Small, strange projects where algorithms get room to surprise us.
Where ELI5 gets too complicated.
A language model with the memory of a goldfish and the confidence of a man explaining crypto at a party: it remembers exactly one word, and it bets on the next one anyway. Watch it learn nursery rhymes live, every single number on display, right up to the moment "one potato, two potato" breaks its tiny heart.
Explore Word Nursery →The Little Language Model, now with two word memory.
Same baby, now with a room where each word can gossip about the one next to it. Give or take five billion babies in a trench coat, that's how the grown-up models do it. 🧥 Watch it crack the rhyme that beat its smaller sister, audit the maths behind any bet, and weigh it on the growth chart against GPT-3 (it loses, but it's trying).
Explore Word Nursery: The Room →Wobblers evolves random blobs of nodes and muscles into walkers. Nothing is designed, just blind mutation plus survival of whatever gets farthest in 10 seconds.
Explore Wobblers →I love it because it keeps surprising me: it cartwheeled, skated and skipped its way through every loophole in my physics before I knew the loopholes existed.