Bienvenue sur mon blog
Reflexions sur la technologie, le developpement et l'innovation
Articles récents
Meituan trained a 1.6 trillion parameter model without a single Nvidia chip
A food delivery company just broke the central assumption of US AI export controls. Meituan released LongCat-2.0 today: a 1.6 trillion parameter Mixture-of-Experts model, open-sourced on Hugging Face, trained end-to-end on a 50,000-chip cluster of domestic Chinese silicon. No Nvidia. No AMD. No workaround. Just homegrown hardware running the full pipeline from pre-training through inference. That distinction matters more than anything else in this announcement. DeepSeek’s V4-Pro from April, the
Lire l'articleSouth Korea pledges over $1 trillion for AI and semiconductor dominance
South Korean President Lee Jae Myung just unveiled an investment plan that makes every other government chip subsidy look like pocket change. The total: more than $1 trillion. Over the next decade. Into AI infrastructure and semiconductor manufacturing on Korean soil. For context, the US CHIPS Act committed $52 billion. The EU Chips Act allocated about $47 billion. South Korea is going 20 times bigger. The numbers The plan has three pillars. Samsung Group pledged 1,000 trillion won ($648 billion
Lire l'articleSomeone dumped 23 zero-day exploits on GitHub with zero vendor warning
An anonymous GitHub account called “bikini” published a repository named “exploitarium” on June 23, packing proof-of-concept exploits for 23 previously unreported vulnerabilities across some of the most widely used open-source software on the planet. libssh2, FFmpeg, Ghidra, nmap, Docker, VLC, OpenVPN, PHP, Gitea, Firefox, c-ares, AnyDesk, RustDesk. By June 27, the repo had 1.7k stars, 360 forks, and 46 commits. By June 28, it was still growing. The README is blunt. “At the time I post these, no
Lire l'articleDeepSeek open-sources DSpark and makes V4 generate 85% faster
DeepSeek just dropped DSpark, a speculative decoding framework that makes DeepSeek-V4 Flash 60 to 85 percent faster at token generation. V4 Pro gets 57 to 78 percent. No quality loss. The speedup is measured against their own single-token MTP baseline, and it is already live in production traffic. This is not a new model. DeepSeek-V4-Pro-DSpark on HuggingFace is the same V4 Pro checkpoint with a speculative decoding module bolted on. The interesting part is the algorithm. Two problems with exist
Lire l'articleThe White House is now hand-picking who gets OpenAI GPT-5.6
Sam Altman told OpenAI employees on Wednesday that the Trump administration asked the company to stagger the release of GPT-5.6. Not delay it. Not block it. Stagger it. The model ships first to 20 government-approved partners, routed through Amazon Bedrock, and then the government decides who else gets in. Altman’s exact words to staff, per a memo obtained by The Information: the government will be “approving access customer by customer” until the preview period ends. A wider release might follo
Lire l'articleAnthropic names Alibaba in the largest Claude distillation attack yet
Anthropic sent a letter to the US Senate Banking Committee on June 10 accusing Alibaba of running the biggest model theft operation it has ever seen. Bloomberg broke the story on June 24. CNBC, Reuters, and the BBC all confirmed the details from the letter. The numbers are specific. Operators tied to Alibaba and its Qwen AI lab created roughly 25,000 fraudulent accounts and pumped 28.8 million exchanges through Claude between April 22 and June 5. That is a six-week window. Anthropic does not sel
Lire l'article