Meeting brief · Sutro #31 · Mon 14 Sep 2026 The competition is public. Nobody outside has entered. Tonight 18:00–20:00, SPC, 380 Brannan. Covers Mon 7 – Mon 14 Sep. Three things to fix, in order: make the board usable, give Andy an answer, book an organizational sprint for tomorrow morning. Shipped: MNIST went public Thu 10 Sep Blocked: PR #70 still conflicting Owed: release to Lucas Cassiano, today 1 Bottom line Last week the research shipped. The group didn't move. 0 Outside entries since launch, 4 days ago 11 PRs waiting on you 8 of Andy Zhang's are 9–11 days old 0.1 h Median merge, your own PRs the asymmetry collaborators see ~100 Minutes before 17:00 Qingqing Mao at 17:00, then the room Sutro became a public competition #30 agreed MNIST-as-transduction. Spec out Tue 8th, public Thu 10th, ranking moved to the pitch-128 multicore grid Fri 11th. Public, but newcomers can't play Every entry on the current spec is yours. PR #70 is conflicting, so main's generator and evaluator still describe the old spec. Organizing is the bottleneck, not research Planning sprint #2 stopped after five lines. Its one real sentence: “We need to listen to get more people involved.” 160 visitors in, nothing out Unique visitors, last 14 days160 New forks since launch0 New issues since launch0 Outside submissions0 Launched Thu 10 Sep 18:11. GitHub, checked 15:05 today. Your PRs merge in minutes. His wait a week and a half. Your own PRs, median time to merge0.1 h Andy Zhang's 8 open PRs, age, no comment9–11 days Same scale — 0.1 h is the sliver on the left. Andy is your most active collaborator. Recommendation for tonight: make the board work for a newcomer, then give Andy an answer and an owner role. Put a 90-minute organizational sprint on Tue 15 Sep so Wed 16 and Thu 17 stay Anthropic days. 2 Next ~100 minutes Four things fit before 17:00 — one of them is load-bearing 1 Resolve PR #70's conflict and merge it The conflict is mnist/README.md only; the 5-line fix is in outstanding-prs. Same commit: set accuracy_targets.json to match the README (still 60/98/98), and list your 96.41% entry in the 8% and 12% bands. If only one gets done, do this one — everything you show tonight depends on it. 2 Tell Lucas Cassiano the release is live One message. You promised him today. 3 Reply to Andy's 11:49 Telegram question About what you tested together in February. 4 Commit the reversible-net bundle Untracked in ~/git/sutro-problems. Label it honestly in the README: activation workspace flat at 19.8 MiB, end-to-end saving 3.8%, because augmentation dominates. 3 Tonight, 18:00–20:00 An agenda for “scaling up / analogue” The doc stub for #31 says only those three words. This fills them. 0–5 Why we're here — for newcomers Moving a byte costs more than adding it: a 1 mm wire ≈ 100 adds, a DRAM read ≈ 64,000 (Dally, CACM 2022). The algorithms were designed for the opposite world. Goal: discover a learning algorithm with a small memory footprint. 5–15 What changed since #30 — credit people by name MNIST public Thu 10th. Tiers: small 67%; medium 2/3/5/8/12% error bands; original 1%. Records: 4×4 matmul 1,316→675 (Juraj Selep); 16×16 66,300→63,819 (Alex Varga); MNIST-small −14% grid / −18% A100 energy (Alex Varga, PR #71). 15–25 One honest result Reversible nets keep activation memory flat — the whole run still saves only 3.8%. Say what the bottleneck turned out to be. Being honest here earns trust in the ruler. 25–40 “Scaling up”: which ruler do we trust? Grid model or A100? In the one matched test (#71) the old single-core model got the A100 time ordering backwards, and pitch-128 has only 2 scored entries. Ask the room: what would make you trust the ruler enough to compete? 40–50 “Analogue”: the second track Thomas Dybdahl Ahle's spicenn2 shipped in 30 h; his video is Mon 21 Sep (#32). Cross-list as a track, or keep it separate? Its joules can't be compared with the A100 ladder. 50–60 Organization: ask the room three questions Who wants to own something · what stopped you from submitting · which day should releases land on. These feed tomorrow's sprint. 