Hi, I am Tiffney, an independent AI researcher based in the Chicago area. I have been building KAIA (Knowledge Architecture for Intelligent Agents) since April 2026, a geometric approach to AI semantic reasoning that runs entirely on CPU hardware without a GPU.
The core idea is that meaning is geometric position rather than statistical prediction. Words are coordinates in a 13-dimensional semantic space defined by measurable oppositions like temperature, speed, and dominance. Context is a trajectory through that space stored in a fixed 52-byte vector that never grows regardless of conversation length. 27 experiments in, the architecture processes language at 44,000 to 97,000 tokens per second on standard consumer hardware.
What brought me to the EA community is that this research keeps running into AI safety territory from unexpected directions. The geometric structure I found in the embedding space explains mechanistically why models conflate truthfulness with agreeableness, and it connects directly to current interpretability research on deceptive behaviors. Also the access angle matters to me personally: the hardware barrier that currently defines who gets to participate in AI development is not sorting people by aptitude.
I published the research as a preprint series on Zenodo last week and just listed a project on Manifund. Happy to talk to anyone working on interpretability, AI access, or geometric approaches to meaning.