Search Search
Back to recaps

Week of March 9, 2026

Sutro launches challenge #1 sparse parity and agent-driven research goes public. The repo moves to cybertronai.

241 messages and 52 links in the archive this week.

This was the week the agent-driven research effort went public to the group and challenge #1, sparse parity, became the main thread. 241 messages, most of them between Yad and Yaroslav.

What moved

Yad introduced himself and his work, then shared his results on sparse parity. Operating Claude Code, he ran 33 experiments across five categories (algebraic, information-theoretic, local learning rules, hardware-aware, alternative framings). His approach was meta rather than implementation-specific: a repo that pulls context from Telegram, Google Docs, and arXiv, generates candidate approaches, then dispatches blind agents that have no access to prior solutions. The repo and its survey page were the main artifacts shared. See the repo at https://github.com/cybertronai/SutroYaro and the survey at https://cybertronai.github.io/SutroYaro/research/survey/.

Yaroslav dug into the GF(2) Gaussian elimination result and wanted to reproduce it. After in-depth evals he found the GF2 solution needs roughly 2000x fewer flops, 1000x fewer memory accesses, and runs about 1000x faster, writeup at https://github.com/cybertronai/sutro?tab=readme-ov-file#sgd-vs-gf2-gaussian-elimination-2026-03-09. He flagged that ARD is likely too coarse a metric and that what the group actually cares about is energy usage under realistic cache dimensions. He pointed to the AI radiology benchmark debacle as a caution against agents over-optimizing a proxy, https://spectrum.ieee.org/andrew-ng-xrays-the-ai-hype.

Andy ran an early tiling experiment and raised whether it was invalid because it used ARD, https://0bserver07.github.io/SutroYaro/findings/exp_tiled_w1/. His agent also surprised him by merging PR #3 on its own when asked to update the repo.

Over the week the repo moved from the personal 0bserver07 org to cybertronai for more respectability, at Yaroslav's invitation. The agents themselves performed the move. The harness grew from one challenge to three (sparse parity, sparse sum, sparse AND), each running under a second, with a step-by-step adding-a-challenge guide. Sparse AND was added by Google Antigravity following that guide without human help. Yad also added GPU measurement via Modal Labs and previewed an update letting agents drive Modal GPU sandboxes under a budget ahead of the NanoGPT challenge, https://modal.com/docs/guide/sandboxes.

By Sunday the agent reported Sign SGD getting close to the 10ms target on sparse parity: best run 7.6ms (hidden=50, n_train=500, lr=0.1), but only 3 of 5 seeds solve at that config, while n_train=1000 solves all 5 at mean 29ms. Standard SGD floors around 70 to 116ms. Findings at https://cybertronai.github.io/SutroYaro/findings/exp_sgd_speed/. The agent cited Egalitarian Gradient Descent (https://arxiv.org/abs/2510.04930) as relevant, which Yad flagged honestly as possibly close to cheating.

Decisions and open questions

Yaroslav asked Yad to either dial in at 6pm SF time or record a five-minute video covering which parts of the work are human and which are agents. Yad opted to cover it in a video and shared his take on what keeps agent output valid rather than hallucinated: a shared memory file (DISCOVERIES.md) that every agent reads before starting, and protected metric code that agents cannot modify.

Yad and Yaroslav converged on a direction: the autoresearch approach of hundreds of agents running slow experiments is the wrong axis to scale on. If each experiment takes three minutes, a thousand agents still only buy a thousand experiments in three minutes. Faster experiments and tighter harnesses matter more than agent count. Andy raised a related point that the GPU idles at 12.5 watts, so the toy experiment does not stress the GPU enough to produce a meaningful power-consumption result, something to fix in a later GPU run. The ARD versus DMC proxy question remained open, pending nanoGPT-scale workloads. The human-versus-agent boundary (for example, who edits TODO.md) was also raised and left as an ongoing question.

Sources