Week of June 8, 2026
Monday's session takes up Tilde's two-matrix update, learning-rotations ships, and the week's thesis becomes multiplicative updates.
38 messages and 9 links in the archive this week.
A shorter week in the archive so far, but it carries the clearest new direction in a while: treat learning as equation solving rather than minimization, and prefer multiplicative updates because of the memory wall.
Monday's session
The June 8 meeting moved 30 minutes later for a venue conflict at South Park Commons and ran small, a few people in the room plus a Meet link. Yaroslav's agenda item was Tilde Research's two-matrix update, which he called a promising direction. Andy brought a techxplore article claiming order-of-magnitude energy savings from asynchronous AI and asked who knows the area. No meeting notes have been committed to the docs yet; this entry is from the archive alone.
Learning-rotations, and the multiplicative thesis
On Wednesday Yaroslav published learning-rotations, with a rendered site, as an example of the wins available outside gradient descent orthodoxy: its best method is one he describes as too weird to derive from gradient descent. Asked whether the insight is exploiting known problem structure, he agreed, and sharpened it: look at learning as an equation-solving problem rather than a minimization problem, which opens a different literature and toolset.
The observation he flagged as the carry-forward: the best update in that series is multiplicative rather than additive, and matrix multiplication suits GPUs better than addition because of the memory wall. A multiplicative update for the transformer, if one exists, would be promising. That lands directly on the group's joules-per-intelligence thesis.
The Almanac ships
This site went live during the week: cybertronai.github.io/SutroAlmanac, with the repo public. Yad built it as the group's index: monthly and weekly recaps, the challenges, meeting notes, and the visual tour, every claim traced to a source.
One more seed from the week: out of an exchange with Cosmin about re-presenting classic papers, the idea surfaced of adapting the LeCun spec draft into a nesterov-problems spec, which would be a third wave-build catalog if the group takes it up.
No decisions were recorded this week. The carry-forwards are the Tilde two-matrix direction from Monday and the multiplicative-update question.
Sources
- Tilde Research two-matrix update, Monday's agenda item.
- Asynchronous AI energy article, Andy's question to the room.
- learning-rotations and its rendered site, the equation-solving example.
- Sutro Almanac and the repo, this site.
- Telegram archive, week of June 8, 2026, paraphrased rather than quoted.