![Génération Do It Yourself - [SNIPPET] Andy Yen - Who will win the AI race?](https://podcasts.happyscribe.com/rails/active_storage/representations/proxy/eyJfcmFpbHMiOnsiZGF0YSI6NzAwMDksInB1ciI6ImJsb2JfaWQifX0=--cfafdd5ee8e8d11846cfef0d07b5e1ccb3c006bb/eyJfcmFpbHMiOnsiZGF0YSI6eyJmb3JtYXQiOiJ3ZWJwIiwicmVzaXplX3RvX2ZpbGwiOlszODQsMzg0XSwic2F2ZXIiOnsicXVhbGl0eSI6NzgsInN0cmlwIjp0cnVlfX0sInB1ciI6InZhcmlhdGlvbiJ9fQ==--6548c1d9ae417f005796959830b9dc128fe7a7e1/517cf746-773f-429d-a6dc-32de3310e20d.jpg?locale=fr)
Cet épisode revient sur la course mondiale à l'intelligence artificielle, entre les États-Unis, la Chine et l'Europe. L'invité, Andy Yen, défend l'idée que la stratégie américaine de concentration des moyens et la concurrence chinoise dessinent deux trajectoires différentes, et il estime que les modèles ouverts aujourd'hui peu coûteux pourraient disparaître vers 2030.
Les échanges abordent aussi la question d'un éventuel plateau des performances, le rôle des pouvoirs publics et l'exemple du CERN, qu'il propose de reproduire pour l'IA à l'échelle européenne. Les auditeurs découvriront les arguments avancés par l'invité ainsi que les objections soulevées par l'animateur.
Dans cet épisode Matthieu Stefani, Andy Yen
Cette transcription de l'épisode sur la course à l'IA avec Andy Yen a été réalisée avec la technologie HappyScribe, utilisée pour convertir un fichier audio en texte.
Les publicités de cet audio diffèrent du texte. Touchez la phrase que vous entendez pour suivre.
Synchronisé : le texte suit maintenant l'audio.
We are the average of the people we spend the most time with. My name is Mathieu Stéphanie, and I invite entrepreneurs, athletes, and artists to raise my average and yours. Just before we begin this episode, I'll switch to French for a brief moment to thank my sponsors.
Well, I think people sometimes ask me who I think is going to win the AI race. I don't know if you have an opinion as well, but— and also what can Europe do in AI, right? There's— and there's some controversy around Mistral seeming to pull back from AI and, you know, a bit losing the race.
But the race, I would say Nvidia.
Well, okay, depends on what race, right?
Yeah, yeah, but I think at the end, the one who's going to win the AI race is the— the—
The—
The—
The—
The—
The—
The—
Comment sur le marchand de pelles, de pioche, de—
Yeah, yeah, yeah, I know what you mean.
Yeah. You know, I actually did the math, and something like half of our AI money revenue goes directly to Nvidia.
Yeah.
It's— it's— it's actually mind-boggling. But I don't think that situation will last. I think they have a maybe 2-year lead, but eventually other chip makers will enter that space. It's too lucrative not to have competition.
AMD is going— is getting much better, I've heard, right?
I don't think it's going to be them. I think it'll be one of the startups. AI, we can say for conclusively that Europe has lost the race, unfortunately. Now, it doesn't mean we can't catch up in the future, but today, you know, we have lost. So then it's a question of, is it the Americans or Chinese that are going to win? And the way to think about this question is to analyze the American approach and then the Chinese approach.
The American approach is, I'm going to take billions— no, maybe trillions— of dollars, and I'm going to funnel it into 3 or 4 of my biggest companies, and I'm going to hope they can win. That's the American approach. The Chinese approach is, I'm going to fund 50 companies, 50, 60 companies, and then they are going to fight each other to the literal death.
Yeah, that's what they did in the— in the EV, electric vehicles, or solar panels, or—
Yes, exactly.
Yeah, many industries.
Yes. And at the end of this bloodbath, I'm going to have 3 survivors who are going to be my champion.
Yeah.
If you consider these two approaches, the Chinese one is far more— is far more effective at innovation, execution speed, and velocity. The American one actually is going to be slower. So I think the mistake that the U.S. made is they didn't do antitrust. They didn't break up the tech companies. And the Chinese did do that. And this is a structural difference that, in my opinion, gives the Chinese a quite strong advantage in the AI race today, and very probably puts them in a position to win.
So this is, I think, you know— and the problem with Europe is, once the Chinese have won, the open weight models that we rely upon today to power most of our infrastructure is going to be gone. So the cost of AI is going to skyrocket when we're forced to pay much inflated, you know, API prices to the Americans and the Chinese. Because the lower cost situation that we have today is an artificial situation created by the dog-eat-dog, you know, vicious competitive cycle of the, you know, Chinese tech industry.
