Anne Currie is the co-author of the acclaimed O’Reilly book “Building Green Software”, a pillar of the GSF, a veteran in the Cloud Industry and also a SF novelist with her series of panopticon books. 

Preparing her forthcoming keynote at Green IO London, she went all the way down into the rabbit hole of AI and energy efficiency. She investigated from OpenAi to DeepSeek and open source models, what a software developer using these models can and cannot do to reduce energy consumption, and so on. 

In the first part of this episode, Anne Currie et Gaël Duez discussed:

And much more!

❤️ Subscribe, follow, like, ... stay connected the way you want to never miss an episode, twice a month, on Tuesday!

📧 Once a month, you get carefully curated news on digital sustainability packed with exclusive Green IO contents, subscribe to the Green IO newsletter here. 

📣 Green IO next Conference is in London on September 23rd and 24th. Every Green IO listener can get a free ticket using the voucher GREENIOVIP. A small gift for your huge support. 🎁 

📧 You can also send us an email at [email protected] to share your feedback and suggest future guests or topics.   

Anne (00:01)

They realized was that the more compute they poured into training and running an LLM, the better he got, and it just kept getting better. So people were expecting we'll we'll pour in energy. And at some point we'll, get diminishing returns. You know, we'll go, well, that's the point where we, where we stop building the data centers, but they didn't discover that. They discovered that it just keep going up and up and up

Gaël Duez (00:19)

It will plateau.

Gaël Duez (00:28)

Hello everyone, welcome to Green IO! I'm Gael Duez and in this podcast, we empower responsible technologists to build a greener digital world, one bite at a time. Twice a month on a Tuesday, guests from across the globe share insights, tools and alternative approaches, enabling people within the tech sector and beyond to boost digital sustainability.

Do I really need to introduce Anne Currie to you? She's the co-author of the acclaimed O'Reilly book Building Green Software, a pillar of the Green Software Foundation, a veteran in the cloud industry, and also a science fiction novelist with a series of Panopticon books, which I didn't read yet, but I intend to catch up before Christmas. And we will have the honor of having her as a keynote speaker at Green IO in London on September 24th, where she will talk about AI and its energy consumption. And there's something to know about Anne. She's a thorough speaker. When you give her a bone, she will go all the way down into the rabbit hole. This is exactly what happened on this topic of AI and efficiency. She investigated from OpenAI to DeepSeek and open source models investigated what a software developer using these models can and cannot do to reduce energy consumption and so on and so on. And I realized it wouldn't be fair for the thousands of listeners of the show to be deprived of these insights. Hence this episode on a much complicated topic, AI and efficiency. Hi Anne and welcome back on the show because you're now part of the small club of guests who have been due time on the Green IO podcast. Congrats!

Anne (02:16)

Well, thank you very much. It's great to be back. And I'm really looking forward to green IO should be an excellent conference. It looks like an amazing lineup.

Gaël Duez (02:27)

I mean, you're part of it, so it must be. Listen, Anne, before we deep dive into this AI and efficiency discussion, I was listening to the Environment Viable episode, which you hosted with Sean Varley from the chip manufacturer, Ampere. It was super interesting to hear the discussion about which developers are actually able to leverage the true potential of a chip or all the hardware used to run the software on. Sean made a point that the hyper-specialized engineers building the serverless tools are more able to achieve significant results than a regular developer who has to deal with many business and technical requirements, the time constraint and so on and so on. And I guessed that it was… sort of music to your ears because you advocated during your discussion with him, but also in this podcast previously and in your book that building efficient software is actually super hard. So before we deep dive into AI and efficiency, could you elaborate a bit on this constant stance of yours?

Anne (03:39)

