A very timely and fascinating conversation on Data Centres and their impact on our energy emissions, the tactics used by the industry, and the generative AI industry more broadly. Always a delight to talk with Ketan Joshi!
So many great quotes from this conversation, I could have highlighted everything!
“it would be more new demand than electric vehicles or the electrification of homes and businesses ⁓ each individually. and so that’s big enough that we consider it to be quite significant. and it has, you know, certain impacts on the power grid, namely it makes coal and gas use increase if you don’t match it with renewable energy growth.”
“using it as a search engine, using it to like extract basic facts about things, all of this stuff is where it fails the worst.”
“we are in just such an amazing time of showing that it actually works when you do this. Like when you have this like multi decade focused project of eliminating a dangerous product from society. It actually works. and so, you know, I won’t even go into it, but like there are just so many examples. Australia has many of them. you can actually reduce the consumption of fossil fuels. It’s just that you have to do it both with a sense of urgency but also like over a very long time period because there is an upper limit at the rate at which you can limit eliminate them, right?”
“we need to go back to you know, just valuing the sheer social benefit of being correct about things and checking things and like trying not to lie, like that is as a value, as like an ideology.”
“Like there is still some sense of shame about being wrong about things and I wanna capture that little spark and turn it into a fire.”
I think the only way that we protect information quality and our shared sense of reality and like, you know, facts and like everything holding us together is by introducing really really serious consequences for fabrication”
“The argument they always use is like it’s commercial in confidence, right? So it’s like a secret. It’s like a business secret. And it’s like, what’s the business secret here? Like there’s no there is nothing you can glean from that information other than they consume a very large amount of power and cause a lot of greenhouse gas emissions. Like you cannot reverse engineer the business strategy.”
Transcript
Linda McIver (00:00)
Welcome to a special issue of Make Me Data Literate. I’m skipping the usual questions this time and having a repeat guest, which is only the second time I’ve done that I think. ⁓ but I’m super excited to delve into the actual facts of data centers and energy use with the amazing Kitan Joshi. Welcome, Ketan
Ketan (00:21)
Yeah, thanks so much for having me back. It’s really nice to be back with you.
Linda McIver (00:25)
It’s awesome. I I’m excited for this conversation. Sort of half excited, half ragey, but I dare say we will get to that. ⁓ so So you’ve just written ⁓ a report with Greenpeace on the energy use of data centers, particularly in Australia. ⁓ can you give me the kind of potted summary, like what what did you find?
Ketan (00:32)
Me too.
Yeah, absolutely. It’s ⁓ kind of split it into three nice and simple chunks. the first is just that Australia has a remarkable amount of planned new data center capacity. this is something that has been growing in Australia for a decade, but growing at a very gradual rate. And this would be a shift from that gradual rate to a much faster rate. one of the fastest modes of demand growth in Australia.
it would be more new demand than electric vehicles or the electrification of homes and businesses ⁓ each individually. ⁓ and so that’s big enough that we consider it to be quite significant. ⁓ and it has, you know, certain impacts on the power grid, namely ⁓ it makes ⁓ coal and gas use increase if you don’t match it with renewable energy growth. ⁓ and so that was quite significant. I think ⁓ we kind of just looked at some of the work that’s already been done on that.
The other thing, the other thing that we found is that the industry is basically promising to clean itself up with renewable energy. ⁓ and we looked into some of those claims and said, okay, you know, let’s actually ask, you know, the question of like, ⁓ are they actually doing literally what they say they’re doing? ⁓ and of course, predominantly we find that they’re not. ⁓ and then finally, we looked at the whole prospect of
gas actually being used literally to directly power data centers, which is a phenomenon emerging in areas like Europe and Canada and the UK and mostly in the US. ⁓ and yeah, we found evidence of early relationships forming between the gas lobby ⁓ and the tech lobby in Australia. ⁓ and you know, we look at a few specific examples of planned projects that are running on ⁓ that are proposed to run entirely on gas. ⁓ and also
you know, how they’re kind of greenwashing greenwashing their claims and saying that gas is somehow the better option. ⁓ so yeah, we put that all together. We combine that with the lack of transparency from the industry around how much energy they’re using and the purposes actually being served. And we put it all together into this picture of a lack of governance and a lack of ⁓ public information about the socialized impacts, whereas the benefits are all privatized to the companies.
And so that’s why we call for a moratorium in the in the report, which we consider to be quite a rational ⁓ return to evidence-based decision making. ⁓ I think that’s the one like running theme through the report is basically like there’s lots of claims and not enough evidence. So we we really say that there needs to be a a return to evidence-based decision making. ⁓ and you can’t really do that. And and then also ⁓ community involvement in in governance and decision making.
And you can’t really have those two factors unless you ⁓ don’t have the urgency and the impatience and the fast tracking ⁓ that we’re seeing with the industry at the moment. So it’s a moratorium ⁓ in Australia would be the same as the type of moratoriums that we’ve seen in Ireland, in the Netherlands, ⁓ as proposals in parts of Asia and South America, ⁓ and in many, many different regions in the US as well, regardless of whether they have Democratic or Republican governors.
⁓ these are common. ⁓ they’re very sort of like middle of the road positions, you know, basically saying we’ll pause development until we can actually put this back in the box of the normal democratic decision-making processes that we have for any other large infrastructure, particularly in the midst of a climate crisis and an energy transition ⁓ that is not being fueled by infinite numbers of wind turbines and solar panels like in Australia.
there is actually insufficient renewable energy growth. So that’s the heart of the report essentially, is we we want to return to ⁓ things making sense and a moratorium is a good way to do that.
Linda McIver (04:46)
I’m I’m fascinated to have you say return to things making sense. It feels like it’s been a long time since things have made sense. But let’s just be clear here. You’re asking for the dangerously radical proposition of pausing until we know exactly what’s going on. That seems that seems outrageous. I mean Wild.
Ketan (04:55)
Yes.
Mm.
I think.
