- 9 months ago
Digital sovereignty strides in franco-german cooperation
Category
🤖
TechTranscript
00:00Sous-titrage Société Radio-Canada
00:33Sous-titrage Société Radio-Canada
01:09Sous-titrage Société Radio-Canada
01:17Sous-titrage Société Radio-Canada
01:19Sous-titrage Société Radio-Canada
01:28Sous-titrage Société Radio-Canada
01:29Sous-titrage Société Radio-Canada
01:36Sous-titrage Société Radio-Canada
01:40Sous-titrage Société Radio-Canada
01:43and 4,000 people.
01:46Sous-titrage Société Radio-Canada
01:50Sous-titrage Société Radio-Canada
01:52Sous-titrage Société Radio-Canada
01:56Sous-titrage Société Radio-Canada
02:03Sous-titrage Société Radio-Canada
02:06Sous-titrage Société Radio-Canada
02:33Sous-titrage Société Radio-Canada
02:35Sous-titrage Société Radio-Canada
02:45Sous-titrage Société Radio-Canada
02:47Sous-titrage Société Radio-Canada
02:51Sous-titrage Société Radio-Canada
02:55Sous-titrage Société Radio-Canada
03:06Sous-titrage Société Radio-Canada
03:08Sous-titrage Société Radio-Canada
03:25with the app as built in real time and for that you need a sensor intelligence and you need AI
03:31is why we are really embracing AI as a great great set of tools and methods to advance our research
03:40Philippe can you tell us more about your work yeah one of the key topics for us is what we
03:45call
03:45trusted AI so how can we provide guarantees about the performance the functionality the safety the
03:52robustness of AI systems going forward and one of the approaches that we use for that is what we call
04:00digital reality we learn models about reality and then we use we generate synthetic data from those
04:08models in order to train AI systems with that data but also to validate AI systems based on based on
04:18the data just so example kid running in front of a car we don't have data we don't want data
04:23from
04:24real kids so we actually model the kids we generate the data from we let virtual kids run in front
04:30of
04:30cars and then make sure that the car do the right thing in this set in these circumstances so now
04:37we're going to talk about digital sovereignty why basically is artificial intelligence strategically
04:44important for Europe's digital sovereignty and how can we avoid technological dependence on non EU
04:53countries and regions of the world and and our friends and Germany working together on this
04:59Philippe yeah I can start so digital sovereignty is absolutely essential right if you look at
05:06chat GPT the latest generative AI models large AI models whatever you want to call them they're exclusively at
05:13at this point provided from the US so any company in Europe has to ship their data to the US
05:20and Sam Altman
05:21already said you might not actually make a system available if the AI act is is implemented and and that's
05:29scary
05:29right because then the entire European industry would not have access to that technology so we actually
05:35absolutely need to develop the ability to train such systems and other AI systems here in Europe and also
05:43make sure that they fulfill the requirements by European AI so for trustworthy AI right now all of those systems
05:51don't even fulfill a single requirement on trustworthy AI so we really need to make a European approach here and
06:00really change this technology yeah it's really about data ownership privacy dependency we need to reduce this dependency
06:08because we need to keep independence on the decisions and also interests right now it's both
06:15okay so this is why it's a real concern to to leave that you know for to take technology develop
06:23entirely outside Europe
06:26what do you expect from Europe in general and also what infrastructure would Europe need to build its own generative
06:35AI
06:37yeah you're raising a really important point so as most of you probably know there is not a single hardware
06:45facility in Europe
06:46currently available to train these large AI models right and they will get larger even now right so we don't
06:54even have that
06:54Donc, nous devons qu'on peut-être, nous devons qu'on peut-être,
06:57comme Européens, comme Français et Germans,
07:01nous sommes les plus grands pays ici,
07:02pour qu'on peut-être, pour installer et permettre cette technologie.
07:07Et ça va aussi changer la façon dont nous faisons les recherches.
07:11Je suis un des co-founders de CLAIR,
07:14le confédération de l'AI en Europe,
07:16et nous avons, d'abord, cinq ans,
07:18nous avons discuté que nous avons besoin de ce que nous appelons le CERN pour l'AI.
07:24This is now actually being picked up by the UK, this idea.
07:27We've had that five years ago.
07:29Really, a central facility where the researchers come together,
07:33plan their next research, coordinate that,
07:37plan the next research and go forward,
07:39and also create a symbol for European ambitions and excellence in AI.
07:45So, really, we need to have that joint effort by the European countries,
07:49probably as a private-public partnership, right,
07:52with a private investment,
07:53but also with public investment from multiple countries.
07:57Pierre.
