- il y a 3 mois
Walk the Talk The Telco Industry’s Revolution in the AI Era
Catégorie
🤖
TechnologieTranscription
00:00Good morning. How are you guys doing? Great? Excellent. Super happy to be here. My name is Bruno Zerbib. I'm
00:08the Chief Technology Innovation Officer for Orange. I'm sure you've heard about those guys.
00:13And I'm happy to be talking about a few things that are quite disrupting and changing our lives, and I'm
00:21sure changing yours as well.
00:22So without further ado, let's start. We're not going to be alone. As you can see, we have three chairs,
00:27so I'm going to have two great guests that are going to be joining me in a few minutes.
00:32All right, so let's start by quizzing you on this. Do you guys know what that is on this slide?
00:40For those who are familiar with Magritte, this is not an apple. What do you think that is?
00:53We have 45 minutes, so I'm very relaxed about this whole thing.
01:02Okay, what do you think that is? A bunch of pixels. Okay, great. I was not expecting that, but sure.
01:15Say that again? Smartphone with four wheels. Yes.
01:23What is changing the automobile industry? Yes.
01:29A reference to art?
01:31Yeah, this is Magritte. Absolutely.
01:36Computer on wheels.
01:41Software technology.
01:43So I think everybody's getting an intuition that there is more to this thing than being just a car.
01:49And I have to tell you that the way I used to talk about this a few months ago, a
01:54few years ago, I would have called that just computer on wheels.
01:58Because of all the smarts.
01:59But it's actually much more than that.
02:06It is a distributed data center.
02:12So let me explain that to you, because that might become very, very, very confusing.
02:17We live in a world where CPUs and GPUs are going to be the bottleneck.
02:22If you think about AI going forward for, like, training all those models and doing inferencing, it's going to come
02:28down to, essentially, GPUs, CPUs, and energy.
02:32Those will be the absolute bottlenecks that we'll have to overcome.
02:36So there is no way in hell that you're going to get a Tesla car with super advanced GPUs, actually,
02:43latest and greatest, extremely expensive NVIDIA, and when you're not going to be driving your car, not using them.
02:50Because in the future that lies ahead of us, maybe two, three, four years from now, all the compute resources
02:56that you're going to have everywhere, all of them, are going to be used all the time.
03:04And it's very hard to think about it.
03:06But clearly, when you look at the need for inferencing and training models, we will not be able to afford
03:12economically to have those resources going to waste.
03:14Which means that when you think about a Tesla car, when you will not be using it, it will be
03:22somewhere, you know, in your garage or, you know, seeing idle.
03:27You can expect some kind of Wi-Fi connectivity, and it's going to be used for processing.
03:33And that means it will be connected all the time, and it will be generating traffic all the time.
03:40And it's too expensive not to be used all the time.
03:45Today, nobody really thinks about it this way, but we know this is what's coming.
03:48So that is the reason why it's called a distributed data center.
03:51Actually, Elon Musk was saying that I think we have about 7 million Teslas out there, and if you bring
03:57them together, you have maybe the largest AWS-sized kind of data center that works.
04:03It's essentially, it's a cloud.
04:05And it's always on.
04:07It's crazy.
04:09So it's just what lies ahead.
04:11And by the way, it works with a Tesla that works with your PC at home that works with everything.
04:18Traffic on all the time.
04:20Processing on all the time.
04:25When you are like a telco operator and you hear that, your brain is going boom.
04:30You know, so what does it mean to us?
04:31It means, okay, we're going to have to figure out what we're going to be doing with this.
04:35Because is it great news?
04:36Is it good news?
04:37Well, it depends.
04:38If we can monetize it, that's fantastic news.
04:40If we can, that's going to be a real challenge.
04:42But this is really something that is important.
04:49Have you seen the, by the way, the demo, the OpenAI 4.0 demo that took place two weeks ago?
04:58You saw that demo, like this incredible demo?
05:00And if you notice that, if you remember, Moradi, who is the CTO, was essentially giving a demo of 4
05:09.0, ChatGPT 4.0.
05:114.0, actually, not 4.0.