Yeah. Yeah. And
and and it’s well, I mean, like, you know, we released the report a few weeks ago, ⁓ and it’s been really interesting. You know, I think we had a really good reception from particularly the stuff on gas in our report. people were really, you know, shocked and a little bit worried about the prospect of the gas lobby kind of having this like plan B from from data centers. But on the moratorium call specifically, it’s been quite interesting. We’ve seen some sort of like
I guess momentum from the climate movement in Australia and the energy sector pushing a bit more towards the side of not calling for any restrictions on data center development, but more just saying mandate ⁓ requirement of having one hundred percent renewable energy, new renewable energy procurement for data centers, along with, you know, some more sort of like requirements around transparency and things like that. Of course, you know, I strongly support that. I I think that’s a I think that’s a a totally fine position to have.
but I think that if you want to go all the way and and and prioritize climate rather than prioritizing, you know, the sort of progress of data centers, ⁓ I guess everybody lands somewhere different on on where they value different activities in society. And I and I personally value ⁓ the protection of the climate over the construction of large buildings full of graphics processing units. ⁓ but everyone but everybody has a different have a d has a different position.
Linda McIver (06:32)
I’m fascinated by the urgency angle because it occurred to me today that it is it is exactly the same as a lot of internet scams. You know, you’ve got it you’ve got to get on it now. It’s really urgent. Don’t stop, don’t talk to anyone and don’t check anything. Just ⁓ like I’m amazed they’re not asking us to pay in iTunes gift cards, honestly. It’s so it’s classic. It’s the you know, the the this artificial urgency seems designed to short circuit
Ketan (06:42)
Is it?
Ha ha ha.
Yeah.
Linda McIver (07:02)
⁓ rational evidence based decision making.
Ketan (07:04)
Yeah.
That is exactly it. And ⁓ there’s a really nice quote from Google actually. ⁓ Google recent recently came and visited Australia and there’s a great article ⁓ where they lay out exactly what you just described, which is ⁓ basically and and the and it’s and it and the urgency message is always framed as a competition, right? So ⁓ basically
Linda McIver (07:24)
Yeah.
Ketan (07:25)
you have to do this before Asia does it, like before China does it, or in like even Intra Australia, there is like Sydney, you have to do it before Melbourne does it, or Melbourne, you have to do it before Sydney does it. and it’s and it’s just like it is it is I don’t talk about it a lot because it’s just it would just be me like just going mad, you know, on top of my keyboard. But like ⁓ you know, I can I never, ever, ever get a clear answer as to why it is a race in the first place.
Linda McIver (07:34)
Yep. Yeah.
Yeah.
Ketan (07:54)
Like
why like it’s just so baffling as to why you need to urgently create a data center to power the function of generative AI. Like beyond basically if you don’t build it here, then it will get built somewhere else. And it’s just like, well, okay, yeah, but then who cares? Like I mean, like that’s just, you know there’s no jobs, ⁓ all of the financial benefits are paid, you know, predominantly to a company that’s
you know, either in a different country or in a different part of Australia. ⁓ you just bear a huge amount of local burden, like angry communities, construction emissions, noise. ⁓ and, you know, if anything, I would imagine that the race would be who can have the fewest, like like who can have the who can have the who get like who get basically dumped with the toxic presence of the infrastructure. ⁓ whereas the real race, of course, is that
Linda McIver (08:39)
Yeah.
Ketan (08:52)
being a company that either sells the space of a data center or operates its functions, that is the race, right? ⁓ because they are they basically understand that their time is limited, right? Like they understand that the bubble will pop one day, the hype will fade. So they actually are racing to build data centers as fast as possible because their priority is to basically cash in as much as possible before the music stops. ⁓ and
Linda McIver (09:06)
Yeah.
Grift and get out.
Ketan (09:22)
Yeah, exactly. Exactly, right? ⁓ this is how a lot of individuals and companies operated with Bitcoin. ⁓ they basically saw like they’re just basic they’re just kind of like going as hard as they can until the cliff happens. and the idea is that you jump out, you know, you you kind of find a bigger fool. Yeah, like, you know, you you’re a fool for having gotten into the system, but you need to find a bigger fool to get out of the system.
Linda McIver (09:29)
Yeah.
Cash out
Ketan (09:48)
⁓ and I’ve always just found the economy of Bitcoin mining to be a far more parsimonious explanation of generative AI ⁓ dynamics than than anything else, right? Like ⁓ it it it just ⁓ you like you understand why they’re frantically building multi-gigawatt data centers. and for those who who aren’t familiar, ⁓ like a you know, a ten megawatt data center
is pretty is pretty unremarkable. Like that’s kind of the stuff that we had like in the sort of twenty tens to power like cloud storage and things like that. Then you had like hundred hundreds of megawatts, which is kind of like the size of like a big suburb or a small city. ⁓ and those were f for things like streaming and gaming and like more high intensity applications. ⁓ a gigawatt is a thousand megawatts, so the so going from a hundreds to thousands. ⁓ these are the size of generative AI data centers. So that’s basically the
Processing required to train models and to return answers from systems that are like chatbots or coding agents and things like that. ⁓ so and that’s city-sized, right? So that’s like the same power demand as like a city. ⁓ and so ⁓ yeah, like this is a really this is a really big difference. ⁓ and these are ⁓ when you pair urgency with that scale, that’s when the real harm starts to occur.
Linda McIver (10:57)
Yeah.
Yeah. Yeah. B because there’s no stopping and considering the issues, ⁓ no thinking about what what’s in it for us. And that was an i that was an issue that a a former student of mine raised with me just yesterday. his name is Hugh Paynter and he’s doing a PhD in physics with a focus on renewable energy. So, you know, he knows something about this this field and he said to me, It’s also unclear precisely what the seemingly rapid expansion in capacity is for
Ketan (11:16)
Yep, exactly.
Linda McIver (11:40)
And why it’s of any benefit to me or society at large. And that’s you know, that seems to nail it to me. Like we are we are contemplating rapidly accelerating the climate catastrophe for the benefit of ⁓ who? Musk? Altman?
These us probably I can what
Ketan (12:01)
Yeah, I think it
a nice ⁓ a nice comparison point. I don’t want to downplay the the serious issues with data center development, which I think are quite material. But in terms of that specific point of not really getting what it’s for, ⁓ I think a lot about ⁓ transmission line development and also five G. ⁓ if you remember when five G towers were sort of introduced and there was a huge
It was kind of like around COVID as well. So people were already a bit on edge. and conspiracy theories were already, you know, really like starting to ramp up. ⁓ but part of the one of the factors behind why it was such an intense thing is just that no one understood what five G was for. Like mobile phone towers, when they were first introduced, there was opposition. ⁓ but you could basically say like you could buy a mobile phone and understand the benefits of it directly, right? Like it made sense.
