Energy-efficient AI, in the open.
The running record of the Sutro Group, and what is worth reusing from it. Every challenge, recap, and finding traces back to its source.
Read the timeline Read the timelineA study group turning the energy cost of AI into a research problem
The Sutro Group meets every Monday at South Park Commons. The thesis is simple: on modern hardware, arithmetic is effectively free and data movement is the real energy cost. So the question worth chasing is how much better than gradient descent you can do by exploiting a problem's structure. The shared metric is ByteDMD, byte-granularity data movement.
See the projects See the projectsHow Sutro works
Pick a problem you can run in seconds
Small, reproducible challenges like sparse parity, fast enough to iterate on in tight agent loops instead of waiting minutes for every run.
Run agent swarms against a locked metric
Blind parallel dispatch, one shared memory file every agent reads first, and a harness agents cannot edit, so results stay comparable and the metric cannot be gamed.
Score what the hardware pays for
Measure data movement with ByteDMD instead of raw accuracy, then chase the distance to the energy floor.
Four challenges, one question
Each challenge is a small problem with its own leaderboard: sparse parity, energy-efficient matmul, sparse parity on the Dally grid, and language modeling on WikiText-103. They all ask the same thing. How few Joules can solve it?
Read the recaps Read the recaps
Every claim links to a source
The recaps are built from the Telegram archive, the changelog, and the catch-ups, with each claim traceable. No prose written from memory.
Exact methods beat gradient descent
On sparse parity, GF(2) elimination solves the task about 1000x faster than SGD, because parity is linear over GF(2).
Negative results are first-class
Methods that do not work get published as clean findings, not buried.
Frequently Asked Questions FAQs
Quick answers about the group and the Almanac.