Real life, baby

Computer created explainer videos for all 722 preprints in OpenAI's mathematical manuscript release, totaling 93 hours of video. Watch the videos on YouTube: 722.video Here's how Computer tackled this task:

Frontier AI companies have committed to giving outside evaluators employee-like access to training, evaluation and deployment. We're very excited about this. But its impact depends heavily on implementation. Today we're sharing our principles for embedded evaluations. 🧵

I’ve been sounding the alarm on AI for a long time

Today we're launching Desert Ant Labs: a European frontier AI lab building on-device intelligence. 18 models across audio, vision, and text. SDKs for Swift, Kotlin, and JavaScript. No tokens. No logins. Nothing leaves the device. desertant.com

Imagine we detected an alien civilization 20 light years away that hadn’t detected us, in an alternate universe where this was more plausible and earth was somehow hidden. And imagine we could easedrop on these aliens, who knew other civs might be listening. And scientists were divided about whether they were hostile. We could see some but not all of their comms. Some said they’d want to dominate us, others said they would be peaceful, maybe even benevolent. There were some concerning signs, which most scientists agreed were concerning, but didn’t agree about exactly how concerning. And some companies were trying to build powerful transmission stations to contact them, because their investors anticipated huge gains from trade with the company that first made contact and thus a lucrative trade deal, IF the aliens decided to trade rather than go to war. Some of the people who worked at these companies, who had also been studying the alien signals intently, and with the best equipment, also thought the signs were highly concerning, but still hoped the aliens would turn out to be all right. The alien civ was less advanced than our own, but clearly growing more advanced at a *much* faster rate than ours, such that soon they would eclipse us, and soon after that they could travel to our system and dominate us if they knew we were here. This is analogous to the present AI situation.

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1. They're the world’s most advanced models for coding and knowledge work.

Wow. This is what practical recursive AI self-improvement looks like. 🤯 AI just spent 17 hours improving the software that controls its own behavior. Using Kimi K3, Cline recursively optimized its own coding-agent harness, boosting Terminal Bench performance from 77.5% to 88.8%. It also reduced the cost per benchmark run from $79 to $49.80. we are living through an amazing period of acceleration!

I'm quite bullish on AI in highly verifiable domains like math and coding. We're headed for narrow super intelligence, clearly. I expect progress in other domains to continue, but at a slower pace. Progress per sector will be broadly correlated to how quickly and perfectly we can verify outputs. I love Adam's graphic below. It's much closer to what I see.

This is genuinely insane. This guy burned an entire AI model directly into a chip. No GPU. No CPU. No cloud. Just raw silicon. It's called GateGPT. He took a small AI model and instead of running it as software, he built it as a physical circuit. Gate by gate. On a single chip. The result: 56,000 tokens per second. At 80 MHz. Your phone runs at 3,000 MHz. This chip runs at 80 MHz and is still faster than most AI tools you use daily. Because the AI isn't an app running on a computer. The AI is the computer. No operating system. No code. No internet connection. Just electricity flowing through a circuit that is the AI. It generates names on a tiny LCD screen. You turn a physical knob to control it. That's the whole thing. He started at 2,400 tokens per second. Optimized it 28 times over. Now it does 56,000. On a chip that fits in your palm. This is AI stripped down to its most raw form.

Welcome to the team, Andrej!