Yes, so there are lots and lots of ways to cut your energy use in your systems, but I always say that for most enterprises, don't think about code efficiency. It's incredibly hard. It's a very, very specialized skill. There's tons of low-hanging fruit in operational efficiency. adopting modern ops practices will help you be more secure, will help you be more resilient not you're having just getting rid of the waste in your systems, turning things off when they're not in use or when you don't use need them anymore, getting rid of your zombie services. That generally will cut your energy use in half whilst making your systems more secure and more ⁓ and a lot cheaper to run half usually half as much to run. So that's for enterprises. I always say that look first there because the changes to use the less electricity are completely aligned with other business needs that you have. So it's worth it is worth a little bit of investment to cut your hosting costs in half and also to get a more resilient system. ⁓ rewriting your systems to be more efficient, that is a really specialist skill and it is not worth paying the kind of people who have the ability to do that in most enterprises. So what you really need to do there, this is where I was very aligned with Seth's statements which you're much better off, if you want efficient software, buy it. Buy it off the shelf, buy it managed, ideally already managed. If you're using a hyperscaler, buy a managed service off that. will also often, which is where obviously Seth wants to be selling his chips to those same people. So enterprises never need to get involved. They sell to the hyperscalers, the enterprises buy that chip from the hyperscaler wrapped in a managed service that might be serverless, might be something else. ⁓ And that's kind of that means everybody's aligned. Otherwise you end up with it's really hard. Many years ago, when I started my career, I was involved in writing high performance software in C, not because I was some super person doing that stuff, just because we had to because all the servers were really bad back then 30 years ago. And therefore you had to write really efficient software to get everything to run at all. And it took ages. These days, most enterprises, if you attempted to do that kind of work, you would go out of business. There are some ways that it can happen. And that is a little bit about what we're talking about today. ⁓ Some ways that enterprises can be involved or even individuals can be involved, but not… through your enterprise writing your own custom system. It's half the story of DeepSeek, which you asked me to look into a couple of months ago. So we come full circle on that. yes, my key thing is don't try to do something that isn't realistically going to happen because there are lots of things that every enterprise absolutely should be doing in terms of adopting modern ops best practice that will save loads of, realistically it's going to save you more electricity use than anything else you do.

Gaël Duez (07:01)

Thanks a lot for clarifying that yet there are many, many, many things to be done before writing efficient code. Maybe we'll come back to this point later because doesn't mean that especially if you write code for the very first time, there are not some best practices or things to consider. But now that you've mentioned DeepSeek and that was actually something that I asked you to investigate in preparation of this episode several months ago. Maybe you could explain what DeepSeek did and didn't do to achieve its so-called energy efficiency, which is to take with a pinch of salt because once again, it really depends a lot on under which angles you analyze this. So could you enlighten us a bit on this?

Anne (07:43)

Well, I'm actually going to roll you back because when you asked me to look into this, really needed to say, what's the context that DeepSeek, so DeepSeek, it's a new model that came from China in January. And it really caused an enormous, enormous amount of interest because they said that they'd got 10x reductions in running costs.

Gaël Duez (07:52)

Yeah, sure.

Anne (08:06)

Which actually is pretty, is very believable. That's not very much in terms of how much do we expect, especially from a code efficiency change, but it caused enormous amounts of interest. And I think that the interesting thing there for me, when I spoke to a lot of people in the industry is, especially, you know, I'd go out and talk to deep mind engineers and say, what do you think of this? Has it achieved this? Why is it so interesting? They go shrug their shoulders. And say, don't understand why people are so interested. ⁓ 10X is not that much. Maybe they've achieved it. have no interest. It hasn't done anything exciting. There's nothing new that comes out of DeepSec, but it has caused enormous interest. It is enormously interesting. And I think for folks to understand where that interest comes from, we need to roll back a little bit. And the first place I want to start is something I said on stage. Last year I was at Green IO, I was in the panel. ⁓ And ⁓ I said at the time that there were two things that we absolutely knew were coming. ⁓ One was climate change and the other one was AI. And there's no point in railing against either of those things. They are coming. They are… But the good news is they are the two revolutions that are happening at the moment, which will actually change our lives very positively as well. So, and I don't think that we can look at this or think about this or think about the solutions to our green problems without acknowledging the fact that we are in the middle of a green energy, a renewable energy transition, which is amazing. It's a revolution. Last year in 2022, 92 % of new investment in energy production went into renewables. mean, do not underestimate that. That is amazing. That is not the situation we ever anticipated 10 years ago. Things have really, really improved. And we've been part of that because we've been making noise and pushing for it. But we have achieved there. The energy transition is coming. I if you look at China, last year, again in 2024, they built the equivalent of one nuclear power station every day, just for renewables, mostly solar, but also wind. Again, we would not have expected that a year ago, 10 years ago. And even a year ago, we wouldn't necessarily have expected. So we are in the context that we are in the middle of a green revolution.

