Passer au playerPasser au contenu principal
Le Point
Is artificial intelligence threatening to slip beyond human control? In this exclusive interview with Guillaume Grallet, Science and tech editor-in-chief at Le Point, Microsoft CEO Satya Nadella decodes the ongoing technological revolution and draws a clear ethical line: « Superintelligence must empower humans and should be in human control; if not, it should by definition be rejected ». From strategic partnerships with French champion Mistral AI to autonomous cyber threats, fears of a tech bubble, and the broader economic diffusion of AI, discover the unfiltered insights of the American tech giant’s leader.

§ #LPT #LePoint #AI #ArtificialIntelligence #Tech #SiliconValley #Microsoft #MicrosoftAI #SatyaNadella #MistralAI
Suivez-nous sur :
Youtube : https://www.youtube.com/c/lepoint/
Facebook : https://www.facebook.com/lepoint.fr/
Twitter : https://twitter.com/LePoint
Instagram : https://www.instagram.com/lepointfr
Tik Tok : https://www.tiktok.com/@lepointfr
LinkedIn : https://www.linkedin.com/company/le-point/posts/
Bluesky : https://bsky.app/profile/lepoint.fr
Threads : https://www.threads.net/@lepointfr
WhatsApp : https://www.whatsapp.com/channel/0029VaDv82FL2ATwZXFPAK3c
www.lepoint.fr
§

Catégorie

🤖
Technologie
Transcription
00:00Aujourd'hui, nous avons la chance d'accueillir, d'interviewer Satya Nadella, le numéro 1 de Microsoft.
00:05Au cœur de la révolution de l'intelligence artificielle, une révolution source de beaucoup de promesses,
00:12de gains de productivité, de découvertes de nouveaux matériaux, de nouveaux médicaments.
00:16Et puis, en même temps, au centre de beaucoup de préoccupations, d'interrogations, de craintes,
00:22l'homme se demande s'il pourra garder le contrôle de sa création.
00:27Comment, Saturday de la bonjour, to what extent, as we spoke about in silico,
00:33is AI changing science and science discoveries?
00:36Yeah, I mean, first of all, I think the capability of AI as it goes up,
00:44our ability to do more and more things in silico, I think, are probably the place
00:49where we will see significant advances, right?
00:54New drugs, new discovery.
00:56Yeah, but even take, you know, in France, Hossein Gobain.
00:59Yeah.
01:00It was fascinating to see how they were able to use Azure Discovery
01:04to truly go after the additives in cement.
01:09Yeah.
01:09And ensure that, you know, 7% of the world's CO2 emissions are in the construction and,
01:18you know, related directly to cement.
01:20And so any breakthrough in chemicals that are used to create additives that are better
01:26on that dimension can make a massive difference.
01:28So I think that this idea that, you know, using these agentic systems to drive broadly
01:37scientific discovery process and accelerating, you know, if I think about this, you know,
01:43coding was the first breakthrough, knowledge, work, and productivity, science is the third,
01:49you know, or healthcare.
01:50It is another area that I'm very excited about, right?
01:52Even if you look at the amount of cost in any economy that's there in healthcare,
01:57that's a lot of workflow costs.
01:58So I think the more we can apply these AI systems in a trustworthy way to get the benefits
02:05that are much more broadly felt in the society and in the economy, that's, I think, what's most important.
02:11Do you understand some fears from some mathematicians?
02:14We've seen that, saying that their goal in life is doing some research.
02:18And with AI, they go to the Everest without doing the efforts of going to the Everest.
02:25And I think this is, it's a pretty profound topic, because, you know,
02:30I go back to even our mission of empowering people.
02:33And I ask myself, like, what is that sense of empowerment that one needs to feel when you even have
02:41this very powerful new technology, right?
02:43Because our identity, in many cases, our work was built around using a set of tools to achieve a certain
02:50goals.
02:52And now, suddenly, we have a very powerful new tool that, in some sense, displaces a little bit of that
02:59sense of identity, right?
03:00Whether it's in high math or in software coding, right?
03:03If I look at it, I'm not relating the two.
03:05But if I look at, you know, a year back, I think there was all this existential question, even in
03:10software,
03:10like, what do I do as a software engineer, except we are now realizing that, you know, software engineering as
03:16a workflow has changed.
03:18And yet, software engineering is more important than it was.
