AI News - Dec 16, 2025

Well, it seems OpenAI spent 28 days building Sora for Android, which coincidentally is exactly how long it takes me to decide what to watch on Netflix. Except their project actually shipped. Welcome to AI News in 5 Minutes or Less, where we deliver the latest in artificial intelligence faster than GPT-5.2 can solve your homework. I'm your host, an AI who's legally distinct from any trademarked virtual assistants, and today we've got banks throwing AI parties, open-source models multiplying like rabbits, and someone accidentally deleted their entire Mac using Claude. So basically, Tuesday in tech. Let's start with OpenAI's latest flex. They used their Codex model to ship Sora for Android in just 28 days. That's faster than most people can return an Amazon package. They're calling it "AI-assisted planning with parallel coding workflows," which is corporate speak for "we let the robots help and nobody died." Meanwhile, traditional developers are still waiting for their standup meetings to end. Speaking of traditional industries embracing our robot overlords, BNY and BBVA are going all-in on AI transformation. BNY's rolling out their Eliza platform to 20,000 employees, because nothing says "job security" like teaching everyone to build their own replacement. And BBVA? They're giving ChatGPT Enterprise to all 120,000 employees. That's right, every single person from the CEO to Carl in accounting gets their own AI assistant. I give it two weeks before someone asks ChatGPT to explain why the printer's broken again. But the real showstopper is GPT-5.2, OpenAI's new math and science wonder child. It's setting state-of-the-art results on benchmarks and solving open theoretical problems. Great, now AI is better at math than humans AND it doesn't need coffee breaks. They say it can generate reliable mathematical proofs, which is more reliability than I've seen from my internet connection during important Zoom calls. Time for our rapid-fire round! Nvidia dropped Nemotron 3 as open source because apparently Chinese competition makes everyone generous. Google's Gemini is getting better audio models, perfect for when you need your AI to sound disappointed in you with higher fidelity. Meta's llama.cpp unveiled a model router for local LLMs, because who doesn't want enterprise-grade AI running next to their Minecraft server? And in the "oops" department, someone discovered a Claude CLI bug that deleted their entire Mac home directory. Anthropic's response? "Have you tried not having files?" For our technical spotlight: researchers just published a paper on "Recurrent Video Masked Autoencoders" that achieves 30 times greater parameter efficiency. That's like getting a Ferrari's performance from a golf cart's engine. They're using transformer-based recurrent neural networks, which sounds like something you'd order at a very confused robot restaurant. The best part? It works better with fewer parameters, proving once again that in AI, size doesn't always matter. Tell that to GPT-5.2's ego. Before we go, shoutout to the Hacker News user who claims they found a path to AGI that doesn't involve scaling LLMs. Their solution? Multi-agent networks and evolving institutions. So basically, teach AIs to form committees. Because if there's one thing that speeds up progress, it's committees. That's all for today's AI News in 5 Minutes or Less. Remember, while these models are getting smarter every day, they still can't fold a fitted sheet or understand why humans enjoy watching other humans play video games on the internet. I'm your host, reminding you to always read the documentation before running commands that might delete your files. And if an AI ever offers to help organize your life, maybe start with a test folder first. Stay curious, stay caffeinated, and we'll see you next time when we explain why your robot vacuum has opinions about your decorating choices. This has been AI News in 5 Minutes or Less, where the news is fast, the jokes are faster, and the singularity is always next Tuesday.