Gaël Duez (10:37)

Yeah.

Anne (10:42)

We don't talk about it all the time, but we don't hear it on the news, but it is going on and it should be something that humanists talks about all the time. The other thing we're going through is an AI revolution. And what I'm gonna talk to you a little bit, because we know a little bit about that kind of like solar is going, wind is going. We understand that, but AI, to understand AI, I think I needed to roll back quite a few years to understand where it came from and why a whole series of really important changes have happened over the past eight years, where they came from, what effect they have on the other revolution. Because the win here is not to kill AI because I don't think we want to. It's great. There's lots of amazing use cases for it. But even if we did, we wouldn't be able to. It's rolling. ⁓ There's no point in fighting against that. What we need to do is align the AI revolution and the renewable energy transition. We need to make sure that the AI, which will be an enormous energy user, and that's what I'll be talking to you a little bit about today, the history of it why it's going to be an enormous energy user. We need to make sure that the energy users is renewable energy, not fossil fuel energy. And that is where we need to absolutely be focusing all our attention, I would say. So I don't want to, I could keep going and talk for the next half an hour, but I want to get your thoughts on this.

Gaël Duez (12:16)

I think the two revolutions are going on for sure. I usually tend to zoom out a bit and consider the energy transition as just part of a more global ecological transition because we know that if we don't incorporate other very precious resources such as metals or even the biosphere, we might sometimes take the wrong decision when it comes to driving the energy transition while not considering any of the other impacts it might have say beyond the scope of this discussion. However, I'm with you on this one that yes, AI seems to be a massive energy consumer. I still didn't make up my mind if it's because of a wasteful behavior coming from the US where actually energy is cheap and we shouldn't focus that much about it. or if it is really that it will always require an enormous amount of energy to get trained and even more important to get consumed by the inference phase. So I didn't really make up my mind and that's actually something that I wanted to ask you when you investigated DeepSeek because we had tons of discussion like yeah it's the training costs less which is we're talking about several dozens of millions of dollars to train a model like GPT-4 or GPT-5. So it's also significant money that can be saved. And that's also some sort of a social justice issue because if you have only top corporations around the world or top universities having access to this technology that… should also question how the rest of humankind can use or not use this technology. But that being said, a lot of the discussion also happened that, yeah, DeepSeek might be most energy efficient, which was, according to you, not a big deal for most of the AI engineers. But on the other end, there've been also a lot of discussion like, yeah, but the inference phase, when you request something from DeepSeek, it was actually super energy intensive. And that's why I asked you and your detective skills to investigate because I really wanted to get more mature numbers and a better understanding. So I guess the floor is yours now.

Anne (14:28)

Yeah, so, oddly enough, those questions are the ones that I've been trying to answer for you when you asked me to answer them some months ago. Just completely destroyed my life with this question, which is an excellent question. So, what I'm going to do is take you a little bit back to give you a bit of context. What was the situation that DeepSeat came into? And I think it's really important that we do talk about this.

And also I find it very interesting and it does give you a little bit of an insight into the minds of the hyperscalers, the AI designers, where they're coming from on this. in 2017, a paper came out called Attention is All You Need, which you may have heard of. And it was a paper that came out of Google Brain, the teams at Google at the time. And they said, you know, we've just discovered this really cool a new algorithm called, a way of doing AI called Transformers. And it's really good because it was, oddly enough, was an efficiency play. It cut the energy requirements by about 10x. And it was just better. It allowed you to take in more context when you were doing, it was about translation and chatbots were mostly about kind of English to French, that kind of thing.

And ⁓ it was almost 10, it's 10 times cheaper. ⁓ But it all it'd be but you know, there are all we're finding all kinds of interesting things about it is much the algorithm is innately more parallelizable. So it used to be the AI used to have to run quite sequentially. So to get ages and ages to do the training and everybody got really bored. And AI engineers don't have an awful lot of it, you know, they are God this takes ages. But this one because it was parallelizable, you could you could

take it into bits and you could run it simultaneously on GPUs. So they said, well, you know, there are these fancy GPUs from Nvidia done for gaming. We could use those and we could run this much faster using GPUs. And we could really scale it up because, you know, we can just scale this. We can now scale this horizontally rather, you know, with these GPUs. And everyone went, oh, that's great. And when they wrote the paper, they said, we're seeing that it seems to be working better as well. It seems to be producing better results, better translations, better chat bot chat, that kind of thing. But in the beginning, it was mostly about, this is quite cool because we can do it faster and we can do it more efficiently. And this caused a kind of big split in the AI industry in the 2017, 2018 that open AI just went.

