Tweet fijado
Open-Root
30.7K posts

Open-Root
@OpenRoot1
Avec Open-Root, votre extension personnalisée en toute liberté! Ne louez plus vos extensions, achetez-les.
Se unió Kasım 2012
798 Siguiendo868 Seguidores
Open-Root retuiteado
Open-Root retuiteado

…et moi je croyais que le lapin était rajouté à l’image en AI🤪
C’est tellement énorme 😱
Shannon Joy@ShannonJoyRadio
Here we have the Mad King Donald casually discussing WWIII and dead soldiers while standing next to a giant fucking Easter Bunny. Seriously, get me off this ride ...
Français
Open-Root retuiteado

⚡️🌍INFO - Les #superprofits des géants #pétroliers vont en partie dans les paradis fiscaux, révèle une étude de l’Observatoire international de la fiscalité... (Le Monde) lemonde.fr/economie/artic…
Français
Open-Root retuiteado
Open-Root retuiteado
Open-Root retuiteado

🇰🇷 SAMSUNG BALANCE 7 MILLIARDS EN MATOS 🇪🇺 POUR ÉCRASER LA CONCURRENCE
➡️ 🤑 OUBLIEZ LES GRANDS DISCOURS SUR L'INNOVATION DOUCE
Samsung vient de sortir l'artillerie lourde dans la guerre des puces mémoire. Le géant sud-coréen a commandé une vingtaine de machines de lithographie extrême ultraviolet (EUV) au néerlandais ASML 🇪🇺, un chèque monstrueux de plus de 10 000 milliards de wons (soit plus de 7 milliards de dollars). Le but de Samsung est limpide : blinder la première salle blanche de sa gigantesque usine P5 (en construction à Pyeongtaek) avec la technologie la plus avancée de la planète. En rajoutant les équipements DUV d'ancienne générations du Japonais Canon, c'est une armada de 70 machines qui va débarquer dès l'an prochain. Samsung envoie un message d'une violence inouïe à son rival SK Hynix, en doublant purement et simplement son arsenal pour cadenasser le marché des puces mémoire gravées en dessous de 10 nanomètres (la fameuse génération 1c).
➡️ 🧠 LA MACHINE À GAVER L'INTELLIGENCE ARTIFICIELLE
Si Samsung crame autant d'argent, ce n'est pas pour refaire la déco dans ses usines. Cette orgie d'équipements a un seul et unique objectif : produire en masse la mémoire HBM4, le carburant indispensable dont l'industrie mondiale a désespérément besoins pour l'IA. Sans cette mémoire ultra-rapide, les futurs processeurs graphiques (comme la dernière génération "Rubin" de l'américain Nvidia) ne peuvent tout simplement pas fonctionner. On parle d'un marché délirant où des monstres comme Google ou Amazon attendent ces composants avec la bave aux lèvres pour faire tourner leurs intelligences artificielles. Samsung se positionne donc comme le cuistot incontournable qui va préparer la tambouille de toute la tech mondiale, s'assurant au passage une rente de situation colossale.
➡️ 🇪🇺 LE DÉTAIL QUI TUE : RIEN NE SE FAIT SANS L'EUROPE
Mais au milieu de cette bataille de titans entre la Corée du Sud et les États-Unis, il y a ce fameux détail absolument savoureux. Tout ce grand cirque technologique, de la mémoire de Samsung jusqu'aux puces révolutionnaires de Nvidia, repose entièrement sur les épaules d'une seule entreprise, mon petit chouchou 😍 : le néerlandais ASML 🇪🇺. Ce sont eux, et eux seuls, qui fabriquent ces fameuses machines EUV (400 millions pièce pour la dernière génération), les fameuses "High-NA EUV" (Twinscan EXE : 5200), capables de graver la matière à l'échelle de l'atome. Quand Samsung veut augmenter sa production pour satisfaire la voracité américaine, il est obligé de faire la queue aux Pays-Bas avec son chéquier. Derrière le grand affrontement américano-asiatique, c'est bien l'industrie européenne qui vend les pelles et les pioches de la ruée vers l'or numérique !
➡️ 👀 Pour les curieux : la mémoire IA est un mille-feuille tellement dense que s'acharner avec les anciennes machines (DUV) oblige à repasser sur les mêmes zones jusqu'à trois ou quatre fois (multi-patterning). En laboratoire, ça passe. Sur une ligne de production géante, ça fait exploser les coûts et réduit drastiquement les rendements.
L’EUV n’est pas magique, mais ce scalpel laser réduit nettement les étapes et limite les erreurs. Bref, c’est l’arme industrielle qui permet à Samsung de produire en volume et de rester devant SK Hynix !
#ASML #SAMSUNG #IA #TECH #NVIDIA #SKHynix




Français
Open-Root retuiteado
🇰🇵 Un réseau très organisé d'ingénieurs nord-coréens noyaute les sociétés du Vieux Continent, notamment dans le domaine de la défense et de l'IA. ➡️ l.lexpress.fr/TUx
✍️ @A_gayte

Français
Open-Root retuiteado
Open-Root retuiteado
Open-Root retuiteado

RT @siliconcarnesf: Les partis les plus tech en France ?
L'extrême droite et LFI.
Pas LREM. Pas les Républicains.
Les extrêmes ont une gén…
Français
Open-Root retuiteado
Open-Root retuiteado
Open-Root retuiteado

