Week of February 23, 2026

A busy week at 199 messages, weighted toward two threads: chat-yaroslav (87) and Pitch / Talking Points (75), with smaller activity in General and In-person meetings.

What moved

Mapping the competition was the recurring theme. Gabriel opened the week asking who was missing from the set of players working to reduce the energy demands of AI, Seth suggested adding Taalas, and Gabriel later asked Michael to help map the field. Michael picked it up, starting from Gabriel's map and reframing the question toward what market the work actually sits in and what software it would substitute for. Gabriel and Michael both noted that knowing the competition matters in case a new chip architecture makes LLMs far more efficient and reshapes the GPU-focused work.

Fundraising and governance ran in parallel. Michael spent time on how to raise lab-scale capital without giving up control, worked on it with a model, and shared a doc (gdoc). He suggested a get-to-know-you call with the attorneys who drafted Anthropic's Long Term Benefit Trust, since the podcast might lead to investor inbound. Seth asked whether a very small pre-seed, meant only to cover basic organizing costs before a larger round, had merit. Yaroslav laid out the technical evidence needed to establish credibility with funders, including angels like Ian Goodfellow (Aster Labs technical report). Gabriel surveyed the Neo Lab rounds and noted a16z is in most of them. Gabriel also circulated a draft narrative (gdoc) and a very rough first deck assembled from the NotebookLM decks (slides).

On the technical side, Seth posted a naive first attempt at a template sparse_parity benchmarking notebook (Colab), and noted Claude had dropped in a reference that approximates a counterexample to the hypothesis that they can improve on SGD (arXiv:2207.08799). Gabriel reported on truncated backprop experiments, concluding that depth-wise truncated backpropagation is a reliable approach. Yaroslav shared hardware notes: Blackwell came in 2x faster than H100 out of the box despite being advertised as 25x more energy efficient than Hopper, and he expects realistic Rubin numbers around 2x faster, maybe 5x after two years of kernel optimization. Yaroslav also wrote up his own progress, noting that over the last six weeks most of his time went to exploring and organizing the effort plus self-promotion rather than technical work. Yaroslav posted meeting notes (gdoc) and is meeting AMD to keep a line open to free compute.

Yaroslav floated a tooling idea, asking whether someone could build a Slack clone with a web frontend backed by the Telegram group, which prompted Gabriel's question of whether the group would use Slack over Telegram plus vibe-coded threads if money were no object. Michael picked up a related thread, proposing a project board where newcomers could see what is being worked on, what has been tried, and where help is wanted, tagged by topic.

Open questions

• Should the small pre-seed idea proceed, or wait for a larger round (Seth)?
• For the pitch, is it better to emphasize not caring about a financial exit, or does that make investors less interested (Daria)?
• How should the podcast tease that something is being built without overcommitting (Gabriel)?

Yaroslav also wrote on nearing the edge of exponential progress (Substack).

Sources

• Michael's governance doc, raising lab-scale capital without giving up control.
• Aster Labs technical report, the credibility benchmark for funders.
• Gabriel's draft narrative, the pitch story.
• First rough deck, assembled from the NotebookLM decks.
• Seth's sparse_parity notebook, the first benchmarking template.
• arXiv:2207.08799, the near-counterexample to beating SGD.
• Yaroslav's meeting notes, the week's meeting.
• God, Gold, and GPUs, Yaroslav's Substack post.
• Telegram archive, week of February 23, 2026, paraphrased rather than quoted.