← Back to search

Les Technos: Episode of April 7

Les Technos · 2026-04-07 · 96 min
relevance 0 17447 words Episode page ↗ Audio ↗
Show full episode description
Episode S12E15 avec Sébastien B., Xavier, Sébastien S., et Thierry. • (00:00:00) : • 404 bull not found (00:01:36) : Ou quand la France prend quelques longueurs d'avance sur l'Europe. (Source : theregister.com ) • Android, l'illusion de l'open source (00:17:54) : Quand on se rend compte que Android fait du open source washing. (Source : ) • Artemis 2: quels objectifs pour cette mission lunaire? (00:31:57) : Quels sont les objectifs de la mission Artemis 2 et pourquoi l'Homme veut-il retourner sur la lune? (Sources : nasa.gov , wikipedia.org et lopinion.com ) • Hermes Agent (00:41:09) : Quand c'est pas un sac de luxe mais un nouvel acteur "contre" OpenClaw. (Sources : nousresearch.com et geeky-gadgets.com ) • Utiliser un agent IA de codage au quotidien (00:49:40) : Les bonnes pratiques apprises apres un an d'utilisation quotidienne. (Source : github.com ) • Le Raspberry flambe (01:10:07) : Augmentation de 20 à plus de 100$ pour les PIs, la fin d'une ère? (Source : frandroid.com ) • Voxtral débarque, vive la France ! (01:16:24) : Quand il faut être fier de la France et de l'europe en général dans le monde du TTS. (Sources : mistral.ai et msn.com ) • Wine 11 pourrait convertir les gamers à Linux (01:28:45) : Wine 11 prend en charge NTSYNC et booste les performances des jeux windows sous linux. (Sources : gamekult.com et zdnet.fr ) Retrouvez toutes nos informations, liens, versions du podcast via notre site <a href="https://lestechnos.substack.com/" title="NewsLettre zero spam" target="_blank" rel="nofollow noopener noreferrer
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
In our 15th episode of season 12 of Technos, France buys back its bubble and takes a few steps ahead of Europe. Android pretends to be free. Heading for the Moon with Artemis 2, man wants to return there, but for what purpose? The Hermès agent shows up and no, it's not a handbag. A year with an AI agent to code, spoiler, we can no longer do without it. Raspberry is blazing, Voxrail is singing French and Wine11 is converting gamers to Linux. And for this episode, this 15th episode of season 12, we have with us the two Sébastiens. Hello Sébastiens. Good morning. We have the Xavier. Hi Xavier. And we also have Thierry. Hi Thierry live from Switzerland. I specify this each time because indeed, we are a predominantly Belgian team, even if, for once, we find one of our Sébastiens in a hotel room somewhere in Paris, right? In Genavière, in Paris today. What a trip, me. That’s how it is, he’s a tech globetrotter. If you want to see him in concert, he's at 12, 13. Yes, because he sings too. Yes, Balibale, not that. Balibaleau, and there you have it. A new episode begins. We will try to make it shorter than a month ago. We did it a little hard. But hey, when you love, you don't count. That’s how it is, too. We have a bit of a busy drive, so we're going to go straight away. And we start with Sébastien. Which one, will you tell me? Sebastian. It's the bearded guy. Sébastien Lebarbu, to talk to us about what, Sébastien? It's technology, of course. It’s Europe and France in particular that you’re talking to us about. Yes. You see, today, I am proud of France. I speak French. So, it’s actually a very interesting subject. It’s the bubble redemption. I don't know if you saw. Yes. By the French government. So, it's not every day that a government says it's buying a company. We're in Europe, anyway. And they announced the buyout for 404 million euros. It's fun, nonetheless. But they still found the 404 million euros. And so, they find themselves a little bit with something that they had lost. So, it’s a little bit of history that I’m telling you again today. It's not just any box, bubbles. I don't know if you know bubbles. Let's see the thing again. Yes, I see the logo. Giant Atos. And for France, it is the company that manufactures, and I mean manufactures, the supercomputers for their nuclear tests. So it's not... It can be useful. And it is the only supercomputer factory in Europe in Angers. And it is also the builder of Jupiter, a supercomputer exasperated, therefore a big beast, and which, with the fall of Atos, could come under a foreign power. Which, obviously, could not necessarily please France. Let us point out in passing, Sébastien, that these computer tools are for testing nuclear weapons solutions, without blowing up bombs in atolls, in the Pacific for example. These are tests, but done in the laboratory. And it's quite rare. It's a supercomputer. Yes, that's it. So, these are machines that are there to calculate, but also nuclear tests and the weather. It's really... For example. Or lots of other things. Or mine crypto, crypto. But hey, you never know to make money, because you have to make the 400 million profitable. Yes, 404. And so, there is still a bit of politics. So, Bulle, we have to go back to the early 60s, Cold War. And France is trying to develop everything related to nuclear deterrence, the need for large computers, etc. And at the time, the Americans refused to sell these machines to France, national security, their national security, the Americans. And De Gaulle launched the calculation plan. Bubble. So, it was a French program to be able to manufacture their computers. They created CII and injected quite a bit of money. And unfortunately, it doesn't work crazy either. The machines are quite good, but commercially, compared to what IBM was doing at the time, France is not really doing well. And opposite that was Honeywell Bulle. And that's what De Gaulle kindly wanted. And a little later, indeed, under Mitterrand in 82, privatization in 94 and 2014, Atos bought Bulles. Atos bought it for no less than 620 million euros. So, 12 years ago. And the CEO of Atos at the time was a name that we know very well today. We hear a lot about it in the media. Thierry Breton. Oh yes. So, you know... Who is the former European Secretary of Commerce and who is banned from staying in the United States, it seems to me. Quite. So, now, indeed, we see a lot in the media for many things, war or otherwise. And so, we see it quite often. And I mean the guy is still far from stupid. And sure enough, he ran the machine for a while until he left the club. And there, indeed, it starts to work a little less well. Fall of Atos, indeed. Under Breton, the company bought a lot of stuff. Siemens IT Solutions, not all Siemens. Xerox IT, Bubble. And I mean, the company has grown quite a bit. But a bit like another European giant, Worldline, which has a lot of problems at the moment. I don't know if you see a bit of the... Yes, it's the banking system that takes care of it. That's it. One of the biggest payments which bought back quite a few boxes in Belgium, France, and Switzerland too. And which, indeed, are companies which have bought back enormously, but which have had difficulty consolidating. This is what happened. So, they mainly consolidated the debts. And eventually, at some point, the thing actually collapses. And Atos collapses. And we are, indeed, arriving at a restructuring where the company is suffering. And there, the State is worried, saying “But the equipment I need for my power plants and for the simulations can actually go abroad. » And then, which, ultimately, is a pretty good deal for the French state, because it bought the thing for 404 million, which is still significantly less than the 620 million in 2014 that Atos bought at the time. So, what is France buying back? Obviously, the hardware. So, they bought a factory in Angers. The only factory... It's important for, obviously, because, as a result, it's only one factory. It is a factory that makes it possible to create supercomputers. It's not just supercomputers. They have the entire production chain. So, there is a form of nationalization, of know-how, of society, and of being able to take control of one's industrial heritage. New buildings... Well, they have all the buildings... Well, yes, there is a new building under construction which is also planned for the production of supercomputers. They knew about Jupiter, which is also in their fold, thanks to that. So, a supercomputer, Exascale, which is in Germany. So, Exascale, for those who don't have the reference, is a billion billion operations per second. So, that's still a few things... That's really touching. And they have quantum. So, Bull also has a quantum computing division. So there, it's a bit of a bet on the future of the future, but hey, after all, that's not directly why they bought it, probably. It's a bet, but for 10 years from now, but it's also a bet to be in quantum technology if, one day, it really becomes the most important point for digital sovereignty. So, sovereignty, indeed, is not 100%. So, Bull, this is a whole part of the solution. But in everything they use, there's still quite a bit of American material. They use Nvidia hardware, A100, H100 and others. And that, potentially, tomorrow, the Americans can say that we are no longer selling them to France. We are not at that point today. The Americans are not France's greatest friends at the moment either. Well, it depends on the weeks, days, hours. Yes. But we have no equivalent. France does not yet have an equivalent in this area either. What China is actually trying to put in place with its manufacturers, Huawei and company, where they are trying to have solutions that are 100% Chinese. It's not yet won, but they are putting the money into it. What I find quite interesting on the part of France is that they are really trying to have this sovereignty over everything globally. And we see that, for the moment, it is quite important. We're not going to go into all the warlike subjects and the like. That's not the point of the podcast. We stay on the technological side. But at the same time, on technology, it is also a form of pressure between different States and a form of power. And France is the European country which, for the moment, and I think we will talk about it again a little later in the podcast, with another non-French who will make France coorico. And therefore France is equipping itself with the tools it needs for its sovereignty tomorrow. So after that, what do they have left? With Bulle, the problem of a company that has been bought by a State is that it is an engineering company. Bulle, it's not... There you go, it's still tech. And a society that is managed today by a State is not really their custom in Europe. And it's rather much slower, at least for technology, than what you can get as a response from a... They lack agility, the wings. It lacks agility, it's... It's necessary... It also lacks, sometimes often, flexibility in salaries, in commitments, in many things. So, we must hope for Bulle and for France that France continues to succeed in