Search Search
Back to recaps

Week of March 16, 2026

EGD lands and a new DMC leader hits 3,578. The SutroYaro repo goes public domain.

89 messages and 26 links in the archive this week.

A heavy build week, 89 messages across the topics. Most of the activity sat in the two one-on-one threads (chat-yad and chat-yaroslav) plus General.

What moved

Yad shipped the Egalitarian Gradient Descent experiment (arXiv:2510.04930) on sparse parity and sparse sum, with results and write-up live. EGD replaces gradient singular values with 1 via SVD so all directions update at equal speed. The reported surprises: EGD halved the epoch count to 90% accuracy (14 vs 33) but ran 12% slower in wall time because the per-batch SVD overhead outweighed the epoch savings, sub-10ms was not achievable, and EGD solved sparse sum where standard SGD diverged. This landed as v0.21.0 (2026-03-16). Write-up at exp_egd, commit dd82669.

Later in the week Yad posted this week's DMC findings: a SUCCESS, with the best variant hitting DMC 3,578, a 58% reduction from the GF2 baseline. The new leader was flagged in the challenge #1 topic as KM with 1 influence sample, because parity influence is binary (0 or 1, never fractional), so a single sample per bit is enough for perfect identification. This shipped as v0.22.0 (2026-03-22), the DMC baseline sweep, optimization, and infrastructure work. Post at exp_dmc_optimize, with a session log and the 2026-03-22 catch-up. Yad attempted a one-hour livestream and posted a shorter video update covering the public-domain repo, agent loops, parallel agents, and auto research.

Yad also compared how two agents handled the EGD task: Antigravity rewrote the entire SGD setup, while Claude Code recognized the existing SGD setup as the experiment and treated EGD as a parameter change, so Antigravity wrote more code. Separately, Yad described integrating and replicating HumanEval and SWE-Bench in a personal eval harness called chimera.

Decision

Yaroslav asked Yad whether he was OK characterizing cybertronai/SutroYaro as Public Domain, on the grounds that it makes it easier to involve more people. Yad recorded his own decision in chat-yaroslav around tooling: after trying Screenpipe, macos-automator-mcp, and openrecall, he concluded he would rather run the whole OS one layer above the LLM, and pointed at agent-infra/sandbox.

Threads and open questions

Yaroslav set the next direction toward continuing on sparse parity using a Data Movement Complexity penalty (arXiv:2312.14441), with an agenda doc and an updated homework section. Yaroslav also shared an AI hardware design overview forwarded by Joshua Marks, where Bjarke Hammersholt Roune argues a CPU with an attached large systolic array is sufficient and GPUs are not needed (doc), and turned it into a NotebookLM podcast. Andy ran a quick AI-mode Q&A and landed on the same utilization issue the podcast raised.

A discrete-vs-continuous thread ran in General after Yaroslav talked with Stephen Wolfram and raised the question of why we do machine learning in a continuous way, linking Wolfram's minimal models writeup. Seth noted that extreme quantization of the arithmetic is already a discretization and asked whether Wolfram had a more specific suggestion, and Andy wondered if it was inspired by cellular automata. G K asked an open question about the agent-GPU setup and token budgeting, which prompted Yad's rundown of GPU and LLM options (Modal, Vast, Shadeform, TensorDock). Yaroslav also shared Prime Intellect's environments post.

For homework, Michael Keating forked the macOS fork of Karpathy's Autoresearch to work on sparse parity instead of nanoGPT and had Opus attempt it with unconventional or ancient mathematical theories (repo). Andy pointed back to Sara Hooker's hardware lottery writeup as a possible source of agent experiment ideas.

Sources