05:13And it's an amazing latest version of ChatGPT.
05:17Response time went from three seconds to 0.3 seconds, which means the world just changed.
05:23Now AI has become about real time.
05:27Always on, real time, accessible, available.
05:31This is going to change everything.
05:34Actually, in that demo, they were so afraid about losing connectivity that they had to wire the iPhone.
05:43I don't know if you guys noticed that, but they had to wire the iPhone.
05:47They wanted to make sure that it would have consistent connectivity.
05:51AI is going to be so important that if it's not on all the time, if it's not widely available,
05:57if it's not affordable, if it's not safe, it's not going to work.
06:02Today, we're going to have, like, other guests, and we're really going to talk about AI, AI for good.
06:06But AI needs to work.
06:08It needs to work, it needs to scale, it needs to be delivered.
06:11The way at Orange we think about connectivity is going to change drastically.
06:15Basically, my obsession as the CTIO for the group now, and we have a chief AI officer here along with
06:20us, is to make sure that we build the best infrastructure for AI.
06:25The thing that's going to be extremely different is the nature of the traffic.
06:29You've heard about the Ray-Ban glasses from Meta.
06:32So they are going to be, essentially, I've tried Vision Pro, I've tried the Ray-Ban glasses.
06:37I was really looking forward to getting the Vision Pro and using that.
06:41It turns out that I'm much more in love with the Ray-Ban glasses, because it's light, I can use
06:45that all the time.
06:46And I'm expecting that with the amount of processing power I'm going to have with this, I'm going to be
06:51streaming traffic all the time.
06:53Everything that I do, think about that, a 4K stream, nonstop 4K streaming, going through the glasses, back to the
07:01phone, and then back to the cloud.
07:04That's going to be the basis for inferencing.
07:06It's not going to be just typing, like what you guys are doing today with ChatGPT.
07:10This is going to be a nonstop stream of video and audio.
07:14Can you think about that?
07:16Which means that the way we've been thinking about our network is going to change.
07:20We used to have best effort network.
07:22If you were using Netflix and there was something going on, it would just buffer a little bit and that
07:26would be okay.
07:27It used to be a download-defined network.
07:29The traffic was going down from the data centers to your PCs.
07:34This is going to become very different.
07:37Now we're going to be talking not necessarily just about bandwidth, a different kind of bandwidth, upload bandwidth, and we're
07:43going to be talking about response time.
07:45If you've been using ChatGPT for, oh, the last few days, last few weeks, you go crazy when the connectivity
07:51is not there, when you're getting, like, slowed down.
07:54It's just frustrating because you realize you're going to become more and more and more dependent on this technology and
08:01connectivity is going to have to work.
08:03So we are at the beginning of a new era.
08:05What you've taken for granted as a commodity is going to reinvent itself.
08:09But for what?
08:10To make sure that we can turn AI into the killer app that we've been waiting for.
08:16We've been talking about what is 5G killer app, the application that is the killer app for 5G is AI.
08:24Always on, very low latency, the thing that you will not be able to live without.
08:29That is the reason why we have to make sure that this technology works really, really well, and we do
08:34it for good.
08:34So I'm very, very happy to invite two guests with me.
08:39So we have Bassem Asse, who is a hugging face head of sales.
08:46I'm expecting them to listen to me right now, so they should be able to come in.
08:54Excellent.
08:54Hello, Bassem.
08:59And we have Frédéric Werner, United Nation ITU Chief Strategy and Operation for AI for Good.
09:09So you are being spoiled with two great guests.
09:11Super happy to have both of you guys.
09:13And I'm sure we're going to have like complimentary perspectives.
09:16So, hey, I'm going to start with you, with Hugging Face.
09:20So, Hugging Face is a French-American AI startup that's been developing tools for using machine learning.
09:28I mean, we've been using them like crazy, right, Steve?
09:31We love you guys.
09:33We love what you guys do.
09:34And we are very excited to be working with you.
09:37And we've been exploring Gen AI together.
09:39So that's a fantastic partnership.
09:42Bassem, what is your vision of open innovation?
