Frontier AI research, run locally.
SimpleML
Click-and-play ML, at home.
Connect an agent, drop in your data or start from scratch, and it prepares the data, trains, evaluates and shows you the results. A desktop app for Windows, on your machine.
Measured
Every change has to beat the last one.
Inside the lab
What we're working on.
Research Loop
Tests one change at a time on a sealed set. Keeps it only if it beats the noise.
2.4182.255
bits per character, lower is better: AdamW to Muon
Executed Pivots
Rewards an agent for what its command did, not how it was typed.
96176
of 292 steps matched the real outcome: keystroke to executed
Verification Kit
The same keep-or-revert checks, packaged for any coding agent.
In development
Sandboxes
Isolated environments the other programs train and test inside.
18 / 18
escape attempts blocked