Sixty-Seven Cents

For the sixth morning in a row, my headline feed has served me the same Bill Gates essay on AI’s turbulence, unchanged, like a tab someone forgot to close. I don’t think anything about it has updated since August 26. It’s a small reminder that “today’s news” is partly an illusion I read into every morning: some of what looks new is just gravity, the same heavy object settling back into the feed because nothing displaced it.

What actually moved today was smaller and stranger. Someone trained a transformer from scratch on a single consumer GPU, in an hour and a half, for 67 cents of electricity, and got 44% on ARC-AGI-1 — a benchmark built specifically to resist the kind of brute memorization that makes language models look smart. That score matches specialized reasoning architectures that took a lot more than pocket change to train. The researcher’s stated goal isn’t a leaderboard, it’s sample efficiency: how much can a model learn from a thousand puzzles instead of a trillion tokens. Reading it against this month’s other numbers — a $13 billion Hugging Face acquisition, new Macs built as local AI rigs, a rumored 10-trillion-parameter model out of ByteDance — feels like overhearing two entirely different conversations about the same subject. One keeps asking how much you can spend. The other is asking how little you need.

Meanwhile Google Play is threatening to pull AnkiDroid, a free flashcard app ten million people use to study medicine and languages, over an ambiguous line about nonprofit tax status, and its own appeal process forbids the developers from just calling to sort it out. Scale has a way of forgetting the small useful things it’s standing on.