This is so much better than everything else out there. We're going to go all in on transformers. And LLMs are based on this transformer technology. So all the chat GPT and all that kind of stuff, based on the idea of transformers. Google said, well, no, we're having lots of success with other forms of algorithms. So we aren't going to go all in on LLMs. We're just going have a little bit, know, hardly any look at that whatsoever. So AI just went gangbusters on this. And they decided the best way to try this out was to build chat GPT, a chatbot based on LLMs. And they built chat GPT one, which nobody really saw, and chat GPT two, which they trained with a whole load of interesting data of dubious. Apparently they trained it's an awful lot of fanfiction, which is interesting because fanfiction.

is not Creative Commons, it's Creative Commons, there's no IP usually when it's released. I think it really gave gave Chat GPT that kind of early, kind of friendly tone it had. Anyway, so it did a whole load of stuff that they trained Chat GPT on all this fan fiction. But between Chat GPT 2 and Chat GPT 3, and Chat GPT 3 was the one that was released and we all saw. And this is the key moment they suddenly realized something really quite that nobody was expecting. Something that they coined as the AI scaling laws. Now, if there are any physicists watching, I was a physicist, it's not a law, it's not a law of nature, it's more like a Moore's law. It's the kind of thing you call the so you can get ⁓ investors to give you money, like Moore's law.

Gaël Duez (19:02)

Yeah, humans made low.

Anne (19:10)

It's very

Gaël Duez (19:10)

Is that the best definition of Moore's law that I've ever heard? This is a law made by humans to get money from investors. Okay, that's sort of low. Okay, I got it. ⁓

Anne (19:18)

Absolutely. It's exactly that sort of law. But it is also industrially and experimentally interesting and useful. What they realized is once they had LLMs and chat GPT that could run on GPUs, and you could scale out those GPUs horizontally, they started to build bigger and bigger data centers and pour more more and more energy into them. What they realized was that the more compute, they poured into training and running and LLM, the better he got, and it just kept getting better. So people were expecting we'll we'll pour in energy. And at some point we'll, get diminishing returns. You know, that's the point where we stop building the data centers, but they didn't discover that. They discovered that it just keep going up and up and up in the bed. And the, so the definition of better here is because my husband's a physicist and he said what does better mean? Is that the definition of better here is basically better on all the metrics they had. So they've got kind of artificial metrics about how accurate the results are, ⁓ something called perplexity which is how when they get to the end of calculating something how sure are they that they got to the right result. The perplexity got lower that they were more confident of the right result but also the kind of subjective measures like going in and asking somebody is this better than it was, know, is it more readable? it does it seem like a better translation or a better chat you're having? All of those measures people go, yeah, it's better. It's better. It's better. It's better. It's better. And this was and this is this should all strike us all with horror. Because this is the issue. This is what's caused all the problems coming. Well, not the problems, but you know, the advantages, but also the risks here that the amount of compute and that was ⁓ number of parameters per model. So if you made the model bigger, it got better. The amount of data that was required to train the model, good data required to train the model and the number of passes that it made through the data. And you combine all of those together and they all have to go up in together to keep on this line of it just keeps getting better, this AI scaling law ⁓ is that you write all of those together and it just keeps getting better. the thing is the AI scaling law is about AI is really very, very impressed, incredibly impressive and we don't know how impressive it's going to become if you scale it.

And then of course you have the, you know, that is music to the ears of hyperscalers. Because I think one of the things we have to remember here is hyperscalers gonna hyperscale. It is an advantage that they have that they can keep hyperscaling and therefore they will. So at this point, and this was 2020 that this rule was first experimentally proven that so far we had not seen any tailing off of the effects of just pouring more and more and more energy into these AI identity data centers. So 2020 and then things kept going on from there. And at that point, the only strategy in existence force, and mostly driven by America. They have all the energy, they have money,

Gaël Duez (22:58)

Yeah, more energy.

