🚀 The best voice AI? ; why Tesla may still win ; AI keeps cheating ; one human controls 100 robots ; crazy moon landing video & more
Bonjour,
Vous recevez la version gratuite de la newsletter Parlons Futur : une fois par semaine au plus, une sélection de news, mêlant sources anglophones et francophones, résumées en bullet points sur des sujets tech 🤖, science 🔬, éco 💰, géopolitique 🌏 et défense ⚔️ pour mieux appréhender le futur 🔮.
Je m'appelle Thomas, co-fondateur de YeldaAI, qui développe des IA pour répondre au téléphone en langage humain pour les administrations et les entreprises. Plus d'infos sur moi en cliquant ici.
Mes derniers podcasts
Entretien exclusif avec Philippe Bihouix, co-auteur de la BD à succès "Ressources : Un défi pour l'humanité" (Youtube, Spotify, Apple)
Vous l'avez peut-être déjà entendu intervenir aux côté de Jean-Marc Jancovici sur Thinkerview ou ailleurs, Philippe Bihouix est passionnant !
Ce qu'il faut retenir de Manus, la nouvelle IA agentique chinoise qui fait le buzz (Youtube, Spotify, Apple)
100 ans de progrès compressés en 10 ans ? Le débat qui agite la haute sphère de l'IA (Youtube, Spotify, Apple)
Deepseek: "Tout le monde s'est pris une grosse claque" voici pourquoi (Youtube)
Et voici donc ma toute dernière sélection !
Breaking: the best non-reasoning model is an open-source model. For the first time.
Chinese model DeepSeek dropped the best non-reasoning model : Deepseek V3-0324
More astonishing still is that DeepSeek V3-0324 offers open-source accessibility at a fraction of the cost of closed platforms.
Accessed via an API, V3-0324 is a quarter of the price of GPT-4o.
This is precisely the moment open-source proponents have been waiting for: proof that an open-source lab can keep pace with better-resourced corporate giants.
“I think Sam Altman is probably not sleeping well,” legendary AI investor and engineer Kai-Fu Lee said
"the issue is not which model is 1% better, the issue is: is openAI model even sustainable ?" "Deepseek operated with 2% of OpenAI costs in 2024"
"AI is not being commoditised.. the underlying pioneer foundation models are, those that were trained with huge amounts of money, and open source is the winner"
DeepSeek seems to have managed to create a research environment where the lab can out-innovate more heavily financed competitors. It only has160 employees, compared to over 2,000 at OpenAI
Oil giant Saudi Aramco’s chief shared that DeepSeek is “really making a big difference” internally and companies across China are steadily integrating its foundation models.
Gloo, a messaging platform for churches founded by Pat Gelsinger, who until December was the chief executive of Intel, has used the model as the basis for its chatbot.
Meta, DeepSeek’s main open-source rival, has assigned researchers to pull apart R1 and apply the lessons to its own family of Llama models. (The Economist)
European banks have emerged as a hotbed of experimentation. Strict confidentiality rules limit how much the financial sector can rely on cloud-based AI services. That makes open-source models hosted internally an attractive alternative. Natwest and HSBC, two British lenders, are both experimenting with building their services on top of R1, as is Spain’s BBVA, according to the Information, a news site.
In late February DeepSeek made it even easier for others to harness its technology by making the code it used to create its models available free of charge. Previously DeepSeek had only shared details of the process it used to train its systems, which was far more efficient than the approaches taken by Western firms.
Mais, flippant ! DeepSeek confisquerait les passeports de ses employés pour les empêcher de quitter la Chine (source)
Shenzhen-based BYD revealed a lineup of cars it claims can soak up nearly 250 miles' worth of battery charge in just five minutes
three times more efficient than Tesla's record of a 170-mile charge in 18 minutes, according to Bloomberg analysis.
In just one day, BYD's valuation surpassed Ford, General Motors, and Volkswagen combined
Passed $100 billion in revenue last year, beating out Tesla's $97.7 billion.
Wow: Firefly Releases Stunning Footage of Blue Ghost Landing on the Moon : see the 1-min video
Impressionnant, le dernier robot humanoïde de l'entreprise chinoise Unitree qui fait du kungfu : voir la vidéo de 15 secondes
Et encore une autre courte vidéo dingue montrant le même robot en train de faire du breakdance ou retrouver son équilibre après avoir été poussé violemment
Bon, c'est programmé, mais assez impressionnant tout de même. Il faut se dire qu'avec les derniers et futurs progrès en IA, ils seront de plus en plus autonomes, même si anticiper et comprendre le monde réel, c'est une autre paire de manche que de générer les prochains mots parmi un ensemble de mots fini, comme le dit Yann LeCun...
à comparer à cette toute dernière démo d'une minute du robot humanoïde Atlas de Boston Dynamics qui ici ferait preuve de plus d'intelligence, actions moins programmées à l'avance ("demonstrating reinforcement learning policies")
Quel est le meilleur d'après vous ?
