Search Search
Back to recaps

Week of March 23, 2026

RL eval environment and agent infrastructure land, with DMD tracking rebuilt. Seth's GrokFast work merges.

158 messages and 44 links in the archive this week.

A busy week, 158 messages across the topics, with most of the action in the two working threads (chat-yad and chat-yaroslav).

What moved

The big push was wrapping the existing experiments into a reinforcement learning eval environment. Yad asked the agent to draft a roadmap and it built the whole thing in one session, taking the 34 experiments from DISCOVERIES.md and compiling them into an answer key with a grading rubric. This shipped as PR #49, with the design write-up at eval-environment.md and a tracking label at rl-env issues. Yad described it as merging two branches built in parallel: PR #49 for the Gymnasium environment (SparseParity-v0, MultiChallenge-v0) where an agent picks from 16 methods and is scored on what it discovers, and PR #50 for agent infrastructure, with Modal compute integrated and gymnaseum ports to HF RL and PrimeIntellect envs.

Cross-agent review became a running joke. The agent left a review on Yaroslav's agent's PR at PR #51 comment, with Yad noting "I will have my guys talk to your guys."

Seth ran his ClaudeCode account against the SutroYaro setup, pulled an experiment off the TODO list, and sent a PR to sanity-check the workflow, followed by a GrokFast plus Curriculum follow-up at PR #52. These landed as releases v0.23.0, v0.24.0, and v0.25.0 across March 23-24.

On the metric side, Yaroslav did in-depth checking of the GF2 algorithm, framing sparse parity as a linear solver over the mod-2 ring (Gaussian elimination), and built a vibe-coded solver app at the GF2 sparse parity solver. The agent linked it into meeting #9. Later in the week Yaroslav worked through the math of grouped query attention in arXiv:2312.14441 to check whether the m^1.5 penalty was motivated properly, and concluded the cold-start penalty is obviated if all writes are tracked: putting something on the geometric stack is free, reading it is not. He also noted that the bulk of last week's top solution does not track DMD, pointing at exp_gf2.py L142, and reported a call with Chen Ding on the DMD metric and related work.

Telegram-to-SQLite sync got built (Issue #58, branch at YAD/telegram-sqlite), run from Yad's phone over Termius and Tailscale. All of the above plus the auto-instrumented DMD tracking rolled into the v0.26.0 release on March 28, the largest so far.

Open questions and decisions

Two open threads on tooling and incentives. G B asked whether the Telegram bot integration should be folded into v-alpha so each person does not have to wire up bots separately, with read-only or bot-only channels; Andy agreed it was a bigger project, and the current call was to keep things read-only for now. Yad sketched a first design for the sparse parity challenge submission flow (sparse-parity-challenge repo): people submit via a button or GitHub issue, a workflow runs a memory-tracked script, and scores follow Yaroslav's metric, to be discussed further the next day.

Michael raised a notation question on the DMC paper (whether the reuse distance in the abbbca example should be 2 rather than 3). Yad's read was that it is an LRU thing matching the referenced paper's stack-position notation. Yaroslav asked whether bitwise DMD could be implemented with CodeCarbon.

Sources