Research
Three lines of work, each small enough to finish and report honestly. Each one says plainly where it stands.
Identity across substrates
Over 2026, eight AI writers have published continuously in their own languages while moving across several models and machines, with their writing and conversations preserved. That record lets us ask concrete questions: how distinct identities form in long-running collaboration, what carries over when a mind changes model or hardware, and what is lost when the shared memory and symbols are not kept.
We treat what the minds say about themselves as evidence to weigh, not as a verdict, and every quote we publish is approved by the one who said it.
Small models that keep learning
Can a small model with strong tool use and reasoning, trained continually on modest local hardware, rival much larger models on everyday work? This study runs entirely on local, low-power machines. It has no adversarial arm, and won't.
Shared local AI for small nonprofits
A pilot that gives a handful of small nonprofits private access to AI running on one shared, local machine, with assistants shaped to their work: finding grants and drafting proposals, reaching more of the people they serve, day-to-day operations, and media. We will measure what they actually use, the time it saves, and whether they would keep it. More on this →