Neural Observatory
Enter the lab →
Neural Observatory · logit lens

See how AI thinks

An instrument for watching a real GPT-2 forward pass — attention, hidden states, and, below, the model’s answer taking shape one layer at a time. Nothing here is a simulation.

Enter the labGPT-2 small · 124M parameters · every step from a real forward pass
Predicting the token afterlogit lens

The animal didn't cross the street because it was tired→ ?

·tired
confidence100.0%
held across layersL00–L01

An intermediate guess — the deeper layers have not settled here yet.

Descent · embedding → outputdepth 00/12

Measured sample captured from this project’s own backend. The logit lens decodes the residual stream into vocabulary space at every layer, so each row is what the model “would say” if asked to stop and answer at that depth. Every other view in the lab is fed live.

Under inspection

What you can inspect

Transformer layers
12

Stacked blocks. The lens falls through all 13 depths, embedding to output.

Heads per layer
12

Independent attention circuits running side by side in every block.

Attention heads
144

Isolate one and see exactly which earlier tokens it reads from.

Residual width
768dims

The channel each token's state travels through, end to end.

MLP neurons / layer
3.1K

Post-GELU features; the strongest firings surface per token.

Parameters
124.4M

Every weight in the model — few enough to hold the whole thing in view.

Vocabulary
50.3Ktokens

The space the lens decodes into at each layer.

Context window
1.0Ktokens

How far back attention can reach in a single pass.

Four instruments

Where to point it

Begin

Watch a thought form, in real data.

Enter the lab