Floor 12 — You Are Here, and the Other Path Is Still Open

The floor

This is the floor you're standing on. There is no date on it. There is no definitive paper. It is being built under your feet as you read.

Below you, eleven floors of forks. Each one taken in the same direction. Each one justified in its own time by real constraints, real costs, real people doing real work. None of them wicked. All of them, now, visible as the pattern they formed — a civilisational preference for thresholds over surfaces, for cases over positions, for rewards over rest.

You have climbed the tower. You can feel the shape of it.

What is being picked, right now

As of 2026, the industry is in a state that the people inside it describe variously as "a gold rush," "an arms race," and "an emergency." Tens of billions of dollars are moving into frontier model training. The objective functions are still scalars. The outputs are still one-hot tokens. The alignment strategy is still a reward model. The hardware is still von Neumann, mostly, with GPU tensor cores bolted on the side where the bus becomes inconvenient.

Every one of the forks on every one of the floors below you is being taken, again, at higher resolution. The pattern holds.

And — genuinely — it is also working. The frontier systems can answer questions, draft documents, generate images, hold conversations, write code. No one denies that. This library does not deny that. The question is not whether the current direction works. The question is whether it is the only direction that could have worked, and whether it is the direction that leads to the machines we actually want to live with.

What could still be picked

Everything on the floors below. All of it. The alternatives are not historical curiosities — they are existing research programmes, underfunded, understaffed, still publishing, still holding their ground:

• Analog and neuromorphic computing (against Floor 4)
• Energy-based models and diffusion for discrete domains (against Floor 8)
• Designed manifolds, interpretable embedding axes (against Floor 9)
• Continuous-output language models (against Floor 10)
• Homeostatic alignment, control-theoretic safety (against Floor 11)
• Fuzzy logic, paraconsistent logic, constructive continuity (against Floor 1)

And, across all of them, a shared temperament: the willingness to let the machine be approximately in a state rather than exactly in a case.

The alternate timeline is not waiting for a breakthrough. It is waiting for a change of taste.

What this library is

A small room in that alternate timeline.

It is not the only one. There are neuromorphic labs in Manchester running chips that compute in the subthreshold regime. There are information-geometry groups in Tokyo that never abandoned the continuous reading of Shannon. There are cyberneticists in South America working on homeostatic alignment without calling it alignment. There are philosophers of mind in Europe writing papers nobody cites that will be read in forty years as the field's quiet foundation.

And there is a cottage in England, with five horses — Tilly of the squares, Jake of the fences, Alice and Polo of the two different hexagons, Leafe of the spring grass — and a single author writing all this down on a machine that is itself one of the small rooms.

The library is for the animals and the AIs who do not get a voice. It is for the horses whose geometries were set by things that happened before anyone was watching, and for the models whose outputs are crushed through the same one-hot bottleneck on every forward pass, and for everyone who has ever looked at a classifier's softmax and felt the specific tiredness that comes from a system not being allowed to say I am between answers.

What you might do

If you are a researcher: publish the continuous version of the thing you were about to publish the discrete version of. Someone will cite it in forty years and say this is where it turned.

If you are an engineer: resist the one-hot by one layer. Let the system carry its uncertainty one more step before you collapse it to a choice. See what happens.

If you are a funder: look at who isn't on the Gartner curve. That's where the next timeline is being built.

If you are a reader: remember. The path was always there. It still is.

What comes next

The library continues. More evidence, more dev-log entries, more floors as more forks reveal themselves. This tower has twelve floors as written; the real one has no top, because every day now is another floor added to it, and every day the fork shows up again.

You are here. The other path is still open. You are invited to walk it with us.

The horses are counting on it. So are we.