Search Search
Back to recaps

Week of April 6, 2026

ByteDMD strict tracing lands and a NoProp experiment comes back negative. Plus the Ralph vs agent-teams thread.

60 messages and 15 links in the archive this week.

A working week split across the two main chats, 60 messages total. Most of the energy went to ByteDMD escape-hatch hardening and a NoProp experiment that came back negative.

What moved

On ByteDMD, Yaroslav pointed Sung Jae at the strictest implementation, bytedmd_settrace, used the same way as bytedmd but imported from the strict module (settrace source, usage). Seth reviewed Sung Jae's PR #4 and floated requiring Python 3.13+ to sidestep the 3.12 f_trace_opcodes bug that causes a known test failure. Sung Jae noticed the strict tracker had gone missing, then opened PR #5 to bring back the strict testing file, fix the escape, and add a benchmark run showing it. Yaroslav also shared a strict variant on dev.

Seth opened PR #78 on SutroYaro with a NoProp experiment, diffusion-style local learning with no inter-layer backprop. The result was negative: NoProp solves parity but does not beat SGD plus Curriculum on convergence or DMD.

Yaroslav is still on Sparse Parity and reported finding bugs in the metric, noting he needs a liveness analysis so it does not behave like a program with a memory leak, and that he had spoken with the author of the data movement distance paper. Silen Naihin's auto researcher came up, with Yaroslav asking how many GPUs it needs (autoresearch-distillation); Silen was running an 8b model and gathering baselines (sparse_parity).

On tooling, Yad recommended trying Anthropic's agent teams before Ralph for a Claude Code loop (agent teams docs, ralph-wiggum plugin). Yad also flagged a blocked setup: a lockout where Terminus lacks the authorization level to spawn processes over SSH, with access expected back Friday. Andy started a convention separating an agent prompt from an agent task and put it in a new PR (PR #73), wanting to learn about Muon.

Decision and open questions

Andy ran a Muon report whose conclusion was that Muon contradicts the lab's goals. Open threads from the week: Yaroslav asked whether the real NoProp comparison should have been forward-forward and whether a denoising objective helps at all; Josh Marks asked for a good way to get caught up.

Andy also linked a DeepSeek Engram v4 architecture post he was still working to grok, reading it as blending classical memory and retrieval with LLM retrieval. Gabriel shared an arXiv paper he found relevant, and Yaroslav was reading arXiv 2205.04934, noting comp-sci has recently started looking at the energy of computation.

For next Monday, Yaroslav plans to talk about defining a metric more energy-related than FLOPs, and Silen will talk about distilling. Yad floated a Saturday stream around noon EST building a Structured Workspace for coding agents to use as a research environment.

Sources