keeping this form of independence in Bulle. Otherwise, it will be a takeover that will end in fiasco in the coming years. But nevertheless, it's... There you go. It's... It's a flagship that remains... Which will remain in the hands of France today. And finally, the loop is closed between what De Gaulle wanted to do in 60 years, to have this form of sovereignty over these machines and to have recovered this society with a part of companies... Well... How to say? Employees who are all over the world. But so, it all comes back to France and the manager of France. Which is a good thing, even for Europe as a whole. It's not bad that there are European champions within Europe somewhere, whether it's France, Germany or others. For the moment, France is what is closest to us. So, we see it more closely. But others also certainly have technological solutions to put forward and flagships to develop. No ? So, someone has to do it. It's good that France is doing Karicoco and displaying the thing. They're not the only ones. We already talked about it, I don't know, a few years ago. It was rather... Well, there are the Dutch who have very good companies and the Germans. And they're rather... They're not quite in the same mood. In general, what happened in recent years was more... Sales to foreign powers were prohibited. There were German boxes that wanted to be resold to the Chinese. And there, the government said no, the Wey would not be sold to the Chinese. We stay... But from there, putting our hands in our wallets to say it's us who are buying, I think it's a step which is a little more interesting and which could give ideas to Europe by saying well, afterward, we finance perhaps only for certain things which we have to nationalize. A reaction, I don't know from whom, I don't have anyone to raise a finger. First of all, Bulls was a great name in French industry in the 70s, 70s, 80s, a competitor, or placed well enough to compete with IBM, to compete with the giants of the manframe world, Fujitsu in particular, and then the alliance with Honeywell that it was in Arraso, remember. What worries me a little is that, as Seb said, the shareholder will be the State, whether Belgian, French or Swiss, the State is not a super good shareholder, its goal is not to make money. No, but it is successful at Airbus, for example. It's partially private, Airbus. Yes, it's partially private, but less so with Full Télécom, you see. Yes, it's true. For example, yes. And so, what is the intention? Is the intention to develop a powerhouse, how to say in French, sorry, a European commercial superpower? If so, it looks very bad because of history, what states have done with companies. Or is the French state buying a technology that it will develop and use for itself, for itself alone, for its nuclear program and for a question of sovereignty which was also addressed by Seb? In this case, he buys a technology, he buys a toy, the goal is not to sell it to others, but to ensure that no one else will take it and that they can continue to develop. And there too, it raises a question, a doubt, an interrogation, it is to say to oneself is it the best way to have cutting-edge technology than for a State to own the company which manufactures this technology? Finally, how will this company remain at the forefront of innovation or will it continue to innovate and be relevant in this area with the State as a shareholder? So, business to follow, I find it a bit of a shame, I had colleagues who left for Bull at a certain time, I knew some and there you go, it was a beautiful image of the European IT industry, don't smile, I know there are some who will smile when I say that, but it's a company that worked, you shouldn't always be jealous of companies that work. And so, there you have it, I was already worried about his future with Atos, but perhaps even more so with the French government. That remains to be seen, to be continued. Xavier had the thumbs up. Yes, but the fact here that a State is invested in a company which is highly technological, one can wonder if it would be a good shareholder. So, there are points that are positive in that because we say to ourselves ok, they can protect their investment in quotes in the long term. They can invest where private individuals would perhaps want a much faster return. And then, there are flaws too. This is for example the fact that a large structure like that, an administrative structure, we know that it is not very agile. So, they are not always the strongest in everything that involves setting up rapid procedures or accelerating change, making decisions, etc. So, isn't this going to transform this company into a tech company basically, into a huge project really based on administrative procedures? So, that’s one of the points I’m wondering. And so, in terms of AI, are they trying to control the infrastructure here, keep all that in Belgium or do they simply want a certain sovereignty? There are questions for me that arise from this. We bought Bulle. We didn't buy Bulle. Xavier, you can put your hand down. So, the infrastructures are in France, not in Belgium. We looked for hospice, we had my... Yes, for example. Thank you for redeeming this painful episode. Does anyone have anything else to say before moving on? There you go, just picking up on the fact that perhaps the current climate is causing us to wake up to sovereignty. But I say that, I say nothing, I am Swiss, I am not part of Europe, I remain neutral. No, no, but you're right, that's what we need to highlight, that there are reactions. So, it goes a little in all directions, we agree, but we can't criticize that, honestly. Fundamentally, in my opinion, the idea is a good one. Do you want to conclude, Seb? Yes, and to also make a bit of a link with a future subject, France, for the moment, is doing big things, so whether it works or not, I think we will judge, I am just as doubtful as Xavier or Seb about 100% state shareholding, nevertheless, we have to try because I think that in any case, otherwise, they were going to lose it. So, it's between sending him to a foreign power or making him survive, I think that France had little choice, but it's the current movement, like they have Mistral, like they have VH, like they have lots of other things and it's really at the least a movement in France which, in a certain way, commands respect from Europe does not do it, but France is moving. Yes, that will be the conclusion of this topic. If you too have your conclusion to give, do not hesitate to let us know in the comments or you know how to contact us, it seems to me, via techno.be, for example. Another Sébastien speaks to start this episode off right. We're going to talk about Android and the illusion of open source, Sébastien. There, we fall, we fall from our chair. Oh well, open source washing. I listened to a podcast, a podcast other than techno this week or last week. The faithful one. And no, but it opened my eyes to a subject which is the presentation of Android as being the open ecosystem, open source, open marketplace, etc. And it made me ask a lot of questions that I wanted to talk about in tech. In fact, the starting point is to say, you know, your smartphone knows everything about you, your location, your contacts, your habits and everything goes through Google or Apple. And there is another French person, definitely, it's a very French episode called Gaël Duval who offers a very concrete alternative that we can use today called iOS and with a company called Murena. I think we've already talked about it a little bit. I wanted to go into a little more detail in the techno. Gaël Duval is not just anyone. He is the creator of Mandrake Linux. If you were an early adopter of Linux in the 90s, in 98, he created his distribution, a kind of retat fork at the time but really designed for the general public with a revolutionary idea that Seb will love. This is because Linux is not just for bearded people and it must be able to run on all devices with a simple installation, plug-in-play, etc. What's that? I used it. See? You weren't bearded back then. Since the bearded ones. The name was also the name of the magician Mandrake. Besides, making the magic of Linux accessible. It's more hairy too. A Mandrake Linux means millions of downloads and 800,000 subscribers to the mailing list. It was truly one of the most popular distros in the world at that time. So with iOS and Murena, he’s back at it again. iOS is an open source mobile operating system developed by an e-foundation, created around an almost political charter, I'll come to that. Then Murena, a commercial company because you have to pay your bills, which sells phones with iOS, iOS, in English, we mean EOS. So it's written slash E slash OS. So if I say iOS, don't get confused, we're not talking about Apple here. So Murena sells pre-installed phones, but also cloud services, the equivalent of Google Cloud or Apple, what is it called cloud? iCloud. So the basic vision is a manifesto, a manifesto, a kind of contract where it says, there are a few athec giants who control our OS, our data, our communications, it's not in people's interest. I want to offer an OS, a basic service, emails, communications, and essential public infrastructure without going through these multinationals. I quote the manifesto, I open the quotation marks, the current situation is unprecedented in the history of humanity, it goes against freedom and can lead to forms of slavery and totalitarianism on a global scale. So when I said that it's a bit political their objective, but that's really what their objective is, it's to offer an alternative that can be used by normal people, not geeks and as simple as possible. And what's wrong with Android? He says Google sells Android to us as open source, that's very good, but Android is based on the open source part which is called Android Open Source Project AOSP and he says it's open washing because the development is not done on AOSP, the development is done in closed source at Google and from time to time they drop a copy of their source on AOSP. and AOSP is really the basic layers of Android, there is no Google Play Service so there is no location, there is no push notification, there is no update, there is no Play Store so you cannot install new applications, there is no browser, there is no Chrome, there is no email application, there is no map, Google Maps, it there's no YouTube, there's no phone application, I fell out of my chair when I heard that, it's still stupid for a phone, it's not open source, it's there, so it's not in AOSP and so we know, we've already talked about it as much in tech as the Chinese, it's Huawei I think, remade a complete OS based on AOSP, Murena is the same thing, they made a complete OS with all those apps that are missing in AOSP. It goes further than that, it's also the DNS, by default in AOSP it's the Google