I've spent most of my working life on other people's computers — bank datacentres, then bank clouds, machines I could use but never own, behind change windows and approval chains. In March I bought a Mac Studio, an M4 Max with 64GB of unified memory, and put it on a desk in Pune. For the first time in a long while, the most capable computer in my life answers only to me.
It runs models locally, which is why I bought it. A thirty-billion-parameter model, quantised to four bits, loads in seconds and writes faster than I read. Code models, chat models, Whisper on old recordings, diffusion models turning sentences into images, the machine staying so quiet I check whether it's actually working. It is. CPU and GPU sharing the same memory means a model that would want a serverside card just fits. The fans don't bother.
What surprised me wasn't the speed. It was what disappeared — the meter. Every cloud call carries a small voice asking whether it's worth the money, and I hadn't noticed how much that voice was shaping my curiosity until it went quiet. Now I'll run a prompt forty times to watch how temperature changes the failure modes. I'll leave something generating overnight, because why not. Waste turns out to be a precondition for a certain kind of learning, and renting compute had quietly priced waste out of my habits.
Computing keeps doing this loop — mainframes you petitioned for time on, then the PC on the desk, then the cloud pulling everything back into a cathedral. This feels like the desk winning another round. Nothing on this machine trains anything; it's a reading room, not a factory. But a reading room is enough. Most of what I understand now about these models, I learned by being alone with one, late, asking it one more thing.