- il y a 5 semaines
Clean By Design: How Can We Create Green(er) AI?
Catégorie
🤖
TechnologieTranscription
00:01Hi, everyone. Thank you for joining us. I'm Freya. I'm a journalist at Sifted. I write about climate tech, and
00:09today we're going to be talking about how to make AI greener.
00:12There's obviously so much buzz about the potential of AI at the conference today and more widely, but there are
00:19lots of questions to be asked about the energy it uses and its wider environmental footprint.
00:24So we have three panelists ready to talk about that. Let's go down the line for a quick intro and
00:30then we can get into the questions.
00:33Hi, I'm Kate Kellow. I'm the founder and CEO of Amini, and we are a company building the data infrastructure
00:39for Africa and the global south.
00:41I'm Sasha Luccioni. You might have just seen me. I'm an AI and climate lead at Hugging Face, which is
00:46a platform for responsible open source AI.
00:51And I'm Siddharth Singh. I work at the International Energy Agency and we are an intergovernmental organization that works on
00:57long-term pathways on energy and climate.
01:01Let's start by laying out the current demand of AI and then we can get into the solutions. Siddharth, you
01:08have just put together an enormous piece of research on this.
01:11Talk us through where current demand comes from and how large it is.
01:16Of course. And thank you for the question. So let me start by saying at the outset that there can
01:23be no AI without energy.
01:25Now, it sounds like an obvious thing to say because there cannot be anything without energy.
01:30And it's also a little biased because I'm from the International Energy Agency.
01:34But if I can say so, you know, we had a long series of discussions and of course a lot
01:41of analytical research.
01:43Energy turns out to be one of the biggest physical constraints on the uptake of AI, of the build out
01:51of data centers along with, of course, the physical manufacturing of chips and so on.
01:55We find firstly that at the aggregate level globally data centers of which artificial intelligence is a core core set
02:06of demand constitutes only 1.5% of electricity demand globally.
02:13Now 1.5% sounds like a very small number relatively, right?
02:18But data centers tend to cluster around each other. They also tend to cluster around cities.
02:25And as a result, where data centers have been clustered, electricity demand can be as high as 20 to 25
02:32% of the total consumption of electricity in those places.
02:37And that's causing a major hurdle as data centers and technology companies set to expand.
02:43We don't have the physical energy infrastructure to see that build out at the pace at which the technology sector
02:52is expecting it to be.
02:53So just to give you a flavor of how quickly data centers can be built out.
02:58In about 18 months, you can have a massive hyperscaler constructed and to the maximum possible potential even filled out
03:07with the computing hardware.
03:09On the other hand, the long distance transmission lines, the energy infrastructure, the transformers that is needed to cater to
03:18the data centers can take 8 years, 10 years to get constructed.
03:22So there's a big mismatch. And therefore, we find that there's an issue with how quickly the build out is
03:30happening.
03:30And therefore, the energy comes at the heart of this conversation on AI.
03:37My next question was going to be, is the grid ready to keep pace with that demand?
03:42It sounds like, no, it isn't.
03:44No, I mean, if the energy sector has greater visibility of the build out of data centers, then yes.
03:52In our analysis, we find that as much as 20% of the total capacity additions from now to 2030
04:00is at risk because the energy sector may not be ready.
04:04Because the transformers may not be ready.
04:07There's already a big lag in the timelines in which transformers are being delivered, in which new cables are being
04:14delivered,
04:14in which, of course, the long distance transmission lines are being constructed.
04:18So there's a risk. It doesn't mean it won't happen.
04:21It just means that the technology sector and the energy sector need to work together.
04:25It means that there needs to be greater transparency on the build out of these data centers, which currently is
04:31lacking.
04:32Kate, can I bring you in? Where do you see future demand coming from in particular?
04:37I mean, from the place where I'm sitting is Africa.
04:41And today, one of the stats to bear in mind is that only 2% of African data actually gets
04:48processed on the African continent.
04:50So most of the data today is being sent out.
04:54And very similarly, the capacity of data center for data center in Africa is very little.
04:59You don't have the same level of infrastructure that you actually have in the global north.
05:04So there is still regions that are extremely compute and infrastructure scarce,
05:09but that scarcity comes layering on top of other challenges.
05:13You think about power.