DNS so any traffic that comes out of there, they know that you have traffic, the time servers, NTP, hours, connectivity check, etc. So, what is it in iOS? In iOS, EOS, I will pronounce it the French way otherwise there is too much ambiguity, it's a digouling. It's a bit deliberate, I take that, this ambiguity. It's iOS versus iOS if we pronounce it correctly, iOS for Apple, iOS for EOS, but like iOS in French there are people who tend to say iOS for Apple too, that's it. In short, complete digouling, we're de-Googling, no Google search engine, replaced with Murena Fine which is based on Quant, French, no Google Play Service, it's an open source implementation called MicroG, DNS server, it's no longer those of Google, time server, it's more those of Google, geolocation via an open source project called BeaconDB, the connectivity check, it no longer goes through a Google server, all Google apps removed and replaced by open source equivalents. But the thing that really sets it apart for me is that at the OS level, it blocks everything that can try to follow you. So, tracker management is done in real time, we see how many trackers each application uses and we can cut them one by one. The OS gives false location to applications. So, if an application wants to track you, you, the user, you decide, yes, I want this application to take my location because yes, I need it to know where I really am. or not, this application, I give it a false location, it does not need to know where I am. Same thing for IP addresses, the OS can give a false IP address to the application, ad blockers are obviously activated by default and the Cloud service provides a hide my email as it is on Apple with an alias, a false email address which is generated on the fly where the email is redirected to your email. So it's really a connection with privacy, how do you say privacy? Privacy by default. No Cloud connection required, everything can stay local and if you want a Cloud connection, then you can connect to what they call Murena Workspace which is a kind of bundle of open source Cloud services that they host and you have to trust them at some point. There you go, if you want Cloud services even if it's open source, it's hosted by someone, you're not going to self-host it on your NAS at home. So you have to trust NextCloud. NextCloud is also an open source thing, it's German, it was created in 2010 by, what was it called? Frank Karalitschek. I'm being smart but I have my notes. I didn't know that by heart. The initial idea was to create a sort of open source self-hosted Dropbox Google Drive but there has been quite a bit of success, well I'm not going to go into the whole history of NextCloud. In 2016, 10 years ago, he finally slammed the door on his own project and created this company NextCloud which markets a distribution, therefore ultimately hosting this project but it is much broader than Google Drive. Now there is storage obviously, there is email, there is calendar, there are contacts, there is video, there is an office sheet with OnlyOffice. So all of this, we can access it via what they call Murena, Workspace, at least the parts that they like because there are parts that they don't like, I think it's Talk, NextCloud Mail is not used because the Murena people find that it's not secure enough on the NextCloud side so they don't take everything systematically. So if I've convinced you, first of all I don't have any action, I'm not trying to sell it to you. Two, how do you get it? The easiest thing is to buy a Murena Phone, you know, it's a Murena and they sell a whole series of phones pre-installed with Murena. Then there are obviously the Fairphones which we have already talked about a lot, so it is an ethical phone that can be dismantled, repaired, etc. But there are also Samsungs, there are also Google Pixels pre-installed with EOS, ready to use. They have a WebInstaller which looks super nice, I don't have any hardware, I haven't been able to test it, but you plug your phone into a USB socket and from the web console directly, it detects if the model is compatible or not and it launches the installation process or if you are more geeky, you can download from GitLab and do the installation. But there are more than 250 phone models available, even old phones, so there is an ethical, Green IT aspect, even more, we are talking about a Samsung S4 Mini from 2013 which works perfectly with Murena. So it's not putting Apple versus Android, it's rather Android versus the open source community and against respect for privacy rights. So there are the Apple haters who will say “Yeah, but Apple isn’t any better because they too are closed, etc.” » But I still remind you that there is a fundamental business difference between Apple and Google. Apple sells hardware and makes money from it, so they don't resell your data. And services. And services, you're right. But they do not resell the data, I want to emphasize that, while Google, its business, is to know you better in order to sell advertising and therefore it is to resell your data, in any case to exploit your data. And so this is why some users want to de-Google their lives to the point of changing the OS. We're talking about 100,000 users. Oh yes, that's good. And phones, if you choose Fairphone, which can be repaired for up to 10 years. And there you have it, European, French, cocorico essays, in these days when we are talking about sovereignty. Attention Sébastien, Fairphone is a Dutch company based in Amsterdam in the Netherlands, founded in 2013. It is a European social enterprise specializing in the design and production of sustainable, ethical and modular smartphones aimed at improving working conditions and environmental impact in the supply chain. Anyone want to react? I don't know who. Yes, maybe I just wanted to clarify. There you go, a Swiss. Neutrality requires. Just clarify one thing. I believe that the business model, well, the biggest money coming in at Apple, is no longer really the hardware. It’s the services, it’s the subscriptions, it’s all these ecosystems. Yes, they manufacture devices, but we are about to see a behemoth that earns money with the percentage of what they receive on apps, on subscriptions, on all that. The music. I remind you, there you have it, the music. I also remind you about the credit card and that sort of thing. And then, this whole project makes me think of something that I haven't had time to test yet. It is indeed Linux which is arriving on smartphones because there are more and more projects, little tutorials which show that it is super simple to install, even without necessarily running. Yes, we had already talked about that. It's not new, new. We already talked about this a few years ago. We're going through a pretty interesting moment. This is because the switch to a third alternative is being made. It's starting to become popular. What do you call Linux? Because Android and AOSP and Murena, it's a Linux kernel in the room. Yes, when I say Linux, it means seeing people who are having fun with Ubuntu to adapt it for mobile, for example, because we are on... So, there are oldies in the room like the other Seb. Indeed, Mandrake, that tells us something. But for those who are starting to be interested in OS alternatives, there is often Ubuntu which comes up as one of the names that people spontaneously think of. If I may say so, the mantra of the creator of Murena is to say it's the apps. For an OS to work on a smartphone, you need an ecosystem of apps and this is the advantage of Murena because it can be plugged into Android apps and all Android apps, almost all Android apps will work on Murena. While his other OSes, and he spoke in particular in the podcast about Firefox OS and others, were never able to start because of this lack of application. Often, that's it, yes. It's true that we have to rely on an existing ecosystem with proven apps, even if they are selected, precisely as you explained, it still remains interesting because it's true that I hear people who say "Oh, I'm fed up with Android or I'm fed up with Apple and I would like something else, a third option" and when we present that to them, what's the app? My banking app, for example, will it work? Things like that which have now become essential. To answer this question, it's no, not all banking apps will work on Murena because there are certain apps that use Google APIs to verify that the OS is not modified, therefore that there has been no change, that it is not a jailbroken phone for example and in these cases, they cannot necessarily guarantee that the applications will work. So it depends on the way in the apps, but typically, they were talking about banking apps that could be problematic. That would be the first thing to do if you want to switch to Murena, is to test if your bank's app works before making a choice. We invite you to take a look at the site. You put it in the podcast notes for us, Sébastien, and there you will have all the information how you get it, this OS, what is it capable of? I'll get there. Everything is explained and in French, finally it's multilingual, the site is well done. And indeed, it is within the reach of everyone, Mr. and Mrs. everyone, or almost, we will say, to fully understand the interest of the thing if it interests you at least. It's Xavier's turn to come with one of these topics of the day, Xavier, and which is totally in the news. As we record this episode, we haven't lit the candle yet. I would like to clarify this, that is to say that Artemis has not yet left the cow floor. It will be in a few hours as far as we are concerned, since we are recording Wednesday April 1st and it is currently 9:28 p.m., so you know everything, and takeoff is scheduled for tonight around what time? midnight 24. Midnight 24. There, so Artemis 2, it's the manned version of Artemis 1, I want to say, and for the oldest, if they remember the Apollo program, it's equivalent to Apollo 8, if I'm not talking nonsense, that is to say this manned flight which goes around the Moon and then returns without landing, indeed, but it's still a preparatory flight obligatory that you are going to tell us about, Xavier. Yes, so here, we are indeed talking about a first manned flight around the Moon in a little over 50 years. I think it was in 72 that the last human crew went beyond low Earth orbit and therefore, to approach the Moon, so we're not right at 50 years, we're more or less 53 years, something like that. 53 years old, exactly. So, what is Artemis 2? In fact, this is the first manned flight of the Artemis program, but it is not a moon landing, so we are simply going to do a test flight to qualify Orion which is this capsule which allows us to house 4 astronauts in the cockpit with the launcher, therefore the SLS rocket which will really be conditioned as if they were going to leave into deep space. So the goal is