But once that phase is passed, we're very screwed. And my prediction is it probably happens around 2030. It's kind of my guess.
2030 times more expensive than we're expecting.
No, no, 2030. So in about 3 years from now, it's probably when that transition— when that phase of Chinese competition finishes.
Okay. 2030. Okay, sorry.
So— so if we don't get our act together in the next 3 years, I think Europe is pretty screwed.
All right. Interesting. Scary. The thing is that some people ask themselves that. I mean, where do we get the money from? From the models? I mean, people say here, on my mic, and everywhere, that the models are more or less the same now. I mean, they're getting to, like, a plateau somehow. I mean, well, if you had said that a year ago, you would have been very wrong. And probably within a year we're going to say we were— I was very wrong.
Yeah.
But they have, like, 3, 4, 5 months difference. I mean, it takes them, you know, ChatGPT was nowhere in April, and now they're back.
Yes.
And a year ago Claude was nowhere, and— and so.
Plateau is the wrong word.
Yeah.
It's not that they are plateauing and that improvements are getting less and less. It's that the competitors are very close together. It's a very tight pack, but it's still improving quite rapidly. So it's not a plateau. At least— at least not yet. Now, it will eventually plateau. It's inevitable. But we are not really at that phase yet with the current technology.
But then, I mean, is there a world where we can benefit that the fact that people are at more or less at the same place? How can we get our act together?
Yeah.
I mean, the question you— I mean, can we—
Yes.
Get back in the game? And how can we get back in the game?
Well, we benefit today by having Chinese models available for cheap.
For free?
For free even.
Yeah.
And why are Chinese models doing this?
Because they want us not to pay American models, right?
No, no, no. It's— it's not like a centralized, you know, it's not like there was a Communist Party meeting where they sat together and said, "Oh, let's dump on the Americans and kill their pricing power." That's not the thinking. Imagine you're in a space where there's 50 companies competing, and you know only 3 will survive in the next 3 years. And you want to be one of the survivors. What you need—
Well, there's one way you're the best, or the other way you're the cheapest. Or if you're in the mid—
No, no, no.
Yeah?
Not even the best or the cheapest. What you need is you need users.
Yeah.
And you need to become famous. The way to be famous and known and to get lots of users is to open your model, to make it open source so that anybody can use it. That is how you get attention in an ecosystem where there's 50 players, and you stand out to secure funding, resources, mindshare, and attention to survive. So open weight models today, it's not a strange strategy. It is simply a result— it's a consequence of the current phase of their competitive dynamics in their internal market.
But this is going to change in the next, you know, 3, 4 years, because these labs begin to consolidate. The weaker ones go to irrelevance, run out of funding. There will be, like in every other Chinese market at the end, probably 3, 4, 5 big players. And at that point, the incentive to open everything, that's going to disappear. So I think open models, open weight models, probably we have another 3 years, possibly. And then— and then that's over. And then Europe, we're left paying inflated, you know, API rates to the remaining ones who are now going to have to make money.
Wow.
Yeah, it's a— it's a bit scary. So what can we do? Well, I go back to history. I'm not a historian, but I'm— maybe this is the Chinese part of me. It's a very keen, you know, study of history. This is not a new challenge for Europe to be caught between two superpowers on a new emerging technology that is very difficult to do, that we don't find the ability to compete from the private sector or even from any, you know, country putting resources behind it.
The last time this happened was actually in the late 40s, early 50s, around nuclear. In the nuclear age, and France was, you know, quite good at this. You had the Soviet Union and the U.S., the two nuclear superpowers, and nuclear was unlimited free energy that would change society forever. That was— that was the pitch. What did the Europeans do in the 50s? Well, they created CERN in Geneva.
That was a pan-European effort to pull our collective resources and intelligence that no single member state was able to do on its own into one centralized organization to become a global leader in nuclear and— and this new, you know, field that was opening.
And you have to be careful what you're saying, because, I mean, you're going back to that Zoom protocol you were talking about earlier on. I mean, CERN is a public research.
Yes.
You think we have to go on public research to finance a public research for AI?
I think what needs to be done is we have to create a CERN of AI.
All right.
And there are— and by the way, there are no frontier labs in Europe that we would be competing with. It's not like, you know, this CERN is going to put 3 or 4 other businesses out of business. It's not like Zoom, you know? In the Zoom case, there's companies like Proton, there's dozens of European companies doing that, right? But nobody's doing frontier AI. It doesn't exist anymore. The industry has been decimated. It's been killed off. Mistral tried, and even they would admit that they were not at the frontier anymore.