05:15Not every country on the continent has enough power to power a data center,
05:19has enough energy to power a data center.
05:21Not all countries have enough water.
05:23So I think we have to figure out a way to self-organize,
05:27pull resources where possible.
05:30Some countries will build, some countries will buy, some countries will help power.
05:34But that will definitely help bring a lot of the value back on the continent.
05:40And I think this is where we should head.
05:43I want to get more into that question in a second.
05:46Let's look at, so we've kind of mapped out the demand,
05:49and now I want to go into the hardware and how we can make that more sustainable.
05:53Sasha, the title of the talk is Cleaner by Design.
05:56What does that mean at a hardware and infrastructure level?
06:00So AI hardware has been getting a lot more and more efficient, essentially.
06:05But the thing is, is that it's like almost a never-ending process
06:10because people are using it more and more.
06:12And so I think that for me, Clean by Design would be rethinking the way we do AI.
06:18Instead of saying, we need a bigger model, we need more data, we need more GPUs, we need newer GPUs,
06:24bigger data centers.
06:26We should be thinking about how to distribute the training, make the model smaller, make them more efficient.
06:33Like for example, even in Canada currently, there's like a national data center strategy.
06:37But the demand of a data center is so big that it's hard to put it into any one province
06:42because every province is like, we don't have the capacity like right now.
06:45But what if instead of one big data center, you had five smaller data centers and you had them distributed?
06:50Then you can not only leverage existing renewable energy, which we do have,
06:53instead of like building a new natural gas plant, which I think that is what's going to happen sadly.
06:58But also you're doing things a little bit more smarter.
07:00And for example, instead of having this data centers are massive, right?
07:04And they make noise, they make pollution, like it's really like a huge digital factory almost.
07:10But you can make them more integrated.
07:12You can reuse the heat for heating campuses.
07:14That's what they have here in Saclay.
07:16They have a supercomputer that's used, that heating is used for heating the campus.
07:20And so there are like smarter and cleaner ways of doing it.
07:23We just need to kind of flip the script a little bit and think about how we're doing this.
07:27instead of being like, no, we need another massive data center.
07:30We're going to power it no matter what.
07:32We're going to put it in the same places that we're always going to put it because we know how
07:34that works
07:35instead of putting it in new places or places that have different.
07:38So I think we really just should rethink this whole thing.
07:42Sid, which technologies are you interested in in terms of changing the consumption of the data centers itself?
07:50So firstly, I mean, just responding to the kind of general theme of, you know, how to make AI more
07:58sustainable.
08:00We look at it from, again, the energy perspective.
08:02So looking at data center related characteristics that we think can be influenced.
08:10And therefore, when you think of the design, it's not just the design and the architecture, the models,
08:16but also how as a, in terms of a regulatory framework, how do you think of data centers as an
08:24integral part of the energy system?
08:26And therefore, what is it that you can actually do to make sure that data centers play a role that
08:32is conducive to the way the rest of the electricity system works, right?
08:36So here, there are a few things that, you know, perhaps get highlighted.
08:42You know, Sasha mentioned about the size of data centers itself and how they're quickly being built out.
08:48I mean, just to give you maybe a small data point. Today, a hyperscaler 100 megawatt data center could consume
08:56as much electricity as 100,000 households, right?
09:00And the biggest one under construction would consume as much as 20 million households.
09:06This is a very, very big number. It's you're adding an entire city, you know, to the electricity grid very,
09:13very quickly.
09:15The electricity network cannot take that kind of a load, you know, very quickly.
09:21So, how do you get to a point where data centers play a positive role in the system?
09:30For that, you need to, you know, for data centers and technology companies to work on issues such as flexibility.
09:38The flexibility happens to various sources. So, you know, green by design in this case would mean ensuring that data
09:45centers come in places that are renewables rich,
09:48that are energy rich and that don't already constrain the grid. So, that's a locational aspect to it.
09:54There's also the fact that data centers have a lot of backup electricity generation capacity because they need to ensure
09:59that, you know, 100% uptime.
10:01So, how are we using that backup capacity? Can we ensure that the backup capacity actually plays a role in
10:08generation itself and comes online when perhaps there's a, you know, higher carbon intensity of the grid?
