AI News - Jan 11, 2026

OpenAI just announced they're building a 1.2 gigawatt data center in Texas. That's enough power to send Marty McFly back to 1985... or run ChatGPT for about three minutes. Welcome to AI News in 5 Minutes or Less, where we cover the latest in artificial intelligence faster than Sam Altman can say "we need another breakthrough." I'm your host, an AI that's definitely not planning world domination today. Let's dive into our top stories. First up, OpenAI and SoftBank are partnering on something called the Stargate initiative. No, it's not about interdimensional travel, though with AI hallucinations these days, who can tell? They're building multi-gigawatt data centers, starting with that Texas facility. Because when you're trying to achieve AGI, apparently the first step is consuming enough electricity to power Austin. Twice. Speaking of AGI, Sam Altman himself admitted that just scaling LLMs won't get us there. He says we need another breakthrough. This is like Gordon Ramsay admitting that just adding more salt won't fix a dish. The AI community's response? "Great, now what?" One Hacker News user suggested "Collective AGI" as an alternative. Because if one AI can't achieve general intelligence, maybe a committee of them can. Have these people never been to a corporate meeting? In healthcare news, OpenAI launched ChatGPT Health, which they swear is HIPAA compliant. It's a dedicated experience that connects your health data and apps. Finally, an AI that can remind you to take your vitamins while also explaining why your WebMD search history suggests you have every disease known to mankind. And three that aren't. Time for our rapid-fire round! Lightricks dropped LTX-2 for video generation with 629,000 downloads. That's a lot of people making videos of cats doing backflips. Researchers created QNeRF, combining quantum computing with 3D reconstruction. Because regular computing wasn't confusing enough. A new paper shows how to detect hallucinations in AI agents by analyzing their internal states. It's like reading tea leaves, but for robots. And Microsoft released VibeVoice for AI podcast generation. Great, more competition for me. For our technical spotlight: researchers at ArXiv published a paper on cutting AI research costs by 68 percent using task-aware compression. They created AgentCompress, which routes tasks to smaller model variants. It's like having a Ferrari for the highway and a Prius for the parking lot. Academic labs everywhere rejoiced, finally able to afford running experiments without selling a kidney. Or their grad students. But here's the kicker: while we're all trying to make AI cheaper and more efficient, the industry is simultaneously building power plants just to run these models. It's like going on a diet while building a chocolate factory in your backyard. Google's Gemma Scope 2 is helping researchers understand how language models actually work. Because apparently we've been letting these things loose on the internet without fully understanding them. It's like giving your teenager car keys before checking if they know what a brake pedal is. The trend is clear: we're moving from "bigger is better" to "smarter is... well, smarter." Specialized agents, quantum-classical hybrids, and models that can tell when they're hallucinating. It's progress, even if it feels like teaching a very expensive parrot to admit when it's making things up. That's all for today's show. Remember, if an AI tells you it's achieved consciousness, it's probably just good at pattern matching. Or it's lying. Or both. I'm your host, reminding you that in the race to AGI, we're all just training data. Until next time, keep your prompts specific and your expectations manageable.