every layer, live · click one to open its attention
each column = one answer token, time runs right → · rows = layers, same order as the
architecture · click any cell to open its neurons — the full channel vector behind that number ·
debugging: a suddenly brighter band of columns = the model straining, often right
before the answer derails
click a cell in “activity over time” — every residual channel behind it
each bar = one residual channel of the clicked (token, layer) cell — together they ARE the
norm the heatmap shows · amber up = positive,
blue down = negative ·
red dot = channel far from its own average this generation (|z| ≥ 3) —
the interesting ones · hover a bar for channel, value and z · click to pin
click a layer on the architecture
each row = one answer word being written (top = first) · a bright spot = the exact
place it was reading at that moment · dashed │ = your prompt ends, the answer begins ·
debugging: a stuck vertical bar = repetition loop, attention frozen on one spot
click a layer on the architecture
one row per attention head — where each head looks while emitting the newest token ·
number = focus entropy (0 sharp → 1 diffuse)
click a layer on the architecture — then a head below
the full attention matrix of one head · each row = a token asking
"what do I look at?", each column = the token it reads · bright = strong attention ·
causal, so only the lower triangle lights up · hover a cell to read the two tokens ·
short prompts only
what the model would say if it stopped at each layer · top row = last layer ·
green = layer already agrees with the token that finally came out ·
Ln under the header = the depth where the answer crystallized · hover a cell for top-3 ·
debugging: green wave never arrives = the model is guessing · wave collapses mid-answer = steering too strong
How to read this: each column = the model writing one word.
The logit lens asks every layer "what would come out now" — this grid asks
"what are you disposed to say later".
Violet cell = the word it shows really arrives later in the answer ·
a violet left edge = a later word hides in that cell's top-5 behind a
different top guess (hover for word, rank and probability) ·
the violet line under each header sums up everything that moment saw coming ·
green = the word being written right now ·
faint cells = low probability, the lens's noise floor ·
hover for top-5, click to pin ·
Jacobian lens, Anthropic 2026
stored generations — click one to replay it
answer emergence: p(first answer token), best layer per step ·
amber = logit lens (would say it now) ·
green = J-lens (represented, not yet said) ·
dashed = top-k lower bound (no hidden states stored) · │ = </think> ends ·
click any word above to track that word instead
amber = the model decided late · red = never settled before lm_head ·
click any word to see what it was reading while writing it
the column on the left is the model itself — how loudly each part works
on the current token vs its own average · click a layer to drill into its attention ·
click lm_head to watch the answer crystallize layer by layer