60+ Hack / pairing Pair each newcomer with a runnable starter entry. 4 Decide in the room Three yes/no calls, each unblocking something Unblocks 5 PRs Do verified schedules that don't beat the record belong in matmul/? One ruling closes Andy's #49, #51, #52, #55, #56. Close #60 and #61 too — #61 would publish Telegram message IDs. Unblocks the ruler Do we freeze the constants table for v1? Pitch 1.0 / 1.2 / 1.22 µm and wire speed c/160 vs c/120 have all been used. Freeze one table; put the Manhattan-distance correction on the public design page. Unblocks newcomers Pick a release day Proposed: a spec or tooling version lands Monday before the meeting, never mid-week. 5 Say these out loud Four things stated wrong in earlier rooms A micron is a millionth of a millimetre — it's a thousandth. (said by you at #30) A 0.5 ps step is 2 GHz — it's 2 THz. (Alex Varga's slip) An add is worth “5 µm” of movement — Dally's figure is ~10.5 µm (20 fJ ÷ 1.9 pJ/mm; fix the Dally Heuristics page) “Numenta merged with half of Esperanto” — wrong story. Jason Yosinski accepted tonight; correct it with him first. 6 Last week Mon 7 → Mon 14 Sep, day by day Mon 7 Sepweek 37 08:02 call with Thomas Dybdahl Ahle on analog MNIST Alex Varga's two 16×16 records (14:53, 16:53) 18:00–19:52 #30, 9 people Check-in: “some reluctance getting into the zone for the sutra group” Tue 8 Sepweek 37 Grid-redesign dossier + the metrics/IR doc Andy asked for — promise kept Coffee with Mark Saroufim (megakernels) Lucas Cassiano wants to compete; release promised for today Dictated check-in: “the goal of the competition is to organize the effort” Wed 9 Sepweek 37 Thomas Ahle agreed to a video for Mon 21 Sep Coffee ask to DensityAI — no reply Most of the day on the Cerebras dropout paper Thu 10 Sepweek 37 18:11 — the competition goes public IR choice and GPU transpiling handed to participants; PRs #64–#68 are your baselines Andy's agent self-merged a PR; you asked for a revert and got it Fri 11 Sepweek 37 Pitch-128 model (#7) becomes the ranking model; medium tier split into error bands Apex FPGA demo with Hasan Unlu and Devrim Yasar Emails to Bill Dally and Ronny Krashinsky — Dally away until Tue 22 Sep README split into “for humans” and “for agents”; posted to SPC Slack and Geek Club Sat 12 Sepweek 37 Alex Varga's #71 merged; old model moved to “historical” Backprop memory-wall page published Breakfast with Jason Yosinski (Apical) Sun 13 Sepweek 37 18:06–20:58 reversible-net runs — the first measurement behind the no-backprop bet Natalia Vassilieva (Cerebras) critiqued the grid model Mon 14 Sepweek 38 · today Telegram: “we'll talk about the MNIST competition” Blog-post note on the competition's degrees of freedom #70 still conflicting at 15:05 The week cost 7–8× the agreed floor Energy / Sutro work, last week~35–40 h Kateryna Peters's floor5 h Displaced Anthropic time on Wed 9, Sat 12 and Sun 13 Sep. 7 People Who's owed what — top four decide tonight's tone Person Who they are You owe Andy Zhang Most active collaborator; wrote the dally-eval Rust/GPU scorer. 8 open PRs. The matmul/ ruling; approve dally-eval #1 and #2; reply to his 11:49 question Lucas Cassiano Wants to compete The release — promised for today Alex Varga 16×16 record; MNIST-small PR #71 Coffee you moved to “next week”, i.e. now. Book it Hasan Unlu Apex Compute (ex-Tesla Autopilot); arranged the FPGA demo Invited verbally but not on the invite — add him Juraj Selep 4×4 matmul record (675) Credit tonight Thomas Dybdahl Ahle spicenn2 analog MNIST; joined Telegram Confirm the Mon 21 Sep video slot Mark Saroufim GPU MODE, megakernels — “I can try!” on a single-kernel MNIST Check in; Boris Ginsburg ↔ Jared intro nudge Jason Yosinski Apical (formerly Numenta); possible real-chip access. Accepted Correct the Esperanto story Natalia Vassilieva Cerebras; listed six ways model 3 differs from real hardware Follow-up email (paper