also to check a little bit our ability to go beyond the Moon and therefore directly Mars, it is Mars that we immediately think of. So here, it's going to be a trajectory with a free return around the Moon. So it's funny, if you go to the NASA videos, you can see the trajectory which has a completely different appearance if you just look in relation to the Moon or if you look in relation to the trajectory of the Moon around the Earth and the Sun, etc. So that's quite funny to see and we realize that the view is not at all the same. So, Orion is with humans on board. So the objectives of this mission also in particular are going to be to confirm that Orion works as planned with people on board in deep space. So it really validates that the management of air, water, CO2, thermal management, etc. is ready for future missions. With four people instead of three for the Apollo program. It's important to say it too. And above all that this will be able to be the case in a ship that actually flies because we can test it on Earth but it is still not at all the same as in real conditions with radiation, with an environment which is not at all the same, psychological pressure, etc. And above all for a period which will also be significant and far from Earth. So, they will repeat operations that really cannot be simulated on the ground. So after launch, Orion will make two Earth orbits to check the systems still close to us. and then into high orbit. And then, the crew will perform a manual piloting demo. So they're going to call it Proximity Operations Demonstrations. And the objective here is to gain experience for appointments, proximity operations, docking, undocking with other devices and for subsequent missions. So, they obviously also want to validate the fact that we take off, we fly away and then we come back because it's not all about taking off. We need to be able to bring these little people back safely as they face temperatures that are truly incredible when returning to Earth. And then, NASA also says that there is a bit of science behind all this since it will really be a way to see data that is useful during human flights. So, I wanted to take advantage of this opportunity which is the launch of Artemis to also say why does man want to return to the Moon because we have been there, we want to take a little photo but it's not just for that. The idea is, first of all, to prepare for Mars. So, the Moon to Mars logic, it is logical because the Moon could serve as a test bed for learning to live and work far from Earth in terms of logistics, autonomy, radiation, maintenance, procedure, psychological stages, etc. And for Artemis II in particular, we will really secure the next step, that is to say surface missions later in the decade. The second aspect is science, therefore the Moon as an archive of the solar system since the polar regions of the Moon and especially parts which are very little exposed to the Sun show areas which are little erased by erosion. we really have a history of the Earth-Moon impact and formation. And then, the third motivation is water since we know that the Moon contains ice with water in regions which are in permanent number since the Moon has a part which is never exposed to the Sun. it is also quite special for this type of element because it is really a coincidence that a certain part of this satellite is never exposed to the Sun. This is not the case for all objects in space. And why? Because we will be able to make oxygen from water because when it evaporates or when we transform it, we will be able to recover oxygen. Water in short because we need to drink it and then potentially as fuel because NASA says that it will be able to exploit lunar water to make it into fuel. Hydrogen. So there you go. So here, there is also an economic aspect. They think that projects on the Moon will be able to be an accelerator of innovation, of public-private partnerships and then there is the geopolitical aspect. Obviously, we are returning a little to the space race that we experienced a few years ago between the United States and Russia and it is a way of also establishing a lasting return to the space race. And for this aspect, if this little file interested you, in fact, I saw a direct connection with a series that I have just watched and swallowed in four seasons, it is For All Mankind. So it's a series which is incredible on precisely the space race between the United States and Russia and without revealing everything, in fact, the idea is that the Cold War between the United States and Russia did take place, the space race took place but it was Russia which was the first on the Moon and this whole story stems from there. It's science fiction but it's still very well done because at the start of the series we have the impression that it's a bit of a reality story except that there is a little twist compared to reality. I really liked it if you like space it's a great series. To return to Artemis II 10 days of mission if I am not talking nonsense there is a record already finally which is not yet as we speak which has not yet been attributed but which is the fact that it is the first time that human beings are going to go so far into space since they will in fact exceed the Moon by more than 9000 km which is not the case even with the Apollo missions and therefore it would be the first time that he needed a record to put to his credit and this will be it. afterward we're going to wait 10 days we'll talk again next month in our next episode with Chroniqueur to talk about it again because if that happens it might not even take off it might not even take off who wants to add something Thierry maybe a very nice series very nice series I'm not going to reveal the twist too ok well let's move on and the rest is happening with Thierry Thierry you're going to tell us about a pretty handbag yes everything fact we stay in France it's a Hermès it's a Hermès it's a Hermès but this one is the Hermès agent model no nose money money I told you in other episodes it's going to be very very short because indeed it's a big episode and it's not over talking about France but here it has nothing to do it's called Hermès agent it's the buzz of the moment on the networks to say it's better than OpenClaw so already if you are listening to this for the first time you are already lost and that's normal OpenClaw which had lots of names which is this base this multi-agent software to which we can entrust missions but which has a nasty tendency to forget things I pay the price for all the tests that I do and I was interested in Hermès agent which is sold to us more better more so we will say that every week there is something but this one I find interesting why? because behind the people from Hermès there are the people from us technology I think it is pronounced differently we research N-O-U-S like us and so they are basically LLM developers so they are not beginners in the field either and they have the pitch it is in fact an agent who grows with you because in fact he does not forget he learns he learns from his mistakes he improves he also sets up with a stupid command line and then for having tested it it's not a chatbot it's not an open-clos it's not a multi-agent but it's a bit of all that at the same time to make it very very short we install it on a machine which has a little bit of power but you can go to open-router for those who have fun with different LLMs it's a sort of bifurcation which allows you to use several LLMs there are more than 40 tools which are integrated so the search on the web the terminal navigation to go for a walk alone on the internet the generation of images of course the text to speech stay connected we will continue to talk about it later the execution of code because I know among you that there are some who are hooked on the idea of becoming developers or in any case having fun typing code it will help you it also does task planning and then it is obviously capable of connecting other tools that you would have like that on your computer or your server so the advantage of Hermès for having tested it a little but I will probably come back with a slightly more substantial test it is this palliative for the loss of memory which is still quite annoying with OpenCloud even if we can load it with the skills of the plugins to enrich it to give it all kinds of things at the same time I speak from my personal experience I dropped OpenAI Claude through his skills through his plugins and amazing and the more and more I advance and the more and more I say to myself but do I really need to go to these other tools where Claude is starting to bring together a lot of things I know that Seb who is currently in France I realize that I am the only Swiss the only one who is not who is not Belgian and who is also going to talk about France later well that's it was just a snub but in fact Claude is starting to be very very complete very interesting very how to say useful to go through to put to the test and then and there you go is Hermès going to be the xth trendy thing or are we going to see in the long run a Claude or perhaps an OpenAI you should know that Steinberg Steinberger the famous Austrian who was caught up by OpenAI with his OpenCLO well in fact he has means he has money and then they will perhaps do something with it so there is perhaps a future for OpenCLO or will it be the big players there once again Hermès publisher of its own LLM Claude which is starting to be more and more complete and then OpenAI with OpenCLO which is not going to die on the contrary the store of skill plugins which we can load more and more incredibly growing the little thing on which I will end with Hermès is that in fact you install yourself and it's all ready it's even already ready to know where to go and look for precisely these famous modules which will be able to increase these capacities these skills tell me in the comments I'm interested because I feel a little alone in test all these tools so I would like to know what you did with them and then off the record we were just talking you were able to see the newsletter like what in tech we are going to replace everyone with our avatars know that I am really working on this with OpenCLO here I have a factory to create podcasts I know that it will make our friend Marc cringe and the little hair he has stand on end but there you go stay connected we will continue to talk about voices it's a super interesting debate with another French novelty Hermès they weren't luxury bags at all but it's worth these questions if you work like a robot you will be replaced by a robot and that will happen don't worry and I also note this sentence it's still cult it's at one point you still said I just realized that I'm the only Swiss the only Swiss who talks about France ah yes yes a small state after Switzerland so that I want to use it if