A neural agent learning English from scratch – and now with a friend, Kogli. No rules, no vocabulary list — only hunger, curiosity, books, a teacher’s voice and someone to talk to. This page shows its brain at work, live.
Every milestone is a test against a baseline: with and without the spoken words, and against acting at random. It only counts as stable once it holds in three tests in a row.
NikolAI lives in a small world of five places — hut, orchard, forest, river and workshop — in real time, with its current brain. Everything it hears comes from its teacher; everything it says, it has learned by itself.
Ask it something – “Where are you?”, “How are you?”, “What do you have?”, “What do you want?”, “What did you do?” – or say anything. It answers with what it learned from reading and from its teacher. Give one of the teacher’s instructions and you are the teacher: if it follows, it learns.
NikolAI reads what a child hears and reads by about twelve – some 100 million words, all from people: parents talking to toddlers, conversations, children’s books, film subtitles, Simple Wikipedia (the BabyLM corpus). It keeps guessing the next word – nobody tells it what words mean. That reading brain is its hearing and speaking. The question: does it then understand instructions with words its teacher never used?
From a start that doesn’t give the answer away (several things to choose from, or the hut with four ways), 4 steps: with the sentence, without it. The teacher never says these words. Control: the same sentence with a made-up word (“Blick the stone!”) – if that works just as well, it only reacted to the noun.
Left: the words NikolAI says most – bigger means more often, the redder, the hungrier it was when it said them. Right: every word has a place in its brain (a vector of 256 numbers); words it uses alike lie close together – once as it knows them from books, once as it uses them in its world. Type a word to see its neighbours – or take the whole map into Obsidian.
Download as Obsidian vault →
One note per word, linked to its closest words; open the graph view.
NikolAI has never seen a picture. Every thing in its world has real properties – colour, size, shape, surface, parts – and it only learns them by seeing, taking, carrying and eating. From the heard word alone, its brain predicts what to expect and how sure it is. Each picture overlays several guesses drawn from that expectation: where it is sure, they agree and the picture is sharp; where things vary (apples: red, green, yellow), the guesses differ; where it has no experience, there is only fog.
Every half hour: how it pictures an apple right now.
Success within 15 steps, 30 trials each — with the sentence, without it (same situation), and acting at random. Marked rows: never trained, only tested.
A big and a small one lie on the floor (or it carries both), with other things around, 10 steps, 30 trials. Each thing is asked for both ways – “big …” and “small …” – from the same starts. When it takes (or puts down) that thing: how often the right size? If it ignores the size word, it gets one of the two right at most. “Understood”: both at least 70 %. “Noun only” = which size it takes for “Take a stone!”. Marked rows: the teacher never says “big stone” – NikolAI heard “big” and “small” with apples and branches, but “stone” only as “small stone”.
Imagined size from the words alone (0 = tiny, 1 = huge):
In the workshop, far from the apple tree, 10 steps, no instruction: does it say the right word on its own?
Hungry in the hut, the apple tree is empty, one apple lies somewhere else. 4 steps — two are enough if it understood.
From the heard word alone: how far are its expected looks from what the thing is really like? “Average” = just guessing the average thing.
NikolAI stands where an apple lies; its friend Kogli is hungry in the hut; no teacher. Does Kogli get the apple within 10 steps? “Cut” = they can’t hear each other.
A random situation, then the teacher asks. Exactly the true answer? “Guessing” = always giving the most frequent answer.