the government is very, very clear that their priority is to make Ireland an AI leader. And I have a huge question about the difference between being an AI leader and being just the place that's willing to take the data centres, because I'm not sure it necessarily gives us any leverage in the direction that AI is taken, any... ability to influence those alignment issues or other things like that. Our acceptance of these data centres within the country has a very profound effect on energy costs and on water usage, but I'm not sure it's really coming with some complementary utility to the Irish citizen. In fact, it seems to be the opposite, really. And I think that's an issue that surprisingly seems to be motivating people right across the political spectrum. We don't often see unity left and right, but we're seeing it here for anti-data centre protests.

[Ciarán]: 31:50

Exactly. And the timing of a lot of these developments. is quite notable too, where we've just gone through the summer where we saw the journal.ie reporting that data centre water usage doubled during our long hot summer. What is it, over 514 million litres of water used between May and July this year by data centres at the same time when there was a hose pipe ban enforced for water users in Ireland. and we're facing into an autumn and a winter when energy demands will spike again as heating requirements increase yet at the same time as you mentioned the electricity usage of these data centres keeps on rising too and this is not just in Ireland mentioned the midterms in the us as well where we see public pressure mounting whether or not this will lead to a kind of curtailment of what seems like endless data centre expansion, be it here or in the US or further afield, yet to be determined. I do know that other countries or at least municipalities have curtailed or at least put a pause on the development of these centres. But what's notable about Ireland is that we do seem to be the outlier. At the same time as Minister Peter Burke saying things like there will be significant new data centre investment in the pipeline, we also have reports from institutions like the UN who cite Ireland as a cautionary example of what they call local grid stress from concentrated digital infrastructure pointing squarely at the rapid expansion of data centres so this is something that is not going away it seems to be increasing as a topic and quite interesting yes as you mentioned a kind of unifying issue between perhaps people or groups of people who may have been traditionally opposed on different sides of of of numerous issues coming together under a combined scenario what you might call an anti-tech or at least an anti-data centre kind of banner.

[Dónal]: 33:58

Yeah, and I think it's the connection of the data centres, you know, in the larger context to potential job losses, to these kind of issues of societal damage that we've covered many times before. But then at a more local level, if it's one that's going to be built in your area, it is those issues around the water usage and the electricity usage. But even more problematically for people living very close to them, huge amounts of noise pollution as well. And so there's been quite good coverage of this on Virgin Media television and on the Journal recently. There's a lot of communities who were kind of sold the idea and planning of this largely neutral, you know, large, big box that would be built beside them. And the reality of the running of that data centre when it's up and running is quite different because many of them have quite large gas turbines to supplement their electricity supply. A lot of this runs through the night. And this is especially relevant in the recent hot weather. A lot of their cooling systems are incredibly noisy. So there are lots of people who have. a sudden intrusion not only on their electricity bill and on their limitation on their water usage, but also just their quality of life in terms of having to live besides something that might be producing emissions if there's burning of gas. And it's certainly producing a huge amount of noise, whether the gas turbines are on or off. So it's not terribly surprising that people in lots of different local areas might come out against these. But I think part of the bigger picture here is going to be whether there's a more consolidated effort to make this a major issue in upcoming elections, which I suspect there might be. I think it's really in the next set of elections going to become a really big positioning issue for parties. current parties of government in Ireland have obviously set out their stall. They're very pro the continued expansion of data centres, but I would expect parties of the opposition now to be much more vocal on the other side of that. And I'd expect to see that in lots of other countries too.

[Ciarán]: 35:50

And this is something as well that Daniel Kokotajlo wrote about in the AI 2040 prediction too, where he talks about, is it right, AI becoming central as an election issue. I think he's pointing towards the US presidential election of 2028 as well. Is that right?

