Emmanuel Bacry

Yeah, before answering strictly speaking the question, I would like to comment on some of the things that were said. First of all, I think we are focusing a little too much on LLMs and chatbots. AI is not just that.

So we don't have to forget. I mean, the whole buzz in the whole world is about that, but we should be careful not just, you know, being part of this buzz. It is clearly a big thing.

I'm not saying, I'm not trying to minimize it. And it's a big, I mean, with plenty of applications and it will transform quite a lot of things. But AI is also a lot of other stuff.

So when we are saying about the fact that we need a huge amount of data, this is really about LLMs, a huge amount of compute, it's also about LLM. So we have to be careful a little bit about that and talk also about other AIs that could have huge impacts. I mean, if we talk about health, I know health very well, there are huge impacts on health without LLMs. That's the first thing I wanted to say. Second thing, maybe it's not politically correct what I'm going to say, but what you are saying, Melanie, about China, the fact that, I think it's you who said that, I'm sorry, maybe it's not you, but when the information shaping, the search information, there is one thing that I like in the, I'm being a little bit provocative, OK, but there is one thing that I...

And OK, you are saying that when we are doing search information on DeepSeek, it is clear that there are plenty of problems and so on. But because we know China very well, but it is totally biased what we get with ChatGPT, with Claude, with Gemini and so on, that it is more subtle and nobody knows exactly how it was trained. I don't know which one is the most dangerous.

It is not clear for me. It is an open question. I don't have the answer, but I think both are kind of dangerous and not only the one which is openly biased and we know very well the bias.

So I think we should be also careful about that. And so now I come back to the, and yeah, sorry, one more point. I think one of the very big challenges is also, and you mentioned, I don't remember who, deployment.

In deployment, there is one challenge which is extremely important and very hard, is the evaluation. The evaluation, there are plenty of challenges about evaluation, not even talking about LLMs. On LLMs, it is a nightmare to evaluate. But even if we talk about regular AI, evaluation in real life is a very big problem.

And again, please, I will take an example in health because a lot of things are very clear in health because it is the center of so many problems and so many touches everybody. So when we deploy a medical device that has AI in a hospital, of course it is certified. It is FDA cleared or in Europe it is certified and every country has this certification.

But certification is kind of a still ideal world where you have a closed space and you test specific things. When you deploy in real life, when you deploy in a chain of process where human beings interact with the AI, everything can go wrong, even if the AI is very good. But you have also to evaluate the whole chain of process that will change because you change the one brick, one step by an AI model.

And there are examples in hospitals where some AI were deployed, very good level AI, and where the number of deaths increased dramatically. So there are many challenges about that. Now, sorry, I come to your question now, Jean-Marie.

So your question, sorry, it was about the fact that do we need a lot of data or can we have... Yes, yes. So, yeah, as I said this morning, but nobody is sure.

So maybe some people in this roundtable would say, but it is likely that DeepSeek and Qwen used what we call distillation of the big American models. So, of course, they didn't need as much data and it costs less money because they basically, they used the big models to take all the knowledge of these models. Now, would it have been possible really to do a DeepSeek or Qwen if ChatGPT, Claude and Gemini didn't exist?

It is not that clear. I don't have a clear-cut answer about that. But I want to...

And about the quantity of data, this is necessary for LLMs. This is very clear. It is necessary for LLM. But not at all, not in the same size for many other AIs.

Yeah, but that is different from the question of Jean-Marie, which was the training with the quantity of data. But I totally agree that it seems like they have found a way of running it in a cheaper way. And when you look at the researchers in AI today, everybody is looking for models that are simpler and that do as efficient LLMs as what we have today, but in a simpler manner so that the energy needed to run would be much less.

And there was one announcement, I mean, almost every month there is such an announcement of a big improvement about that.

I thought you had a reaction, No, no, positive reaction. I agree. And I do believe in small models more than the big GPT models, where nobody sees a business model in these big models.

Everybody's losing money with these models. And they are too generic. They are supposed to know everything about everything, which is stupid.

It will never reach this goal. This is very clear. So they keep, you know, they keep adding layers and layers on top of it so that they repair something that they see and they repair something else.

But we should totally change by design these things. And so I do believe absolutely in much smaller models that you train on specific data, local data, specific to a small population and with a purpose, a specific purpose, trying to answer specific questions. This is clearly has a lot of applications and a lot of value.

So I totally agree with Audrey.

Let me rebounce on the rare diseases. Rare diseases are interesting because we are, I mean, we were several to say that we need local systems, because local systems are potentially more neutral than, I don't know if I believe in a neutral AI, I don't think that will ever exist as a neutral human being, then they don't exist, okay, I'm not saying. But okay, we need local models to limit the bias and so on.

Now with the rare disease, we have a problem because if you just use the local data, you don't have enough data. So you have to manage in some way. So data becomes very important, and interoperability, quality of data becomes a major issue.

