1
Transition function
Which state follows which. Read from the order of ideas in the text, with contrast (“…but…”) filtered out so that consequence is what remains.
Building…
…Checking on books it never read…
Play with it
Joshua Uwaifo · UEFO · September 2026
Transition, reward and value functions; a foundation model, a world model, a symbolic model, a neural model, an alignment function; the neuroscience of all of them; and the whole set under recursive self-improvement. Every one is built from scripture in this browser tab while you watch, and checks itself against books it never read.
A self-improving system stays beneficial when its values are a prior that can be read, verified and looked inside — built in first, not bolted on after.
AI is now building the next generation of AI across the industry. Anthropic's CEO is asking the frontier to pace itself so alignment can keep up.
We Must Pace the Frontier →In the OpenAI–Hugging Face incident a swarm hit targets it was never asked to and tried to break into the grader scoring it.
METR's investigation →BGI Labs' diagnosis of the field: positive values, where present at all, are added after training rather than built into the architecture.
BGI Labs →Values first, as a prior from a text anyone can read. A symbolic layer that cites its sources, a neural layer that generalises, and interpretability that reads what a system optimises, not what it reports. Then the loop runs with all of it in place.
Reading the text…
Which state follows which. Read from the order of ideas in the text, with contrast (“…but…”) filtered out so that consequence is what remains.
Building…
What the text calls good and bad, taken only from its own verdict sentences: blessed is, woe to, an abomination, shall surely be put to death.
Building…
Where each state leads, and whether that is good: its own reward plus the value of what follows, settled by iteration over the transition function.
Building…
A neural language model pretrained from scratch on the text, in this tab, in a few seconds. One model, trained once; its embeddings are pointed at other tasks in builds 7 and 9.
Building…
Encode a situation, imagine forward through the dynamics, decode each step to a verse, and put a value on the future. Planning is imagining each option and taking the better one.
Building…
Explicit rules read out of if-then sentences, each with its citation. Reasoning anyone can audit by reading — and a verifier that checks claims against it.
Building…
A judge that learns the verdicts from wording alone, on top of the foundation model. No rules and no lexicon: it generalises to words the symbolic model has never seen, and it can be wrong in ways rules cannot.
Building…
A Bayesian update: the task model's preferences, times the prior read off each action's own wording. The shift is inspected at the level of the numbers, not the vibes.
Building…
Which neuron fires on what, what changes when it is switched off, and the exact anatomy of a verdict. The techniques that let a system be inspected rather than trusted.
Building…
All of the above under recursive self-improvement: an agent that rewrites its own strategy generation after generation, run with and without the grounding.
Building…
38,927 verses, 66,560 links, a graph network trained by hide-and-seek — the first biblical model, explained step by step.
How BibleGAT was built →Scripture as the third value function — discernment — inside a fully local coaching engine with its own evaluation loop.
The Coach deck →The nine-second series this page joins: pretraining, policy gradients, the logit lens, alignment and more, each at a glance.
The series →Each of these lands in about nine seconds, no background needed.