SYNAPSE
WAKES UP.
Move your cursor and the network feels it. Every dot is a neuron, every line a connection deciding whether to fire. This whole page is one HTML file, built by Claude Code from a single prompt.
Intelligence became
a counting problem.
A language model never learns rules. It reads almost everything ever written, compresses the patterns into billions of weighted connections, then predicts one token at a time.
parameters in the model that made the world notice, back in 2020
tokens of context a frontier model can hold in one thought today
needed to build the page you are reading right now
It thinks by guessing
the next word.
Watch the model weigh candidates for the next token. Thousands of times per sentence, each guess reshaped by everything before it.
invent p=0.41
sampling…
Every word looks
at every other word.
This is the whole trick behind the transformer. To read one word, the model scores how much every other word matters to it. Pick a word and watch where its attention goes: "it" knows it means the animal, not the street.
One knob decides
safe or strange.
The model produces a score for every possible next token. Temperature decides how sharply those scores turn into a choice. Turn it down and the model repeats itself; turn it up and it starts to wander.
Seventy years,
four sentences.
The question
Turing asks "can machines think?" and proposes judging by conversation, not by mechanism.
The method
Backpropagation lets networks learn from mistakes. Small ones, for now.
The architecture
"Attention is all you need." The transformer reads everything at once, and scales.
The coworker
Models stop suggesting and start doing: reading codebases, running tools, verifying their own work.