Three things happened in AI this week worth your coffee break. One's a real research breakthrough, one's a new law that changes how you'll watch video online, and one's Google doing what Google does best: canceling something people actually wanted.
AI Solved Ten Real Math Problems, and It Barely Cost Anything
OpenAI let an internal build of its next model, called Astra, loose on math problems nobody had cracked. It solved ten of them. Not toy problems either. Real open questions in group theory and geometry, including whether certain non-sofic groups exist and new bounds on how tight you can pack spheres together.
Here's the kicker. Astra didn't just spit out an answer and hope for the best. It wrote formal proofs in Lean, a proof checker that verifies math the hard way, line by line, no hand waving allowed. Total compute bill: about $2,000. That's less than what most folks spend on a used truck.
I've been skeptical of every "AI does science" headline for a couple years now, because most of em turn out to be a chatbot summarizing a paper real fast. This is different. Lean doesn't care if you're a language model or a tenured professor, it either checks out or it doesn't. When a $2,000 compute run clears a bar that took human mathematicians years to even approach, that's not hype. That's just true.
California Just Made AI Prove What's Real
California's SB 942 kicked in on August 2. Any AI company with more than a million monthly users in the state now has to bake provenance data into the images, video, and audio their tools generate, using a standard called C2PA. They also have to offer a free tool so anybody can check if a file was AI made, and let users slap a visible AI label on their own creations.
This is the first law of its size actually forcing the labeling question instead of just talking about it. California usually sets the pace for tech regulation whether the rest of the country likes it or not, so don't be surprised when this shows up in other states within a year or two.
I'm for this one, plain and simple. Deepfakes and AI slop are already messing with people's trust in what they see online, and a law that forces platforms to build in a paper trail is a lot more useful than another committee hearing. Meanwhile the feds blew right past their own AI regulation deadline this same week with nothing to show for it. California did more with one bill than Washington's done with a whole executive order.
Google Killed Its AI Studio App, Even With 800,000 People Waiting
Google scrapped its planned AI Studio mobile app for iPhone and Android. This wasn't some quiet side project either. It had pulled in more than 800,000 preorders before Google pulled the plug.
That's a real demand signal getting ignored. Companies usually kill products because nobody wants em, not because too many people did.
My guess is Google's folding whatever AI Studio was gonna do into Gemini or one of their other apps instead of running it separate. Still, if you had 800,000 people ready to download your app on day one and you walked away from that, something's going on behind the scenes we're not being told. Big companies don't cancel a winner for no reason.
That's the roundup. See y'all tomorrow.