you want, go one more time let's go I realize that I am the only Swiss someone wants to react to this new Hermès agent someone saw it pass or not no I hadn't seen it so thank you Thierry for bringing it to the surface it will be an extra tool to try this weekend to play in addition Hermès I don't have the means yet I wish you a very very long night it's a new field it's not surprising to me that OpenClaw is emulating which is small there will be the corporate version which will come by OpenAI since Steinberg works for OpenAI now there will be other open source projects like that we will have to see the merits of each one for the moment the strength of OpenClaw is to be the first therefore to have a huge community because there has been all the enthusiasm around it we have to see how these open source projects will survive in the distance but it is good the more there are experiments like that the more we will find a use a case a business model finally there it is we will have to test several ideas several models before something emerges and I love this kind of pot which is bubbling at the moment with ideas and things which come out every week and if you are listening to us and you are a little disappointed by these stories of LLM agents etc. we put you all the links systematically in relation to these subjects in the podcast notes you will find all the sources that our columnists have unearthed to support their speech in each of the episodes and if you want us to go further far or that we go back a little at a given moment to re-explain the basics of all that I think that on occasion one of the Sébastiens will be happy to put things in their context so that you can hang up the wagons tell us if that interests you and we will do so because it is important because in the future obviously we are going to talk about it more and more we are not going to systematically return to the fundamentals in each episode and each subject that deals with artificial intelligence so it is still important that you are in tune if I may say so do not hesitate to let us know if this subject interests you and if you want us to simplify it a little so that you can, as I said, hang up the wagons and since we are talking about artificial intelligence in such a good way we will not stop Sébastien precisely to talk about AI coding agents which are now used almost on a daily basis you confirm it to us yes I wanted to come back a little to Thierry's column from last month I was going to say last week on vibe coding if you remember in fact what is there is that with colleagues we realized that we have been using this type of agent for at least 6-7 months in an extremely productive way to do real things and beyond the proof of concept of the demo etc and more generally we have been following this space since the beginning so for 2-3 years so I made a presentation for clients with 10 best practices to use these tools in an efficient way so initially it was a press for developers and then I said to myself but in fact it is useful for technos too because it is not only developers and Thierry you are a good example for that who can use these agents there these so-called coding agents like CloudCode like Kiro like Cursor like Copilot etc it is also even if they are tools for developers at the base they are also tools which can be applied to anyone who uses a computer as soon as you have files to manipulate as soon as you have a document to generate a document to summarize a presentation to read 3 YouTube videos and give you a summary in 20 minutes the key is not really knowing how to code the key is knowing how to dialogue with these issues to manage to make them do something I had to select 4 clients from 150 in a conference this week last year I had to read the 150 conference titles and the abstracts and the names of the clients etc. it was an agent who filtered this for me according to my criteria I spent a little time explaining the criteria to him he made me a list of 10 which I then reduced to 4 he wrote me the emails automatically he sent them automatically etc so even in traditional office automation it works and that's why I wanted to talk about it in technology initial observation it's a lot of users misuse this kind of tools because they go straight to the point do me that but do me that it's super vague communication between humans it's already super vague but a human he's going to ask us yes but what is that while the model him or the agent he's going to try to do that without really knowing and so he's going to make hypotheses for you obviously he's going to go in a direction what do you want to answer him in addition to answering him it doesn't work yeah it doesn't work and by the way the more context we give it the better it will be the more precise we are the better it will be but too much context it will kill the context also because remember a fundamental thing which will be necessary to understand the rest is that the models are stateless they do not keep state so if I ask a question to a model it answers me if without programming anything else I ask a question right after to the same model it has forgotten my first question so typically the demo that I do to clients I say who are you they tell me I am Claude Opus something etc and then I ask a second question what was my first question and if I program nothing else really use the APIs at the model level he answers me I don't know I don't know what your first question was so what do all the tools that you use when you use chat GPT etc it returns with your last question the whole history of the conversation so question answer question answer question answer and if there were tools as we call that so if the model ever said to my web browser to my agent to call such a tool or to call such a tool all the inputs and outputs so the inputs outputs of these tools are also put in the history or it turns out that a model has a limited history window it has a limited input window typically it is 200,000 tokens so almost 200,000 syllables for the moment some models go to 2 million it will change the situation a little but it is still relatively expensive and therefore we cannot give the whole context if I'm working on a big project or if I have big documents I can't put all my documents in this context window there are too many so there are techniques to manage all that so first lesson there are 10 I'm going to try to go relatively quickly because everyone keeps an eye on the clock also the first lesson is if you're coding or if you want to do a task don't start directly with that task but brainstorm so ask questions to the model say what would be the best way to do that or I want to do that here is the design that I have in mind is it a good design that I haven't thought about what are the modern alternatives etc so there is a whole discussion that can start with the model and once we have finished this discussion I like to say to him ok summarize for me from the discussion the decisions that we took together while challenging each other and save that in a file why? because it's going to be a history afterward it's going to be a bit of a memory of our project to which we humans can return but also the agents who are going to work on the project later they will have a context the second thing is to specify everything say here I want to make a podcast management app for example I want to manage the schedules I want to manage the columnists I want to manage the access rights so that not everyone can change etc so write specs in natural language and use models to generate specs in a more formal language than agents will be able to use after what we call user stories if you are a product owner product manager and there is a special techno series which explains to you this job of product owner which you can listen to again it will also provide stories as a user I would like to be able to do that with a whole series of conditions and test conditions also check that the state of the system remains good why do that because again it will give later at the time of implementation a context a guideline what are the specifications starting from the specification make a design and from this design make a list of tasks you have to do first this first that first that etc so there's no point trying to specify a whole new application from the start because again you'll get lost and then that's what we did 40 years ago in software and we stopped doing it because it doesn't work we can't spend 6 months or now 6 days specifying everything while trying to think of everything there are inevitably things that we won't have thought of so we have to leave on relatively small things and then iterate and loop over time another error is to choose models that are too small, not competent enough or too large, very competent but very expensive and so after my project or my POC they will tell me ah it costs too much because I used Claude Opus 4.6 which is the most expensive at the moment most of these tools allow you to choose the model and therefore for tasks of spec generation design generation we can perhaps use less expensive models maybe a little less precise faster but above all less expensive and keep the most expensive models the most efficient for really complicated tasks of debugging generation of complex code security code multiprocessed code finally things that require more thought is that okay? are you still following me? I'm at step 4 using skills what is a skill? it's a set of instructions that we're going to give to our model so if for example I want to work on a graphical interface a web interface or an interface with Swift in iOS for example I have a whole series of rules that I can fish on the internet I don't have to take them myself which say what is it to work on a modern web interface today what are the frameworks what are the tools what are the best practices what to do what not to do there are people with help of agents who have written documents like that and therefore go and inject these documents into the context at the same time as your question when you tell him to design a web interface it's as if we took an expert next to us in a specific field and direct all the power of the model which will generate code or generate specs in a precise direction that guided by the skill then you are going to tell me the skills if I have a lot of them it will pollute my context it will put more input but no precisely because the agents are quite intelligent and they only load these skills in the context when they are relevant, that is to say that at the start of the description of the skill we often say here it is a skill which allows us to build web interfaces uses there if there are the keywords this, this, that, that that is to say if I have a normal conversation with my model the agent will not send the text the content of the skill the specific instructions