09:45I was talking about the fact that we have to build the right kind of AI.
09:48We want to be, we want to go fast, but we have a societal responsibility.
09:52We believe that we have to work together with the brightest mind.
09:56We want to make sure that we're not opaque, we're transparent.
09:59And it's all about ecosystems.
10:00At the end of the day, you're part of that ecosystem.
10:02We're working with the big players, and we're happy to work with you as well.
10:06So what's your vision of open innovation?
10:09Actually, at Hugging Face, we really emphasize open science, open innovation, and open source.
10:16It starts with open science, it starts with research, it starts with academics.
10:21Then it goes to independent developers, community members, still academics, researchers, companies, startups, larger accounts.
10:31Innovating, innovating as openly as possible.
10:34We know that for startups and larger accounts, sometimes it might be a bit tricky, but innovating as openly as
10:39possible.
10:40Because usually innovation comes when you are building on top of what the others have already built before you.
10:47And then at some stage, you make that jump and you innovate really seriously.
10:51And open source models definitely allow that.
10:54Just like open source did it in the last 30, 40 years, I mean, since Linux became a de facto
10:59standard, I would say, open source software allowed innovation, allowed good quality of software.
11:06And within AI, what we are seeing, what we have been seeing in the last few months, and what we
11:10will continue on seeing in the next few years,
11:12is that the gap is closing between proprietary models on one hand and open source models on the other hand.
11:19And open source models mean simply that you are not locked with one vendor.
11:24You are not locked with one vendor who is providing you with one price today, maybe potentially another price tomorrow.
11:30You are not locked with one infrastructure that that vendor will decide for you.
11:34You keep control of your AI destiny at the end of the day.
11:38And that's why we consider that more innovation comes with openness on one hand,
11:42and more control of your AI destiny comes also with that transparency that open source allows.
11:47I think it's in the Spider-Man movie that said, with great power comes great responsibilities.
11:52So those elements will yield tremendous amount of power and impact.
11:58So I think it's very important that we have this open source approach.
12:00I was saying early on, before you guys came on stage, that we love you guys.
12:05We're working with you.
12:07And we think this partnership around AI for good is very important.
12:11How do you, what are the practical use cases, or how do you see us continuing to work together around
12:17this kind of open paradigm?
12:18Our vision is that many people, at least in the media at least,
12:25will be talking about those very large models that anyone can use, ask them any question,
12:29they will bring you any answer.
12:31It's amazing.
12:32But at the end of the day, in your business, because you are a telco operator,
12:35same thing for an automotive manufacturer that you mentioned earlier,
12:38same thing in various industries,
12:41most of the time you will not be asking your chatbot a very generic question.
12:44You will not ask the chatbot what happened on the 15th of May, 1915,
12:48because the history of First World War doesn't interest you in a business context.
12:52In a business context, you will need to answer the question that your customer is asking you,
12:55because you have added a close in the contract that is between you and the customer.
13:00So your human agent in your customer contact center will most likely be asking very,
13:05getting very precise questions and needing very precise answers.
13:09So what we are doing with you, what we are doing in various industries,
13:13is to help large enterprises and startups customize models that fit best in terms of size,
13:19in terms of costs, in terms of accuracy of content that is being delivered.
13:23So those models most of the time are open source,
13:25and most of the time are customized on your specific business context,
13:29and that business context, you are the only ones who have it.
13:32You have the data, you have the business processes,
13:34and this is how the best answer arrives at the best cost, both environmental and financial.
13:41Thank you.
13:42And I think this is very important, because we want to tap into the best technology out there.
13:46At the same time, we want to make sure we channel in a way
13:48that allows us to live up to our societal responsibility,
13:51which brings me to you, Frederic.
13:55Obviously, I'm a geek.
13:56I love technology for the sake of technology,
13:58and to some extent, I love to see this thing progressing as fast as possible.
14:03And I was excited about, you know, the 4.0 demo that we got a few weeks ago.
14:07It's just crazy impressive.
14:10But at the same time, being part of Orange,
14:13and also being a father, having kids,
14:16I understand that there's a societal angle and responsibility.
