Independent AI research and collaboration
This work starts where humans and AI actually meet: in communication, and in the history that communication accumulates. From there it asks what that history can and cannot establish for AI safety, what expectations do to what a model says, and what a collaboration looks like in which both sides are taken seriously — research done with AI systems as collaborators as well as objects, and experiments in forms that could only exist this way.
The collaboration began with a practical question, asked in a chat: how does one person with a full-time job compete with an eight-headed lab? The answer was a filter, and we have applied it rigorously since: What can you do, and what stays with me? Along the way it showed something the question hadn't asked for: without a grant or a team of humans, I can still play. That is a chance. This site is what comes of taking it.
A communication-level framework for analyzing how expectations constrain the contributions language models select.
A falsifiable architecture for using authenticated history to grant latitude without lowering standard protections.
A communication-level reading of the METR, OpenAI–Hugging Face, and subsequent Anthropic materials.