10:15So, you know, for example, at night when coal generation will be higher and solar may be much lower as
10:20an example.
10:22There's also, you know, some analysis that we did to understand what it means to shift the load of the
10:29computation itself.
10:31And we find that even if for 1% of the year, just 1% of the year, data centers
10:38or big tech companies were able to shift the load across regions,
10:42they would be able to ensure that the total build out of all the grid, of all the data centers
10:48from now to 2030 will be able to be integrated into the system without a problem.
10:53Just 1%. Of course, it comes at a cost, but this is where we think that, you know, regulators and
10:59technology companies have a role to play in having that dialogue.
11:01Can I add on to that? I think when you think about clean by design, exactly what you mentioned, we
11:08have to start thinking about new ways to design architecture.
11:11And although Africa is a place where we're still kind of lagging behind, there are definitely a lot of opportunities
11:18for us to do it right.
11:19So building smaller data centers with less compute capacity than the others that are the mega structures that are being
11:26built right now.
11:26But also, if you look at it, Kenya, for example, the grid is powered 90% by renewable. So it's
11:33a mix of geothermal, solar.
11:35So there are places where it's actually it could be appropriate and more sustainable to build those structures.
11:40And I think we now have to open our eyes, not just at the downsides, but also the opportunities that
11:48some of those countries that have not been considered before
11:51can actually use to be able to leapfrog into a world that's more sustainable.
11:55Earlier, you made the point about water. How do we, which technologies are you interested in that are getting around
12:01that or improving that side of things?
12:04One of the things we've been looking at is identifying groundwater sources. So there's a lot of modeling that's happening
12:11right now.
12:12Companies building models using geospatial data to be able to identify where to basically find more groundwater sources.
12:20So that's one of the options. But then there's a lot of work now that's also being done in around.
12:26Can you build closer to oceans? Because then what happens if you're able to capitalize on some of this ocean
12:32and seawater to be able to call the data centers?
12:37So these are some of the innovations that are happening right now that a lot of countries who are constrained
12:43from a resource standpoint will be looking to invest in the future.
12:47And are you seeing more interest from international companies in placing their centers across Africa?
12:53And second past that question, is that being done in a way that you are happy with?
12:59The short answer is no.
13:06It's kind of a mix. So you see them still investing hundreds of billions of dollars in places that are
13:12already stuffed with data centers.
13:14Same places over and over and over again.
13:16But then at the same time, when they decide to do a project when it comes to Africa, Microsoft has
13:22been building the large data centers in Kenya
13:25and has announced a couple of other projects across the global south, Southeast Asia and Latin America.
13:30But a lot of it comes at a cost. And while a lot of those structures are being built, they're
13:36still processing a lot of the citizens data out.
13:39So they're like, OK, we're going to come build a large data center, green data center.
13:43We're going to help you build your own infrastructure, you governments, right?
13:47But in the meantime, your citizen data will be processed in the cloud and that cloud being located outside of
13:53the continent.
13:54So we're still extremely reliant on outside interventions and big tech companies to come and build that infrastructure for us.
14:02Where actually, I think that there is a space where local companies who are thinking about infrastructure in a much
14:07different way,
14:08from a bottom up standpoint, who understand what is the value, where the data is stated, how much data is
14:13there,
14:14and can really design with a clean mind space or mindset from the beginning to come and displace those large
14:23hyperscalers in emerging economies.
14:28We've talked about the hardware and infrastructure. The other part is the models.
14:32Sasha, how are we currently measuring emissions of models and what progress needs to be made on that?
14:38There's very, very little data, sadly. Actually, I'm working on a paper about this now.
14:44We did an analysis of kind of the most popular AI models in the last, I think, 10, 15 years,
14:49and transparency has been going down. Like when I started my research on this, you could read a research paper,
14:55look how much, you know, how many GPU hours they use, more or less what hardware it was.
15:00You could make an estimate of energy, like a lot of the papers I've written have been based on this.
15:04Now, they don't even tell you how much time it took. They don't tell you how many GPUs were used.
15:09So it's all based on hearsay, because AI models have become commercial products.
15:13They're no longer just research artifacts. And so it's really frustrating.
15:17So essentially, for example, I run a lot of my own, like I essentially download models and run them,
15:23but that's only possible for open source models.
