Search Search
Back to recaps

Week of April 27, 2026

Challenge #2 energy-efficient matmul opened, and the hinton-problems v1 catalog shipped all 53 stubs.

48 messages and 21 links in the archive this week.

A focused week, 48 messages, split between two private threads and the new challenge group.

Challenge #2: energy-efficient matmul

Yaroslav opened challenge #2 with the problem posted to cybertronai/sutro-problems/matmul on April 30 and asked Sung Jae, Seth, and others to try it. Seth sent PR #1 and Sung Jae followed with PR #2 the same day. Yaroslav left review comments and asked whether agents could read the PR and implement them directly, and later asked Anastasiia, Seth, and Sung Jae to have their agents add a short report to each submission. His higher-level framing is in a Google Doc.

hinton-problems v1 shipped

Yaroslav asked Yad about the state of his agentic loop, since he wanted reference implementations for all of Hinton's learning problems and had only encoder-4-2-4 done. Yad posted a minimum spec as issue #1 and planned to deploy waves of ten implementations at a time once an unrelated release finished. By May 2 he had 33 of 53 stubs open across six wave PRs, and by May 3 hinton-problems v1 was shipped: all 53 stubs in main across ten wave PRs (#32 through #41), with a RESULTS.md catalog in PR #42 reporting 27 stubs that reproduce paper claims and 25 partial. The site is live. Yad also stood up a related repo for ICLR efficient-learning and matmul literature at iclr-lit-builder, not yet moved.

ByteDMD and tooling

Yaroslav noted that ByteDMD is a useful heuristic but may not actually be a lower bound, and after weeks trying to connect it to energy without success, the next step is to skip the bounding and tighten agent evaluations by forcing an explicit IR and scoring energy on Bill Dally's 2d grid model. He posted a current-bounds summary doc and the simplified explicit communication model. He asked about alternatives to Claude Code, and Andy walked through his Nix-based setup for running multiple agents in parallel, linked from his home-manager config.

In-person meetings

Armins joined for a first meeting and asked about the topic and pre-reading. Yaroslav planned to cover having agents implement nine energy metrics across many algorithms, with pre-reading at arXiv:2205.04934 and the CACM model-of-computation piece, plus the ByteDMD grid experiments.

Open question

Yaroslav's question about whether ByteDMD can be tied to energy as a lower bound is still unresolved, with the explicit-IR plus 2d grid approach proposed as the way forward.

Sources