Good morning. The hottest AI model of the week is Le Chaton Fat, a supposed Claude Fable 5 killer with impossible benchmark scores and no evidence it exists. Naturally, this did not stop tech people from treating it like the second coming of AGI. The whole thing started after Mistral renamed Le Chat to Vibe, because apparently that was all the internet needed to invent a fake French kitten with 30 trillion parameters, 1000 meows per second, and maximum chonk. Now the meme has fake launch posts and real executives wondering if they missed the next frontier model.
Microsoft is replacing its AI coworker with China’s DeepSeek. Copilot Cowork was supposed to be Microsoft’s trusted enterprise AI, running premium models from OpenAI and Anthropic. Then people started using it like an agent, and the bill got so bad Microsoft had to pull the plug on unlimited use. Now Cowork is shifting to usage based pricing while presenting DeepSeek as the cheaper model. Microsoft insists DeepSeek isn’t a security risk because it’ll run on Azure and keep customer data inside Microsoft’s cloud. The irony here is that they’ve been warning us against replacing corporate American workflows with Chinese AI. Then as soon as trust got expensive, they went bargain hunting in Beijing. (Axios)
Researchers at Stanford found a cheaper way to scale AI agents. Most multi-agent systems rely on an orchestrator to hand out work, collect updates, and tell agents what to do next. That works until the orchestrator becomes slow and expensive. The Stanford team built a system where agents grab tasks from a queue and write verified progress into shared context. Later agents can see what worked, what failed, and what evidence matters before wasting money chasing the same dead ends. This approach cut costs roughly 50% per task and outperformed the best baseline by up to 10.5%. AI agents got cheaper once they stopped making every worker learn the same lesson the hard way. (arXiv)
Stop hallucinating and get real data
Everyone knows LLMs hallucinate (usually with confidence). You.com’s guide shows how AI grounding connects models to trusted data so teams get more accurate answers and fewer fiction projects.
This playbook covers:
- Why grounding beats RAG alone
- How to build audit trails
- Open vs. closed platforms, and what it means for your next model switch
👀 closer look
The government is using AI to babysit nuclear weapons. The White House launched the Genesis Mission in November 2025 to push AI deeper into nuclear weapon management. DOE already controls America’s nuclear arsenal and owns some of the world’s most powerful supercomputers. Now officials want AI running nuclear simulations, spotting radioactive threats, and steering dangerous experiments in real time. The same government that refuses to regulate AI promises the software won’t get the launch codes. Great, but launch codes aren’t the problem. The real risk is handing AI control over nuclear decision making before anyone even thinks about pressing the button. Washington turned nuclear weapon management into a science experiment and hopes nobody notices. (Sci American)
DOJ says xAI’s gas turbines are vital to national security. When the NAACP sued xAI for running unpermitted natural gas turbines at its Mississippi data center, the Justice Department quickly stepped in to defend Elon. DOJ argued shutting down the turbines would threaten America’s security because Grok supports Pentagon work on classified networks. Residents say the turbines raise the risk of asthma and heart disease in communities already burdened by pollution. So the fight is no longer just about whether Elon can run dirty power without permits. It is whether a chatbot with military customers gets to make pollution patriotic. (Wired)
The great American brain drain is starting to look like a fire sale. The US spent generations becoming the place the world’s best scientists came to build the future. Now Washington is ending $1 billion in grants and pushing more than 10,000 PhDs out of federal jobs. In a Nature poll, 75% of US researchers said they’re considering leaving. Countries around the world are circling, eager to welcome scientists fed up with uncertainty. Canada committed over $1 billion to lure our talent, while Europe is selling stability, faster visas, and academic freedom. The US probably won’t notice until the next cure, battery, chip, or defense breakthrough gets built somewhere else. America built the world’s best science engine, then gave away the keys. (Sci American)
fun stats
💬 60%. Enterprises seeing more traffic from AI search engines. Meanwhile, 86% of shoppers still read AI answers with one eyebrow raised.
👓 $2,195. Cost of Snap’s new AR glasses, which let you play games, watch videos, and pull up AI info on what you see. That’s about 6 pairs of Meta Ray-Bans.
👎 180. Investors that passed on voice AI startup Bland during Y Combinator. It has since raised over $100 million and now automates 3.5 million complex phone calls a week.
🔻 $39 billion. OpenAI’s net loss in 2025, up from $5 billion the year before, according to audited financial documents obtained by AI skeptic Ed Zitron.