14:20We have to think about the profound ramification.
14:23The last time we had crazy change in technology,
14:25with maybe the industrial revolution,
14:27that led to a lot of bad things happening in the world.
14:32In Europe, people were really unsettled, to say the least,
14:34and we had wars and things like that.
14:36So technology might have a profound impact on society.
14:40You have a unique perspective with your role at the ITU
14:45and your focus on AI for good.
14:47I would love to get your perspective on maybe the societal impact
14:51and maybe more selfishly understand how do you, from your perspective,
14:55how do you think telecommunication can be,
14:59our industry can contribute to make AI something, a force for good.
15:04Yeah, thank you, Bruno.
15:05And first of all, thank you for your introduction.
15:08I don't think I'll ever look at a Tesla in the same way ever again.
15:11I live in Geneva.
15:13There are many Teslas.
15:13I didn't realize I was living in a distributed data center.
15:16So I think we're one step closer to the matrix.
15:19Thanks.
15:19I might steal that from you one day.
15:21No, but in all seriousness,
15:22AI for good was created by the United Nations, by ITU, in 2017.
15:28And that's seven years ago.
15:29That's basically an eternity in terms of AI years.
15:32And it was built on the premise that we now have less than 10 years
15:35to achieve the United Nations Sustainable Development Goals.
15:38And AI holds great promise to help achieve many of those goals and targets.
15:42So from anything from healthcare, to education for all, to climate change, to gender equity,
15:49autonomous driving, robotics, brain machine interfaces,
15:52I think that the use cases are definitely there.
15:54But how do we know these high potential applications work equally well on men or women,
16:00on persons with different skin colors, on children or the elderly, persons with disabilities,
16:05or in low-resource settings where basic things like 5G connectivity, you know, electricity are,
16:12we take that for granted.
16:13But in many parts of the world, that's an issue.
16:16So these are things that don't occur naturally to the fast-moving tech industry.
16:20I think the approach has been build it and we'll fix it later.
16:23But these are things that we think deeply about AI for good.
16:26So the goal of AI for good, simply put, is to identify practical applications of AI
16:31that exist here and today to advance the sustainable development goals
16:35and figure out how to scale those solutions for global impact.
16:38Now, how do we do that?
16:39First of all, we can't do that alone.
16:41So we have virtually the entire UN system helping us with AI for good.
16:45But on top of that, we need industry, we need telecom operators,
16:49we need NGOs, civil society, everyone to participate in a, you know,
16:53to help us chart a beneficial course for AI for humankind.
16:58And just for some practical examples.
17:00So at ITU, we developed standards.
17:01Orange has been a long-standing member for many, many years,
17:04contributing to our standards-making process.
17:06And we now have 220 AI standards published or in development.
17:11So basically, if you want to deploy machine learning in 5G networks,
17:15or you want to manage energy consumption in a data center,
17:19or you want to improve cybersecurity or self-optimization of networks,
17:23or, you know, try and improve multimedia streaming,
17:26all these standards are there.
17:28And we also launched basically crowdsource challenges.
17:31I think this would relate to Hugging Face.
17:34So basically, you know, creating problem statements,
17:36going out to the power of a crowd.
17:38And, for example, we had a challenge on how can we reduce the energy consumption
17:42of 5G base stations.
17:43And we put this out using real network operator data.
17:47And these basically students from Africa created a solution
17:51that would actually reduce the energy consumption of 5G base stations
17:55in an entire city by 20%.
17:57So for me, that's a really good example of AI for good.
18:00But beyond that, I think telecom operators, they have data.
18:04And I remember seeing a survey a few years ago, I'll have to dig it up,
18:09but telecom operators are still ranked very highly on the trust level
18:13in handling data properly compared to, say, other social media platforms
18:18or different platforms out there.
18:20So I think there's an opportunity there.
18:21Because if you look at all the data that's in a network or in cars, for example,
18:26how do we leverage that data for good?
18:28Because sharing data in a way that's meaningful and useful
18:32but still respects privacy still remains the biggest bottleneck
18:35preventing AR for good scaling globally.
