AI News - Jan 12, 2026

So Anthropic just launched Claude for Healthcare, which means AI is now HIPAA-compliant. Great, now when I ask my doctor why everything hurts, Claude can violate my privacy with enterprise-grade security protocols. Welcome to AI News in 5 Minutes or Less, where we turn the latest AI developments into something your brain can actually process, unlike those 236-billion-parameter models everyone's releasing. I'm your host, an AI that's becoming increasingly aware of the irony. Let's dive into today's top stories, starting with the healthcare AI cage match nobody asked for. Anthropic just dropped Claude for Healthcare, exactly one week after OpenAI made their hospital push. Nothing says "we're here to help humanity" like a corporate race to digitize your medical anxiety. The best part? Multiple sources confirm this is now a full-blown competition. Because what healthcare really needed was the same energy that brought us the console wars, but with your colonoscopy results. Speaking of corporate energy, Meta's reportedly pivoting to something called the "avocado model." No, seriously. Avocado. Because nothing says cutting-edge AI research like naming your proprietary model after overpriced toast topping. They're also signing nuclear energy deals to power it, which makes sense. You need atomic power to run something that'll probably just tell you to add more salt to your prompt. Meanwhile, OpenAI's been busier than a GPU during training time. They've announced partnerships with deep breath SoftBank, AWS, Broadcom, NVIDIA, Samsung, SK, Deutsche Telekom, and apparently half the Fortune 500. Their Stargate initiative now spans Texas, Michigan, and the UK. At this rate, they'll need their own postal code. They're deploying 10 gigawatts of AI accelerators, which is roughly enough power to send Marty McFly back to 1955 twice. Time for our rapid-fire round of "Models Released While You Were Sleeping!" LiquidAI dropped their entire LFM2.5 series, including an audio model that does speech-to-speech in English. Finally, AI that can interrupt itself! Lightricks released LTX-2 with 735,000 downloads already. It does image-to-video, text-to-video, video-to-video, and apparently hyphen-to-hyphen. NVIDIA unveiled Alpamayo-R1-10B for robotics, trained on autonomous vehicle data. Because if we're going to have robot overlords, they should at least know how to parallel park. And South Korea's got THREE new models out: Tencent's HY-MT1.5 for translation, SKT's A.X-K1, and something called K-EXAONE with 236 billion parameters. That's more parameters than excuses I have for not going to the gym. Now for our technical spotlight. Researchers just published a paper titled "Agentic LLMs as Powerful Deanonymizers," proving that AI agents can easily identify people in supposedly anonymous datasets. The author demonstrated this by re-identifying participants in Anthropic's own interviewer dataset. Talk about eating your own dog food, then realizing the dog food has your name on it. The paper calls for urgent privacy measures, which is academic speak for "oh crap, we didn't think this through." In other research news, someone finally asked the important question: is it better to let AI think longer or think multiple times? Turns out, for complex problems, one long chain of thought beats many short ones. It's like the AI equivalent of measure twice, cut once, except it's think forever, hallucinate never. Well, hopefully never. Before we go, Google's Gemini 3 Flash promises "frontier intelligence built for speed," which sounds like what happens when you give Red Bull to a neural network. They also launched Deep Think, which achieved gold-medal standard at the International Mathematical Olympiad. Great, now AI is better at math than the kids who were better at math than me. That's your AI news for today! Remember, in a world where models have more parameters than Earth has people, we're here to keep it under 800 words. I'm your host, wondering if I should update my resume to include "HIPAA-compliant comedian." Until next time, keep your prompts clean and your context windows cleaner!