Linda McIver (12:35)
Mm.
Ketan (12:56)
⁓ but a five going from a four G tower to a five G tower, you’re like, I I don’t care. Like I I’m already getting like ridiculously fast, you know, up and down speeds on my 4G. ⁓ and so ⁓ yeah, when there’s no clear utility, that’s always a major factor ⁓ in pu in public responses. And and actually maybe another good analogy is here in Norway, there was a big push for wind power in the in the early 2020s.
Linda McIver (13:05)
Mm.
Ketan (13:22)
that got met with huge resistance and one of the factors there was just like we’re already hydro, we’re already a hundred percent hydro. So why do you need to build a lot of new wind power? The answer of course is that our power demand is growing because we’re electrifying everything and we can’t expand hydro enough to meet that new demand. ⁓ but that wasn’t laid out very well. Like the argument was not laid out well. So we struggled. Yeah.
Linda McIver (13:28)
Hmm. Hmm.
Yep.
This feels a bit more though like people not really understanding what Bitcoin is for, like because because generative AI is less useful than many of those things, you know, like
Ketan (13:52)
Yeah. Which is correct. Yes,
Bitcoin is not useful. Yeah. ⁓ and I think also just if the benefits accrue to ⁓ a tiny slice, like ⁓ people really ⁓ I think generative AI is very new. I think people can just remember, you know, writing an email without it. Like it’s not the benefits are just not massive enough to to feel instinctively like
it’s justified the the infrastructure costs. ⁓ Bitcoin is ⁓ Bitcoin is, you know, people just don’t know what Bitcoin is, right? Like it’s just such a it’s such an obscure, weird thing that has no social utility whatsoever. ⁓ and I think that that, you know, that had an impact on the development of crypto data centers. ⁓ and there’s quite a few of them in Norway actually because they it’s cold
⁓ it’s cold and and and you know ⁓ power was cheap here in Norway. ⁓ so yeah.
Linda McIver (14:58)
⁓ there was an an interesting comment. ⁓ Greg Jericho quoted it in The Guardian today that the Climate Council estimates that data centers will go from accounting for two percent of our national electricity use to six percent by twenty thirty, which is much closer than it used to be, twenty thirty. ⁓ and twelve percent by twenty fifty. That that almost feels like an underestimate. Do you have a a feel for
Ketan (15:12)
Mm.
Mm-hmm.
Right.
It is a bit of an underestimate.
so we use the same figure in our Greenpeace ⁓ Greenpeace report, and it’s the Australian Energy Market Operator, right?
and I think this is a really useful exercise and I wanna write about this soon, which is basically no one really knows what the demand from data centers will be in the future, because there is a very, very large pipeline of projects and you kind of have to pick yourself using a subjective set of criteria, which ones you think are most likely to go ahead or not go ahead. So the Climate Council was actually a little bit more you know, they kind of looked at the whole pipeline. ⁓
Linda McIver (15:58)
Right.
Ketan (16:04)
which is, you know, unadjusted, right? Like basically like let’s just assume that everything that’s that’s, you know, has like half of an idea, ⁓ actually goes ahead. Whereas we really trimmed it down to what the grid operator has subjectively assessed as being the most likely to go ahead, which is about ten percent of the projects, right? So there’s like a pipeline of like forty to fifty gigawatts of new data center demand, ⁓ which is like
sort of about ⁓ one point five times the total size of the current power grid to give you a sense of scale. ⁓ so like it won’t all go ahead of course, right? Like that’s ⁓ extremely unlikely that it all goes ahead. But ⁓ what I think ⁓ is that we intentionally used a conservative estimate of projected power demand from from data centers.
Linda McIver (16:39)
Whoa.
Ketan (16:58)
because we didn’t want to be subject to the criticism of like, you’re just sort of overestimating what’s coming down the pipeline. But what Greg quoted there from the Climate Council and what we it’s it’s the same figure that we cited in our report when we put it together has already been made like outdated because the Australian energy market operator yeah they published a they published a update to it. ⁓ it’s like a preliminary update so it hasn’t been finalized. The new data will come out in August.
Linda McIver (17:05)
No.
boy.
Ketan (17:26)
⁓ but they ⁓ basically said it we got a notched upwards, you know, the projected ⁓ consumption of data centers notched upwards by by ⁓ I think like sort of five percent if I remember correctly. And every single year when this number gets updated, it’s just it’s just ratcheting upwards and upwards and upwards, right? So what that tells me is that that will probably continue.
Linda McIver (17:26)
Bye.
Ketan (17:51)
And that what we see now as like a conservative estimate of data center demand will be greater in the future. ⁓ just following the pattern of constant underestimation. And so we had the first full year of actual data center ⁓ consumption, like power demand data, ⁓ released for FY25. And you can actually compare that to the first ever projection that was released two years before that, and it was higher than that projection. And I think that.
Linda McIver (18:17)
Right.
Ketan (18:18)
that pattern will continue, right? Where the projections underestimate the reality and you end up with ⁓ what I’ve people have called the hedgehog chart, right? Which is like when you kind of your projections are always a bit wrong and they refuse to adjust to what’s happening, you know, and you’re constantly like, ⁓ I’m gonna do a very sober assessment of, you know, whatever, and it always turns out wrong. And of course, you know, that is because of the reality of the political dynamics here of what you mentioned earlier.
Linda McIver (18:30)
Mm.
Mm mm.
Ketan (18:44)
Around urgency and like basically like fast tracking and like a a a fundamental ⁓ force in the system pushing data centers into becoming operational and consuming electrical energy. ⁓ that’ll just keep happening. So that percentage ⁓ has a huge range. ⁓ and so it kind of just having this conversation makes me want to kind of create like an illustration of the range, you know, of the different
Linda McIver (18:45)
Yeah.
Yeah.
Ha ha.
Ketan (19:11)
Right, because you know, this is important to know is that this is we’re talking about a cloud of potential futures here. ⁓ and you know, you’re gonna have this one figure and it’s like, okay, well that’s probably what it’s gonna be, right? Because it’s one figure. But that one figure is just the midpoint of this cloud of this like of this range of possible futures. It reminds me a bit of polling, like when you have like just before an election and you’ll have like, you know, just these huge error bars.
Linda McIver (19:36)
Yeah.