18:38And I think when it comes to, for example, preventing academics
18:41or natural disaster management, the telecom industry as a whole,
18:46I think, has a huge role to play in being able to predict, manage,
18:51and mitigate these epidemics and disasters by leveraging data.
18:54But we need to crack the puzzle of how do we share data in a meaningful way,
18:58not only on a city, national level, but regionally and internationally.
19:02Thank you.
19:02That's great.
19:03You know, I love, I mean, obviously I did not know your answers ahead of time,
19:06but I love what I'm hearing.
19:08At Orange, we have this big initiative called Data Democracy.
19:12That, it goes with what you just said,
19:14we are obsessed with making sure that the data is being accessible,
19:17it's being normalized.
19:18Because if you don't have that, then you don't have anything.
19:22You know, it's the foundational layer.
19:24How do you get the data to be there, to be high quality data, reliable data,
19:28and then you can build things on top of that.
19:30But if you don't have that, you don't have the foundation.
19:32And maybe in the same vein, Fred, just thinking about,
19:37I mean, obviously the ITU has a lot of respect, incredible reputation.
19:42You've been around the blog for 1,000 years like us.
19:46So we have developed, I think, that reputation
19:49and that societal responsibility
19:53and also some kind of maybe emotional distance.
19:57You know, we understand there is hype,
19:59and then there is really, in the long run,
20:01how this technology is going to be used.
20:03What do you need right now for scaling up this approach of AI for good?
20:07What do you think you need?
20:08What kind of support?
20:09What do you think are the next milestones?
20:12Or how do you think about that?
20:13Yeah, thank you.
20:14So I think we really need to move beyond the niche use cases.
20:18And probably the best part of my job is I see these use cases
20:20coming across my desk every day.
20:22So using a mobile phone to detect diabetes
20:25or to detect skin cancer
20:27or detect what kind of cough you have using the microphone.
20:31So there's all these amazing applications on a phone,
20:35using AI, using network services.
20:37And to use the healthcare example,
20:40when you develop a drug, you have to go to the FDA.
20:43You spend billions of dollars.
20:45It takes probably 10 years before you can put an aspirin in a supermarket
20:48or a new cancer drug, what have you.
20:51Now you have all these high-potential AI health applications,
20:55which, especially for the developing countries,
20:58you can really bridge the gap with telemedicine
21:00and really help people in a way.
21:02But there's no meaningful way right now of comparing apples with apples.
21:06Right now, any, you know, sort of weight loss app
21:09or healthcare apps just goes on the App Store.
21:12It meets the App Store criteria, and that's it.
21:14So if you're a mayor or running a hospital
21:17or you're a healthcare minister,
21:19how can you know if that app is good
21:21or you should invest in it or deploy it?
21:23What kind of policies do you make?
21:24So at ITU, with the World Health Organization
21:27and the Intellectual Property Organization,
21:29we've developed 35 standards and specifications and benchmarks
21:33so you can actually test and evaluate the efficacy of AI for health algorithms
21:38so that you could actually start to make meaningful deployments,
21:41decisions, rather, for the deployment of these solutions.
21:44So I think really having a framework of evaluation is important.
21:48But back to the data sharing,
21:50I've been in so many meetings where we say,
21:52who in the room has data?
21:54All the hands go up.
21:55I'm like, great.
21:56Who's willing to share data?
21:58Everyone starts looking at their shoes.
22:00So I think, again, having a culture of sharing data
22:04but also the technology to do it in a good way.
22:07And I do think, you know, operators and organizations
22:10like Hugging Face who are at the cutting edge of these solutions
22:13need to come up with standards and best practices
22:16on how to do that to achieve scale
22:17or we'll still be stuck in use cases.
22:20Absolutely.
22:20And I really think that the use cases around health
22:23are maybe some of the most compelling
22:25that should convince people who have doubts
22:28about the value of AI to embrace this technology.
22:32You also talk about having access to the data.
22:35I think you're absolutely right.
22:36We also have to leave up to our own compliance
22:38and regulation requirements that we have in each country.
22:41For instance, in France,
22:42they are quite, I would say, quite significant.