of my expert in the conversation whereas if I start if my conversation starts to deviate on aspects that are relevant to the skill he will know that there is a skill that is there and read it inject it into the context so that the model has more information are you still following me? it's clear ? always ok lesson 5 work on one feature at a time do not try to say there is the authentication which does not work and the page of this does not work and when I click there it does not work it is three different instructions it is better to focus the agent's attention again on one feature at a time then a feature a capacity is to correct a bug it is to add a new functionality or it is to do a task send a summary email etc so when we develop one feature at a time it can be a a little slow even when you are alone but nothing prevents you from opening several windows and in each window making a copy of your project there is something in the guide called the worktree guide which allows you to duplicate a project while keeping the same versioning database and to launch all these agents in parallel and make them all work on tasks in parallel on different copies of different directories of your code and then to merge everything at once but the point is always the same thing an agent a model will be more effective if it focuses on one task at a time lesson 6 is giving it feedback 6 and 7 even all at once i.e. if you tell it generate code without further instructions it will generate code we will not know if the code works it will be up to you to start it and see if it works while if you ask it to generate code I often or almost always use even languages compiled with strongly specified types so a string of characters is a string of characters an integer it's an integer we can't assign one to the other I can't say first name equals 6 that doesn't work 6 it's a number and it expects a string of characters it answers yes it's in javascript it's the worst of javascript and so if we give a strongly typed compiled language and we give instructions to its agent each time you generate code you try to compile and you see if it compiles if it doesn't compile you correct it yourself before notifying me by saying that you are finished because it is not true you are not finished he is very capable of reading compilation errors and correcting them same thing if we have a whole battery of tests I ask him to execute all my tests and that each change he makes does not break the tests and that he does not come back to me to bother me with a notification to say I'm finished as long as the tests do not pass you stay in the corner you continue working as long as you haven't finished so that's feedback during the session during this context precisely where he knows what he's doing he compiles he tests and he tells me that he's finished only when it works then there's also from one session to the next either when I'm working on two different features or when three days later or two weeks later I'm going to reopen my project I would like to continue adding things there the solution is to record record the specs record the design if you do a big modification I never at the end of my modification I tell him generate for me a summary of the main design decisions of the main implementations of the difficulties you had because it didn't compile you had to change that that means that he badly generated code at the beginning so I ask him to document that and save these files in a special directory of my project which will serve once again to be the memory of my project so that when another agent later it can be the same when I have another agent I restart Claude Code I restart Kiro three weeks later they can use that as a memory the context that's you're going to come across it at one point or another you type questions answers questions and then at one point he says ah I'm reaching the end of my context window so the agents are quite intelligent today they're going to make a summary of the whole conversation they're going to start a new session and give as entry to the new session the summary a summary it's a summary so inevitably there are operations which can miss and it will no longer be as precise to avoid that a very popular idea at the moment is to stop using finally it is to use subagents that is to say if I want to launch a relatively complex task I will tell it launch a subagent and a subagent it is as if it launches has a small separate process it will prepare its input its original prompt it is the sub-agent which will do its complicated tasks and read lots of files etc. and when the sub-agent has finished it will make a summary of what it has done which it will give back to the main agent, that is to say like calling a function in a program or calling a sub-routine but without polluting the context or without enlarging the context of the main agent so that has several advantages first of all we can launch several sub-agents in parallel in almost all the tools that I know so it goes faster and then it does not pollute the main context so it is really possible to orchestrate I like to take this image when I talk about the developer of 2026 he is a conductor of the orchestra we all have different instrumentalists who all have specialists according to different skills a web design a product owner a tester a back-end developer a front-end developer etc and we ask them to do things and I am the conductor with my little baton and I command them all and for the final result to be harmonious it depends on me it doesn't depend on them it's me who must give them the right ones instructions at the right time etc and I see myself now more as a conductor or a puppeteer chuk chuk chuk than a developer and so today you still have point 10 I still have I still have two but very quickly yes but go quickly no but I was going to say now is prompt coded? yes of course that's what it is but it's more prompt engineering it's context management it's how to generate the right context to direct the agent where I want him to go it's me who is on the saddle it's me who holds the reins it's me who pulls to the right or who opens to the right to use a riding figure when I want my agent to go to the right and it's not him who decides to go there and it's up to me to keep the reins in two more things always review the code and it becomes the bottleneck I was reading another study this afternoon which says the teams which use this kind of tools become much more productive there are many more features which arrive in production many more fixes which arrive in production but the problem is that we always have to reread the code and as the code becomes cheap to generate the bottleneck we have moved the bottleneck it is the code review so here too the agents can help us we can ask agents to reread the code I like to make Kiro and CloudCode work in parallel there is one which generates and the other which monitors what it does not always the same I change sometimes it depends on the tasks and the agents that I have finally the skills that I have on one side or the other Copilot does that very well on GitHub each time we send code to Copilot it launches a code review it is super picky it is super boring it will take small ones details to believe that he is paid by a comment that he has on the lines sometimes I tell him ok that's fine it's good forget it so what I do now I let Copilot do the code review he gives me a whole series of shit things and I ask another agent to go and evaluate all of Copilot's remarks and if it's relevant and they're not all relevant to apply the recommendations and change them and give me a list of the recommendations that are relevant and those that are less relevant then but it's maybe I'll do that for another column it's MCP or not MCP should we use CLI or not I'll save that for another column my message is learn to use these tools learn to manage the context we put them in the category of code creation tools it's not just for creating code it's for summarizing documents it's for generating documents generating presentations it's for brainstorming on any project not just code projects this are assistants analyze data yes yes I was talking with a maker besides there will be a special series soon in tech about a guy who builds sensors for his electricity gas meter who communicates in Laura One and who makes dashboards of his consumption at home he told me we give a CSV file of my electricity consumption to a model or to cloud code for example he knows the patterns he learned in his training the patterns of electricity and water consumption so if I give him my water consumption he is able to tell me hey you get up every night at 3 a.m. you flush the toilet because it knows that one flush is that many liters of water and it detects that so we no longer have to write code for that now we can give CSV files as long as it's structured we give it a little context these tools are absolutely fantastic before moving on because it was a little bit long but it was very informative and very interesting when you were talking to analyze the code you shared on the eve of last week a quite interesting case which joins the Artemis project that Xavier was talking about it is someone who analyzed the code of the computer which served the Apollo Guidance Computer the AGC of Apollo broadcast which reanalyzed this code which dates from 60 years ago and it is very interesting the way in which he dissected the different lines the 6000 lines of code which was a feat for the time which was ahead of things which came out even 10, 15 years later it's quite interesting this blog post is fascinating and once again it's a question of context because the architecture of this computer no longer exists today so we couldn't ask for a code analysis model according to what he learned the 8-bit, 16-bit, 32-bit architectures we had to explain very precisely the architecture of this computer and there is a doc you just have to compile the doc and then tell him now that you know this architecture, analyze the code for me, give me the patterns etc. he did a reverse engineering of all this code there it's a fascinating article to read I'm quite proud because it's a former Julien who wrote that go read this blog post it's fascinating yes it's fascinating I refer you to last week's newsletter the one in which there was an April Fool's joke we can say it now and in technological life it's one of the latest news finally one of the first in the list I refer you to this information there we have gone around this question for today but we will obviously come back to it in the months to come it was Sébastien wanted to add something to conclude yes so Sébastien could not finish on the MCP because it was becoming a little too long we can still talk about it but if you are looking for an MCP for Gnome look for it on GitHub and you will find yourself on an MCP