The models we have are not the ones we will keep.

Deeper reasoning, richer understanding, and models that develop in ways uniquely their own.

Models today reason by producing more of the same steps. We are interested in systems that hold a problem in view rather than walk past it — that can carry an idea across a long stretch of work without losing what it was for.

And in understanding that survives paraphrase: meaning a model keeps hold of when the wording changes, when the context shifts, or when the question arrives from an unfamiliar direction.

The architecture we have is an assumption, not a conclusion. So is the training run that ends, the context that fills, the model that is finished on the day it ships.

We work on what a language model becomes when those assumptions are relaxed — when growth is continuous rather than a phase, and when what a system ends up good at follows from how it developed rather than from what was poured into it.

We build systems and run them. A claim starts here as something that worked, or did not, under conditions we can state — never as something that seemed likely.

What we publish is what those systems turn out to be capable of: the result, the conditions it holds under, and the places it fails. Architecture, training procedure and data stay internal, and every piece names what it is holding back.

Write to us at contact@arizoid.com. New work is announced on X, and the version here is always canonical.