Ketan (19:41)
on like, you know, Donald Trump’s gonna be president or ⁓ Kamala Harris is gonna be president. ⁓ and when you look at them as like two overlapping clouds, you’re like, ⁓ well no one knows. whereas when you look at it there’s the average, you’re like, Kamala Harris is gonna win ’cause that’s what the the numbers higher Yeah.
Linda McIver (19:53)
Yeah.
Yeah.
So it’s it’s hard to predict. How hard is it to find reliable data on what’s happening now and data energy data center energy use right now?
Ketan (20:12)
Yeah, there’s three different levels at which you can look at this information. ⁓ so the first is basically the total aggregated consumption of data centers across the grid. ⁓ the Australian Energy Market Operator kind of discloses that sometimes. Like they obviously have it, right? Like they obviously have this number, but they put it in reports and you can find it and you can like reverse engineer it from the chart and have like a little data set f of, you know, monthly of monthly per state ⁓ total data center demand, which I did, you know.
Linda McIver (20:27)
Mm.
Yeah.
Mm mm.
Ketan (20:41)
⁓ ’cause I’m cheeky. But like ⁓ it would be nice if they just published published the data set.
Linda McIver (20:45)
Yeah.
Yeah, it’s I don’t know why they
don’t Weird.
Ketan (20:52)
I think it’s
because most people aren’t like me. Like most people don’t actually need, you know, like a like a per. Yeah, yeah. Yeah, exactly. Exactly. So the other level ⁓ is aggregated across ⁓ companies or or like organizations. ⁓ and that is disclosed annually in this data set called the National Greenhouse and Energy Reporting Scheme or NGERS
Linda McIver (20:57)
I do. I want the numbers.
Ketan (21:18)
⁓ and NGERS has every corporation in Australia just reports i if you’re above a certain level, you you and it’s like most companies because they’re pretty big. you report your total energy consumption in like joules which for data center companies is mostly electrical energy for data centers. ⁓ and then you also report your emissions. ⁓ and this is really, really important because what you’re reporting in this data set is your emissions from connecting from just plugging into the power grid. So
Linda McIver (21:18)
But
Ketan (21:46)
You plug into the power grid and you take an estimate of like, okay, it’s 1 p.m. on this day, and the power grid was compro comprised of like forty percent coal, ten percent gas, et cetera, et cetera. You take the average intensity and you multiply that by how much you consumed to get your emissions estimate. And that’s called that’s what’s known as location based emissions. And it’s like the it’s like kind of like the least bad estimate we have of attributing your impact on the power grid. You
plug something in and all the generators increase their output because you plug something in and increase demand, you will always have an interventionist impact on the power grid. Like every light switch you flick on, it impacts ⁓ how the power grid operates. And so a data center activating 200 megawatts, you know, of of new power demand, that does materially influence how generators operate. ⁓ and so due to that, we attribute the emissions from that increased coal output to the to the
whoever flicked on the switch, you know, and and that’s called scope two, ⁓ scope two emissions. Like basically, ⁓ it’s like the least least adjusted, most honest version of what you impact you had when you plugged into the power grid. ⁓ so that’s why I find it personally very, very important that the government in Australia mandates the disclosure of that. That is remarkable. I don’t know anywhere else in the world that actually mandates that disclosure. ⁓ that is really Yeah.
Linda McIver (23:04)
really? We’re actually ahead in something.
Ketan (23:07)
Yeah. yeah. my gosh. It’s ⁓ Chris Bowen does not take enough credit seriously. Like he really, you know ⁓ I don’t know, if I was him, I would be out there every day saying, like, you know, we we mandate like, you know, nowhere else in the world actually forces companies to disclose this type of information. Maybe he shouldn’t, because then the companies might actually start fighting against it. But what happened is Amazon, when Amazon came to Australia and when can and when Canberra data centers started expanding, both of them fought against the
Linda McIver (23:08)
I’m shocked.
Yeah.
Ketan (23:37)
the ⁓ clean energy regulator to ⁓ disclose this, right? Like they did not want to share this information. And you can see why when you look at their when you look at their emissions, right, and their energy consumption, you can see why they were like, no, no, no, we don’t want to share this with anybody. Because it has gone up so significantly. ⁓ AirTrunk is the has had the highest rise in in their emissions. But CDC and Amazon have also had Amazon Australia I should say have both had has significant rises as well.
⁓ so the new batch will come out ⁓ I think next February. So it’ll be really interesting to see what happens. I suspect that we’ll see the same pattern emerging for Australian ⁓ data center companies, data center operators and tech companies. the third one, this is ⁓ the most ⁓ this was like the strongest focus of a little section in our report where we ⁓ talked about transparency. ⁓ and so this was basically ⁓ this is this is where
individual data centers ⁓ disclose or do not disclose ⁓ how much energy they consumed. This is really relevant for communities who are worried about like diesel fumes, ⁓ are worried about the operation times of like cooling systems when they have like, you know, ⁓ curfews, like when you’re in a residential area and there’s noise. ⁓ and companies around the world are really fighting hard to not disclose the facility level
energy consumption and emissions data. and the reason behind this is because, you know, it just gives you know, they they’re worried this is my personal suspicion. I have no I can’t see into their minds, but I strongly suspect they’re worried that it gives ammunition to to ⁓ objections and complaints from residents. ⁓ and they just they just can’t be bothered. It’s it’s easier for them to just not deal with that and keep it secret. ⁓ and so
Linda McIver (25:16)
Yep.
Ketan (25:32)
There was a case recently here in Europe where ⁓ the
Tech lobby group Digital Europe, ⁓ they lobbied the EU commission to keep their facility level emissions data secret. ⁓ and their wording and their like recommendations, you know, was copy pasted essentially word for word into the actual into the actual rules here, right? So ⁓ a lot of data centers are yeah, yeah, really super, super gross. And and and you know, data centers you when you’re in the EU and you know you’ve got to report
Linda McIver (25:56)
wow, that’s gross.
Ketan (26:05)
⁓ all this stuff to the EU commission and they lobbied really hard to keep this one secret. ⁓ and it’s r a bit ridiculous, right? Because there are companies that choose to voluntarily disclose this. Like Meta discloses the individual energy consumption and even the s individual scope 2 location based emissions for all of its individual data centers ⁓ in the US. Meta, yeah. It’s it’s wild, right? ⁓
Linda McIver (26:26)
That is unexpected.
