Back to Sutro Almanac

Records

The current best result on each challenge, dated and traced to the page that states it.

Each challenge scores one cost metric, and the record is the best result the group has measured on it. Every number below is taken from a challenge page or a monthly recap and links to the page that states it.

ChallengeMetricCurrent bestHolderAs ofSource
#1 Sparse parityByteDMD, byte-granularity data movementDMC 3,578 (KM-min, legacy element-level metric)Not namedMarch 22, 2026March recap
#2 Energy-efficient matmulDally 2D-grid cost model67,821 (16x16)Not namedMay 14, 2026May recap
#3 Sparse parity on the gridDally 2D-grid cost modelNo record stated yetNone yetJune 2026June recap
#4 wikitextGPU energy in Joules (NVML) at fixed accuracy and wall-clock timeNo record stated yet (baseline 54,784 J)None yetJune 2026June recap

Challenge #1: Sparse parity

The best stated mark is DMC 3,578, set by KM-min, a single-sample Kushilevitz-Mansour variant, in the first DMC baseline sweep on March 22, 2026, 58% below the GF(2) baseline (March recap). The metric has moved since, so the mark carries a qualifier.

  • March 16, 2026: Meeting #9 replaces the raw read-and-write count (ARD) with Data Movement Complexity (DMC), which weights each access by the square root of its reuse distance.
  • March 22, 2026: the first DMC sweep reorders the leaderboard. GF(2) wins on DMC even though KM wins on ARD, and KM-min takes the new best at DMC 3,578.
  • April 2026: Seth Stafford finds the top entry doing bit-packed GF(2) work the element-level tracker never saw; counted honestly, its cost moves from about 45K to 81K. ByteDMD becomes the primary metric on April 14 (April recap).
  • April 23 to 25, 2026: the floor-gap survey measures KM-min at ByteDMD 268, which is 3.8x the read-floor of about 70, and GF(2) at ByteDMD 101,501.

As of June 2026, KM-min and GF(2) elimination lead, both still scored under the legacy element-level metric and not yet re-measured under ByteDMD (challenge page). Whether oracle-query KM-min counts as a fair reference floor is an open question, since it is not benchmark-submittable. Submissions run through sparse-parity-challenge, with entries as recent as May 12.

Challenge #2: Energy-efficient matmul

The 16x16 record on the Dally grid stands at 67,821 (challenge page). The May recap dates that mark to May 14, 2026 and does not name its holder; the challenge page calls it mid-May. The progression:

  • May 5, 2026: 68,452, Sung Jae Bae.
  • May 13, 2026: 68,392, Cosmin Negruseri.
  • May 14, 2026: 67,821, holder not named in the sources. Cosmin Negruseri was running a Codex-driven auto-research loop that produced a steady stream of small PRs in this window.

Lower bounds are an open hard problem: an agent-generated bound was wrong on inspection, and AlphaTensor could not bound 4x4 either. Submissions go to sutro-problems, the matmul/ directory, where the scorer is locked.

Challenge #3: Sparse parity on the grid

No record stated yet. The challenge launched May 8, 2026 and scores the Dally 2D-grid cost model in sutro-problems, the sparse-parity/ directory (challenge page). What the May recap and the challenge page report so far: precomputing intermediate XORs plus bit-packing works well, while tiling does not help at this problem size. A 50%-accuracy target variant has been added.

Challenge #4: wikitext

No record stated yet. The reference point is the baseline: modded_nanogpt runs at 54,784 J and 0.7285 character-accuracy in 322.7 s, with GPU energy measured through NVML on Modal (challenge page; the numbers are reported in the May recap). A forward-forward submission reaches about 0.39 accuracy at roughly 10x fewer Joules, below the baseline's accuracy. The challenge got its own repo, cybertronai/wikitext, on May 12, 2026, with Armins leading. The open scoring questions are whether CPU package energy (RAPL) should also count and whether cross-entropy becomes a separate scored track (June recap).