I do think, as I said this morning, that when we talk about AI revolution, AI revolution is before all a data revolution first. And it's not easy to operate the data revolution, meaning setting up quality data with agile governance, but with privacy, if it's personal data, and interoperability, and so on. It is extremely hard to operate.

So there is a unique initiative in Europe, I don't know if it will work to tell you the truth, but clearly it's a unique initiative on health data. I don't know if you heard about the European Health Data Space. So it's a new regulation that has been voted at the same time as the AI Act, and that tries to homogenize all the rules about governance, access, quality of data.

And so today, I mean in a few years, every data producers, health data producers will be required to share the data for public interest research. Otherwise there will be fines to those who don't want to do that. So we are talking about private data producers and public data producers.

It's a major change, really a major change. I'm at one of these, I'm working at one of these national platforms called Health Data Hub, which is the national French health data platform. We play a major role in this thing.

It's a unique, I think it's really a unique initiative. I hope it will succeed because this data revolution was not solved by any country. You could have companies that are very good at data and so on, but at the national level, nobody solved this problem of data, from what I know.

Okay, maybe I don't know some of the countries. And what is interesting is that since the publication of this regulation, now a lot of countries are looking at Europe, saying, ha ha, it is interesting, there is a new model for health data. Let's look at it because maybe there is something to take.

And now before this regulation, I was going all over the world and trying to set a partnership and now people come to me. So Audrey, actually the Ministry of Health, they came, 28 people came to France to visit us and we are going to sign a partnership. And we do that with Japan, we do that with South Korea, with Israel, with the US also, with Quebec, with so many countries, more and more countries.

And just to mention one pretty original maybe partnership, and it's always about data. The bottleneck for health, for operating AI on health, like partnerships, international partnerships, the main bottleneck, but like way beyond everything else, is the problem of data and finding data that are matching and that you can make them work together. Once you have that, the rest is easy.

It's easy apart from the fact that if you are working transnationally, you have governance problems and you have access problems. So what we are doing, I think it's a nice project, I like very much, with India, and it was announced by Modi and Macron in the last two summits in New Delhi. India has developed an architecture for a securitized space that is so well securized for privacy purposes that you can put it anywhere and they are very confident about the fact that privacy is preserved.

And as a proof of concept, it's open source, it's called DIPA, Data Empowerment Protection Architecture. And as a proof of concept, what they are telling to us, to France, they said, OK, if you take it, if you build such an infrastructure, such a secure space, well, we would be ready to send to you any health personal data from Indian people. So we are building this thing and we are going to run five use cases.

So we are going to take personal health data from India, we are going to bring them in France, put together some French data and doing AI on that. So that's one. I think that's a very important project because around data, governance is a major issue, major issue.

I can comment on a few questions. And maybe I'm not going to answer all the questions, because some I don't know how to answer. But I would like to comment on the first one.

I'm not going to give an answer. I'm sure Melanie or Audrey will have a better, will have an answer. I don't have an answer.

But I just want to warn, to do one warning, because there is something that I see too often, a mistake that is done when we're talking about rare minerals and the whole cost of the supply chain of the AI. We have to be careful today. I mean, that's at least when we're talking about what's happening today, we have to be careful to differentiate the impact of AI and the impact of digital transformation.

Because the fact that we have a phone, this is digital transformation. It is not an impact on AI. So we have the fact that we have cars with electric batteries.

This is not AI. This is digital. And this has a huge impact on ecology and energy and so on and climate.

It has the digital transition. Today, the proportion of this digital transition, the proportion of AI in this digital transition impact is very small. It's a small part.

It is increasing exponentially. So I'm not saying that it's OK and we should not care about that. But I think it is important to say so because we mix everything.

Today, when we talk about artificial intelligence, it's such a buzzword that you include everything in it. And when you do some streaming of a video, this is not AI. The AI is just for recommendation system.

But after that, and you need clouds and stuff like that for that. So this is just a remark on the first question. And for the second question, I just want about the inequalities between the countries.

Yeah, the data is very, very important for that. I mean, as I said this morning, the fear is that we are building basically a Western AI and the most terrible thing that would be to just deploy them everywhere in the world. That would be really, really bad.

And it would generate a lot of inequalities and so on. So we have to be very careful. And of course, there are inequalities between the countries, because even in terms of data, because some countries, a lot of countries haven't started collecting data.

And I think this is the main asset. And I help a French agency, public agency in designing deployment of AI in sub-Saharan African countries, and helping them making their AI. So the main bottleneck, again, is data.

I always say that in Europe, the day where we would have high quality data sets, we would have a huge impact on public health without AI. There are so many things to be done without AI. Then with AI that are on the shelves, ready to be used.

Then with the latest models of AI. But this is very, I think, important. And very, very quickly for the last question, the UN question.

You understood I'm a data person, so I would advise UN to focus on the data, access, governance and stuff like that. I think that's a major. And I know I would have some colleagues at the UN doing the regular lobbying for other matters.