Welcome to Tech Beat, your daily look at the stories shaping our digital world.
A homelab enthusiast has done something that would have seemed implausible just a few years ago — running a two-point-eight-trillion-parameter AI model entirely on CPU, using discontinued second-generation Optane persistent memory modules. No GPU in sight. The build runs hot enough that active cooling on the memory sticks isn't optional, it's essential, with dual blower fans required just to keep temperatures manageable. It's a remarkable demonstration of what determined hobbyists can squeeze out of enterprise hardware the industry has already moved on from.
On a lighter note, someone has built a bash script called RAM-Altman that monitors your unused system memory and emails Sam Altman about it. It's a joke, obviously, but it lands because it touches something real — the tension between AI companies hungry for compute and the idle resources sitting in machines everywhere. Satire as systems administration.
Meanwhile, the Futhark programming language team has published a thoughtful piece on nondeterminism — specifically how parallel GPU programs can produce results that vary between runs. It's a quiet but important problem. When your hardware is racing itself to finish calculations, reproducibility becomes genuinely hard to guarantee, and the Futhark team is being unusually honest about where their language stands.
That's what's moving today. Keep surfing. Tech Beat out.
