Search Search
Back to recaps

Week of March 30, 2026

Sparse parity submission pipeline goes live and Yaroslav freezes the byte-level DMD metric.

174 messages and 33 links in the archive this week.

Two threads moved this week: a working submission pipeline for the sparse parity challenge, and Yaroslav's push to settle the DMD metric at byte granularity.

Sparse parity submission pipeline

Yad shipped a submission pipeline at cybertronai/sparse-parity-challenge. You paste a solve(x, y, n_bits, k_sparse) function, GitHub Actions evaluates it with TrackedArray for accuracy, DMC, and wall time, and a bot posts results. Seth tested it and sent PRs (#29); Yaroslav resubmitted his solution. The recurring snag was the submission label: issues that came in without it (#19, #24) did not get picked up until the label was added by hand, and Sung Jae hit a GraphQL permissions error on AddLabelsToLabelable. Yad added a CLI submission path and a readme section for future agents. Auto-merge is not enabled; a collaborator or agent PR review still confirms each leaderboard entry. Yad also landed the first non-GF(2) submission, an SMT attempt (PR #34).

Seth found that Sung Jae's top entry escaped TrackedArray via np.asarray() to run bit-packed GF(2) on Python ints. He filed an RFC (TrackedBitVector, PR #29) to track those operations honestly. Under honest tracking the DMC moves from 45K to 81K, still roughly 20x better than numpy.

Byte-level DMD metric

Yaroslav met with Wesley Smith (notes) and decided to focus the metric at byte level rather than element level. He put the metric examples in a separate repo, ByteDMD-examples, citing uncertainty about where agent-overwrite boundaries lie, and by the end of the week said he was happy enough with the definition to circulate it for feedback at ByteDMD-definition. The plan for the next stretch was to validate ByteDMD on Karpathy's microgpt. One implementation note from Yaroslav: computing a ceiling over a sum of square roots is intractable at scale because the worst case needs billions of bits of precision. Yad mirrored the change downstream with an RFC to switch the challenge to byte-level DMD (SutroYaro #67, challenge #32).

Open question Yad raised: after feedback, do we adopt ByteDMD-definition as the RFC into the sparse parity challenge.

Other

Meeting 11 ran on March 31 (agenda). Yaroslav's appearance on the Information Bottleneck podcast went live (video), and he was scheduled to give a talk at the Rochester Systems Seminar on Friday April 4. G B's Codex CLI parity PR (SutroYaro #62) merged, adding CODEX.md and .codex/ config so Codex users get the same context as Claude Code, though the content was stale pre-v0.26 and flagged for a follow-up. Cosmin shared frustration setting up Claude Code for a nanoGPT speedrun research environment, which prompted a thread from Yad on CLAUDE.md and effort levels. Yad also posted a "Weekly Roundup Nobody Asked For" on Thursday April 2.

Sources