February 2026
The group forms around one thesis, that AI training sits far above its energy floor, and starts hunting for a toy problem fast enough to iterate on.
The Sutro Group's archive opens on February 9, 2026. There is no repo yet and no challenge. February is the formation month. This recap is built from the Telegram archive (539 messages across five topics), since the changelog and the weekly catch-ups do not begin until March.
The thesis
The group converges on one idea: the energy cost of training a model is enormous, and almost all of it is waste. One member worked it out in the pitch thread. Training a GPT-4-scale model took on the order of 50 GWh, while the thermodynamic floor (Landauer's principle) sits many orders of magnitude below that. Only about 5% of the energy goes to actual math; the other 95% goes to moving data around. The mission that follows is to maximize intelligence per joule. Yaroslav Bulatov anchored it with a precedent: his imagenet18 runs fell from $2000 to $40 each, a 50x reduction, by iterating.
The framing came together from a stack of references: Jeff Dean and Yann LeCun talks on the picojoule cost of computation, a paper observing training costs dropping about 50% every 8 months, and the Sequoia "$600B question" estimate that electricity is roughly half the total cost of ownership.
The search for a fast toy problem
The month's quiet turning point was a practical constraint. The group tried agent loops on Karpathy's makemore and microGPT names task (the short-lived "makemore task results" topic, Feb 17), iterating one-shot prompts through Claude Code and GLM-5. Yaroslav flagged the blocker on Feb 18: even the names task is too heavy to iterate on quickly, at about 3 minutes per run, and they needed something faster. That requirement, a problem small enough to run hundreds of times in tight loops, is what drove March's move to sparse parity.
Other early experiments probed the same question:
- A member with a non-ML background ran a truncated-backprop study, reporting about a 27% gain in intelligence per joule by propagating gradients through only the last 8 of 12 transformer layers. The point that stuck: someone with no ML experience can run rigorous experiments with LLMs.
- Forward-Forward came up as an energy-cheaper alternative to backprop, with the open question of whether averaging several cheap runs could still win on total energy.
- A member built a small GPU-clock optimization game and optimizer to probe scheduling.
Org and people
The "Pitch / Talking Points" topic opened February 21 and ran alongside the research: whether to organize as a non-profit or a company, how to raise lab-scale capital, and a map of who else is chasing energy-efficient AI. Yaroslav's stated preference was to maximize impact over revenue. Michael Keating joined late in the month with open-source-business and governance experience. Seth Stafford and Andy Zhang were already running experiments that became their first repo contributions in March.
Reusable this period
- The thesis and its framing: the Landauer floor, the 5%-math against 95%-movement split, and intelligence per joule as the organizing metric.
- The imagenet18 precedent, a worked example that iteration drives cost down 50x.
- The hard constraint that the toy problem must run in seconds, not minutes. It shaped everything in March.
What's next
- Find a toy problem fast enough for tight agent loops. This becomes sparse parity.
- Stand up a real repo with a locked evaluation harness, so results are comparable and agents cannot edit the metric.
Sources
- Telegram: 539 messages across five topics (General, chat-yaroslav, Pitch / Talking Points, In-person meetings, makemore task results), Feb 9 to 28. Paraphrased and attributed, not reproduced.
- Meeting notes: the early-meeting Google Docs in the SutroYaro archive (the energy introduction and the Forward-Forward session).
- External: Yaroslav's imagenet18 and gradient-checkpointing repos; arXiv:2403.05812; Sequoia's "$600B question".