They're not working on models anymore?
No, no, not really. So we need, I think, a CERN of AI.
What about Ami Labs? It's different? Yann LeCun?
I mean, is it French or is it New York? I'm not sure. Maybe you should ask him. Maybe you should ask him and see what he says. And that's— and that's, you know, I think. But Europe's success has always been this sort of, you know, joint venture, starting from government approach. How do you think Airbus was created? Very similar story. And today it's— it's a, you know, one of the dominant players, right? Maybe the dominant player, at least in commercial, you know, airline.
This is the European way. It's always been done this way in the past. And I'm not saying that CERN should, you know, the CERN of AI should go and build applications and build all these other things. No. What we need to do is create an open weight, open source, frontier model available for all of Europe's businesses. No patents, no IP, just simply technology, a foundational technology on AI that we make available to all of Europe's businesses to build upon, in a way that is independent from American and Chinese technology, and pulling our resources.
If you take all the money that Europe has available that could be invested into AI, actually we have enough to be competitive. But today it's fragmented across 28 different countries, and small projects all fail one at a time because they don't achieve any scale. So this is, in my opinion, this is the way that we do it.
Very interesting.
We create a CERN of AI to tackle the exact same problem that we had 70 years ago.
Well, I was not expecting that from the time when these American-based in Switzerland saying we have to create a public research center for AI, kind of.
Yeah.
Yeah, in Europe. Yeah.
But I came from CERN, so I see that it can work.
You've seen it.
And, you know, Proton came out of CERN. So what that proves is if you build a foundational technology, you will then have all the spin-offs that come out of that, and that will help to revitalize our economy.
But plus, there is a kind of a other philosophical thing coming in, because, I mean, the way Americans think, we've been talking about that. The way Chinese think is not the way we think.
Yes.
And—
We cannot copy their approach.
Americans think through private companies.
Yes.
And, I mean, I don't know that our kind of, I don't know, in the AI era, uncontrollable. Chinese think through kind of private controlled by public.
Yes.
That we don't really agree with on freedom.
Yes.
So that would at least solve something that is more of a kind of a not private. Well, it would be public somehow, but.
Yes. But also, I'll give you an example, right? First of all, governments are very bad at picking winners. The track record of picking the right company that's going to win is not very good. Also, the resources required are beyond, you know, any one European country. So you cannot— and let's assume that we think Mistral is going to be the one. You cannot really go to Germany and say, "Hey, I'm going to take Germany public money, and I'm going to stick it to Mistral." Not going to work. But a CERN of AI that is pan-European, that makes the IP available to all European countries, that the Germans can back, that the Polish can back, that the French can back, that even the Swiss can get behind.
So this is a model that could actually work, given the unique context that Europe sits in.
You know what? Sorry.
And by the way, I'm not— I'm a business person, so I'm the, you know, last person to say government intervention. But what is the role of government? The role of government is to act when the private sector has failed. And today we can say, without any ambiguity, that when it comes to building AI in Europe, the private sector has more or less failed. And it is time, I think, for this type of vision, this type of, you know, research institute to be created in Europe.
And perhaps in Geneva, because Geneva was picked for CERN for a reason.
I was about to say, you know what, like, just to start this little war. Yeah, you know what, I agree, but as far as it's in France.
Well, the reason CERN is not in France is the same reason that. There was a reason for that.
Yeah.
You wouldn't have gotten some of the other funding that you needed if it had been in France. Now, the compromise, of course, is that CERN is half in France, by the way.
Yeah.
The border.
The accelerator is in— yeah, yeah.
It's half and half, right?
Yeah.
The border literally crosses, you know, the CERN campus. And I shouldn't say this, but it was kind of used like a smuggling route as well.
Oh, great. Yeah.
You didn't have to go through the border check.
Well, you could smoke on the French side, maybe. I don't know.
Well, you could buy your groceries in Saint-Germain.
Ah, yeah.
You could enter through the.
And go through the CERN.
You could enter the French entrance.
Insane.
And then come out the Swiss entrance.
Pretty good for— yeah. Okay.
Yeah. I won't say I did it myself, but.
Yeah.
I know of people who did it.
Notre nouveau film « Europe : il va falloir arrêter de subir 🇪🇺 » est dispo sur YouTube : https://taap.it/rewrosY
To listen to the full episode, type "#576 - VO - Andy Yen - Proton - The Modern-Day Story of David and Goliath" on your listening platform.
Hébergé par Audiomeans. Visitez audiomeans.fr/politique-de-confidentialite pour plus d'informations.