a correlation is two vectors that agree
papers say "r = 0.1" as if anyone could picture it. here it is as colours: two rows of numbers, blue for low, amber for high. when the rows rhyme, r is high. then we put it on a brain — where an AI's predictions meet real recordings, and r = 0.1 turns out to be a lot.
01two vectors
each cell is one measurement. top row: what a model predicts. bottom row: what was measured. the dot cloud is the same 48 pairs, one dot per column — hover a column to find its dot.
02the ladder
same predicted row every time; only the agreement changes. below ~0.3 your eye finds nothing. at 0.5 the colours start to rhyme. you need ~0.9 before the rows look alike.
03now put it on a brain
an encoding model reads the story a person hears, turns every word into an LLM embedding, and predicts the response at each point of the cortex. every point gets its own pair of vectors — predicted vs measured over time — and so its own r. paint all of them and you get a map. hover the cortex.
04why r = 0.1 is a lot
one spot at r = 0.1 can't tell you which sentence someone heard. but a decoder doesn't look at one spot — it compares the whole brain pattern against every candidate. each of D independent dimensions adds a little evidence, and the right answer pulls ahead like r·√D.
05where the best systems stand
encoding (AI → brain) is reported as r; decoding (brain → meaning) as accuracy against chance. a few landmarks, checked against the papers.