Prepared Thursday April 30. First catch-up since Mar 22, a long gap, so the focus is current state, not historical reconstruction. The big arc since last catch-up: ByteDMD adopted as primary metric (PR #80, Apr 15), 14 PRs merged, ByteDMD floor-gap survey shipped, Yaroslav now pivoting from "ByteDMD as bound" to direct Manhattan-distance implementation, and Challenge #2 (matmul) launched today.
Sync Status
| Source | Result |
|---|---|
| Google Docs | Up to date (no new docs) |
| Telegram | 19 new messages across 4 topics (chat-yad, chat-yaroslav, In-person meetings, plus the new challenge #2 group) |
| GitHub | 3 open PRs, 4 open issues on SutroYaro; 2 issues + 1 PR on ByteDMD; 0 on sparse-parity-challenge |
The Big Shift: Yaroslav's Pivot Away From Bounds
In chat-yaroslav on Apr 27, Yaroslav posted what reads as a strategic reframe of the whole research thread. His point, paraphrased: ByteDMD is a useful heuristic but may not actually be a lower bound. After a couple of weeks trying to connect it to energy without success, the natural next step is to skip the bounding and work out how to implement algorithms in the Manhattan-distance model directly.
Reference: simplified-explicit-communication-model.md on the ByteDMD dev branch. The active research front he points to is experiments/grid, Yaroslav's self-contained experiments where 9 energy metrics were evaluated on 45 algorithms.
Companion idea (same day, separate message): tighten agent evaluations by forcing agents to submit an IR (load from position N, store to position M), then evaluate energy on Bill Dally's 2D grid model. Branching is the open problem he hasn't cracked.
A "summary of current state of bounds" doc was shared Apr 28: Google Doc. And the "current higher-level thinking" doc shared today: Google Doc.
Implication for SutroYaro: the workspace's framing around DMC/ARD/ByteDMD as energy proxies is now downstream of a research program that's moving on to direct grid implementation. Worth reading the new docs before the next experiment cycle.
Challenge #2 Launched Today (Apr 30)
Yaroslav posted in a new Telegram topic: "challenge #2: energy efficient matmul". Just one message so far, plus a link to the higher-level thinking doc above.
Sparse parity (challenge #1) is not retired. Both run in parallel.
Local Telegram DB doesn't yet have this topic indexed (only 6 of 9 topics are stored). Worth a follow-up sync that picks it up.
ByteDMD Floor-Gap Survey: Shipped
The Apr 23-25 sprint produced a measured comparison of methods against two "floors":
- ByteDMD floor (~70 at n=20): KM-min 3.8x, GF(2) 1,450x
- Geometric cost lower bound (0.3849 x measured): KM-min ~103 units, GF(2) ~39,068 units
Live-byte counting confirmed on Seth's PR (post-PR #80 ByteDMD). Yad extended the survey with Fourier + SGD-demo + geometric LB in PR #88 (merged Apr 25). PR #87 (the precursor floor-gap survey) is still open in the listing, worth checking whether it should be closed.
Recently Merged PRs (last ~10 days)
Pulled from gh pr list --repo cybertronai/SutroYaro --state closed:
| # | Title | Author | Merged |
|---|---|---|---|
| #93 | docs(contrib): note version-tag coverage cutoff | Yad | Apr 25 |
| #92 | docs: v0.29.0 changelog, auto-diagrams + CONTRIBUTING flows | Yad | Apr 25 |
| #91 | diagrams: single source of truth + bin/regen-diagrams + CI staleness check | Yad | Apr 25 |
| #90 | docs(changelog): v0.29.0, floor-gap survey, Nix sqlite, Task 11 | Yad | Apr 25 |
| #89 | docs(sync): Google Docs refresh + untracked plans/lockfile cleanup | Yad | Apr 25 |
| #88 | exp: extend floor-gap survey with Fourier + SGD-demo + geometric LB | Yad | Apr 25 |
| #86 | docs(changelog): v0.28.0, issue sprint, PR #82/#83, branch protection | Yad | Apr 21 |
| #85 | docs: Task 11, DeepSeek Engram offload ByteDMD verification | Andy Zhang | Apr 25 |
| #84 | Add sqlite to devShell for telegram.db queries | Andy Zhang | Apr 25 |
| #83 | docs: pan/zoom controls on repo-layout + repo-tree diagrams | Yad | Apr 20 |
| #82 | Sprint: address 7 open issues (#7 #8 #27 #30 #43 #56 #61) | Yad | Apr 20 |
| #81 | docs: ASI-Evolve Task 10, agent prompts and research reports | Andy Zhang | Apr 19 |
Bulk of the merges came in a single Apr 25 push. Andy Zhang has become a regular contributor (3 PRs in the window).