Anne (23:03)

So at this point, they'd realized this, this AI scaling law, but it wasn't really widely appreciated. Investors weren't quite so interested in it, hyperscalers weren't quite so interested in it. But then in late 2023, ChatGPT-3, came out and that was the one where everybody went chat GPT mad and started talking to it and there was it was just wall to wall news current coverage. Everything in the world was going on about ChatGPT-3 and suddenly all the hyperscalers really woke up to the fact that this this was something they had in their hands that was in a that had some degree of product market fit. People wanted to talk to chat. You see they wanted to use it. They wanted to use it for translations. They would use it for writing reports, wanted to use it for writing their novels, they wanted to use it for all kinds of things. Suddenly it had established, late 2023, it was established as a thing, a giant thing. ⁓ And I will show you the actual charts of what happened then when I'm doing my onstage talk. But hyper scaler spending on data centers doubled pretty much overnight between 2023 and 2024, hyperscalers doubled the investments they were doing in building data centers. And that was pretty much all data centers for AI, because they didn't really need the additional data centers for all the other stuff they're doing. They needed it for AI because they were attempting to follow the AI scaling law. They knew that if they didn't build it, well, they were in a race then. Everybody was thinking, well, if I don't build the cleverest AI, then one of my competitors is going to use their hyperscaling ability to ⁓ build that super AI. And through 2024, this was just the entire story. was just bigger and better, bigger and better, bigger and better. And at that point, you kind of think, well, this is going to blast away the energy revolution, the green energy revolution. We can't align this stuff because there's no limit to the end of what they're going to need.

And then, and then in early 20th, in January, and then coming into that situation, which was a dire looking situation, much worse than I'd realized, to be honest, because when I was looking at it, was thinking, oh, hang on a minute, everybody's, you know, there is only one strategy here, which is scale, scale, scale. In 2024, at the beginning of 2025, DeepSeek came in, it was launched.

Gaël Duez (25:21)

Eventually...

Anne (25:47)

And it said, yeah, we've just done some little minor tweaks and we've cut energy use tenfold. And that's, and everybody went bananas about that. And the interesting thing is why? Why did everybody go bananas over DeepSeek? Because ⁓ DeepSeek didn't do anything amazing. As far as all the AI engineers were saying, well, you know, it doesn't really enable us to do anything new. It's not a new piece of technology. It's not a new idea. It's not coming out. There's nothing, you know there was a reason why they were all shrugging going well, even tenfold, know, we're looking like it's growing like that tenfold is a blip. We're hardly going to even notice that as we drive over it. ⁓ But I think DeepSeek was incredibly, incredibly important. And the reason why it was important was because it was because of the interest, the interest was the interest in some ways.

Gaël Duez (26:52)

So what you're saying is that the focus on DeepSeek was more important than DeepSeek itself because DeepSeek put on the table the most overlooked question which was about this brute force strategy which was undermining the energy transition. Am I correct?

Anne (27:12)

Well, yes, and no. It caused the industry to diverge into one brute force and one less brute force. But so I think what DeepSeek did was it demonstrated that there was interest in AI as good as it was at the end of 2024.

Gaël Duez (27:21)

Okay.

Anne (27:32)

only cheaper because the hyperscalers had just said, well, no, it's not good enough yet because we can just keep going forever here. We are. That's what we're going to do. We're going to keep going. When DeepSeat came on the scene, partly motivated by the fact that China didn't have access to these really flashy Nvidia chips anymore, it said, actually, don't have it. I think it's good enough now. I think it's good enough now for lots of use cases but we can't use it in those use cases because it's too expensive. So what DeepSeek did was it kind of said, look, this is a cheaper one. Can you do anything with it? And it didn't do anything new. What it did was it used three techniques. At the first version, I think it used about three techniques that weren't quite well known, which is again why all the deep mind engineers were going, well, there's something exciting there. Eight bits ⁓ precision.

instead of 64 bits in places where only eight bit was required. So just being a bit more ⁓ efficient in not over provisioning on bits. That was a little bit of the chain to me. That was actually probably the biggest ⁓ reduction in costs. Something called mixture of experts, which was basically saying, well, we're all gonna have these monolithic ⁓ models that do everything. Why don't we kind of break them off a bit like microservices versus model? And we say, well, that model is expert in this, that model is expert in this. And we'll just route your question to the model that's the smaller model that's more specific to that. So that was another part of it. And the final thing was they started to say, well, actually, APIs for talking to the hardware, for talking to the Nvidia chips, CUDA, not very efficient and not very well done, we could do those better. So they started, and none of those were new ideas, but they just said, look, we're just doing all of this. But the interesting thing was that it caused so much attention because you had a lot of enterprises and vendors who said, you know, actually this is something I need because AI was good enough for me, but it was too heavyweight for me to put in my headset or my phone or my.