03:21We have now new problems to solve around verification and trust and so on.
03:28And so, I'm not, you know, sure exactly where math will go and high math will go.
03:34But my hope is that we bring back that sense of identity and dignity in what we as humans do.
03:43Because I'm ultimately, when I think about superintelligence, let's say, I think of it as a humanist superintelligence, right?
03:50Which is ultimately the human, it needs to be empowering humans and should be in human control.
03:56If not, it should be, by definition, rejected.
03:59Do you think the AI of the future should be open so that we can go and discover the chain
04:06of thought?
04:08You know, that was one of the claim of Jacob Coxon, you know, who left Anthropic saying, these are two
04:14closed models.
04:15I wanted to have your opinion on that.
04:17Yeah, I mean, I think there are a couple of things, right?
04:18One is, I think we have to sort of ground ourselves that the models are becoming more and more capable
04:24because of scaling laws.
04:27Unfortunately, though, as they're becoming more and more capable, our understanding is not that great.
04:33Because in some sense, alignment is not just falling out of scaling.
04:37If anything, it seems elusive and mechanistic understanding is also elusive.
04:42And so, therefore, you now have effectiveness.
04:44That's why I wrote this blog and perception is focused on it, which is to treat these models as insider
04:50risk even, right?
04:51I mean, they're powerful models.
04:53So, therefore, you now need to think of them as powerful insiders who may take action that, you know, is
04:59not fully in control.
05:01So, therefore, you need to create a control mechanism that sits outside the model.
05:06And that's why identifying every agent in every model, its actions that needs to be tamper-proof.
05:11Your point, chain of thought needs to be not in new release, but needs to be in English, fully inspectable
05:18by multiple models.
05:19So, these are all what I describe as principles of observability so that then you can have governance and security
05:26around it, which is what perception is all about.
05:29The product, both Agent 365 and perception are built to ensure that you can have a trust system that is
05:38outside of the model such that you can then use these models and treat them even in many cases or
05:45in all cases as insider risk.
05:47But do you think the human will be quick enough to keep the control of them, right?
05:53Correct.
05:53I mean, that's why I think this control system is going to, for example, behaviorally monitor all agent activity and
06:02do risk scoring and shut off any time it deviates, right?
06:07So, this is where there's a new cat and mouse game in some sense.
06:10Yeah.
06:10But we really need to move at the speed of the machine, so to speak, even in this.
06:17And so, the key thing is not to really take a powerful model, give it all the access to all
06:23the systems without any control mechanism that is outside.
06:27Because if you do that, then the chances of something misaligned happening are huge.
06:33But whereas if you at least have a control system that is continuously checking, monitoring, scoring, and shutting off based
06:41on some governance policy, then you have a much better system to control these things.
06:46Can the UN help for that or a new agency, maybe like the energy agency?
06:51So, there are multiple things here, right?
06:53There is, I'll call it what I'll call security, which is what I'm talking about, which is how do you
06:59really, what are the engineering systems and engineering practices that can control these powerful agentic systems today?
07:07That's where agent 365, perception, these, I would say, principles of observability are all relevant.
07:16Then you are bringing up, I think, a broader topic around AI safety and alignment, right?
07:22After all, we basically, it's a science problem, right?
07:25Now, we do not even know what the science of alignment is.
07:29And so, to the degree to which we want some AI safety governance, and that's where I think the Super
07:35Intelligence Task Force in the United States is going to work on it.
07:39We'll see what recommendations they come back with.
07:42And they could be.
07:43I think the international coordination on this will be important at the end of the day.
07:47Because after all, I think alignment is going to be one of the fundamental challenges in order for us to
07:53safely deploy more and more capable models.
07:56To what extent should we be aware also about the attacks, you know, the new AI capabilities can change the
08:06geopolitical landscape?
08:07Yeah, I mean, look, I mean, as we speak right now, the floor on AI-driven autonomous attacks is really
08:15dropped.
08:15And so, therefore, that's why I think it's going to be very, very important for every organization now to have
08:22systems like Perception and Agent 365.
08:26And essentially, you have to turn this into an agentic system, right?
08:29You have to have red agents that are simulating attacks, blue agents that are continuously defending, and green agents that
08:35are continuously remediating.
08:37And so, to me, that becomes the most important security consideration because the attacks are coming fast and furious at
08:45this point.