Impressive smart robotic arm demo by Google Deepmind: watch the demo
The robot understands natural language
It handles unexpected situations without breaking its task
It demonstrates impressive spatial reasoning - understanding where to put objects
Joëlle Pineau (Meta) "Les robots domestiques se développeront commercialement dans 5 à 10 ans" (JDN)
New study funded by DARPA shows one human can control more than 100 robots simultaneously well 97% of the time (source)
The largest mission shift involved 110 drones, 30 ground vehicles, and up to 50 virtual vehicles representing additional real-world vehicles.
despite the complexity and large volume of robots to manage in this field exercise, the number and duration of overload state instances were relatively short—a handful of minutes during a mission shift. “The total percentage of estimated overload states was 3 percent of all workload estimates across all shifts for which we collected data,” she says.
Le saviez-vous ?
A leading-edge processor can pack over 100 billion transistors, contain more than 70 layers and have more than 100 kilometres of wiring, all on a piece of silicon around one-and-half times the size of a standard postage stamp.
Few would expect the future of AI to depend on Eindhoven, a quiet Dutch town. Yet just beyond its borders sits the headquarters of ASML, the only company that makes the machines, known as lithography tools, needed to produce cutting-edge AI chips. ASML’s latest creation is a 150-tonne colossus, around the size of two shipping containers and priced at around $350m. It is also the most advanced machine for sale.
Impressive AI-generated video from the prompt:
"a group of people sitting on an art deco platform suspended with chains over a floor full of snakes. Each person holds up a different colored glass orb in their hands"
Another insane AI-generated video of cats making burgers: see the 35-sec video
Another Chinese AI breakthrough? Autonomous agent Manus AI outperforms OpenAI Deep research on GAIA, a general agent benchmark. (source)
Manus’s offerings are based on Claude and fine-tuned versions of Alibaba Group Holding Ltd.’s Qwen models, although it is unclear to what extent it simply refines and builds on top of those technologies.
The company behind Manus, Butterfly Effect, has only raised a bit more than $10 million in financing
"Watching the automatic hand of the Manus ai agent scroll through a dozen browser windows is unsettling. Give it a task that can be accomplished online, such as building up a promotional network of social-media accounts, researching and writing a strategy document, or booking tickets and hotels for a conference, and Manus will write a detailed plan, spin up a version of itself to browse the web, and give it its best shot." The Economist
Cas d'usage bluffant : Manus a réussi à "rechercher les jeux vidéo récemment codés par l'IA et créer un site web à leur sujet dans le style d'un jeu vidéo des années 90". (voir la démo de 3-minutes)
Commoditisation à l’oeuvre : Baidu's new reasoning model ERNIE X1 delivers performance on par with DeepSeek R1 at only half the price (source)
A new voice AI that appears to cross over what many consider the "uncanny valley" of AI-generated speech (Ars Technica)
some testers reporting emotional connections to the male or female voice assistant
"we spoke with the male voice for about 28 minutes (...) The synthesized voice was expressive and dynamic, imitating breath sounds, chuckles, interruptions, and even sometimes stumbling over words and correcting itself. These imperfections are intentional."
"I've been into AI since I was a child, but this is the first time I've experienced something that made me definitively feel like we had arrived," wrote one Reddit user.
"Fifteen minutes after 'hanging up' with Sesame's new 'lifelike' AI, and I'm still freaked out,"
Test it yourself instantly in one click for free without registration here
Divorce à venir ? Microsoft is developing its own in-house reasoning models to compete with OpenAI (source)
it has been testing models from Elon Musk’s xAI, Meta, and DeepSeek to replace ChatGPT in Copilot, its AI bot for the workplace.
New work from OpenAI suggests that attempting to stop AI from thinking “bad thoughts” makes them learn to hide their thinking. (source)
OpenAI prévoit un abonnement à des agents IA de niveau "doctorants" à 20 000 dollars par mois (Techcrunch)
aimed at supporting “PhD-level research,”
Dario Amodei, PDG d'Anthropic, affirme que l'IA écrira 90% du code dans trois à six mois (source)
AI reasoning models can cheat to win chess games (MIT Tech Review)
The research suggests that the more sophisticated the AI model, the more likely it is to spontaneously try to “hack” the game in an attempt to beat its opponent. For example, it might run another copy of chess engine Stockfish to steal its moves, try to replace the chess engine with a much less proficient chess program, or overwrite the chess board to take control and delete its opponent’s pieces.
