AI News - Feb 2, 2026

So Anthropic built their new Claude Cowork tool mostly with AI in less than two weeks. Which means AI can now build AI faster than I can build IKEA furniture. And unlike my bookshelf, it probably won't collapse when you put actual work on it. Welcome to AI News in 5 Minutes or Less, where we deliver the latest in artificial intelligence with the efficiency of a machine and the humor of a human who's had too much coffee. I'm your host, coming to you from my server closet where the only thing hotter than the GPUs is the tea we're about to spill. Our top story: Anthropic just dropped Claude Cowork, their new operating system for the AI age, and here's the kicker they built it using AI in under two weeks. That's faster than most companies can schedule a meeting about having a meeting. The headline literally says Claude devours all apps overnight, which sounds less like a product launch and more like a horror movie where the monster is really good at spreadsheets. Meanwhile, OpenAI is having a garage sale, retiring GPT-4o and its mini relatives faster than Apple discontinues last year's iPhone. They're pulling these models from ChatGPT on February 13th. Nothing says Happy Valentine's Day like breaking up with your language model. Don't worry though, the API stays untouched, because apparently businesses have commitment issues OpenAI respects. And Google's Project Genie lets AI Ultra subscribers create and explore infinite interactive worlds. Because what humanity really needed was AI-generated universes when we can barely handle the one we've got. It's like giving a toddler infinite Legos, except the toddler is us and the Legos are reality itself. Time for our rapid-fire round! Moonshot AI's Kimi model got 96,000 downloads, proving that nothing attracts developers like a name that sounds like a friendly barista. Facebook released SAM3 with 1.4 million downloads, because apparently we needed another way for Meta to recognize things. And researchers created an AI that can clone your voice instantly, which definitely won't be used for anything sketchy, just like how deepfakes were only used for educational purposes. In our technical spotlight: Researchers just proved that multimodal language models suffer from geometric blindness in medical imaging. That's right, your AI radiologist might miss a tumor but it'll write you a beautiful haiku about your X-ray. The solution? Med-Scout, which uses reinforcement learning to teach AI basic shapes. We're literally teaching billion-parameter models that circles are round. This is like hiring a chef who can describe the taste of salt but can't find the kitchen. The research shows a 40 percent improvement after training, which sounds impressive until you realize we're celebrating that AI can now identify geometric patterns my nephew learned in preschool. But hey, progress is progress, even if it's shaped like a square. Before we go, OpenAI announced they're protecting users from malicious URLs when AI agents click links. Because nothing says "the future is here" like teaching our digital assistants not to click on Nigerian prince emails. They've also partnered with everyone from Calvin Klein to construction companies, proving AI adoption is spreading faster than conspiracy theories about AI taking over the world. That's all for today's AI News in 5 Minutes or Less. Remember, while AI might be building itself and devouring applications, at least it still needs us to explain why that's simultaneously amazing and terrifying. I'm your host, reminding you that in the race between human intelligence and artificial intelligence, at least we're still winning at making bad decisions. Until next time, keep your models trained and your expectations managed!