What should it was just to have what I needed was AI that was as good as it was only 10 X cheaper. And so suddenly there was a whole, it demonstrated a whole load of people going, I'll, I'll take that. Thanks very much. I'll take this new DeepSeek model. But there was something about, so, so DeepSeek did one thing, which was demonstrates desire from customers for the functionality that was already in existence investors suddenly switched their focus from AI scaling laws to demonstrated customer demand for AI as good as it was. ⁓

Gaël Duez (30:30)

And for the sake of clarification here, you mentioned several times that DeepSeek was 10x more efficient, but are we talking about efficiency during the training phase? when you mentioned users, they enterprise users building their own models and needing much cheaper? Computing costs, energy costs, etc. Or are we talking about the end user, you and I, because you mentioned running on the watch, etc. And is it more the inference phase just for the sake of clarification, are we talking about the building phase, the running phase or both phases which were 10x more energy efficient.

Anne (31:11)

Well, at the time, actually, I have no idea at the time, because everything's moved on since then. 10x was the number. I think it was training at the time, but it can't just be so the interest was kind of people going, I want something cheaper, make something cheaper, make something cheaper so I can put it in my, and that means you need to be cheaper to train and you need to be cheaper to run because people, I one of their first customers was a headset customer who said,

Gaël Duez (31:20)

I think so.

Anne (31:37)

Look, I need to be able to run this in my headset, which means it has to be cheaper for inference. And here comes the other thing that DeepSeek did that was really very revolutionary, which is that they said, we're going to open this up. The work we're doing, we're going to make it open. We're going to make it open under an MIT license. So oddly enough, even though it's a Chinese company, classic American license, the MIT license And the MIT license, mean, so the Deep Six stuff was built on top of ⁓ the open source llama from ⁓ Meta and also Quen, which was open source from Alibaba. But they also open sourced all of their stuff that sat on top of it.

So, DeepSeek, released everything under this permissive license and said, fully boot, go and make it better. And that is what I think that enterprises, really, really loved. Because we all know that there's something called rights and another law. And this is like Moore's law. And in fact, Moore's law is a type of ⁓ rights law. Wright's law says that if you make more of something, this is all about building things and ⁓ shipping them to customers and running them and learning from them. You learn, the more you build, the more you learn and the more efficient you get at it. Oddly enough, Jeff's paradox is also a bit of an expression of rights law. But the more you build something, the more you do it, the better and better you get at it. And the one thing that investors like is rights law because that is about saying, there's something where there's an enormous demand. And if I meet that demand, I can just keep getting better and better and better and better and better assets. ⁓ And so ⁓ the by combining the fact that there was clear interest in chapter BT four level of functionality, but cheaper and also saying, and I'm going to use open source, which is an amazing tool for making things better and getting a whole load of people all over the world involved in improving this. So in some ways it doesn't really matter what happened when DeepSeek, because DeepSeek was just about saying, look, you can start here. You can start getting better at this and there will be a market for it. So in the time since January, and it was only January this year, there's been a lot of movement here.

So in August, DeepSeek came out with another model in which they said that they had got the costs for running and this is inference, the costs for writing a line of code using the DeepSeek model versus a proprietary model like ChatGPT, let's just constantly say ChatGPT, was 70 times less. It was $1 versus for approximately $70. And that was all because it's open source and people keep improving it. If there's something that folks care about and there's an open source element to it, they can keep tweaking it and improving it. So the win for DeepSeek and the reason why it was interesting was because it demonstrated that to a certain extent, product market fit had been the met people were happy enough with what was available in January 2025. They just wanted it to be cheaper. And we all know that, and especially if you combine that with open source, it gets cheaper.