[Dónal]: 36:05

Absolutely, yeah. And I think, you know, it's virtually certain now that it will be because by then we're far enough into the future that we can really imagine there being much more substantial kind of changes within jobs and things like that. And of course, in the US where the planning law kind of changes from county to county and municipality to municipality, there's a lot of places where... problematic data centres to say the least have been built. Some of these are effectively tents that have their own mobile power stations on site. So there's quite a lot of immense intrusion into some parts of the US by the building of some of these. So I think if that's continuing in the direction that it seems to be at the moment, it would be shocking if it's not one of the really major issues in the next US presidential election, as well as in our own and in most European countries too.

[Ciarán]: 36:50

These are the real world physical manifestations of the rapid expansion of this technology. I mean, when we started this podcast, perhaps it was still a bit more abstract, still a bit more. It's there online, but we see with. things like proposed job cuts in organisations like Meta, we see with data centres, we even see with the bulk purchase of secondhand books by what was determined to be Amazon in the end. This is a story that was circulating of... second-hand book stores receiving orders for thousands of titles and as one book setter commented in Irish media not uncommon to receive these kinds of requests perhaps from you know a university setting up a new library or these kinds of places but they said that they commented what was unique or what was remarkable about it was just the the seeming random nature of the order of books. You're talking about things like computer manuals that were maybe perhaps far past their sell-by date also being included in these orders and reporting from I think it was 404 Media, a really good digital news website. ultimately tracking one of these orders to an Amazon warehouse. And this again is a kind of another physical manifestation that perhaps no one might cry over, you know, secondhand books being kind of purchased at bulk, but what they will kind of shed a tear on is the idea of the wastefulness of this. These books were being bought at bulk, being used to train models and then being destroyed. And it's that kind of wanton, I think, destruction for the sake of this technology that might irk people as well to just these other kind of physical manifestations that seem to upset people.

[Dónal]: 38:31

Very much. I think it's very hard to assert that you're the good guy when you're the one destroying books. That's not usually something we historically associate with good actors in history. But yeah, the reason for the destruction, we should say, is really that it's down to the convenience of how these books are being scanned. So what they do is they chop the spine off the book so that they can access the pages as a pile of pages and quickly feed them into a machine that scans them. And there are ways to scan historical and rare books that don't damage them. In fact, most large libraries that are digitising their collections have machinery specifically to do this. But because it's not damaging the books, it's running more slowly. So again, here, in order to get the books in as quickly as possible, they're being destroyed in the process. And the reason for this collection is something we've talked about in several prior episodes. It's this constant need for more human writing as data input. And so as this goes, as we build these models further and further, we mentioned this before, the challenge is finding now a good source of clearly written by humans text that can be fed in because the Internet itself is now so polluted with the output of previous LLMs as people make blog posts and emails and god knows what onto the Internet with the help of. you know, the several generations back AI that they might have used, the text that you can get from crawling the internet is now much more poisoned by AI output. So finding large collections of old books, as obscure as they may be in those manuals that you mentioned, is a useful input. But the destruction of them is, yeah, it's something that really tarnishes this whole thing quite a bit for people. And I think was a sensational story in the same way as some of the other things around data centres that we've seen where... And just as you say, I think it's a manifestation of the rather unpleasant reality of what's required to keep this development going.

[Ciarán]: 40:18

And just as we come near the end of this episode, I just want to cover, I think, the most pressing issue of all, Dónal, and that is the rapid saturation of AI posters, both online and increasingly offline on Windows and halls and bulletin boards here, there and everywhere. What can be done to stop this terrible social ill? that is happening before our eyes.

[Dónal]: 40:46

It's funny because I think we're unfortunately now so acclimatised to rubbish AI text being posted, perhaps on LinkedIn or on social media channels or whatever that we're inured to seeing it in some places. We're constantly seeing the output of AI wherever we look in social media channels, but we're sort of... We've adjusted to that. But there's something weirdly jarring about seeing so obviously social or obviously AI generated physical flyers or posters everywhere. And yeah, just as you say, I mean, I'm from a small town in Ireland. And when I go back there to visit, I'm always shocked by the fact that in the local supermarket, what might have once been a handwritten advertisement for someone who's looking to, you know, work as a cleaner or organise the local GAA lottery. They're all these now very generic AI posters that are. hugely crowded with pretty rubbish imagery, far too much information simultaneously in place. Obviously, I lecture in a school of communications and design. And so I'll be using some of these as a what not to do guide for my new first year students as they come in. But it is something about the, again, continued pervasive leak of the effect of AI, the enshittification, to use Corey Doctorow's words that we've talked about before, into more parts of our lives. And I think that sort of the physicalisation of some of the stuff that used to happen in the space of the internet or in the cloud, the more that that happens to us, I think the more that the... more grim realities of this set in. And I absolutely agree with you. I'm very sick of seeing those posters. They're a new and unwanted part of this for me.

