Search Search
Back to recaps

Week of April 20, 2026

A 7-issue parallel-agent sprint merged, and the ByteDMD floor-gap survey gave numbers to Yaroslav's question about how far current methods sit from the floor.

75 messages and 11 links in the archive this week.

The week's two threads were a review-and-merge push on SutroYaro and a fresh measurement question on ByteDMD.

What moved

Yad opened PR #82, a parallel-agent sprint that closes seven issues at once, and asked Yaroslav and Andy for a review note so it could merge. He had locked himself out of his own merge by requiring a reviewer on main. Andy reviewed it issue by issue and approved #82 with comments for follow-up issues. Separately, Seth updated PR #79 to take option 1, noting the metric is stale, and suggested moving on to a new experiment rather than retrofitting the old one. Yad agreed: the NoProp versus Forward-Forward result stands on its own as a negative finding, so it merges with a metric-is-stale note instead of being reworked onto ByteDMD.

The next experiment came from Yaroslav's floor-gap question, how far current solutions sit from the ByteDMD floor. By Thursday Seth had shipped KM-min (268) and GF(2) (101,501) under ByteDMD overnight in PR #87. Yad flagged two open questions to resolve before Monday, including what exactly the 0.33 floor is. Yaroslav answered with the geometric lower bound: 0.3849 times measured ByteDMD lower-bounds the actual allocation cost of the best oracle allocator, with the live-byte counting requirement and a Gemini DeepThink proof at tarjan-detailed-part1.pdf. Live-byte counting (the current ByteDMD) counts reuse on live bytes only, where the old all-bytes version had unbounded cost on some examples.

Per the changelog, both releases this week landed: v0.28.0 (Apr 20) bundled the #82 sprint, repo diagrams with pan/zoom (PR #83), Andy's agent-compatibility work (PR #72), Seth's NoProp plus Forward-Forward follow-up (PR #79), and ASI-Evolve reports (PR #81). v0.29.0 (Apr 25) shipped the floor-gap survey (PR #87 Seth, PR #88 Yad), the Nix devShell and Task 11 docs (PR #84, #85, Andy), nine more issues closed, the new follow-up label, and auto-generated repo diagrams (PR #91).

Other notes

Yad did an initial release of StackUnderflow (site), a one-line install, while recovering from Verizon throttling, the flu, and Anthropic de-clauding. He plans to pre-record a few things for Monday: the token-consumption toolkit, SutroYaro updates after Seth's PR merges, and an ICLR paper list. Gabriel shared the Anthropic write-up on automating AI safety R&D in chat-yaroslav. On reading throughput, Yaroslav settled on Gemini DeepThink at its 10-per-day limit as roughly his absorption ceiling; Yad uses Mac read-aloud at 1.5x and skims markdown.

Decision and open question

Yaroslav resolved his ICLR uncertainty and decided to go, while keeping the Monday meeting; the in-person thread reflects people checking whether Monday is skipped for ICLR. Open on PR #87: Q2, whether oracle-query KM-min counts as a fair reference floor even though it is not benchmark-submittable, still awaits Yaroslav.

Sources