Inside the Machine
What AI models actually do, drawn from their real weights, activations and samples rather than from illustrations: a network grokking, a prediction locking in layer by layer, a search that did not help.
The Click Grokking, from a real training run test accuracy jumps from ~11% to 100% as beads snap onto four rings
Pieces
7 in this seriesRiver of Thought Logit lens 4,624 tokens from 60 short passages through every layer of Qwen3-1.7B; one thread locks onto “Rome” at layer 21
Condensation Diffusion model a model we trained dreams a jellyfish into a nautilus, in 30M points
Minds I Loop pack 13 silent loops of attention, latent space and evolution
Search, honestly measured Negative result two-model beam search scored 24/40 against majority voting’s 25/40, at about 4.7× the tokens
The Garden of Forking Paths Sampling 3,072 answers to one question: all distinct, five ideas
Guess the Next Word, Make It Say “Rome” Games bet against a real model’s odds; rewire heads to force an answer