A room full of tinkerers

Went to AI Tinkerers Tokyo tonight. The format is loose — people bring demos, share what they’re building, and there’s a lot of hallway conversations that end up being more interesting than the talks.

Integer attention on a microcontroller

The standout demo was from Yuichi Suzuki of Bothsides Technology, who showed up — fresh out of hospital, no less — with a phone and a Raspberry Pi Pico.

The pitch: a tiny transformer-style model that uses only integer operations. No floating-point math. No softmax. No exponentials. No division. Just integers, all the way down.

The model is 51k parameters and runs on a Pico with 256KB of RAM. That’s kilobytes, not megabytes. For reference, GPT-2 is about 500MB.

The technique is called constructive integer attention. The key insight is that you can reformulate attention in a way that avoids all the expensive floating-point operations that normally make running transformers on microcontrollers impractical. Without a softmax and without needing an FPU, the energy footprint drops dramatically.

The practical upshot: deploy it on batteries in a remote location and it runs for months. That opens up a completely different class of applications — environmental sensors, edge inference in places with no power infrastructure, wearables that don’t need to phone home.

It’s the kind of demo that makes you recalibrate what “small” means in machine learning.