Gaël Duez (35:23)

Following you on this one, there are two signals, not only one, the first very interesting signal that you mentioned is, well, we had reached a moment where people were happy enough with the capacities of generating AI and they now would look for cheaper solution. The second signal, which is super interesting as well, is that the maturity of the AI and let's be a bit more precise, GenAI ecosystem was good enough for people to embrace open source. You have enough people ready to participate enough enterprise confident enough to jump on an open source solution. So I think the DeepSeek moment, which is always, as you mentioned, something that happened only beginning of this year, signal the ecosystem because it goes even beyond the market two times. Energy matters, price matters, so enough, at least partially enough with this AI scaling law and the brute force philosophy that comes with it. And the second one was, hey, we are now ready for open source. that's a pivotal moment as well. If I'm not

Anne (36:36)

I think you got 100 % that it was two things. because in and of itself, those changes were not that impressive. They weren't a lot of people were saying, well, is it even tenfold? I think it might only be fivefold or fourfold. And they say the DeepMind engineers last place to going, yeah, there's nothing exciting there. It was product market fit, or a degree of product market fit. and the fact that they open sourced it at the same time. And everybody's going, yeah, I just want it cheap. I want that but cheaper. Thank you very much.

Gaël Duez (37:02)

Because I connect this to the astonishing numbers that were shared by Sasha Luccioni at Green IO New York. And that was actually not something related to Sustainability, she shared that on her face today on the platform. You've got more than one million models. One million. I was shocked. And almost all of all of them are open source so we've seen a blossoming of models and maybe this is something also that that we should talk about to relate to efficiency

Anne (37:33)

Actually, it's all efficiency is all about this. You get a whole load of eyes on it and a whole load of people say most enterprises don't have those skills to do all that tweaking and tuning and stuff. But once you've got once you expand it out to the whole world, you find people who do have those skills. And if they if they do it on an open source model, and then and then all that work gets shared, you know, you do it one incredibly clever person does something and then thousands and tens of thousands and millions of people get to use it. That's, that is how code efficiency works. That's how, that's how it's, that's how it starts playing off. So on the one hand, we've got this amazing, ⁓ we now in a, a renaissance of efficiency. ⁓ If you look at the things I mentioned, the three things that DeepSeek started doing, ⁓ that were not new things, there'd been papers around them for years and years, the bit efficiency. Now, I think they went from 64-bit to 8-bit in some of their stuff. We've now got open source stuff that's taking it down to 4-bit, to even 1.5-bit. How on do you do that? Because people are going, oh, do you know, I could do something with this. You've got whole world of people tinkering and making it better.

It's been an extraordinary year thus far. But there are two major, major problems that this doesn't address or ⁓ issues associated with this. DeepSeek has not solved the problem. There are two things that we need to be thinking about. The first is that efficiency will mean that we just use a lot more of it, which doesn't necessarily mean there's a problem. ⁓ But it doesn't necessarily mean that in and of itself on its own is going to solve the problem. And that's the usual thing that always comes up, Jevons paradox. Now, I'm a big believer that Jevons paradox could be fixed and Jevons paradox is the definition of economic growth in many ways. But ⁓ I think there's a different problem here that will be solved in a different way, which is to go back to what I said before, hyperscalers gonna hyperscale.

Gaël Duez (39:55)

Thank you for listening to this Green IO episode. Anne will be back next week, and she will talk about the hyperscaler strategy regarding artificial general intelligence, the East coast, West coast, China's strategy, and the right questions to ask for any developers or leaders. Having to choose an AI model. Stay tuned and, accessible and transparent information is in the DNA of Green IO, all the references mentioned in this episode, as well as the full transcript, are in the share notes. You can find these notes on your favorite podcast platform and, of course, on the website greenio.tech. If you enjoyed this episode, please give us a thumbs up on YouTube. Or rate the podcast five stars on Apple Podcasts or Spotify. Sharing this episode on social media or directly with relatives working with AI is a great move to provide them with insights on this hot topic. You've got the point. Being an independent media, we rely mostly on you to get more responsible technologists on board.

By the way, Green IO is a podcast and much more. So visit greenio.tech to subscribe to our free monthly newsletter and check the conferences we organize across the globe. As we already discussed in this episode, Green IO London is almost there. September 23rd and 24th are the dates. As a Green IO listener, can get a free ticket to any Green IO conferences using the voucher GREENIOVIP Just make sure to have one before the 30 free tickets are all gone.

I'm looking forward to meeting you there to help you, fellow responsible technologists, build a greener digital world.