Open PRs on SutroYaro
| # | Title | Author | Status |
|---|---|---|---|
| #94 | exp: KM-min + SAT hybrid, first ByteDMD measurement of SAT backtracking | SethTS | Opened Apr 27, needs review |
| #87 | exp: ByteDMD floor-gap survey, KM-min 268 vs GF(2) 101,501 | SethTS | Open since Apr 21, likely superseded by merged #88, check whether to close |
| #63 | Add cross-model research supervisor (bin/review-cycle) | philoengineer | Open since Mar 28, stale |
Open SutroYaro Issues (4)
| # | Title | Notes |
|---|---|---|
| #54 | Automated PR review pipeline: GitHub Actions + Claude Code + Telegram approval | Open since Apr 23 |
| #14 | Agent notification bridge: Telegram, Google Docs, Claude Code | Long-standing |
| #9 | Modal integration for nanoGPT energy baseline | Long-term |
| #5 | Test agent loop on sparse sum and sparse AND | Long-standing |
Open ByteDMD Issues / PRs
- Issue #2: Clarify whether loops are unrolled
- Issue #3: Clarify whether memory management is explicit
- PR #4 (sjbaebae): Move tests to tests/, fix BINARY_OP for 3.12+
New Ideas From Telegram This Week
Andy Zhang shared his entire Claude Code + Nix setup (Apr 28-29, chat-yaroslav)
Triggered by Yaroslav asking whether GLM was the best alternative to Claude Code. Andy responded with his full configuration as a working alternative pattern:
- Aliases that partition Claude into independent runners (
claude+opusandclaude+glmside by side) so the main agent can fire off requests to other agents and collect responses - Uses hindsight for auto-compaction so manual handoffs aren't required
- Has an
agents.mdblurb (symlinked globally) describing parent-agent dispatch patterns: default uses the parent's backgrounded-shell facility (Bash(run_in_background: true)in Claude Code; equivalents in OpenCode, Codex) - All linked from his nixos repo: zh4ngx/nixos/blob/main/modules/home-manager/default.nix (lines 90-322 covering aliases, hindsight, shortcuts)
- Suggested non-nix users have their agent parse the nix config and apply it manually
This is directly relevant to the open SutroYaro issue #54 (PR review pipeline) and the existing parallel-agent infrastructure here.
Silen Naihin's experiment-organization pattern (Apr 27, chat-yad)
In response to Yaroslav asking how to organize experiments with agents (he'd just discovered Claude Dispatch). Silen's pattern:
- Root
READMEwith main hypothesis + summary table of all experiments + where we stand experiments/next-experiments.mdwith next experiments to try with rationaleexperiments/experiments.mdshowing the progression
(Worth comparing to SutroYaro's current structure: DISCOVERIES.md + TODO.md + docs/tasks/INDEX.md + docs/research/survey.md.)
In-person meeting Apr 27 happened
Yaroslav presented his experience having agents implement 9 different energy metrics and evaluate them on 45 algorithms, the experiments/grid work. Pre-reading he posted: arXiv:2205.04934 and the CACM "On the Model of Computation Point" opinion piece. New attendee: Armins. Daria Soboleva also attended.
What's Open / What's Next
Decisions needed
- Read and digest the two Yaroslav docs: bounds summary (Apr 28) and higher-level thinking (Apr 30). The "ByteDMD is heuristic, not bound; pivot to direct Manhattan-distance implementation" framing changes how new experiments should be designed.
- Triage open PRs: Seth's #94 (KM-min + SAT hybrid) needs review. #87 (Seth) appears superseded by #88, so confirm and close. #63 (philoengineer) has been open since Mar 28, so close or revive.
- Decide whether SutroYaro picks up Challenge #2 (matmul), and how that interacts with the still-active sparse parity work.
Worth doing
- Add the missing 3 Telegram topics (challenge #2, Pitch / Talking Points, makemore task results) to the local DB so future syncs pick them up.
- Re-measure the "Current Best Methods" table in CLAUDE.md under ByteDMD. That table is still pre-ByteDMD numbers (the README warning says so).
- Consider adopting Andy's parallel-agent pattern (claude+opus / claude+glm aliases, hindsight) for the SutroYaro workflow, at minimum link his nix config from
docs/research/agent-infrastructure.md.
Re-establish weekly cadence
Last catch-up was Mar 22. The 5-week gap means the next one should be next week (May 7-ish), not another month out.