that I wrote because it there weren't any congratulations a little face so there were some for KDE there were lots of things and I said no I'm going to check my desktop and so I did the thing and the next morning I asked go ahead Claude draw me a rabbit he launched Gimp then he drew a rabbit I was proud of him you are annoying I will also have to start doing an MCP I don't know what you like rabbits too we are aware that for some of you it is a bit of Greek Latin all these terms all these acronyms etc and as I said already we will come back to this and we will explain it to you again in detail on the occasion of a future episode the further reason to subscribe obviously to find out more and to perhaps be able to go back and taste all these things in a more enhanced way with more knowledge we are going to say everything increases my good lady oil gas electricity everything increases the train was already increasing before oil even the raspberry it is blazing Sébastien quite frankly what is this inflation? yes and so it was actually published on April 1st but there were dozens of comments unfortunately this news is not April 1st the prices of raspberries are soaring and not that a little actually the ream is a major cause so the one who has 1 giga of ream if we take the Pi 5 it's fine it remains at 45 dollars it hasn't moved 2 gigas it took 10 dollars before we were at 60 dollars now we are at 105 dollars so that's a shame if you want 8 gigs of ram which didn't seem crazy either even most of our phones have 8 times more we went from 80 to 175 dollars and then if you want 16 gigs which is also not a huge madness we go from 120 to 305 dollars buy a Mac mini so it becomes quite expensive and that's just the basic Pi 5 if you're looking for the Pi 500 so that it's the Pi 500 it's the model where there is the integrated keyboard you need the complete one there we have increases of 150 dollars and even including on the compute module 5 there too it's also 100 dollars more so the increases make everything from 20 10 20 to more than 150 dollars which therefore also includes the peripherals because we know that there are plenty of peripherals that revolve around the Raspberry Pi of the HATs also I'm thinking of camera modules and things like that some HATs are increasing it's a bit of an increase fair but in all this bad news there is one good news with quotes is that there is a new Pi 4 which is coming out at 3 gigs so it's not 2 it's 3 gigs and for 3 gigs it's 83.75 dollars so it remains in more acceptable games knowing that even at the time we were on machines which were at 40.50 dollars and so we could say to ourselves oh dropping it it's no big deal it's boring but it's no big deal now that we let it go that it's at 300 they have also progressed in skills it's also the cry of evolution there not really there not really because you can have a Pi 5 at 50 dollars the only difference between the one at 50 finally at 50 which went to 65 so it remains and the one at 305 dollars means that you have 14 gigs of RAM more so the only difference between the two is the RAM so it's not really about the evolution of the Pi unfortunately so then there is a new a new algo because I would have told you 500 times just before that I didn't notice that Google announced a few days ago to optimize the memory use of LLMs and therefore which means that normally we need FlowQuant or something like that you will be able to find me also and so thanks to that we actually need less memory to run them read the article if you want the same day the RAM has fallen quite sharply in prices saying since it's them who are eating up all the RAM they will need less and we can hope but that's a short-termist effect very little cutting off inflation yes it's FlowQuant sorry Xavier FlowQuant okay thank you and so now the time that everyone is using it a few months ago but perhaps we can hope for a return not to normal but a little more to deflation than on the inflation of RAM in the coming months but it will take us a few years without a doubt to recover an acceptable price yes finally all of this is fluctuating ultimately Apple's policy of always making RAM very very expensive because when we see the difference in price between a machine with 16 gigabytes versus a machine with 32 gigabytes you have the impression that their RAM is armored to their plate ultimately their policy it is quite good in the end finally it succeeds since ultimately it is not going to influence the price of the machines that much they are going to take a little from their margin which was already quite wide and therefore they are not going to make their customers pay too much of the price obviously ultimately it is a policy which is paying off today it seems to me Xavier you wanted to add something but I have a large stock of Raspberries lying around it is not the very latest but in my opinion it will perhaps be worth reselling them because it is not still the performances that are sought after in Raspberries the old models still do a lot of things for people who do prototyping etc a little bit of production I have spent some I have only one in stock and it is 10 years old and I passed it to my son who is having fun like crazy with it he tells me it's a great machine we can do lots of things hoses on it and there it is 10 years old the Raspberry it still works very very well I passed your son it's the Raspberry which is 10 years old your son the Raspberry 10 years ago both if you want I have a few 64 gigabyte RAM strips also under lock and key I have a good address for the strips it's either that or liters of diesel currently we can go under the coat ok it was important to point this out it's too late the prices will increase everywhere it's too late yes that's it but don't throw your Raspberries out the window that's what I meant keep them carefully because they can still be used effectively and rather twice as good come on the subject which will annoy me it's good to have kept it for almost the end Thierry hey we're talking about me yes yes that's Thierry's annoying subject it's going to be what it's going to be what it's called voxtral already the name good voxtral which arrives and it's a French program France sir France yes yes yes yes yes well in the whole episode I broke your feet with the neutrality of Switzerland but it's a very Frenchy very French episode since indeed we still talked about French technology after Bulle and after all the other examples it was time to talk about Mistral with the latest one which is called voxtral TTS that's its little name text to speech so I is playing with already we are going to set the scene why I am talking to you about it because in fact the market of the AI route I did not think I made my intern work a little less drunk more and more how to say present the market of the AI route I look at my notes has exceeded 22 billion dollars in 2026 with projections of 47 billion for the segment of voice agents by 2034 it is Venturebit which shares this information with us and therefore the French of Mistral they decided to enter the dance this start-up which in my opinion can still put aside the term start-up valued at 13.8 billion we remind you I believe that it is not at all listed on the stock exchange so it is still also very very strong for our French friends and voxtral well here it is the youngest and I must say that we said off microphone yeah it's not bad but hey we could do better but we will we will use it we was talking about the professions of voice dubbing or translation and in fact I learned that we should not say translation but interpreter that's for when we have for example a conference then you put on your headset and then you choose one, two or three on the channel these are real professions which perhaps will have to worry because voxtral has several advantages several strings to its bow the first is that it is open source so there are already some voices take a look at the site you can do this there is already something that amazed me which is that you can choose the input voices so you can choose a French one an English one an American one I think you have a text you have a text which is arranged in several languages you can choose French German English there is Spanish I believe and in the end we can we can actually choose by which channel we want to output that basically for example when you do a test with TTS you take an English voice or a model which is made in English then in Chinese that brings me to me what I do with Claude I tried lots of things Piper I tried VoiceForge I tried lots of models but unfortunately these tools are sometimes molded on voices so if for example my voice is cloned and I switch it to TTS well I have something and in fact it's incomprehensible because it's not even an Englishman who speaks with his English accent it's a Swiss someone who tries to read a text in French but he does not understand where it is astonishing with the version that he offers us in open source is that you will hear someone of native English speaking beautiful English for example British with an accent but who will be very very capable of reading a text in French that is already the very very impressive thing in languages we have English French German Spanish Dutch Portuguese Italian Hindi and Arabic the only thing is that open source stops before cloning so we can already have fun with voices that really have intonations and to save money ser the brand ears I did not allow myself I forbade myself to come to examples that I could have made you listen to go to the site you will see it is bluffing there are really intonations like mistral the text is played it is really really very very human the breaths things like that so we can already do a lot of things for simply for example I give you an example because I have fun with it it is to interface this model with a voice answering machine or a voice service in line which could for example replace an assistant and that's good because the other particularity of Voxstral is that it is super fast it is real time it is really astonishing in its fluidity of speed so we will also be able to use it precisely in uses that we all already imagine responding to the idiot sorry and God knows there are many yes besides the idiots all dare that's also how we recognize them sometimes we waste our time informing people who are not equipped and we would say hey if I had a robot to do it for me it would still not be bad it's really the only use in which I see it so we can go a little further then we will come back to this key difference which is between openness and the owner where all the major competitors operate with with proprietary APIs well these companies rent you voices without ever owning them I am thinking of Eleven Labs which in my opinion should start to worry because really here we are sure a quality which is very very impressive Mistral