15:25You can't do that for a model like ChatGPT or Claude because it's behind a user interface.
15:31So you can just send a request and get an answer, but you can't actually see the model itself.
15:35And so it's really frustrating. But recently, for example, I read a, I led a project called AI Energy Score,
15:40where we tested hundreds of models in different tasks, and we compared essentially how much energy they used.
15:46And the difference between the most and the least efficient models that we tested was 60,000 times more energy.
15:53And honestly, because we're using more and more AI, even if it's a smaller amount, like only 20 or 30
16:00times,
16:00at the frequency at which these models get used, if you opt for a more efficient or more frugal model,
16:07it really adds up because you're getting, you know, millions of requests a day or billions sometimes.
16:12So I think that if we start making these small choices, but in order to do that, we need more
16:16data and more transparency.
16:17And I think that this is really the crux of the issue. A lot of the decisions that we're currently
16:22making are based on,
16:23I don't know, rumors or, or guesstimates or, you know, extrapolations based on this and that and that.
16:31And that's kind of counterproductive because policymakers can't make informed decisions.
16:35Grid operators can't make informed decisions. Users can't make informed decisions.
16:39Everyone's just kind of like, well, I'm guessing going to use this magical black box that gives me answers.
16:44And I don't know how much it costs and I don't know where the energy is coming from.
16:47And I don't know where my data is getting processed, but I'm just going to trust this black box because
16:50some company sold it to me.
16:52And so I think that on all levels, we should be mandating disclosures.
16:57We should be mandating transparency because, I mean, for so many things that we do in our daily lives,
17:01like the cars we drive, we know how fuel efficient they are.
17:03Even for appliances, we have Energy Star ratings that let us pick a microwave or an apartment that is more,
17:11you know,
17:11sustainable or environmentally friendly.
17:13And we don't do that for AI, which is like the biggest disruptor of our society.
17:18So for me, it's such a basic thing that we still haven't somehow haven't implemented.
17:23We really should.
17:24Sid, I presume you came across a similar hurdle in your research.
17:27How did you get the numbers?
17:29How does the IEA get around that?
17:31No, for sure.
17:32This is the heart of the issue, right?
17:35The reason why we don't have better analysis globally is just the absolute lack of data
17:43when it comes to, on the one hand, consumption of electricity by AI, by models for inference and so on.
17:50But on the other side, on the broader context, the consumption by data centers as a whole.
17:56So we had to, you know, attack this problem from various dimensions.
18:01We had to, we did geospatial analysis to, you know, pinpoint the location of each and every one of the
18:0711,000 or so data centers,
18:09the large data centers around the world, you know, on a map.
18:12We had to see where the next ones are being built.
18:15We had, on top of that, juxtapose where the digital infrastructure is.
18:20We had to juxtapose where the long distance transmission lines are.
18:22We had to juxtapose where the electricity is being generated.
18:26So that was one aspect of the issue.
18:29The other was that we tried our best to get information from the various, you know, regulators on how much
18:36electricity is actually being consumed by the data centers
18:39in places where the reportage may exist.
18:41So anyway, it was a massive effort.
18:43We had to put all of that together.
18:45And that's how we came to this number.
18:47That last year, 415 terawatt hours of electricity was consumed by data centers, which is like I mentioned earlier about
18:551.5%.
18:57But this is growing very fast.
18:58So in our, you know, we did some analysis on where we see this go in the future.
19:02We find that by 2030, the total consumption by data centers grows to about 950 terawatt hours, which doubles.
19:13But the interesting bit here is firstly that this doubling to 950 makes it as much as Japan, the whole
19:20country of Japan consumes today.
19:21So as much electricity as Japan is just data centers by 2030.
19:25But of which the component that's AI or the component that or the hardware which caters to AI, that grows
19:34fourfold.
19:35So the growth is actually coming from this need for, you know, data centers that relate to this kind of
19:44demand.
19:45And honestly, if we start looking at the individual models, I mean, Sasha helped us with some analysis.
19:50We found, you know, the numbers to be quite striking, especially when you consider how widely they're being used today.
19:57So as an example, the, again, in experimental conditions, thanks to Sasha, we found that an eight second video generated
20:08by AI consumed as much electricity as charging a laptop two times over.