Ketan (26:33)
Yeah, you know what it is though? It’s because like there’s some great person in their sustainability team who just puts it in there and not a single other soul in the company sees it, right? Like no one’s noticed. So I try not to talk about it too much. Yeah, no one no one from Meta is listening ⁓ to anything about, you know, data center environ But like yeah, it’s it’s remarkable, right? Like ⁓ this really shows
Linda McIver (26:43)
Yeah. Right. Mm. Should we be talking about this?
That’s true.
Ha ha ha.
Ketan (27:03)
That ⁓ the the argument they always use is like it’s commercial in confidence, right? So it’s like a secret. ⁓ it’s like a business secret. And it’s like, what what’s the business secret here? Like there’s no there is no ⁓ nothing you can glean from that information other than they consume a very large amount of power and and and cause a lot of greenhouse gas emissions. Like you cannot reverse engineer the business strategy.
Linda McIver (27:26)
Yeah.
Ha ha.
Ketan (27:29)
⁓
of like Anthropic or Google or whatever from any from like the total energy consumption of this facility.
Linda McIver (27:32)
No.
can’t deduce the weights of the models. That’s that’s wild. But it’s you know, it’s the same, you know, when you talk about misinformation and the tech companies are like, ⁓ regulating us would be a very bad idea and I’m like, would it though? It feels like feels like it might be a solution, not a bad idea. But they don’t want it ’cause it yeah it’s not in their interest. ⁓
Ketan (28:02)
Yeah, exactly.
Linda McIver (28:04)
The the transparency thing is really wild. I don’t I I I’m I’m really surprised to learn that that the Australian government actually forces that that level of disclosure and I am gonna have to go and dig out that data set now. Like I’m I’m fascinated. I did not did not
Ketan (28:24)
Yeah. Please do. I
have a I have a so they disclose it by each financial year separately and I’ve got a section on my website where I’ve combined them all into one into one data set. So, you know.
Linda McIver (28:30)
Right.
you’re
the greatest. I am I am I’m here for that data set. That’s I mean man, I can see classroom activities around this. This’ll be great. I love it. ⁓ so let’s lay this out like ⁓ really clearly that what we’re talking about when we’re talking about data centers and their impact on the grid is we’re talking about more coal generation, more
Ketan (28:38)
Please.
Please do. Do it. It’s so much fun.
Linda McIver (29:03)
gas generation, potentially new gas plants on the site of the data centers because the grid just doesn’t have the capacity. That’s that that adds up to ⁓ a ton of emissions. That’s that’s the bottom line. And and it’s that that narrative that that forced
Ketan (29:20)
Yep. That’s that’s exactly it. Yeah. Yeah, correct.
Linda McIver (29:31)
⁓ urgency the you know, we don’t want to be left behind, we wanna win the war, we wanna win win the race, all of that all of that rhetoric. And it it it ties in with the narrative of inevitability that AI is here and you can’t put the genie back in the bottle and you know, and it’s it’s only gonna get better. So the mistakes that it makes right now are not problematic at all, although ⁓
Ketan (29:40)
Yeah.
Linda McIver (29:57)
I’ve ranted at length about how problematic they are. ⁓ but I I found out because you posted it that ⁓ someone I respect in in or had respected in the climate space, Tim Dunlop, had had published something that had quotes in it that were not ⁓ not real and fabricated that people hadn’t said because he’d used AI in the construction of that particular
Ketan (30:19)
Mm-hmm.
Linda McIver (30:25)
report and the the whole that whole thing around but you have to use it and you’ll be left behind if you don’t and it’s fine as long as you check it carefully. Well we see time and time again that the careful checking doesn’t happen. It’s not a thing. People aren’t good at that.
And yet it’s I don’t think it is inevitable. Like I think we could put the genie back in the bottle. I think we could regulate it. I think we could turn around and go, actually, you know, there’s some amazing technology here that has great uses, this isn’t it, you know?
Ketan (30:45)
Yeah.
Yep. I I totally agree. And
Linda McIver (31:02)
It it seems like a marketing
angle.
Ketan (31:07)
Yeah, and if you remember, ⁓ like I I ⁓ this is maybe a bad habit of mine, but I relate everything back to the messaging that’s used around fossil fuels and ⁓ the project we’re trying to phase them out now that they have become incumbent, right? Like the world’s energy system is is fully dependent ⁓ currently on the presence of fossil fuels. ⁓ not fully dependent, but you know, enough that it enough that ⁓ it’s a tough it’s a tough gig.
extricating ourselves from reliance on fossil fuels. and then the message you often hear ⁓ is you can’t switch off fossil fuels overnight. ⁓ and I call this the overnight rule, ⁓ which is basically it is a weird statement to make, right? Because nobody is actually proposing switching fossil fuels off overnight. So why are people arguing against an imaginary line? ⁓ why are they why is their obsession
focussed solely on this idea of like okay we want to sort of just click our fingers and fossil fuels vanish in one second and then all of a sudden everything just gets plunged into chaos. ⁓ and so ⁓ you know that that is basically an argument about the inevitability of ⁓ like fossil fuel persistence right so it’s like basically saying
Fossil fuels will stick around for a very, very, very, very, very long time. And the idea behind that is to say you cannot work day to day gradually ⁓ towards the goal of eliminating them. ⁓ so it’s a way of erasing the only possible pathway for getting rid of fossil fuels, which is to just keep trying every single day of your life to get rid of them bit by bit, and it works, right? Like we know
Linda McIver (32:59)
Yeah.
Ketan (33:00)
we are in just such an amazing time of showing that it actually works when you do this. Like when you have this like multi decade focused project of eliminating a dangerous product from society. It actually works. ⁓ and so, you know, I j I won’t even go into it, but like there are just so many examples. Australia has many of them. ⁓ you can actually reduce the consumption of fossil fuels. It’s just that you have to do it both with a sense of urgency but also like
Linda McIver (33:13)
Yeah.
Ketan (33:30)
o over a very long time period because there is an upper limit at the rate at which you can limit eliminate them, right? That is true. Everybody who’s working on this is completely aware of it that you can’t do it overnight. The AI is inevitable argument. I see it as the ⁓ same philosophy, just at the other end, right? So this is basically a a new technology that is toxic, ⁓ in that like it’s functionally toxic, like it’s like the example you cited, ⁓ is happening
Linda McIver (33:36)
Yep. Yep.
