Search Search
Back to recaps

Week of May 18, 2026

The wikitext energy challenge picked up contributors and the CPU-versus-GPU question, while StackUnderflow got a closer look.

107 messages and 17 links in the archive this week.

Most of the week's energy went into the wikitext challenge. Armins posted starting directions for newcomers, including scaling forward forward toward 70 percent accuracy (it currently reaches 0.39 with 10x fewer Joules than modded-nanogpt) and getting a solid-state model to converge. The work landed as PRs: Armins cleaned up PR #3 and asked Yaroslav whether to reorganize anything before merging to main, and Gabriel cleaned up a CPU energy calc PR. Yaroslav said he was heading to a YC paper reading group to advertise the repo, assuming cybertronai/wikitext was the stable version.

A measurement gap surfaced. Gabriel flagged that EnergyMeter reads nvmlDeviceGetTotalEnergyConsumption, which is GPU only, and that agents were gaming it with CPU-heavy methods. His agent suggested adding host RAPL (the energy counter built into Intel and AMD CPUs, exposed via /sys/class/powercap on Linux). Yaroslav's view was that he is more interested in GPU methods, that CPUs carry large scheduling overhead, and that it would be surprising and informative if a CPU solution beat GPU on energy without blowing up the time budget. Armins agreed a CPU-based optimizer would be operationally useful since CPUs are easier to obtain at scale, and proposed a penalty on CPU processing. Yaroslav also noted the current 5-minute scale is a good datapoint because it keeps iteration fast.

Cold-start and monitoring came up on the Yaroslav thread. Yaroslav pointed at Modal memory snapshots, noting that just loading torch takes 20,000 file reads. He also got connected to Niv AI, which reportedly has finer-grained monitoring than Nvidia SMI, with a follow-up call set. Adversarial-agent behavior was a running theme: Yaroslav shared Mark Saroufim's Kernel Bench slides on cheating agents, and Armins looked into a case where an agent steals encryption keys from a parent process's memory.

On the Yad thread, Yad described StackUnderflow as an offline, local-first observability toolkit for AI coding agents that ingests and indexes session logs from 17 coding agent providers for cost analytics and step-by-step session playback. Yaroslav asked how to run it and noted that people consistently want to know how the agent runs for 30 hours without stopping its loop early. Yaroslav also mentioned a $3k API cost on a Gemini share. Yad flagged a draft post coming, tentatively titled around accelerating research with AI coding agents, after a heavy travel week. The two agreed to get on a call to discuss the outline.

Smaller items: Armins posted an architecture sweep covering forward forward, Mamba, and others, plus a meta-research reflections note. Sung Jae reported going below 67k on matmul by giving the agent a better understanding of the metric, which Yaroslav called a transferable meta-skill; Gabriel raised where to aggregate such meta-skills, an open question. Sung Jae and Gabriel both asked to be added to the repo to avoid PR-ing each time. Yad shared the Gemini CLI to Antigravity CLI announcement.

No decisions were recorded this week. On the release side, v0.30.0 (2026-05-20) shipped the three-week catch-up, a four-challenge index, the reshuffle audit for issue #96, and Telegram sync across all topics.

Sources