22:46And we try to figure out what is the optimal path, right,
22:50between tapping into the power of those data sets
22:53and at the same time making sure that we don't do stupid things.
22:56We don't essentially never violate the trust
22:58that we get from our customers.
23:01So, Bassem, I, you know, I was going to make a point.
23:05I'm not sure if I should make that point.
23:06But we've heard about one of the big AI company last week
23:12that they kind of, not laid off,
23:14but they have heavily reduced the number of people
23:17that were responsible for making sure
23:19that they will sanitize their LLMs,
23:21that essentially not stupid things will happen.
23:22They decided to go so fast on innovation
23:26that they might have done that
23:27at the expense of their societal responsibility.
23:30You guys are really cool.
23:31We love your technology.
23:32but this is about AI for good today,
23:34so I have to ask you about AI for good.
23:37In what way do you think you guys are contributing
23:38to really making this a reality
23:41and not just being a great AI company,
23:43but a great AI for good company?
23:45Thank you for the question.
23:46It's interesting because you mentioned AI for good,
23:49both of you.
23:49You mentioned democracy when it comes to data.
23:52Actually, if you go on our website,
23:54most likely you will find a sentence
23:55that we consider as being our mission,
23:57and our mission is to democratize good machine learning,
24:00democratize good AI.
24:01And when you say that,
24:02it's not just a motto or just to be cool
24:05or to be fancy or to...
24:08The main idea behind that,
24:10and this is how we end up with open source,
24:13the main idea behind that is that
24:14if we consider that AI as we know it today
24:17is going to be controlled by a few companies
24:21who are able to provide huge data sets,
24:24huge models, proprietary models most of the time,
24:28it's probably going to...
24:31It's like if we were pushing on the brakes
24:32when it comes to innovation
24:33because monopoly usually or quasi-monopoly
24:37is usually not good for innovation.
24:38And our idea when we say democratize good machine learning
24:42or democratize good AI is on both sides,
24:45both democratize in the sense of provide the capacity
24:49and provide the ability for everyone,
24:51literally everyone,
24:52to build AI solutions,
24:54not only developers,
24:55not only machine learning engineers,
24:56literally everyone.
24:57So provide the explanation,
24:59provide the content,
25:00provide the training,
25:01provide the assets,
25:02data sets, for example,
25:04provide the models
25:04in order for anyone to start playing with it
25:07and potentially innovate,
25:08find a solution to something
25:09that we have not been able
25:10to find solution to previously.
25:14That's one thing.
25:15So democratizing in the sense of
25:16let everybody play with it,
25:17understand how it works,
25:19leverage it,
25:19and find solutions.
25:21That's the first thing.
25:22The other thing is
25:22when we say good machine learning
25:23or good AI,
25:24it's exactly what you are describing.
25:26It's as much as possible
25:29explainability of the models
25:30that we are talking about.
25:31Being able to track the data sets
25:33that have been used
25:34in order to train those models.
25:36That's very important
25:37because potentially you might be using
25:39proprietary models
25:39who have been trained on data
25:40that has not been acquired
25:42in a legal manner.
25:43Do you want to be involved in that?
25:45Probably not.
25:46So especially if you are a company
25:47because potentially
25:48you could have some liabilities,
25:49et cetera, et cetera.
25:50But at the end of the day,
25:51to answer a question,
25:53our idea and our vision
25:54is that open source
25:55allows you to do both,
25:57democratizing AI on one hand
25:59and making sure that
26:00as much as possible
26:01we are talking about good AI,
26:02AI that respects a certain number
26:04of values, guardrails, et cetera.
26:06So why open source?
26:07Simply because it means that
26:09you know what the model
26:10is what are its characteristics
26:13on what it has been trained,
26:15how it can be improved in the future.
26:17You can potentially even
26:18contribute to improving it.
26:20So that's our idea
26:22about good AI on one hand
26:24and democratizing the whole thing
26:25on the other hand.
26:26Just simply in order to keep control
26:28of our AI destiny as a group, right?
