UEFO · Demos · Joshua Uwaifo · September 2026
Nine demos built to land in about nine seconds, with no background needed. Under them, two pages that build the biblical models for real, and fourteen short demos of the mathematics, every number computed from the 66 books.
The single move everything else is built on, what it costs, and how doing it at scale produces a model that can be pointed at many tasks. These three, and the six after them, are illustrations; their numbers are the demos' own, not measurements from the Bible model.
A sentence assembles itself one weighted guess at a time. Click a long shot and watch it swerve somewhere strange.
Open →Genesis 1:1 costs 156 bits under illustrative untrained probabilities and 11 bits under trained ones. Same words, nothing removed, with the arithmetic shown live.
Open →Where the foundation comes from, and why one model trained once can then be pointed at many tasks.
Open →Three ways to steer a model after it can talk: keeping what works, nudging the odds, and building something that looks a lot like character.
Throw darts at a star. Keep what lands inside, toss what misses. The keepers alone draw the shape.
Open →A cannon learns to aim. Its aim curve slides across and sharpens onto the bullseye, shot by shot.
Open →A grey, flat-eyed face warms into an expressive one, one preferred reply at a time. Then it handles something it never saw.
Open →Two ways to read a model's half-finished thinking, and one demonstration of what goes wrong when nobody checks.
Scan up through twelve layers and watch "France" lead at layer 6, then lose to "Paris" at layer 8.
Open →Not the next word but the concepts live right now. Watch them light up, hand off and fade, and see what else fires each one.
Open →Tell a robot to tidy a room. Watch it game every rule you write, and see what each patch misses.
Open →Everything above, built for real from one text: first with a hand-written lexicon, then with no lexicon and no labels at all.
Names, states, verdict words, relationships and values as statistics of the text, each learned until it stops changing. A fair test on six unread books, the Hebrew and Greek read separately and found to agree, and the one result that failed, shown as it is.
Open →Ten model types built from scripture in about five seconds each and checked against books never read, then a self-improving agent that hacks its grader while a grounded one refuses. v2 traced that refusal to the hand-written lexicon, which is why this page is kept as the record of what changed.
Open →Fourteen nine-second demos, from vectors to a GPT trained from scratch on the Bible alone. Every number is computed from the 66 books or from that model; the few outside figures, such as hardware prices, are dated and linked.
Gold and silver point the same way; gold and wicked do not. The Bible's own words as arrows, with the angle between any two exact.
Open →Four numbers move every point at once. Drag the two columns and watch the word cloud turn, stretch or shear with them.
Open →Turn the rulers and every coordinate changes, though nothing moved. A tilted cloud becomes simple in its own basis.
Open →Slide one column onto the other and the plane folds onto a line: different words land on the same point, and no undo exists.
Open →The 66 books' similarities to each other, 4,356 numbers, rebuild to 89% from two patterns and 92% from five.
Open →Every matrix is rotate, stretch, rotate. Keep the biggest stretches and you have the best compression of a 240-word by 66-book table.
Open →After "in", the next word is "the" 35% of the time, because that is how often it was. Spin the wheel and a sentence writes itself.
Open →A word costs log₂(1/p) bits. Genesis 1:1 priced word by word, and why a better model is one that pays fewer bits.
Open →Move over the hills and the arrow shows the steepest way up; step against it and you descend. Slopes multiply through a chain: backpropagation.
Open →Plain descent, momentum and Adam race down one valley; the learning rate decides who arrives. Beside them, the real loss curve of the Bible GPT.
Open →Tokenizer, attention, MLPs and training loop written by hand, trained on nothing but the 66 books, and running in your tab. Type a prompt.
Open →Anthropic's framework: every head has a where (QK) and a what (OV). Watch the heads of a two-layer model on a verse, then test them for induction.
Open →Switch off one head or one MLP and watch the next word change, on the real model, with the three lines of nnsight that do the same.
Open →Pick a model size and a number of passes over the Bible; see the hours and dollars on RunPod and Vast.ai, dated, then the commands that run it.
Open →Reading a model through its features rather than its words.
Tireni's Neuronpedia feature-space demo: features in a global-workspace view. Joins this hub once its standalone copy is in.
ComingA natural-language autoencoder reads the activations behind an answer and writes out what the model was doing. Paper examples and illustrative scenes are marked as such; links to the real thing on Neuronpedia.
Open →