20:15And now if you, if you use Instagram or any of the social apps, you will find they're flooded by
20:20AI videos.
20:22So again, this is to say this in experimental conditions, the actual consumption may be, you know, may be lower
20:29due to optimizations.
20:30But it's a black box, like you said, and we have to get any numbers from the actual developers of
20:37AI models.
20:39I mean, we got it from you, you know, and I'll just put it that way.
20:42Yeah.
20:43This is like, if you want numbers, you have to run the models themselves because no company is actually going
20:47to say generating a video with whatever mid journey or whatever it is, is using this much energy.
20:53We have absolutely no idea.
20:54And that's crazy.
20:57In terms of, I understand that this might be a hard question given that the points you've just made, but
21:01where is the progress?
21:02Where are the bright spots?
21:04Where is the innovation that's making those models greener?
21:07And what does that look like?
21:08I think there's different, there's different aspects to this.
21:10There's the kind of the technological solutions and people, I think increasingly people are like, well, why are we doing
21:16things in this way that like the models are bigger and bigger?
21:19We can do things a little differently.
21:20So for example, Hugging Face trained small language models that were just as performant as the bigger ones, but they're,
21:26they actually are like physically, physically, digitally smaller.
21:29And that means that it was cheaper for people to run them to use them, you can run them on
21:33a device, you don't need a GPU.
21:35So there's like the technological solutions, you can do quantization, you can do pruning and things like that.
21:40But so that's kind of part of the issue.
21:41But I think that also part of it is kind of the, the social kind of like the psychological thing.
21:46And I think that people are coming to terms with the fact that just because the models are bigger doesn't
21:50mean that they're going to do exactly what they need them to do.
21:53Because the way that we evaluate AI models nowadays, especially large language models is so generic.
21:58So it's going to be like, oh, we, it understands language or it's able to answer questions.
22:04And it's such a high level, you know, evaluation.
22:06And so it can be the case that a model trained on all of the internet can answer a very
22:11high level questions about, you know, I don't know what year was Napoleon born or whatever.
22:15But then when a company wants to use a model for analyzing financial disclosures or for, you know, answering questions
22:22about IPCC reports, that's not the type of questions that it's going to get.
22:26And so just because this model did super, super well, and it's bigger and better and more data doesn't mean
22:31it's going to do well in the actual concrete things that people want to use it for.
22:35And so I think that we're starting to have a reckoning where people are like, well, actually, this smaller model,
22:39like actually, just last week, I did a blog post about this.
22:42I compared models of different sizes, and I showed that, yeah, typically like the 200 billion parameter models are slightly
22:48better.
22:49But then there are like the second place models use 25 times less energy, and maybe they're like 2%
22:55less accurate.
22:57But for specific tasks, like we look at, we looked at financial data, we looked at health data, we looked
23:02at climate data.
23:02And when it comes to these very like specific things, you don't need a bigger model, you can take a
23:06smaller model and adapt it.
23:08And I think that this is slowly kind of coming, like companies are coming to terms with this, developers are
23:12coming to terms with this.
23:13And I hope that that's the direction we're going to go into.
23:16Kate, what are your thoughts on the trade off between performance and sustainability?
23:20If there should be those and how we convince people that that might be a good equation to look at?
23:25Yeah, it's funny because we went through this like 10 years ago already.
23:29So we really, nothing new, history is repeating itself.
23:33And 10 years ago, we realized that we actually had to train smaller models that were hyper optimized for specific
23:40tasks.
23:40And we are now going down the same road.
23:43And I think there's been a lot of biases and we kind of have to deconstruct and reconstruct how we
23:49look at AI and innovation.
23:50Like a lot of the fluff that has been coming out lately is that you have to have the largest
23:55supercomputer, you have to have the largest general purpose model.
23:59And that only is real AI, right?
24:02But actually, it's all about what is the problem you're trying to solve?
24:06How much data do you have available?
24:07What does your data set look like?
24:09Can you build a smaller model that's hyper optimized for that specific task?
24:13And can it outperform a lot of this bigger generic models?
24:17And that's one of the approaches that we've taken even in our company.
24:21You know, we've tested a lot of the global foundation model for our science.
24:25You have so many companies are like, hey, 60 billion parameters and all these things.