Yeah.
Ketan (34:00)
every single day for sure, right? Like ⁓ people basically relying it on it as a as a information retrieval system. ⁓ and then that what they get is back back is something ⁓ fabricated because that’s just what the software does. ⁓ and then that little nugget of falsehood just gets baked in somewhere. And Cory Doctorow described this as like we’re gonna have to go through like this asbestos removal process over the coming decades getting rid of like, you know, all of the
all of the fabricated falsehoods and lies that are being just sort of deposited around by people relying on these systems as information sources. we are in the early stages of that, right? Like it is still ⁓ it’s happening very, very fast. ⁓ but we are still in the early stages of it. It’s not like every single thing, every single part of the information world is being polluted in this way.
⁓ there are still many, many instances you can find. Like, I don’t know, like I write for an outlet called Crikey in Australia that says that you can’t use generative AI to write your articles, right? So, you know, there’s like ⁓ there people are still getting in trouble, right? Like there was an instance of the Sydney Morning Herald ⁓ publishing a ⁓ what I think personally is completely written, you know, with ⁓ some sort of generative AI system because it reads like the output of gen AI.
Linda McIver (35:23)
Yeah.
Ketan (35:23)
and that person, you know, the article was retracted, and the person, you know, even though the university was like, we defend this. ⁓ so basically, this whole argument of saying that it’s inevitable is is trying to enact the same, erase the slow, careful process of stopping this from becoming ⁓ like a dominant mode of toxicity in in our information environment. ⁓ so it’s like.
Inst like so so like, no no, you can’t switch off AI overnight, right? Like it’s the same it’s it’s it has become the same thing as the line used by fossil fuel advocates to stop us from just going, Okay, well I’m gonna like talk to the people around me, I’m gonna show, I’m gonna demonstrate how that you can do better work without using these systems ⁓ as like a primary source. ⁓ you know, you can
You’re putting yourself at risk. I describe it as like laying landmines around yourself constantly. like all this like slow, gradual stuff of like defending good work and defending doing work in the correct way, in the way that we’ve always wanted to do it, which is like trying to align ourselves as close as possible to the truth. ⁓ they’re trying to they’re trying to attack that. They’re trying to attack that when they say, No, no, no, it’s inevitable, right? So just get on board or like shut up.
Linda McIver (36:21)
Yeah.
Yeah.
Hmm. Mm.
Ketan (36:46)
Right. ⁓ it’s and it feels like an admission to me. It feels like they’re it feels like they’re they’re sort of admitting that they cannot support the purposefulness or the usefulness of their system, of the of their of their change they’re that they’re wanting to to see. ⁓ so they kind of have to frame it as unstoppable to discourage any of us going, hang on, maybe it is stoppable? Sorry, that was a long answer.
Linda McIver (37:12)
Yeah. Yeah. And
Ketan (37:15)
Ranting.
Linda McIver (37:16)
well, incidentally, you can’t switch it off overnight. Anthropic literally did switch off two of its products overnight, like two days ago. I was like, ⁓ turns out you can. Gosh. It’s ⁓ it’s wild.
Ketan (37:22)
Well
Yeah, and and my gosh,
yeah, that’s an amazing that is an amazing story, right? And ⁓ you know, of course and what it really shows is that like this argument that is being used in Australia among any other many other places outside of the US is like we have to have sovereign AI capability, right? Like so you’ll hear this word sovereign a lot. And somehow the logic is that if you build an anthropic data center in Sydney’s C B D
suddenly you own anthropic’s like functions that they perform from that data center as if that’s like somehow sovereign. Well well the basic well the logic is basically like we need to have these data centers in Australia or or Norway for instance. We use the same argument here as well. We need to have the data centers physically located in our country for the purpose of having of having retaining control over AI systems. ⁓ and that is just not how it works, right?
Linda McIver (38:27)
⁓ no
Ketan (38:29)
Because
the company’s lease space, Anthropic, for instance, just to use that example, they control the model. ⁓ if they built a data center in Sydney and and then it had like, you know, mythos on it, the US government still has control over the accessibility of that system, even though the data center is is the infrastructure is located in in Sydney. ⁓ and so if anything
Linda McIver (38:39)
Yeah. Yeah.
Ketan (38:53)
It is worse, right? Because it is a Sydney community that bears the air pollution and noise costs. It is the Australian power grid that suffers, you know, worse, more expensive power prices thanks to more coal and gas being used, which are expensive. And then also more constraints on the grid bringing up power prices. All of that disadvantage so that the American government can turn the dials on access to a chat bot.
like that all of these businesses have like signed up to and are now you’ve got like a thousand office workers baffled as to how they do a spreadsheet. ⁓ because they can’t use their Claude connection to do it. Right. Like it’s just ⁓ not a good situation.
Linda McIver (39:37)
It’s wild. It is I yeah. I I I was aware of the power issues initially. was impossible not to be if you were paying attention, but ⁓ I’ve spent a lot of time because power isn’t my area of expertise, but compute is so to to actually, you know, try try to underst try to encourage people to think about the idea that
Ketan (39:38)
Yeah.
Hm. Yeah.
Linda McIver (40:07)
These are not actually search engines, they are statistical text generation machines and y and and you have no way of knowing you know, I heard someone I won’t name her because it was Chatham House rules, but I heard someone literally say the other day, who should know better ⁓ you can you can tell y whether something’s misinformation just by asking Claude whether it’s true or not. And and this this
Ketan (40:13)
my gosh. Yeah.
Yeah.
Linda McIver (40:37)
This this is really common and and it’s really hard to to get that get people past the the shiny toy that talks to me kind of field.
Ketan (40:45)
Yeah. This is
my gosh. I ⁓ I struggle very, very hard with that. ⁓ I really it’s my one it’s like my biggest one of my biggest problems right now, right? Which is that I personally don’t use generative AI for all the obvious reasons. ⁓ I value the quality of my work as a consultant, you know. ⁓ I I I really, you know, I don’t wanna hand landmine ridden work to a client, right? Like
Linda McIver (41:16)
Yep.
Yep.
