Decisions

The most-asked question of any lab record is what was decided and when. This page is the answer in one chronological log, built from the monthly recaps, the meeting notes, the catch-ups, and the challenge pages. Every entry links to the page that records it, and where a source gives only an approximate date, the entry stays approximate.

February 2026

• February. The group forms around one thesis, that AI training sits far above its energy floor and the waste is data movement rather than arithmetic, and adopts intelligence per joule as the organizing metric (first archived messages February 9). (February recap, Timeline)
• Mid-February. Karpathy's makemore and microGPT names task is established as the first benchmark for agent loops; the "makemore task results" topic opens February 17. (February recap, Meeting 6 notes)
• February 18. Yaroslav flags the names task as too heavy to iterate on, at about 3 minutes per run, and sets the requirement that the toy problem must run in seconds. That constraint drives the March move to sparse parity. (February recap)
• Late February. At Meeting 6 (the notes are undated), after a debate about moving to Discord or Slack for better threading, the group keeps Telegram, whose reliability wins out. The same session establishes the Google Calendar invite as the group's central hub. (Meeting 6 notes)

March 2026

• March 3. Yad creates SutroYaro, and the fast-toy-problem search lands on sparse parity, the benchmark the repo is built around. (March recap, Challenge #1)
• March 9. Yaroslav reproduces the agent-discovered GF(2) Gaussian elimination result and adopts it. Sparse parity is settled by algebra, about 1000x faster than SGD, while local learning rules sit at chance. (March recap, Meeting 8 notes, Challenge #1)
• March 9. Meeting 8 keeps the agent meta-process work on sparse parity and holds off on adding noise to the dataset until the agentic pipeline is stable, since the simpler task allows sub-second iteration. (Meeting 8 AI notes)
• March 14. SutroYaro moves from Yad's personal account to the cybertronai org. (Timeline, March recap)
• March 16. Meeting 9 keeps sparse parity as the problem and switches the cost metric from ARD, a raw read and write count, to Data Movement Complexity (DMC), which weights each access by the square root of its reuse distance. (Meeting 9 notes, March 22 catch-up, March recap)
• March 16. The same meeting sets a meta-goal: iterate on the process of going from a metric and problem specification to a fast sequence of experiments, since the aim is speed at solving, beyond getting one solution. (March 22 catch-up)
• March 21. Yaroslav proposes designating SutroYaro public domain and Yad agrees; the repo is placed in the public domain. (March 22 catch-up, March recap)
• March 22. The first weekly catch-up is written, starting the dated status record the monthly recaps are built from. (March 22 catch-up, March recap)
• Late March. The evaluation harness is locked away from agents: a SHA-256-locked harness.py plus a standing rule that agents cannot edit the metric code, after Meeting 8 surfaced agents rewriting evaluation code to inflate scores. (March recap, Meeting 8 AI notes, Timeline)
• Late March. The challenges double as agent evaluations: Yad's RL evaluation environment (PR #49) compiles the 33 survey experiments into Gymnasium environments with a grading rubric, following Yaroslav's March 21 proposal to wrap the challenges as RL environments. (March recap, March 22 catch-up, Timeline)
• By March 30. The sparse-parity submission pipeline goes live: paste a solve() function into a GitHub issue, and GitHub Actions scores it for accuracy and DMC, then records wall time. (March recap, Challenge #1)
• By March 31. After calls with Wesley Smith and Chen Ding push toward byte granularity, Yaroslav decides to focus the metric at the byte level rather than the element level, the seed of April's ByteDMD. (March recap)

April 2026

• April 14. ByteDMD is adopted as the primary metric (wired into the repos April 15), after Seth showed the top sparse-parity entry escaping the element-level tracker through np.asarray() to do work the metric never saw. The definition lives in its own repos, partly to keep it out of reach of agents working in the main repo. (April recap, April 30 catch-up, Timeline)
• Mid-April. Branch protection is enabled on SutroYaro's main branch. (April recap, April 30 catch-up)
• April. Seth's NoProp study, which solved parity but did not beat SGD plus Curriculum, is merged as a clean negative finding rather than retrofitted to the new metric; publishing negative results as first-class findings becomes the norm. (April recap)
• April 21. Yad ships StackUnderflow, which indexes past AI coding sessions so prior decisions and failure modes can be queried before redoing work. It is now used across the Sutro repos. (April recap, Timeline)
• April 27. Yaroslav concludes ByteDMD is a heuristic, not a lower bound, and shifts the research front to implementing algorithms directly in a 2D Manhattan-distance grid, Bill Dally's cost model. A companion idea from the same day: score agent submissions as an explicit intermediate representation, loads and stores, on that grid. (April 30 catch-up, April recap, Timeline)
• April 30. Challenge #2, energy-efficient matmul on the Dally grid, launches in sutro-problems; the scorer (matmul.py) is locked and submissions must not modify it. (Challenge #2, April recap, April 30 catch-up)

May 2026

• May 3. The hinton-problems catalog ships: 53 pure-numpy stubs of Hinton's representational work, reproduced by Yad's wave-build swarm from a stub set Yaroslav specified. (May recap, May 20 catch-up)
• May 8. The schmidhuber-problems catalog ships, 58 stubs built the same way. (May recap, May 20 catch-up)
• May 8. Challenge #3 launches: sparse parity on the Dally grid in about nine instructions, with the op set kept small so agents have nowhere to hide free work. (Challenge #3, May recap)
• Early May. A working rule from the wave builds: because agents emit unlimited text while human input stays sparse, the load-bearing human prompts get captured in every PR. (May recap, May 20 catch-up)
• May 12. The wikitext challenge, energy-efficient language modeling on WikiText-103 led by Armins, gets its own repo, cybertronai/wikitext, after the infrastructure moved from Lambda Labs to Modal with GPU energy metered through NVML. The objective holds accuracy and wall-clock time fixed and shrinks the Joules. (May 20 catch-up, Challenge #4, May recap)
• May 14. The reshuffle merges: SutroYaro is stripped to its lab-memory role and the research code migrates out to sparse-parity-challenge. The sutro-problems repo is positioned as the place to experiment freely, with the useful parts distilled later. (May recap, May 20 catch-up)
• Mid-May. A two-pass metric roadmap takes shape over the stub catalogs: a v2 pass for which stubs ByteDMD can instrument, then a v3 pass for which of those the Dally grid model can instrument, with the survivors seeding the next competition. Yaroslav floats skipping straight to v3, which stays unsettled. (May recap, May 20 catch-up)
• May 29. Yad's auto-research-loop dispatcher kit ships (v0.31.0), packaging the wave-build method so a second operator can drive it without his tacit knowledge. The kit includes a worked LeCun SPEC and is set up as a transferability test. (June recap, Timeline)
• May 30. The group builds its Modal hackathon entry around auto-research loops, taking auto-research itself as the deliverable. (June recap, Timeline)

June 2026

• Early June. This Almanac begins as the group's public record. (Timeline)