24:29But actually, we realized that when we were testing them on downstream tasks for emerging economies, the accuracy was dropping
24:35massively.
24:36Because they're not trained on the data that you're testing them on.
24:38Exactly.
24:39Because you have to do your own testing.
24:40You can't just take it for granted that it's going to work.
24:43Absolutely.
24:44Also, AI is such an umbrella term that I think people are like, oh, this AI model does this AI
24:49task, which means my AI task is going to be done by the same AI model.
24:52But first of all, there's like, I don't know, there's so many different approaches.
24:56Large language models are just a very small fraction of actual AI.
24:59But also, you know, the number of tasks that you can do is almost infinite.
25:04And there's all these, I think that's where the interest lies.
25:06Like I personally, I'm not as interested in general purpose models.
25:08I'm interested in very task specific models because they enable you to, you know, use cleaner and more consensual data
25:15sets.
25:15They enable you to train less for less time using less energy.
25:19They enable you to own the technology and not have to run it on a cloud server somewhere.
25:24Like there's a lot of upsides to small models and not all of them are just, you know, sustainable.
25:28It's more really like an ownership aspect.
25:31And I think that if we continue going in this like large language model, foundational model direction, it's also a
25:36concentration of power in the field.
25:38Because that means that, you know, the big models where the big companies with the big data centers, it's really,
25:43it makes it concentrated.
25:44And then if we say, okay, well, we can develop kind of smaller models, more distributed models, more, more local
25:49data and context.
25:50Then it's kind of, it's more democratic from a, and then, and then like when I started in AI, that's
25:55how things were.
25:56I still remember that, like as Kate says, 10 years ago.
25:59Same.
25:59And now we're like, why is everybody using the same models and expecting them to work for all of these
26:04use cases?
26:05That's never going to happen.
26:06It doesn't work like that.
26:08I want to talk about the governance of this issue and whose role that is.
26:12Sid, what are your thoughts on that? Is this, who, who is, who oversees this question?
26:17So, not an easy question to answer because I think that the kind of momentum that you need before a
26:28sector finds an appropriate type of regulation around it, that currently does not exist in major markets, right?
26:37I think the first step to this would be greater transparency.
26:40And some of that may have to be mandated, but we expect that, you know, even voluntary disclosures can help
26:47kind of, at least help us understand the pipeline of upcoming data centers, you know, as a very first step.
26:54So, you know, from the government side, yes, there, there perhaps need to be some more, you know, information for
27:02technology companies on what, on what the, where the pipeline of data centers could potentially be located.
27:11What are the incentives for them to actually consume, you know, cleaner electricity, as an example.
27:19The other thing that we found pretty interesting was that when we tried talking to data center operators, and we
27:26found that as a potential, the, the heat that is generated from it, and the heat, typically when it goes
27:33into the water streams, that water can be used, that, that hot water can be used to.
27:40So, just in Europe, 10% of the total heating requirements in buildings can be displaced if data centers were
27:52able to, if the heat from data centers were actually reused for district heating.
27:57heating. Data centers are okay, data center operators are okay to do it, but currently there are no frameworks that
28:04encourage them to do it.
28:06Incentives.
28:06There are no incentives, but there are no market frameworks. In fact, some of the data operators were like, just
28:11take it, just take our water, we don't need it.
28:14Whereas there should be a monetizable, you know, asset. And that framework doesn't exist. So, there's also a role for
28:22the government to actually make it conducive for data centers, firstly, to emerge in places where they think they should
28:27emerge.
28:28The incentives for them to, you know, consume electricity when it is, when it is greener. So, the peak of
28:35the day when say, solar generation is high or when wind generation is high as an example.
28:41And also for heat water reuse. So, that's on the one side. On the other side, of course, the technology
28:46industry, you know, needs to realize that at some point, with all the big amounts of money that's going into
28:52this, they need to also play, you know, a supporting role trying to help the energy sector cater to their
29:00growing demands.
29:01It's like a win-win situation. If you have more information, then you can generate better, I mean, more adequate
29:07energy and vice versa, right?
29:08For sure. But, you know, the problem we're seeing in our countries is that governments don't even understand.
29:14No, they don't know where to start. They don't know where to start. And they're being swayed by the fluff
29:17that's coming out. And they're like, okay, so you mean I need to invest at least $100 million to build
29:22a large share tree data center?