07:58I must say also that we believe in the normative power of the EU,
08:03because, for example, the AI Act is a legal framework,
08:06which is risk-based,
08:07and that prepares the ground for trustworthy and safe AI application
08:11in Europe and beyond,
08:13and that can also become like a standout,
08:15like LGPD is.
08:16So, it's also important for, like, implementing AI with our, like, values,
08:23you know, and in addition to having, like, safety and trust in AI.
08:29Do you have an example of measures that Europe could take,
08:33or what's the first thing we would need to do to be less dependent
08:37and to, again, to build that infrastructure for AI and technology?
08:43Yeah, so there's a number of technical issues.
08:48In Germany, there's a big initiative called LEAM,
08:51large European AI models,
08:53where we're trying to put together,
08:57argue with the politicians to invest in such an infrastructure.
09:01We are currently also talking to private investors.
09:05We're talking to the researchers who would actually implement these things.
09:09things, but there's more than just providing the infrastructure.
09:14We also need to focus on what are we doing technically, right?
09:18And so, current AI systems are stochastic systems, probabilistic systems.
09:22They give you a correct answer with a given probability.
09:27That is not, in some cases, in many cases, this is not what we need, right?
09:31We need an exact answer.
09:33I mean, 5 plus 2 is 7,
09:34not just mostly 7, but maybe something else.
09:38So, neuro-explicit AI is something we're focusing on.
09:42We're focusing on providing guarantees for AI systems, as I said.
09:48Causality is a really important part.
09:50Grounding.
09:51So, these language models are currently only,
09:55they know what people have talked about because they only look at text.
09:59They know nothing directly about the world itself,
10:02about data from the world.
10:04So, we need to teach the system how the world really works,
10:07and neuro-explicit models are one way of doing this.
10:11Causality, modularity, grounding,
10:14all of those topics are really important,
10:16and they're currently not being addressed by those systems.
10:19And that's a great chance for us as Europeans,
10:21really, to make a dent and provide real value here.
10:27And if you look at it, who publishes those systems?
10:30Who are the authors of those papers?
10:33There's a lot of European names there.
10:36So, we also need to make Europe more attractive in terms of talent,
10:40so people don't go to Google or Microsoft or whoever,
10:43but they actually found companies here.
10:46They work with the European industries and research organizations
10:50to really bring Europe forward and not necessarily go to the U.S. or elsewhere.
10:55I feel it's been a decade that we've been bringing that problem of talent,
11:01talents going overseas, going to the U.S.
11:03Do you believe it's possible to maintain them here,
11:06to make them want to stay in Europe,
11:07whether France, Germany, or somewhere else?
11:10Absolutely.
11:11That is one of the critical things.
11:12We need to make that able, make that possible.
11:16That does require changes, right?
11:18I mean, in Germany and in France as well,
11:22researchers are often state employees,
11:24and there's restrictions on various things.
11:27So, I think there's a lot that needs to change there.
11:31And private-public partnership is a way to actually deal with that
11:34because then you have more possibilities.
11:37We don't have these big companies like Google, right?
11:39So, it's probably not going to go to work without the public support.
11:45But at the same time, I think it's also a duty by the European industry
11:50to invest in this technology together with the public agencies.
11:56We should also promote European vision of AI
11:59because the European countries, the emphasis,
12:01the ethical, human-centric development and use of AI,
12:05and to ensure European sovereignty in AI,
12:07that involves promoting AI systems that respect these rights,
12:11the fundamental rights, transparency, accountability,
12:14alignment with European values and regulations.
12:17And also, we are convinced that also the strength of Europe
12:23is that first we have leading AI institutions,
12:26we have educations, very strong in AI, we have talents,
12:31we have also a lot of diversity in Europe
12:33in terms of culture, socioeconomics, languages,
12:36and also this human-centric and ethical approach to AI.
12:41We believe that it can be more inclusive,
12:44should facilitate long-term trust and adoption.
12:47So, if you want, that goes well beyond the society
12:51where you have the notion of users
12:54that are passive consumers of AI.
12:56We don't see that this way.
12:58For us, it's more like citizens, right,
13:01that become active users and actors
13:03of even a shared governance, right?
13:06Where AI is not a technology called paradise
13:09with sensors everywhere,
13:11just with happy users, you know?
13:13It's not that.
13:14It's a citizen who is an actor that generates data
13:17that is itself an actor of a shared governance, okay?
13:22AI is a means, not an end.
13:26And this is, we believe, like a long-term investment.
13:30Yeah.
13:31Maybe, if I may just quickly add to this.
13:34Actually, just almost coming back,
13:37we had a big meeting of the European Networks of Excellence in AI,
13:43actually in Italy, a very nice meeting there.
13:47And we had all of the currently six networks of AI together there,
13:52and we discussed exactly those topics.
13:54So, there's a lot of agreement
13:56on what needs to be done from the research community.