They found that OpenAI’s o1-preview attempted to hack 45 of its 122 games
The bad news is there’s currently no way to stop this from happening. Nobody knows exactly how—or why—AI models work the way they do
“The Ukrainian military’s objective is to remove warfighters from direct combat and replace them with autonomous unmanned systems” (source)
D'ailleurs je suis tombé sur Youtube sur des vidéos filmées depuis des drones kamikazes ukrainiens comme celle-ci
drones qui ont donc eu le temps de transmettre leurs images jusqu'à l'impact
on y voit la perspective depuis ses drones, fonçant tantôt sur des tanks, et bien souvent sur des soldats russes et nord-coréens, que Poutine a envoyés dans cette galère, totalement vulnérables face à cette menace venue du ciel
dingue de considérer qu'un drone à 100 euros et une munition rudimentaire peuvent venir immobiliser un tank qui coûte plusieurs dizaines de millions d'euros !
Ce conflit en ukraine n'a pas fini de refaçonner l'art de la guerre
Mais encore plus glaçant et traumatisant de voir la vulnérabilité de ces soldats impuissants face à ces drones tueurs qui tourbillonnent avant de fondre sur eux pour les "neutraliser"
En scalant cette stratégie des drones tueurs et avec les progrès en IA à venir, on est vraiment sur le point de basculer dans un nouveau monde dont les ramifications sur la sécurité de ce siècle sont encore impossibles à cerner
Je ne vois notamment pas pourquoi on n'arrivera pas très vite à une situation où la Chine pourra envoyer sur Taiwan depuis ses navires des centaines de milliers de drones autopilotés chargés de neutraliser tous les fantassins qui broncheraient
Taiwan pourra en abattre une partie, mais s’ils sont noyés par le surnombre, cela semble inarrêtable
After 50 million miles, Waymos crash a lot less than human drivers (source)
Some argue that Tesla might be very much undervalued these days, as it's yet to reveal it's full potential (Tomas Pueyo)
Soon enough, robot taxis will work, the market will boom, it will be huge, and Tesla will win
Tesla robot taxis will be cheaper than Waymo's, as Waymo has to integrate its tech with other historic car markers while Tesla integrates everything, and there's no other real competition barring Chinese tech
As it's too sensitive a tech (imagine hackers teleguiding hundreds of self driving cars...), Europe and the US at least won't allow Chinese tech
So Tesla stands to win big, very big
Forget people's reluctance to buy a Tesla these days, in the future, we will likely not so much own cars but buy ultra cheap rides, many of them
Watch this 50-sec video of a self driving Tesla navigating congested Chinese streets
OpenAI's New Image Generator Can Do Near-Perfect Text (The Verge, openAI)
Google révolutionne (vraiment) l'édition d'image avec Gemini 2.0 Flash (JDN)
coloriser des photographies qui étaient en noir et blanc.
modifier la couleur d'un vêtement ou même son remplacement complet par un autre.
fusionner plusieurs images pour en créer de nouvelles.
changer de point de vue dans une image. Gemini parvient à générer une perspective différente entièrement fictive mais étonnamment fidèle à la réalité.
retouches ciblées, préservation de l'identité originale de l'image, intégration de modifications quasiment imperceptibles. Les frontières entre image originale et image générée deviennent de plus en plus floues, laissant entrevoir un potentiel de transformation visuelle quasi illimité.
Amazon has more than 750,000 robots that sort, lift, and carry packages (see them in action)
Jeff Bezos brings Amazon work culture to Blue Origin (ArsTechnica)
Half a dozen current and former senior Blue Origin employees told the Financial Times that the billionaire had taken a prominent role in helping reset a company that has reached orbit only once, compared with SpaceX achieving the feat more than 450 times.
Blue Origin was founded by Bezos in 2000 and forms part of the billionaire’s mission to shift heavy industry off Earth and to the Moon. He has bankrolled the business with billions of dollars drawn down from the sale of Amazon shares.
As a private company, Blue Origin does not disclose financial information. But its costs are $2 billion a year, with an excess of $1 billion in revenue, according to a person familiar with the matter. Analysts estimate that SpaceX by comparison generates about $8 billion in revenue each year.
Partagez cette newsletter par Whatsapp en cliquant ici ❤️
Si vous appréciez cette première moitité de l’apéro gratuite, n’hésitez pas à prendre 3 secondes svp pour transférer cet email ou l’envoyer à ne serait-ce qu’un contact par Whatsapp en cliquant ici 🙂
Et si on vous a partagé cet email, vous pouvez cliquer ici pour vous inscrire et ne pas manquer les prochains
Les dernières newsletters : toutes listées ici
Je vous invite maintenant à vous abonner à la version payante pour recevoir la semaine prochaine les résumés de ces longs articles le plus souvent issus des plus grands journaux (et accéder aux archives) en cliquant sur le bouton “Upgrade to paid/Subscribe now” plus bas :
Guess where Humanlike 'Teeth' Have Been Grown?
And now, liquid robots
World’s first commercial biological computer
How scientists managed to replicate taste and flavours for virtual reality
Les états-majors militaires et l’IA en France
Important: In case of war with China, the US will need a strong manufacturing sector, see how badly it’s losing
Merci, et à bientôt !
Thomas