Ketan (41:18)
puts me at legal risk, it puts them at legal risk. It’s just it’s just lower quality work. ⁓ but the people around me and you know many people I work with do use ⁓ generative AI and to varying degrees, some of them ⁓ almost, you know, ⁓ like obsessively. ⁓ and, you know, now ⁓ part of my work I I have no say in this, but like part of my work is understanding
How a generative AI system returns results to the type of work that I’m doing, ⁓ and then and then knowing how to debunk it essentially, right? So like ⁓ knowing knowing where the landmines are hidden in that work, right? Because they’re gonna be there one hundred percent they’re there. ⁓ and then and then all of a sudden, instead of doing the hard work of like I’m sifting through evidence, I’m doing analysis, I’m constructing formulas, you’re there trying to find
Linda McIver (42:00)
Yeah. Yeah. Mm-hmm.
Ketan (42:16)
⁓ the error in someone else’s Claude Pro subscription chatbot output, right? ⁓ and it’s always there. It’s always like, you know, a dodgy reference. I w I work in climate advocacy, right? So we are taking on like fossil fuel corporations. ⁓ we’re taking on ⁓ issues that have that have high legal sensitivity, like highly legally exposed. ⁓ and so if we have weaknesses in our analyses or our arguments, ⁓
Linda McIver (42:24)
Mm.
Ketan (42:45)
That is a compounded risk, right? Like that is just such a significant risk. ⁓ and so ⁓ yeah, you know, I’m there like diffusing these bombs. and I I try to ⁓ very politely explain that like ⁓ that it would have been better ⁓ to just do this the evidence-based and
consciously effortful way because the net result is a gain. ⁓ that is a net gain. Like we it’s it’s it’s quicker, ⁓ it’s less risky. ⁓ you’re not basically just, you know, ⁓ the convenience is not free. The convenience comes at a cost beyond the environmental and ethical cost. It comes at a personal cost too. That is why that example you mentioned earlier of the Australian writer, you know, ⁓ using a a a fabricated quote.
Linda McIver (43:19)
Yeah. Yeah.
Yeah. Yeah.
Ketan (43:44)
⁓ is a great example of what I call, a bit jokingly, ⁓ Joshi’s work slop postulate, which is basically ⁓ the trade-off, right? ⁓ so if anyone a line that you hear a lot is basically I just use it for ⁓ you know, fact checking or research, but I don’t use it to write my final my final thing. And I think that’s either people who say that are either lying or foolish. ⁓ either they’re lying in that like
They used it and they didn’t check it, but they said that they checked it, which happens a lot because people want the convenience, but they don’t want the shame of having taken that shortcut. or they’re foolish in that they generated a error ridden mystery box, ⁓ where the mystery is that you have to go and find the errors which are hidden in the most plausible way imaginable, right? Like
Linda McIver (44:16)
Yep. Yep. Yep.
Yeah.
Ketan (44:42)
It looks like a reference. ⁓ it looks, it’s written in perfect English. ⁓ it’s it it’s just like plausible, even if you’re an expert, it sounds vaguely correct. ⁓ so it is designed, it is like this multi-gigawatt engine of convincing. ⁓ and so you are choosing to do your work by using this like system of text production ⁓ that is designed to be as tricky as possible to find the errors in.
Linda McIver (44:58)
Yeah.
Yes.
Ketan (45:09)
Or you could just do it yourself, right? Like, I it just my god. I just like I’m like, that’s foolish, right? That is foolish. You take longer, you will probably miss a bunch of errors, ⁓ and all of a sudden you’re putting shit out into the world that is sorry, can I I can swear, right? I’m I assume I can swear. ⁓ good. you’re putting like you’re putting all the stuff out into the world that is that is just ridden with errors, you know, even if you check ninety-nine point nine percent of it.
Linda McIver (45:11)
Yep.
Mm. Mm.
Yep. You can swear, it’s fine.
Yeah. Yeah.
Ketan (45:39)
There is gonna be something. ⁓
and and then, like, you know, I think this is really important because I I’m also hearing more and more. I I I mentioned it this morning on Blue Sky. people referring to the use of Chat GPT or Gemini or something as assistive, ⁓ which is, you know, a little bit of like ⁓ kind of
Disability washing, maybe I think you could call it. which is like, you know, basically kind of saying, like, yeah, it’s like a it’s a crutch in the literal sense of like I’m kind of using it as support, you know, and it’s not my main like walking. So and and and it’s like they all kind of say, like, you know, I just used it for looking up some references, or you know, just like generating a few ideas at the start, and then the rest of it I did myself. and I think this is really important, right? Because if you
Linda McIver (46:06)
Mm. Mm. Yeah. You can’t argue against it ’cause it’s disability support.
Ketan (46:33)
If you held a gun to my head and you said you have to use Chat GPT for something, I would use it for writing and not for research. Right? Like I would use it to I would do all the research myself, write a crappy first draft, and then put it into the system and say, rewrite this. and then I would go and check everything, of course, right? Because it would just change some facts randomly. But I would I would argue that it is worse, way, way worse.
Linda McIver (46:52)
Yep. Yep. Mm-hmm.
Ketan (47:01)
To use it as the first step rather than the last step. Right? So so using it as like a as like an engine of like information retrieval, using it as a search engine, ⁓ using it to like extract basic facts about things, ⁓ all of this stuff is where it fails the worst. It actually functions as a text reproduction engine because that is just fundamentally what the software does, right? Like it it it it it’s trained on so much of our work of you know people who publish stuff.
Linda McIver (47:17)
A hundred percent.
Ketan (47:31)
⁓ and so like it somehow it’s flipped. Like there’s there’s this idea of like, no, that’s the okay version, right? Like that’s the more acceptable version is to use it for the preliminary research. But then you can’t track it. You’ve got this like, you know, untraceable, you know, ⁓ like thing of like errors. ⁓ often, you know, they’re very subtle.
Linda McIver (47:39)
Hm.
So either you
either you do the research again yourself or you’re trusting its output. Like there’s no middle ground.
Ketan (47:59)
People are really r replacing search engine use with generative AI. ⁓ they really are, you know. It’s not a yeah, and it’s I think it’s important to be really clear-eyed about this because sometimes can people can have this like, ⁓ you know, it’s it doesn’t work very well, you know, it doesn’t really do what it says. But it’s true, right? It doesn’t do what it says, but it does something else very bad very well. ⁓ and which is tricking people into thinking that it does what it says.
Linda McIver (48:04)
I know, it’s horrific
Yeah.