links, Mostafa Elhoushi) Vijay Jain Ex-PsiQuantum; asked to join, accepted LinkedIn reply; reading page not sent Suhrud Kulkarni SPC; ex-TD Securities quant. Accepted One bounded problem to own Mitchell Nahmias Sphere Semi; bringing optical-computing people Ask who is coming Bill Dally / Ronny Krashinsky NVIDIA; multicore question sent Resend Tue 22 Sep Rohan Virani, Iván Bogatyy, Islam Newcomers, added to invite/Telegram Welcome by name Cosmin Negruseri · Devrim Yasar · Christian Pehle Competitive programmer · Lucky Robots · neuromorphic Nothing outstanding “Accepted” on the calendar covers the whole recurring series, not tonight — twenty addresses show accepted, so it isn't a headcount. 8 The through-line #28 named the problem, #31 is the test #28 · MON 24 AUG Momentum was fading Promised a benchmark where the algorithm must be learned. Kept #29 · MON 31 AUG Picked the benchmark Scorer and benchmark kept (scoring became the contestant's job). EqProp reading and Eric Frank's memo: no evidence #30 · MON 7 SEP Designed it Spec kept, multicore decided, chip intro kept. Analog open, Andy's project not reviewed #31 · TONIGHT Does anyone else compete? The docs name this as the purpose of these evenings — and it's the test of sprint #1's “people love competitions”. 9 Direction Harden the ruler, then climb the ladder Sprint #1's cube still holds: move one axis at a time. Last week moved the problem axis a long way — sparse parity and matmul to MNIST. Move metric and process next, not problem. AXIS 1 · NOW Metric — the ruler disagrees with itself Re-score #64–#66, #70 and #71 on pitch-128 and publish one frozen constants table. That work is also the substance of the Dally resend. AXIS 2 · NOW Process — the newcomer path Make it work in 30 minutes, with a PR triage rule. AXIS 3 · LATER Problem — climb only after 1 and 2 MNIST-small → medium → original, then sprint #1's target: energy-efficient training of nanoGPT. Not picked: more hardware research Esperanto, Cerebras, optical — it produced people, but no competition entries. Not picked: scaling to nanoGPT now The ruler isn't trusted yet at MNIST scale. Runway constraint ~3 months left. Keep Kateryna's split: Mon–Tue Sutro, Wed–Thu Anthropic, Fri reserve. Overflow goes to one job-shaped ask, not more research. 10 Next four weeks What lands when When Milestone Mon 14 Sep 18:00 week 38 · tonight #31 — board fixed, the three decisions, room input for the sprint Tue 15 Sep 08:30–10:00 week 38 Organizational sprint proposed — the slot is free between “take second phone” and the 10:00 SymHub call Wed 16 – Thu 17 Sep Anthropic blocks. No Sutro work. Fri 18 Sep 11:00 Kateryna Peters #6, last paid session — report the Sutro/Anthropic split Mon 21 Sep week 39 #32 — Thomas Ahle's analog video; proposed v1.1 release with frozen constants and re-scored entries Tue 22 Sep Resend to Bill Dally, with the re-scoring as evidence Sun 27 Sep End of your tooling freeze Mon 28 Sep week 40 #33 — proposed first outside MNIST-medium entry: the success metric for the sprint Mon 12 Oct week 42 proposed planning check — MNIST-original tier open, and decide whether the ruler is trusted enough for a text task 11 Tomorrow, 08:30–10:00 Organizational sprint #1 — 90 minutes, hard stop at 10:00 Objective: an outsider can find, run, submit and get a review within one week, without messaging you. 5 People outside you committed code in two weeks 9 wk Log silence 9 Jun → 26 Aug; 3 skipped Mondays 5 Coordination channels calendar, doc, Telegram, Slack, Geek Club 11 PRs waiting on you the research moves, the group doesn't Min Question Output 0–10 Retrospective: what did sprint #1 produce, what stalled, why did #2 stop at five lines? 3 bullets 10–25 Roles: who owns what besides you? A roles table — ask each person, never assign 25–40 Review contract “Every outside PR gets a human comment within 48 h; one ruling per category.” Turn branch protection on. 40–55 Release and meeting cadence Release day; standing agenda; who takes
