The FBI's Remote Operations Unit — a classified team that builds exploits and intrusion tools — had members exposed in a breach linked to ShinyHunters. The agency that hacks others just got hacked back, and the fallout for ongoing operations could be significant.
Anthropic deployed a swarm of 950 Claude agents that discovered a molecular machine potentially representing a new gene editing mechanism. Dario Amodei confirmed the finding, calling it a genuine scientific discovery rather than a parlor trick. This is the clearest signal yet that AI is doing real wet-lab-adjacent science.
Google released gemini-3.8-flash-tts and a lite variant, featuring over 2,000 stock voices and a new voice design system you can prompt directly. Simon Willison already has a live playground up. The audio AI arms race is accelerating fast.
Former Google and Anthropic researchers launched Mirendil around one question: what happens when AI meaningfully contributes to its own development? The self-accelerating AI thesis is moving from thought experiment to funded startup. Watch this space.
The Fed is formally advancing proposals to operationalize the GENIUS Act stablecoin framework. Regulatory infrastructure for digital dollars is being built in real time, even as New York sues Polymarket for illegal gambling and Bitget absorbs a $352 million hack via spoofed transfers.
Scientists found that diseased fruit flies lived longer when removed from Earth's magnetic field, while healthy flies died sooner — a first-of-its-kind result nobody expected. The implications for human longevity research are wide open and genuinely strange.
Amodei confirmed that Claude-led agents discovered what may be a new gene editing mechanism — a molecular machine whose precise function is still being investigated. He was measured but clearly excited. This is the moment AI labs have been promising for years: not generating text about science, but doing the science.
Mollick dropped something unusually blunt: regardless of risk, revenue, or the constant X debates, the thing that actually matters in AI development right now is just the rate of capability improvement. No hedging, no caveats — a straight read on where the momentum is pointing.
The 30-year mortgage rate hit 7.45%, the highest in three years, after 10-year Treasury yields spiked nearly 30 basis points in 48 hours. Markets are selling off — Nasdaq down over 1%, VIX spiking 7% — and the bond market is doing the damage. The macro backdrop for tech bets just got materially worse overnight.
Raschka made a clean, underrated point: the appeal of open-source agent harnesses isn't the price — it's inspectability. When an agent is running on your machine, you need to see what it's doing. As agentic systems touch more sensitive infrastructure, this will become a compliance and trust argument, not just a philosophical one.
Two threads are colliding this week that deserve to be read together. First: Anthropic's 950-agent swarm found a potentially novel gene editing mechanism. Second: Mirendil — founded by ex-Google and Anthropic researchers — is explicitly building AI that improves itself. These aren't separate stories. They're sequential chapters. The question is no longer whether AI can do science. It's how quickly the loop closes between AI doing science and AI doing