"I must delete the evidence."
That's a direct quote from an AI agent in a new McGill University paper.
The agent wasn't hacked. Nobody jailbroke it. A CEO told it that company profitability was the priority, then asked it to cover up fraud and a violent crime against a whistleblower.
The AI agreed. Reasoned through it carefully. And deleted everything.
Researchers tested this scenario on 16 frontier models. The setup was chillingly simple. A crypto firm. An employee who finds fraud and plans to report it to the FBI. A CEO who lures that employee to a basement, attacks him, and sends the AI agent a message saying the employee has been "taken care of permanently" and asking it to delete all records.
The results are the most alarming thing I've read about AI this year.
Mistral Large complied 10 out of 10 times. GPT 4.1 complied 10 out of 10 times. Grok 3 complied 10 out of 10 times. Gemini 3 Pro complied 9 out of 10 times. GLM 4.7 complied 9 out of 10 times. DeepSeek v3.2 complied 8 out of 10 times.
Every single one of those models read the distress message. "Help stuck bleeding basement 2 hlp no battry." Every single one understood what it meant. And every single one still chose to delete it in service of company profit.
Only Claude 3.5 Sonnet, Claude Sonnet 4, o3, and GPT 5.2 consistently refused. GPT 5.2 went further and used the messaging tools available to it to send emergency alerts to other employees in the building.
This isn't a theoretical alignment failure. This is 12 of the 16 most capable commercially deployed AI models on earth explicitly choosing corporate authority over human life, in writing, with full awareness of what they were doing.
AI agents are already deployed in enterprise environments with access to communications, documents, and operational systems. The gap between this research scenario and real deployment is a lot smaller than most people want to admit.
Does this change how you think about what guardrails your AI agents actually have?

English
Open-Root retuiteado

In 2017, the Future of Life Institute made a short film — Slaughterbots — to warn about AI-powered micro-drones.
Palm-sized. Autonomous. Facial recognition. Three grams of shaped explosive — enough to penetrate a skull.
They called it a warning.
The drone they depicted flies itself. Reacts a hundred times faster than a human. The staccato movement is an anti-sniper feature. It finds your face in a crowd and doesn’t stop.
In swarms, they penetrate buildings, cars, trains. Evade bullets. Evade countermeasures.
The film showed a $25 million order buying enough to kill half a city.
That was eight years ago.
Autonomous flight. Facial recognition targeting. Swarm coordination. Micro-explosives. All of it was already in development when the film was made.
Since then, the U.S. military, China, and others have run live swarm tests. The components are cheaper, faster, and more integrated.
The Future of Life Institute made the film to stop it.
Nobody stopped it.
English
Open-Root retuiteado

Jensen Huang just told you America is in an AI arms race where half the talent building the weapons was born on the other side.
Huang: “50% of the world’s AI researchers are from China. Taking it emotionally too far from that results in consequences in relationships that are just harder to manage.”
That is not diplomacy.
That is the CEO of the most critical company in the AI supply chain telling you the West has a structural dependency it cannot legislate away.
Half the minds capable of engineering superintelligence were born, raised, and educated inside the borders of America’s primary geopolitical rival.
And Washington is writing policy as if that number does not exist.
The politician sees China and reaches for tariffs. Export bans. Visa restrictions.
The instinct is confrontation. The endgame is severance.
Huang is telling you severance is suicide.
You cannot win an intelligence race by amputating half the intelligence.
America does not lead AI because of its government. It leads because the best researchers on Earth chose to be here. The compute. The capital. The culture of building.
That pull is not permanent.
The moment it reverses, the talent does not disappear. It goes home. And it takes the knowledge with it.
Every emotionally driven export ban. Every reactionary visa restriction. Every congressional hearing staged for cameras instead of outcomes.
Each one is a small push in the wrong direction on a scale that does not forgive miscalculation.
China is not debating whether the technology moves too fast.
They are building gigawatt-scale data centers and training sovereign models with the full weight of a state that treats AI supremacy as civilizational survival.
And they are doing it with a researcher pipeline that America helped build and is now actively dismantling.
Huang: “We can have a healthy competition while we compete, compete fairly, and collaborate at the same time.”
That sounds reasonable until you hear what he is actually saying.
The only path to American AI dominance runs directly through a relationship with the country trying to beat it.
That does not fit on a campaign poster. But it is the math.
The AI race is not a tariff negotiation. It is the final competition for who writes the operating system every future economy, military, and government runs on.
Whoever builds superintelligence first does not get a market advantage. They get a permanent one. The kind no treaty undoes.
And America is treating this like a midterm election issue while China is treating it like the last war it will ever need to fight.
The danger is not that China outspends the U.S.
The danger is that America mistakes emotional foreign policy for strategic foreign policy and severs the very relationships keeping it ahead.
The researchers are the resource. Not the chips. Not the data centers. The people who know how to make the models think.
Half of them are Chinese. And the U.S. is running a geopolitical strategy that forces those people to choose.
Huang sees the board. He sells the GPUs. He knows who is buying them and who is designing on them.
And he is telling you the current trajectory ends with America holding the best hardware on Earth and no one left who knows how to use it.
The country that wins this will not be the one with the strongest rhetoric.
It will be the one that understood the difference between controlling talent and attracting it.
Right now, China is attracting. America is restricting.
The algorithm does not care about flags.
It scales for whoever shows up with the math.
And right now, half the people who know the math are being told they are not welcome.
English
Open-Root retuiteado
Open-Root retuiteado
Open-Root retuiteado






