Genesis

A different foundation for intelligence.

Genesis explores how intelligence can be built from separate cooperating parts, and how such a model learns to coordinate on its own.

Kind
Research programme

A different starting point

Almost every model today is built on the same foundation: vast amounts of approximate arithmetic. Genesis began by asking whether that has to be true.

We built Genesis to investigate whether an intelligent system could be organised around separate parts that each do their own work, and whether, given time, those parts would learn to work together.

Watching coordination appear

We trained Genesis in long runs alongside a comparison that differed in a single rule. Where Genesis could adapt, coordination between its parts grew steadily denser as training went on. Where it could not, very little changed.

Probed with different inputs, the same pattern of coordination appeared each time. Something consistent was forming inside the model, and it was forming by itself.

Following the whole path

At a larger scale, Genesis showed us something we had not expected: coordination could collapse and then recover. A model that looks right at the end of training may have travelled a very different road to get there.

That changed how we look at every model since. We judge a model by its whole development, not only by where it finishes.

Built to be repeated

Genesis also became a discipline. Large training runs had to be safe, deterministic and repeatable, so that any result could be run again and arrive at exactly the same place.

The research continues into how a model built this way can hold its coordination steady as it grows.

Cite this record

Forlais Research (2026). Genesis. Forlais Research Library, record FR-P-02.

@misc{fr-p-02,
  author = {{Forlais Research}},
  title = {{Genesis}},
  howpublished = {Forlais Research Library, record FR-P-02},
  year = {2026}
}