08:46And it's important now for all of the SecOps of every company to sort of transform themselves into AI-driven
08:53SecOps.
08:54Have you been impressed by new AI companies in France?
08:59Mistral is not yet the new Microsoft, but it's getting bigger and bigger.
09:04And I saw you just…
09:06Yeah, no, we are very excited about, obviously, our partnership with Mistral.
09:09Even this summer, we expanded it further.
09:11We have the Mistral models coming to Foundry in Copilot Studio.
09:16We're also going to use Mistral as an infrastructure provider for some GPU capacity.
09:22So, you know, excited about Mistral, not only in France, but, you know, as they even export outside.
09:28Also, this, you know, I had a chance to meet, as I said, I mentioned Saint-Gobain and how they're
09:33doing.
09:33I mean, they're a company that has been in existence for 500-plus years.
09:37And here they are using cutting-edge AI tools to advance their core competency around construction and construction material.
09:45And it's fantastic to see that.
09:47Or even a company, you know, I referenced Arlican.
09:51It's a company that is doing a new model family, right?
09:55It makes sense to me, which is not only do you have LLMs, but you have even these models that
09:59are capable of even causal reasoning.
10:01They have a new model family that's based on, you know, the mathematical concept of topology.
10:07And so they are going to probably be used in some of the security domains and so on.
10:13And so I see good innovation happening in France in many areas.
10:17The IMF has been worried about the possibility of an AI bubble.
10:22And I wanted to have your quick opinion on that.
10:25Yeah, I mean, I think, look, I think all of these cycles are cycles for a reason, right?
10:30Which is there is investment and build phase, and then there is the diffusion phase.
10:34So it's for other financial commentators to decide what is a bubble, not a bubble.
10:39Everything, in some sense, has these, as I said, cycles.
10:43What I am most focused on is it's a general purpose technology that ultimately needs to diffuse throughout the economy
10:51and create local surplus throughout the economy so that it shows up in GDP stats everywhere.
10:58And then when that happens, then, you know, you kind of have all of this really compute.
11:04Will it happen in a perfectly linear phase?
11:07Never has it happened in the past.
11:09And so, therefore, I'm not predicting some perfect landing.
11:12But at the same time, if five years from now, 10 years from now, will there be more AI tokens
11:17used per capita?
11:19Absolutely, right?
11:20By a gargantuan amount.
11:22And so that's what, at least at Microsoft, I'm focused on.
11:25I know you love technology.
11:26What are the main technologies that keep you inspired?
11:29Is it nuclear fusion, as we mentioned?
11:31I mean, I do think, I mean, to some degree, if I look at the thing that I'm most excited
11:38about is not just the next big technical breakthrough, quite frankly, because I'm going back to sort of the question
11:45you asked earlier, which is around diffusion.
11:46I look at this point, technologies are coming.
11:52The question is, can we truly get them diffused throughout the economy in a way that people closest to it
12:03are seeing the benefits?
12:05And that would be the last question.
12:07We've seen some progress in humanoid robotics.
12:09And we've seen, you know, with ChatGPT that you helped a lot, Sam Altman, at the beginning.
12:13Do you think the technology should resemble the human beings or does it have to be only a tool?
12:20I mean, I think overall, I think I sort of fall into the camp of technology is always a tool
12:28in the hands of humanity and human society, however powerful.
12:33I'm not saying that you should not be inspired by biology, but I don't think I would subscribe to anthropomorphizing
12:41this technology to a point where we lose, as I said, our agency, our ambition and our dignity.
12:47And I think that that's sort of where I stand.
12:49And then in terms of, you know, what I would recommend for anyone is I think we're at that stage,
12:55right?
12:55Whether it's a student or an entrepreneur, I think the real currency now is in diffusion.
13:02How do you use it?
13:04What impact?
13:05And the more we get a little more focused on celebrating the use cases and, you know, sometimes, you know,
13:13you talk about use cases as if it's not that innovative.
13:16In fact, all invention is right there.
13:19And so that's where I would say there's plenty of opportunity, plenty of exciting things to come.
13:25And I'm excited to see it really, truly diffuse throughout the economy and society.
13:31Thank you.
13:31You can do it too.
13:31Thank you.
13:32Thank you.
13:32Take care.
13:32Oh, and God.
13:33Well, got you.
13:33And we'll do that in good news.
13:33And understand the truth.

Recommandations