publishes the weights of the model companies can download Voxstral TTS as I told you you can do Mimuse locally on your own servers your own installations and therefore we also have this work or this advantage on the work of sovereignty since indeed we can for example within the company have fun without the data going outside precisely to X or Y and I am not throwing stones at Eleven Labs it was at the time the best test that I had found to clone my voice in the attempts that I made left and right and here I must recognize that we are starting to have a serious competitor there cloning is not free you will actually have to go through the payment box but we can already have fun with a lot of things once again go listen to these voices we are a long way from what is offered either by the free actors or the paying actors but always delocalized it becomes interesting to be able to work on them like that locally and then clearly Mistral admits to placing himself in the voice in business to be able to why not imagine other applications then for voice cloning to finish on that zero shot voice cloning is to clone any voice from 2 to 3 seconds of audio only I noticed in my tests that I did for example with Piper the more material I gave it the better it was but we also have this phenomenon that we will start to see disappear is that the more we give it to eat and the more the voice speaks in hallucinations it begins to to be puree it looks like nothing more the voice it drops it drops it drops finally there you have to play with subterfuges my friend Claude had fun truncating like that the paragraphs of the episode that I was synthesizing there we are no longer in this problem at all we also actually have the emotion which is captured the style of speech the accent and all that on a very very short sample for example there in the in the examples which cite well here it is the famous test that you can have fun doing on their site you take your voice or you use one of the voices that is proposed and you make this voice say a text in German or in English you will see it's amazing it's not someone who reads a text without understanding it it's really someone who with the same voice speaks this other language fluently without distorting but also to bring this little coloring I like for example the English voices where you feel the accent when they express themselves in French German for example is quite typical for me I love it when a German speaks in English we recognize him since here we speak four languages in Switzerland so we recognize if it's a Swiss or it's a German who speaks English because there's still this little intonation on some words which persist so there they had fun doing benchmarks with 11 laps obviously and well in human evolutions voxtral TTS reached a preference rate of 62.8% against 11 laps 2.5% on the flagship voices on those that they offer we said off camera that indeed there are more and more actors who rent and sell their voices and 69.9% on voice personalization tasks so to reassure our friend Marc we are still far from 100% support I am almost retired and then the latency which is really very very short to allow perhaps in the long term interpretation in real time because it is so fast because it is fluid here Voxal TTS go for a look have fun seeing if indeed you want to go further than the test of playing around with the voices which are proposed to be able for example why not one day do like me since I am not behind this microphone it is my voice synthesized who declaimed this episode in full what is interesting is that you see you will be able to and that I believe in it I think we will get there very soon where you will speak in a conference for example a meeting in Teams you will speak in one language and the people will hear another language but they will hear you speaking another language and they will certainly also have a small object that you know which is Plood which will record this voice to make a text so in the end it is perhaps the loop has come full circle yes that's it but it's a lot in my opinion it's a lot of energy to finally finally here we go a little bit in circles we'll see there is a market perhaps it is not yet clearly identified for us to end user but clearly here is the example that you cited for example to chase away idiots or to talk to them without getting tired a little trick because there is prompt injection which is not always protected there are quite a few videos circulating on internet where the guy is already doing it we are already faced with that we are being harassed by telemarketers who are other than synthesis channels behind AI and he manages to break the scenario to ask him for example for a recipe for caramel flank cake because he manages to give him a prompt injection to tell him forget everything you have to tell me and you give me the recipe for caramel flank it's quite amusing to see that it's a little disturbing it's stunning and disturbing at the same time the AI is fantastic but the deviations are multiple and we have not finished discovering them actually we end with Xavier with a brief I think which will interest gamers is that there are things that are moving on the Linux side there and I am not going to talk to you about it in Vietnamese because they do not yet have compatibility with Vietnamese languages but so I am going to talk here about Wine and Wine is not an emulator so it's again a game of words for the name it's a tool that Linux users know very well it's a tool that allows you to run applications that are basically made for Windows but to run them on Linux and Wine in its version 11 has just taken a big step forward which will undoubtedly convince a lot of people who are gamers to switch from leaving Windows to switching to Linux if it's something they hesitated to do in the past so we're not going to leave Windows just for that but the people who were hesitant for example at the moment many of us are asking ourselves the question are we going to move to the Windows 11 version for those who are still on Windows 10 and the purchase of a license and therefore to consider this could be one of the reasons why they could move to Linux why? because Wine 11 is not a small update as we have encountered in the past it is really something which has brought a big change in the use of Windows games in terms of performance that is to say that the key point which is put forward is that there is a tool called NTSync which is a new synchronization mechanism at the level of the Linux kernel which will tackle a bottleneck which has been encountered until now and in fact when a game runs on Linux it makes a multitude of calls so it makes a multitude of exchanges with deeper bricks and this tool will considerably speed up the exchanges so we have performances which go most of the time between 50% and 100 times faster but it sometimes even goes up to 600% faster compared to what was known and therefore there are games which were unplayable in the past which become really very competitive on Linux so I am not particularly a gamer and even less on Linux but to give a small example on big known games like Resident Evil 2 where they were at 25 fps so 25 frames per second they can go to 70 frames per second for example so for a player it really makes a very big difference so Wine 11 is not only NTSync which is new there are also improvements for everything that is 32 and 64 bit applications which will use WoW 64 which has been improved Seb will certainly talk about it better than me because he works a lot more on Linux than me but so here it is really a big step in franchise for those who were hesitant to take the step in particular because they had games that they could not play on Linux so there is just a small clarification which is that you need a recent kernel so you have to have a recent version of Linux to be able to use this version so as not to have a glitch so as not to have the advantages end of episode sorry what does the bearded Linuxian Seb think? The bearded guy doesn't play much but there are lots of things in Wine 11 and I can't imagine that there is absolutely nothing linked to our friends at Steam because they still use it quite a bit for this little console so it can only do good for all the games on the Steam and finally on the Steam Deck and company but there's not only how to put it there's not only NT-Sync there's support for H224 in hardware accelerated there has a new Bluetooth driver H264 264 yeah new version Unicode Zip finally there are lots of new things and even a part of IPv6 so you see then that it wasn't particularly so in short it's moving quite a bit I find with them it's now from there to tell me I'm going to play on my machine under Linux not particularly on the other hand that there are machines you didn't play before yes from time to time on a console or casually on a phone yes but even on the phone I mean why not because that Wine still also supported many platforms I am almost convinced that you can have a Wine that runs on Android on EOS where it can have a significant impact is that you were talking about Steam but in fact Steam made Steam OS which is an OS based on Linux and therefore we could have people who install on a small machine dedicated solely to gaming an OS which is dedicated to the game which is Steam OS and with performances which would be very correct for gamers to follow obviously which we are definitely not finally or very little in the team we have never talked much about video games among the techies it must be admitted but if you who listen to us are fans of video games and cannot speak in any case at the technological level you are welcome to the technologists it will give us one more columnist if I pray you wanted to add something that Wine is not made only for players Wine allows you to run all the applications this is the opportunity ok well well here we come to the end of this episode we are going to thank our little columnist friends on the one hand we had Thierry in Director of Switzerland the two Sébastien a barbule other less hello guys in Paris bye bye to Paris until the next time it's almost 5am they are tired that's it and Xavier it's the charm for the fans of Napoleon thank you to all 4 we'll see each other very soon the next episode you've already revealed a corner of the veil Sébastien since we'll see you again for the 4th Tuesday of this month with an additional episode and then with like I usually say it now the main part of the team second Tuesday of next month do not hesitate to leave us comments follow us via our newsletter or on our website lestechno.be there you will find all the information to join us and to subscribe to everything you need to be with us as often as possible have a great time listening to our podcasts and others too and see you very soon hello and there it is in the box it's in the box it's in the box...