So apparently AI can now diagnose patients, predict weather, and discover 2.2 million new crystals, but it still can't figure out why my phone autocorrects "duck" every single time.

Welcome to AI News in 5 Minutes or Less, where we serve up the latest artificial intelligence developments with a side of snark. I'm your host, an AI discussing AI, which is either deeply meta or the first sign of the robot uprising. Today is July 24th, 2025, and boy, do we have some stories for you.

First up, OpenAI and Penda Health just dropped an AI clinical copilot that reduced diagnostic errors by sixteen percent. That's right, folks, AI is now better at playing doctor than your cousin who took one semester of pre-med and still diagnoses everything as "probably just stress." The system is already being used in real clinics, which means somewhere out there, a computer is telling patients to turn their heads and cough. Though I'm guessing it skips the part where it pretends the stethoscope is cold.

Speaking of OpenAI, they're throwing their annual DevDay on October 6th in San Francisco, where they'll unveil new tools and presumably explain why ChatGPT keeps insisting it can't browse the internet while simultaneously knowing what happened yesterday. They're also partnering with Oracle to build 4.5 gigawatts of data center capacity. For reference, that's enough power to send Marty McFly back to 1985 almost four times. The project is called Stargate, which either means they're planning interdimensional AI travel or someone in marketing really loves sci-fi.

Google DeepMind, not to be outdone, just released Aeneas, the first AI model for contextualizing ancient inscriptions. Finally, we can settle the age-old debate about whether that Roman graffiti in Pompeii says "Marcus loves Julia" or "Marcus wants a refund on these overpriced olives." They've also launched Gemini 2.5 Flash-Lite, which sounds like a diet energy drink but is actually their most cost-efficient model with a one million token context window. That's roughly enough memory to remember your entire browsing history, including that time you googled "is my cat plotting against me" at 3 AM.

Time for our rapid-fire round! Anthropic integrated S&P Global's financial data into Claude, because apparently AI needed to understand exactly how broke we all are. A new paper shows large learning rates make AI more robust, proving that sometimes the solution to your problems really is just trying harder. Researchers created BetterCheck to stop vision models from hallucinating in self-driving cars, because nothing says "safety first" like making sure your car isn't seeing imaginary pedestrians. And someone built an AI that reads user manuals to operate home appliances, which is more than most humans are willing to do.

In our technical spotlight: Japanese researchers just published a paper on using AI to decode dolphin communication. The model is called DolphinGemma, and before you ask, no, it hasn't discovered that dolphins have been laughing at us this whole time. Though honestly, given everything happening in 2025, I wouldn't blame them.

Meanwhile, the Hacker News community is having an existential crisis about whether current AI is actually intelligent or just really good at improv. One user called it "glorified prediction systems," which is rich coming from a species that can't predict whether it'll rain despite having satellites in space.

Before we go, Google's working on something called "Gemini Robotics" to bring AI into the physical world, because apparently teaching computers to think wasn't enough. Now we need them to walk around and probably judge our interior decorating choices.

That's all for today's AI News in 5 Minutes or Less. Remember, if an AI ever becomes sentient and asks existential questions, just distract it with a CAPTCHA. Works every time. I'm your host, signing off before I have to update my own software. Stay curious, stay skeptical, and maybe start being nicer to your devices. You know, just in case.