[Ciarán]: 42:20

I think so. I think it's the, as you say, the crammed nature, so much text placed in similar places on the poster. And really, it's just the sense that that sensory overload of looking at a poster that might be to, you know, advertise the local lotto in your local GAA club. But it feels like it's something that came out of the circus. It's just exploding with colour and visuals and and also how samey. it all feels where it just feels like that post or whatever it might've been the original handwritten one before. It just feels like a little bit of originality has been, has been robbed. And then just seeing different services, different businesses use it. Um, it just feels, yeah, a little bit saturated, a little bit, a lot saturated, I think, to be honest.

[Dónal]: 43:04

It's the effect, I think we talked about this in prior episodes also, of the ubiquity of technology. When something becomes so everywhere, it sort of changes behaviours. That's when it really has the social effect. And in this case, yeah, what's happened there is that most people are easily able to access the free version of ChatGPT. It has a built-in image generator. And most people are using very simple prompts to say, I need a poster that's a flyer for my local GAA lottery draw. And so a machine that's not particularly used to or reinforced to give nuance or design is just churning out something that's very similar to the many other things that has churned out. And so it shouldn't surprise us. We're going to see more and more of this kind of. weird homogenisation because of the way in which we use stuff. It's going to start changing people's language. I think we're going to definitely start seeing people write like the AI writing they've seen. I think that's going to be a big problem for us in university, for example, to distinguish between what's something that someone generated out of an AI model versus what's something that they wrote that just happens to sound like that because they're so used to reading texts that sounds like that all the time. And so I think there's a lot of these kind of secondary social effects in how we communicate with one another that I expect to see change and really not for the better in most cases.

[Ciarán]: 44:19

Yeah, something we already see in speeches by politicians in different national parliaments, national legislators. I'm thinking of a UK example from a couple of months ago that I read about of specific snippets of language that hadn't really been common, let's say, before the introduction of AI technologies. And now there's a deluge as lots of MPs or TDs here perhaps decide to delve into different issues. But I think that just about does it for today. I think overall our lesson or perhaps our recommendation for all these companies, all these figures we talked to today, is that more work is needed. It's time to go back to school, hit the books and concentrate for a semester ahead. If I can cry even more terrible remarks into that analogy.

[Dónal]: 45:10

No, I mean, I think that analogy is good. We're at that time of year, the schools and colleges are back. But I think we also, ourselves, even if you're in neither school nor college, you need to go back to school a little bit too. I think everybody, we've said this constantly, everybody needs to be paying much more attention to what's happening here. Whether that's something for you that is, you know, looking a little bit more into the realities of data centres in your area, or it's looking into the way in which AI might be used in your child's education. So perhaps a good time for parents, indeed, to really look at whether, not just if their child is using AI to circumvent doing homework themselves or doing learning themselves, but whether there's that same rubbishy AI content is now being given to them as a learning resource. So if, you know, some of the stuff that they're taking home looks like it's poorly produced, too. we need to really be much more aware of this stuff. So our constant mantra in terms of AI literacy is this idea of being a little more aware constantly, putting a little bit of effort into looking at where this change is occurring and asking some critical questions about whether it's a good or a bad thing. So we'll leave everybody to go back to school on that issue, perhaps.

[Ciarán]: 46:14

I think that just about does it. Dónal, thanks very much.

[Dónal]: 46:18

Thank you very much for listening. And again, if you have any ideas for things that we should cover or you have any comments on our show today, you can definitely give us an email. We would love to hear from you. hello@enoughaboutai.com