Search Search
Back to recaps

Week of March 2, 2026

Challenge #1 sparse parity work and the non-profit funding pitch both moved this week, alongside the project's first agentic run.

125 messages and 24 links in the archive this week.

A busy week, 125 messages across pitch talk, the sparse parity challenge, and meeting logistics.

What moved

Challenge #1, sparse parity. Yaroslav posted the brief, due March 9, framing it as energy efficient learning of parity (doc). He noted that Yad, who is not in the channel and reached him on Twitter, already has an agentic loop framework that organizes and runs experiments automatically (findings). Yaroslav cited Karpathy's dependency-free reimplementation style as a model to aim for (gist) and pointed to Muon, the first optimizer to beat Adam in ten years, as having been found by iterating on a two-second CIFAR run. Gabriel summarized a round of attempts, including AdamW to SGD at LR=0.05 that failed the accuracy gate. Michael spent a long flight working the problem with Claude and reported he learned a lot but did not advance the state of the art on either training energy or agent-driven research (summary doc).

The pitch and funding model. Michael explored open source business models and corporate-funded foundations that exist to commodify the complement, with EleutherAI as a worked example (doc). He flagged that the efficiency framing is not off-limits to climate investors, pointing to Efficient Computer's raise (efficient.computer). Yaroslav added an AI search capability to the Sutro group bookmarks (planning doc) and discussed non-profit structures, citing Gather SF (gathersf.org) and the difficulty of converting an FRO to for-profit (imprint.org).

Decision. Gabriel noted that with roughly 80 percent of Mozilla funded by Google the donor base is too concentrated, and the group should set up a structure that cannot be pressured by a single donor.

Notable

Yaroslav ran his first two-hour technical sprint and annotated his reasoning in detail so an agent could later reproduce or improve on it (sprint doc). After catching up with Lucas, he relayed that energy is not currently the bottleneck, GPUs are, though Coreweave projects energy availability becoming the constraint in about two years (chat). Michael had Claude build the Pebbling Game Yaroslav described, including an API and a basic agent (artifact, repo). Andy started a Telegram clone on Antigravity, ran out of tokens, and moved it to Codeberg (repo). Yaroslav's Thursday reading group covered a paper showing a linear classifier can solve parity (arXiv), though he came away unmotivated to pursue it as an alternative.

Open questions

Seth raised how to get his team collaborating with Claude Code agents in a shared forum for design matters, rather than pasting agent output to Slack. Andy asked whether he could bring two visitors to the next meetup. Michael, after the flight session, asked whether non-experts with access to agents could contribute something meaningful to the problem.

Changelog

Releases landed this week on the SutroYaro repo: v0.1.0 through v0.4.0 (Mar 3, initial environment, MkDocs site, full sparse parity pipeline), v0.5.0 through v0.8.0 (Mar 4, solving 20-bit sparse parity, three rounds of research experiments), and v0.9.0 through v0.11.0 (Mar 7, Phase 2 with 17 parallel agents and the Practitioner's Field Guide, review fixes, homepage refresh).

Sources