Inspect tokens and internal representations
Drowse reads every token it generates. Use the workbench to examine what the model predicted, where an activation sits relative to a fitted concept, what each layer reads and is about to say, and which neurons and heads pushed the token.
Token probabilities and alternative continuations
Click a generated token to open it in the Inspector. It shows the chosen token and the captured alternatives with log probabilities, and you can fork from any of them to test a different continuation. The token strip at the top of each room steps through the prompt and the reply, and the token chosen there is the token every room reads.
Captured alternatives are a limited recorded set. Their count is separate from sampling top-k, which changes how the next token is selected. A high token probability expresses the model's preference in context; it does not establish that a statement is true.
Concept probes and representation geometry
Geometry probes locate activations relative to fitted concepts, subspaces, and manifolds, at every token. The Probes room shows each probe's value at the token and its trace over the conversation, then where the token sits in each manifold's fitted geometry. A probe can also gate steering, so an intervention acts only while its reading crosses a threshold.
Coordinates, distances, fractions, and assignment probabilities have different meanings. A label describes the fitted reference examples. It should not be treated as a direct measurement of a person's beliefs, a model's private intentions, or factual correctness.
The R and S lenses
A lens reads a vector inside the model as a ranking over tokens. The R lens reads what a layer is disposed to make the model say; the S lens reads which input tokens it looks like it has read. Each reads at two places in a layer: the block's output and the residual after attention. The S read is the faithful one early in the model and the R read late, so the Lens room shows each layer through both, weighted by depth.
Lens readings are measurements derived from a fitted instrument, not a transcript of hidden reasoning. The browser uses precomputed lens packs fitted on the same quantized weights it runs.
Neurons and heads
On Gemma 3 and Qwen3 models, every MLP neuron and attention head is something you can read, probe and steer. A token's anatomy lists the units that pushed it; a neuron is one push or silence away, a head one silence. The Units room shows the strongest units at a token or over the reply, a whole layer at once, and a chosen unit's activity, what it writes through the lenses, and what it reads. GPT-2, Pythia and Qwen 3.5 2B Base are built differently and don't offer units.
A unit's reading comes from its weights and its activity on your text. A label is a summary of that evidence, not proof of what the unit does in every context.
Sparse autoencoder features
A compatible sparse autoencoder (SAE) pack exposes learned feature activations. Drowse shows feature strengths at each token, attaches features as probes, and steers along them. Available descriptions come from the matching dictionary; they remain interpretive labels.
Live measurements and historical replay
Readings are captured live as the model writes, and earlier tokens can be read again by replay, with or without the steering that produced them. Newly attached instruments may need a replay computation. Browser users install compatible precomputed SAE and lens packs; training SAEs and fitting lenses belong to the Python edition.