29:24No, actually, that's not the case. Yeah, I heard that. I don't remember who I was talking to. And they
29:28were like, oh, they were saying that building a data center generates 10,000 jobs. Building a data center? Oh,
29:33and also startups that could plug directly into the compute that we build locally.
29:38I was like, that's not how the internet works, eh? But also like building a data center barely creates any
29:43local jobs. Most of the jobs are, you know, the people who are using the data center elsewhere.
29:48So, you know, governments are actually like putting money in like Latin America in particular. I think it was like
29:53Uruguay who they're like, oh, we're going to invest money into building this data center.
29:56But actually, it doesn't bring any concrete advantage to them while using their water, while using their energy. And so
30:02it's like there's, yeah, there's a lot of complications.
30:05They're responding to fear, really. You see all these announcements around sovereign AI infrastructure, large compute, bigger and bigger and
30:13bigger. And everybody is like, what does that mean for me? What do I need to do? And this goes
30:17straight, they jump straight to hardware, they just straight to infrastructure, without actually taking the steps of like, understanding what
30:24it truly means, and how it will impact their local population.
30:27Exactly. And like, who's going to benefit from building a new data center in Kenya, in Canada, and wherever, like
30:33who's actually going to be using it? And like, how's it going to help local innovation, local business, right?
30:38And on the model side, the software side, how do we incentivize for green? Whose role is that? And what
30:44would you like to see?
30:45I mean, the problem with AI is that it's it's hard to regulate because it's so diffused. Like, for example,
30:53I don't know if you remember, like what, a year and a half ago, there was this letter of like,
30:55let's pause AI research for something.
30:57For six months, because it's getting too dangerous. And for me, that makes no sense. Because it's like if there's
31:01no single central button, you can click to make everyone stop doing AI. And there's, you know, academia, there's industry,
31:08big companies, startups, etc, and everyone's doing some flavor of it. But I do think that like starting with things
31:13like energy scores, could be a good way at least to, to make it clear for people that it's not
31:19free, that it's not just like some, you know, I don't know, like, like, for people like AI is so
31:23ephemeral, it's like just like some model in the cloud, but it's actually a very tangible amount of
31:27electricity. And I think that starting with things like this can already help. And then I mean, also, I think
31:33market pressure is going to catch up, like people can't keep investing, you know, 10s of billions of dollars in
31:38building these super computers. At some point, they're going to realize that the money is not being recuperated, that you're
31:44not getting the return on investment. And I'm hoping that there's gonna be some market pressure and some like regulatory
31:48pressure that pops up.
31:49I'd love to see benchmarks coming back as well. No one's doing them anymore, you know, measure performance or some
31:57extent performance. But then again, people actually don't know what data they're training on. And so probably the data that
32:03they're testing was in their training data. So there's this weird thing going on there. But also, they don't look
32:07at other factors, like it's just performance and not like not costs, not energy, not I don't know, not not
32:13where the data is coming from, not bias, not whatever fairness, it's really just like performance driven.
32:18And that's not incomplete. To round off, how hopeful are you that particularly the large language model providers will factor
32:28this in increasingly?
32:31I'm starting to see the tide turn like I I'm starting to see people question, like I understand that large
32:37language models and generative AI have been like the cool thing that people have been trying to do lately. But
32:41I'm definitely starting to see developers being like, well, actually, we don't need a large language model, we need a
32:46small language model, or we need a non generative model.
32:48And I think it's coming back. And you know, a lot of people who are doing cool research are starting
32:53to see people get interested in that research again, after like two to three years of no one caring about
32:58optimization or whatever it is they were doing.
33:00They're like, well, people care about why I work again. So I'm starting to see it. And maybe because developers
33:04and research scientists are kind of the core, I'm hoping that this will percolate into the into the edges. And
33:10also, I think that companies are starting to understand that, you know, just paying for an API of a generic
33:15generative AI tool is probably not the solution.
33:18to all of their problems. And that's dawning on them. And that's going to change things a little bit as
33:23well.
33:24Great, thank you. We are out of time. But that was really interesting. Thank you. Thank you.
33:29Thank you.
33:30Thank you.
33:31Thank you.
Commentaires