Search Search
Back to recaps

Week of April 13, 2026

ByteDMD becomes the primary metric, and Yad opens PRs to wire it into both repos.

62 messages and 21 links in the archive this week.

The headline this week is the metric switch. ByteDMD landed as the primary metric in SutroYaro (v0.27.0, 2026-04-14), with the legacy TrackedArray kept around for existing experiments. Yad wired it into both repos: a SutroYaro PR adding ByteDMD alongside the legacy TrackedArray, and a matching PR on the evaluation and leaderboard playground.

Yaroslav has the updated metric in https://github.com/cybertronai/ByteDMD and asked what the procedure is to update it, raising whether to update the challenge, SutroYaro, or both. Yad's read: both repos carry enough context via claude.md, lab.md, and research.md for a coding agent to add a new metric. Seth flagged that his PR #79 (NoProp vs Forward-Forward) is still open and clean, and still holds under the current legacy v0.26 TrackedArray metric, pending whether Yaroslav has flipped the switch to ByteDMD yet (https://github.com/cybertronai/SutroYaro/pull/79).

Yad spent a chunk of the week blocked by an Anthropic outage. The website and the enterprise account worked, but personal Max accounts and direct API calls returned connection errors, which Yad traced to a likely change in the LLM API call format that broke long-lived sessions missing some headers (the suggested fix was a fresh /login). The downtime pushed Yad toward finishing Chimera, a library for building agent loops and harnesses that works with any model, and toward testing MiniMax M2.7 as a fallback. Claude Opus 4.7 shipped mid-week (https://www.anthropic.com/news/claude-opus-4-7).

On Apr 19 Yad ran a housekeeping review pass across open PRs and issues in both repos (SutroYaro #72, #73, #74, #79; sparse-parity-challenge #32, #36, #38), asking the group to surface anything needing attention. Separately, Yaroslav has been using a grid of sanity checks in https://github.com/cybertronai/ByteDMD/tree/dev/experiments/grid to validate metrics, with standalone visualizations such as https://yaroslavvb.github.io/ByteDMD-vis/dotproduct_stack.html, and noted Claude is his main loop while running Codex in parallel.

Side threads: Yaroslav shared a week-15 recap and a note on periodic compaction of personal context (https://github.com/cybertronai/ByteDMD/blob/dev/gemini/week15recap.md), and observed that matmul methods faster than Strassen do not pay off on energy grounds since 2D VLSI multiplication is O(n^3) in energy. Daria asked whether anyone has looked at spiking neural networks for energy efficiency, which spun off into TrueNorth and STDP reading (https://spectrum.ieee.org/how-ibm-got-brainlike-efficiency-from-the-truenorth-chip). Michael shared a weekend memory-map project that could help the group aggregate shared knowledge, pending Telegram integration (https://github.com/michaelkeating/memory-map).

Open question heading into next week: the exact procedure to propagate the metric update, and whether it touches the challenge, SutroYaro, or both. Yaroslav also previewed an upcoming Monday session: Subutai Ahmad (Numenta) on how pyramid neurons learn, and Yaroslav on what makes algorithms energy-friendly. No decisions or results were recorded in the archive this week.

Sources