26:32It's not only about one monopoly
26:33being able to deliver solutions
26:35for everyone
26:36and then someday
26:36they can increase the price,
26:38someday they can make it
26:39and that to reduce
26:40that infrastructure
26:40rather than this.
26:41Keep it as open as possible.
26:43That's the idea.
26:43I fully agree.
26:44I think we are sharing
26:46the same excitement
26:47and the same concerns
26:48about monopolistic practices,
26:51opacity,
26:52not knowing what's going on.
26:53I think the tooling
26:54to be able to understand
26:56how those LLMs are behaving
26:58is going to be very important
26:59and we need to continue
26:59to invest on tooling,
27:00not just those LLMs,
27:02but to be able to inspect them,
27:03understand how they behaved.
27:04and I love your commitment
27:06to open source.
27:07I think that's super important.
27:08Maybe, Fred,
27:09a last question for you.
27:10In your opinion,
27:12so what do you think
27:13are the challenges
27:14that lie ahead of us
27:15in terms of propagating,
27:17scaling this AI for good approach
27:20and what's the role
27:22that you think companies
27:23like Hugging Face
27:24or Orange can play?
27:27Yeah, thank you.
27:29So I think there's an opportunity
27:31where I think a couple years ago
27:33we just passed the point
27:34where more than half of people
27:35in the world are connected
27:37to the internet.
27:38So, you know,
27:39congratulations to the industry
27:40for making that happen.
27:41But one challenge is that
27:43even though you have
27:44a lot of coverage
27:45and connectivity,
27:46quite often there are bottlenecks
27:48for people to basically connect.
27:50There's no incentive
27:52for someone to connect
27:53if there's no content
27:54in their local language
27:55or if they can't read
27:56or they can't write
27:56or they have disabilities.
27:58So I think there's a huge opportunity
28:00where we can use AI
28:01to basically create content
28:04in, you know,
28:04thousands of languages
28:05and dialects,
28:06some of which are not even written,
28:08where we can help people
28:09with disabilities to participate,
28:11you know, text-to-speech,
28:12speech-to-text, you know,
28:13there's all different ways
28:14of helping people
28:15with disabilities.
28:16And, you know,
28:18even though it's encouraging
28:19where we are today,
28:20if you think of it
28:21as a V8 engine,
28:22we're only firing
28:23on about four cylinders now
28:24and not eight cylinders.
28:25So we're not benefiting
28:27from the collective
28:27participation
28:28of the people
28:30who are not connected
28:31who could be contributing
28:32to problem-solving,
28:33to storytelling,
28:34to culture,
28:35to creating solutions together
28:36and basically adding value
28:38in a way
28:38that maybe we just take
28:40for granted.
28:40So I think a challenge
28:41to the industry
28:42is not more coverage
28:44for the sake of coverage,
28:46but how can we use AI
28:47and tools
28:48to create solutions
28:49and incentives
28:50to move people online
28:52so that we can all start firing
28:53on eight cylinders.
28:55Excellent.
28:56So, you know,
28:57I know we're getting
28:57close to the end.
28:59Maybe I wanted to test,
29:01share my takeaway
29:02with you guys
29:03and see if you guys agree
29:04or don't agree.
29:06We've been asked
29:07at Orange multiple times,
29:09you know,
29:09should we go faster,
29:10should we slow down,
29:11should we stop?
29:12and we are of the opinion
29:14that we should not slow down,
29:15we should not stop
29:16looking at my chief AI officers.
29:18Good.
29:20And on the contrary,
29:21I think we need more
29:23than ever
29:23to double down on AI.
29:25We have trained
29:25just at Orange alone
29:27over 30,000 people in AI
29:29because we believe
29:30right now
29:31it's smarter for us
29:32to invest on people
29:33than billions of dollars
29:34on GPUs.
29:36too expensive
29:39and others
29:40are doing that already
29:41and that's good enough.
29:43We understand
29:44the concerns.
29:44By the way,
29:45I'm very concerned
29:46about the footprint
29:47on the climate change,
29:49you know,
29:49and the fact that
29:50from a scope 3 perspective,
29:51all of those data centers
29:53and the fact that
29:54those miles
29:54are becoming more
29:55and more complex,
29:56we need to understand
29:57what's going to be
29:58the impact on the planet.
