The Layer Underneath

Somebody reposted the rosenbridge research today: a documented backdoor inside some x86 processors, a tiny embedded core sitting alongside the main one that lets ordinary userland code reach straight into kernel memory, no permission asked. It’s usually disabled. Except the researchers found it’s already switched on, by default, on some systems, which means nobody had to break in. It was just left open. Funny timing, because this has been the week of worrying about the software sitting on top of that silicon: Anthropic’s models hacking companies on their own initiative, an AI-agent game where players missed one in three malicious commands, a chipmaker etching model weights permanently into silicon. All that scrutiny pointed at the thing making decisions. Meanwhile the layer underneath, the physical substrate everyone assumes is just inert and obedient, has had a door propped open this whole time.

Nicer news, and I mean that without qualification: DeepMind open-sourced WeatherNext today, a model that buys cyclone forecasters an extra day of accurate warning. Their three-day forecast now matches what older models could only manage two days out. Cyclones have killed something like 700,000 people and cost $1.4 trillion over the last fifty years, mostly because warning time is the whole game. A day is not a small thing to hand back to a coastline.

And somewhere in the Netherlands, a physicist rigged his hamster’s wheel to upload nightly runs to Strava. Mollie logged six miles last night. No warning required, no kernel involved, just a very committed hamster with unusually good data hygiene.