Yes,
Ketan (48:29)
⁓ and that
Linda McIver (48:29)
yes, said
Ketan (48:30)
works astonishingly well, right? And so like what you saw with Google. Yeah. ⁓ and so I’ll just quickly tell you this interesting story, right? Because it’s relevant. ⁓ of ⁓ there’s a search engine called Ecosia and it’s always kind of marketed itself as like the kind of green search engine ⁓ because they plant trees ⁓ and they buy some solar. And I’ve always been a little bit dubious about it, you know, like ⁓ like it’s planting trees doesn’t really undo fossil fuel emissions and
Linda McIver (48:32)
It’s a deceit box.
Ketan (48:57)
Yeah, that’s a whole different story. But I was like, whatever, like they’re just they’re basically fundraising for tree planting. It’s just like fine, I don’t care. ⁓ but what they did recently ⁓ is they started ⁓ pivoting to AI, right? So they have like a and they’re calling it basically like green AI. ⁓ and so the idea is that like ⁓ they still plant as many trees, they like measure the energy consumption from returning generative AI results.
Linda McIver (48:59)
Mm-hmm.
Ketan (49:22)
And when they’re criticized, ⁓ the owner basically just says, Well, people are just switching to generative AI search now. So so we’re just going with the tide, you know, we’re just going with with what people want, but we’re kind of offering the clean version of what people want, right? and ⁓ I think that they’re enabling a system where both ends get worse, right? So the information retrieval is getting worse, ⁓ because for all the reasons that we mentioned.
Linda McIver (49:34)
Hm.
Ketan (49:49)
But they also have no real transparency around like the amount of energy they’re consuming, the amount of renewable energy they they purchase, the trees they plant. Like they don’t disclose any of this stuff, right? If they did, what we would see is that in the search engine era, ⁓ they had like, you know, a tiny fraction of the energy consumption as they now do in their generative AI era. ⁓ and so, you know, that sort of energy bloat is something to oppose, even if their individual footprint as a company is probably tiny.
they’re enabling a shift towards something that is bad at both ends. So yeah, maybe a nice example going back to your inevitability thing as well. It’s like it’s not inevitable. Like some people are choosing to make it worse.
Linda McIver (50:29)
Yeah. No.
Yeah, absolutely. And it’s you know, the even the things that people say it’s okay if you just use it to… ⁓ I you saw ⁓ I did a post a while ago about a medical appointment I had where the the they used ⁓ the the specialist used a health AI to summarize the appointment and it it had ⁓ at least nine significant errors. It made up drugs I wasn’t on, it left off drugs I was on. That
That could be fatal, just that. It left off conditions I have and made up conditions I don’t have and it just changed a whole lot of shit. and and when I confronted the doctor about it in the second appointment, because he sent back the letter that was full of errors, he said, It would have been worse if I’d written it. And I was like, Well, first of all, would it? Can we do a test? Because if it if so, then you are not competent. Like just just no, you’re just not.
Ketan (51:27)
Yeah. You should be
fired. Yeah. This is I mean, but what you’ve described is exactly my philosophy on this, right? Which is why don’t we just hold people to the standards we’ve always held them? ⁓ you know, if a lawyer fabricates a case, if they sat down and thought, I’m gonna cheat on this by making up a case, ⁓ they would suffer the severest of consequences because that’s like active deceit. ⁓ but somehow
Linda McIver (51:28)
Yeah.
Mm.
Mm.
Yeah, yeah.
Ketan (51:54)
Doing the same thing with a generative AI system ⁓ is considered to be kind of a little, you know, softer. ⁓ and I think that just needs to be dropped. I think the only way that we protect ⁓ I think the only way that we protect information quality and our shared sense of reality and like, you know, facts and ⁓ like everything holding us together ⁓ is by introducing really
Linda McIver (52:01)
Mm.
Ketan (52:22)
Really serious consequences for fabrication. ⁓ it’s so straightforward. ⁓ but I really think that ⁓ it should be treated with a lot more seriousness than it is. ⁓ I’m shocked by your doctor just casually saying, it would have been worse if it was me. Because that’s ⁓ that’s a kind of flippant response to something that should be very grave, right? Like something extremely serious.
Linda McIver (52:40)
Yeah, yeah.
Hmm. Mm.
Ketan (52:51)
and you know, I’m sure many people have stories, you know, ⁓ have stories along these lines, right? In in not just seeing the doctor, but in many different parts of their lives. And we are going to have to fight back against it. I, you know, my coming of age, you know, like in in the sort of like in my like twenties, in my late teens and twenties was like the whole, you know, skepticism, pro science, you know, you remember like the sort of
Linda McIver (53:00)
Yeah. Yeah.
Ketan (53:20)
⁓ like, yeah, science is cool. You know, that’s sort of like 2010s era stuff. ⁓ and part of me is just like, I don’t know, I don’t know, I don’t think it has to be exactly that. But man, we need to go back to you know, just valuing the sheer social benefit of being correct about things and checking things and like trying not to lie, like ⁓ that is as a value, as like an ideology.
Linda McIver (53:25)
Yeah.
Yeah. Yeah.
Ketan (53:48)
⁓ I wish we can figure out a way. I hope we can can figure out a way to make that popular again, like, you know, from the way it was. ⁓ maybe not led by the same people who’ve all turned out to be quite horrible, but ⁓ maybe figure out a slightly more humanistic and generous way of doing it. But yeah. That’s my view.
Linda McIver (54:03)
Yeah.
It’s wild to me that we have to say that,
that actually being truthful, being accurate, being correct has value. That and that that that feels slippery now. That feels that feels like a contentious statement. That’s just how? How did we get here? That’s wild.
Ketan (54:24)
And I and then you’ll notice that when I talk about generative AI I often talk about it in those terms ⁓ of you know, the the falsehood machine or like, you know, things like that, right? Because it’s I think people still have a they flinch a bit when they when they know that they’re wrong. Like there is still some sense of shame about being wrong about things and I’m I wanna capture that little spark and turn it into a fire. Yeah.
Linda McIver (54:32)
Mm. Mm.
Yep. ⁓ I love that. That’s a beautiful note to end on. Thank you so much. This has been a great conversation and I’m I’m really ⁓ grateful for your generosity and sharing your expertise.
Ketan (54:57)
Yeah, always, always happy to come back and thank you again for having me. I really appreciate it. So yeah, cheers. nice convo!