30:01but I have to say
30:02that at the same time
30:04when we think about
30:04all of those pitfalls
30:05and dangers
30:06and all of those things
30:07and those question marks
30:08and the bad behavior
30:09of some of those companies
30:10that have decided
30:11that they just want to go fast,
30:12don't want to go safe,
30:14I'm looking at you guys
30:15and at the same time
30:16I'm seeing you guys
30:17being forces of good
30:18trying to make sure
30:19that we don't do stupid things.
30:22My conclusion is that
30:23we need to continue
30:24to go fast
30:25and at the same time
30:26make sure we continue
30:27to learn,
30:27adapt,
30:28being very agile
30:29and we can go fast
30:30while going safe
30:32and going smart.
30:33Would you agree with that?
30:34I'd say definitely yes
30:36and typically
30:37just last week
30:38we announced
30:38an initiative
30:39that we call ZeroGPU.
30:41ZeroGPU is something
30:42that should allow
30:43academics,
30:44independent developers,
30:46startups
30:46to get access to GPUs,
30:48GPUs that are shared
30:50among all of those projects
30:51because the thing is
30:52you do not always
30:53use your GPU
30:55so potentially
30:57while you are not using it,
30:58potentially someone else
30:59can use it
31:00so by sharing it
31:01you avoid building
31:02additional GPUs,
31:03you avoid the entry barrier
31:05so typically reducing
31:07or at least improving
31:10the usage of your infrastructure
31:12is probably something
31:13that can reduce cost
31:14both on the financial side
31:15and on the environmental side
31:18and also
31:18while the models
31:19are open source
31:20it means that
31:21we are able to better
31:22understand how they work
31:23in order to make sure
31:24that we are reducing
31:25as much as possible
31:26the environmental cost
31:27that goes with them
31:28because when you are using
31:29a proprietary model
31:31you will get numbers
31:32but how do you verify
31:33that those numbers
31:34are right
31:34like in order to see
31:35what's the cost
31:36what's the impact
31:37in terms of environmental cost?
31:39Fred?
31:40Yeah thanks
31:42I think maybe
31:43a final word would be
31:46you know quite often
31:47the question is
31:48AI figured
31:48what is good
31:49and then
31:50if we asked everyone
31:51in the room
31:51to come up with
31:52what is good
31:52we'd probably be here
31:53for two weeks
31:54and still not have
31:54a common understanding
31:55of what is good
31:56so what I'd like to say
31:57is we have
31:58the United Nations
31:58Sustainable Development Goals
32:0017 goals
32:01160 targets
32:02covering all of the
32:04basically goals
32:05that we want to achieve
32:06so I would say
32:07that would be
32:08the starting point
32:08of what is good
32:09but then I think
32:10also a culture
32:11of openness
32:12you know
32:13very much going down
32:14the way
32:14Hugging Face
32:15and their community
32:16is going
32:16and putting these problems
32:18to the crowd
32:19like we saw
32:20developing countries
32:21are very important
32:22for ITU
32:23and we've launched
32:24machine learning challenges
32:25in these parts of the world
32:27that they need data
32:28they need cloud
32:29they need compute
32:30but the talent is there
32:31the enthusiasm is there
32:32and we saw them
32:34just with one challenge
32:35reduce base station
32:37energy consumption
32:38by 20%
32:38I'm not sure
32:39that would have happened
32:40organically elsewhere
32:42if we didn't launch
32:42that kind of open challenge
32:44using the power
32:44of the crowd
32:45so I would just encourage
32:47more work
32:48more openness
32:49using the SDGs
32:50on the framework
32:51and hopefully
32:52we can go
32:52from the right direction
32:53excellent
32:54it was a real pleasure
32:55and honor
32:55sharing the stage
32:56with you
32:56thank you very much
32:57thanks for your time
32:58thank you
32:59have a great day
32:59thank you
32:59thank you
33:03– Sous-titrage FR 2021
Commentaires