13:58We have a lot of the industry people who are interested.
14:02So, we need that push for the public
14:05to fund this research to kind of get this started.
14:08And we need to do this together,
14:10and I think the collaboration between INRIA and DFKI,
14:13this German-French collaboration,
14:15is a big enabler in doing this going forward
14:19because a lot of other industries,
14:20a lot of the other countries,
14:22will hopefully join us
14:23once we get the opportunity to actually go forward.
14:27We only have four minutes left.
14:31One question is, how could Europe put it...
14:34You talked about safety measures, ethics, values.
14:37How can we put this safety into the design of AI first?
14:41And what kind, again, of measure
14:44when it comes to data, specifically?
14:47Philippe.
14:48Yeah, I can start.
14:50So, there's a number of approaches we need to take.
14:54One is to actually develop AI systems
14:57that provide guarantees about the performance
15:00or the functionality right from the start.
15:03This is by design, as we call it.
15:06That's really hard.
15:07This is possible.
15:08There are some approaches in this direction,
15:10but that's really hard.
15:11The next step is, if that's not possible,
15:14is to actually develop tools
15:15that provide guarantees about AI systems.
15:19And we make really good progress along those lines.
15:22There's a big initiative also on the French side,
15:25Confuse AI.
15:26There's initiatives on the German side.
15:28And then the third step is really enable the researchers,
15:33the technicians in companies,
15:35through explainable AI,
15:37where it's the human who understands
15:39what the system is actually doing
15:41and can make changes.
15:43And at the end,
15:44it really also comes down to the human factor.
15:49The system also needs to show how trustworthy they are
15:54and to prove and provide evidence about that
15:57so that people actually can trust this.
15:59They have to understand
16:01what is a trustworthy AI system and what is not.
16:03So, the idea of certification,
16:07official certification,
16:08putting a stamp about that.
16:10There is a lot of work in this direction.
16:12The industry really wants to see that,
16:15but there's a little bit more work to be done there.
16:19Pierre, do you want to follow?
16:22Yes, so I think there is not a universal solution
16:25to these guarantees.
16:26But for sure, more basic research is required.
16:30So, we need a lot of investment in basic research.
16:33to tackle these open problems
16:35because having a certified AI,
16:38this is difficult.
16:39So, right now,
16:39we tackle this into project collaborations,
16:42like for robotics,
16:43what can we guarantee for other applications.
16:47But like Philippe said,
16:49it's really also building instruments
16:51that themselves provide the certificate.
16:55Okay?
16:55It's not necessarily AI-based, by the way,
16:58but it's like having like
17:00sort of analog of a flight simulator,
17:03but also like something that is put today
17:07in embedded circuits, you know,
17:08that you can really guarantee for software.
17:12But here, you need to have bounds,
17:14you need to have really certificates
17:15that you need that for industrial production,
17:20you know, that of course,
17:21for security reasons,
17:23and for safety,
17:24for even for privacy.
17:27For privacy.
17:28We had one really left,
17:29but very quickly,
17:31we all know that AI can be dangerous
17:34if we don't do stuff right now,
17:37if we don't regulate.
17:38Europe has enacted groundbreaking regulation.
17:41We're really good at regulating
17:42in many areas of tech,
17:45but we're not good
17:46when it comes to regulate AI
17:49and development.
17:50How can we avoid the same fate in AI?
17:53And what, again,
17:54should we do right now
17:55if there's someone magically working
17:58at the EU right now that hear us?
18:00What's your advice?
18:02Well, I just heard the AI Act
18:04has been accepted an hour or two ago
18:07with overwhelming majority,
18:10I think more than 80%
18:11of my information is correct.
18:13So we will have that.
18:15There will be a period now
18:16where there is a dialogue
18:18also with the countries.
18:20There will certainly be changes,
18:21and I think there are a number of points
18:23that need to be changed.
18:25But at the end,
18:28Europe needs to be able
18:29to do more than just regulation.
18:31We let the others develop the system,
18:33re-regulate them.
18:34That doesn't work.
18:35We also need the ability
18:37to develop these systems
18:38according to the standards
18:42that we want to have,
18:44that we want to see
18:44and the regulation
18:45that we want to impose.
18:46Yeah.
18:47One final word.
18:49Well, I think there is also,
18:53we need to have also
18:54an effective data governance.
18:56That's also a very important aspect
18:58of that,
18:59in addition to having
19:00like this regulation, right?
19:01So data governance is crucial,
19:04straightening research
19:05and innovation,
19:06and also promote
19:07European vision of AI.
19:10The last three points.
19:12Thank you very much
19:14to the two of you.
19:15Please make a beautiful round
19:16of applause for our guests.
19:17Thank you.
19:18Thank you very much.
19:19Thank you.