Learn

Glossary

Every term that gets a dotted underline on the site, in full.

neuron

A neuron is one of thousands of tiny dials inside each layer of the model. It looks at the text so far and produces a single number that says how strongly it reacts. That number nudges what the model writes next, a little or a lot. No neuron works alone: the next word comes from all of them together. In Qwen3-0.6B each layer has 3,072 of them, 86,016 in all, and the ones here are the neurons of the feed-forward block, the part of a layer that follows attention.

layer

A layer is one step in the stack the text passes through on its way to the next word. Every layer reads the running summary of the text, adds its own notes, and hands it on. Early layers tend to react to spelling and grammar, later ones to meaning and to what should come next. The map shows the layers as a stack of sheets, with the first layer at the bottom. Qwen3-0.6B has 28 layers, each with 3,072 neurons.

activation

An activation is the number a neuron produces when it reads a piece of text. A large positive number means it fired hard; a number near zero means it stayed quiet; a negative number means it pushed the other way. The brighter a token is highlighted on a neuron page, the larger its activation was there. Activations were measured by running 2 million tokens of web text through the model and keeping the strongest moments for each neuron. Technical mode shows the raw values; they have no fixed unit and differ in scale from neuron to neuron.

steering

Steering means turning a neuron up or down while the model writes, to see what changes in the text. It is the closest thing the site has to an experiment: if a neuron really is about cats, pushing it up should make cats show up more often. The slider on a neuron page picks a strength, and the text under it is what the model wrote at that strength. All of it was computed in advance, so nothing runs in your browser. Strengths go from -1.0 to +1.0 as fractions of the typical signal size at that layer; 0 is the unsteered model, 0.6 usually still reads, and at 1.0 the text usually breaks down.

logit lens

The logit lens is a way to ask a neuron which words it votes for, without running the model. Each neuron has a fixed direction it adds to the text summary when it fires, and that direction can be read out as a list of words it pushes toward and a list it pushes away from. The lists are a hint, not a verdict: later layers can reshape or cancel the push. For many neurons the lists look like noise, and that is normal. Technically, the neuron’s output weights are multiplied by the model’s unembedding matrix, and the highest and lowest scoring tokens are shown.

residual stream

The residual stream is the running summary the model carries from layer to layer while it reads. Each layer reads it, adds something, and passes it on, so by the last layer it holds everything the model has worked out about the text so far. Neurons write into it when they fire, which is how a single neuron can change the next word. Steering strengths are measured against it: a strength of 1.0 adds a push as large as the summary itself usually is at that layer. In the model it is a vector of 1,024 numbers per token.

token

A token is a piece of text the model reads in one go. It is often a whole word, but longer or rarer words are split into pieces, and spaces, punctuation and line breaks are tokens too. The model never sees letters or words directly, only these pieces, which is why a neuron can react to a word fragment like “ing”. On a neuron page the outlined token is the one where the neuron fired hardest. Qwen3-0.6B has a vocabulary of about 150,000 tokens covering many languages.

featured

A featured neuron is one picked for the full treatment. Every neuron has a page with its strongest activations, but computing steering text for all of them would take weeks, so a subset was chosen. Featured neurons have a steering slider, a label and a written explanation, and they are drawn larger and paler on the map. They were picked for having a readable pattern, so they are not a fair sample of the whole model. 2,996 of the 86,016 neurons are featured.

score

The score is a rough guess at how easy a neuron is to read. It asks a simple question: when this neuron fires hardest, is it always on the same word? A neuron that fires only on the word “the” gets a high score; one that fires on a mix of unrelated things gets a low one. The score is only a hint, and a low score can still hide an interesting neuron whose pattern is about meaning rather than one word. It is the share of a neuron’s top snippets that peak on the same token, so 1.0 means a pure token detector.

histogram

The histogram on a neuron page shows how often the neuron takes each activation value. Most neurons sit near zero almost all the time and fire hard only rarely, which gives the chart a tall spike in the middle and a long thin tail. A wide or lopsided shape means the neuron is busier, or leans one way. The chart is built from a random sample of the measured activations, not from all 2 million tokens. The horizontal axis is the activation value and the vertical axis is the count in each bin.

confidence

Confidence is the high, medium or low mark on each explanation. It is Claude’s own judgement of how well the evidence supported the story it wrote: high when the top snippets, the logit lens and the steering text all point the same way, low when they disagree or the pattern is vague. It is not a measured accuracy; nobody checked the explanations against ground truth. Treat a low mark as an invitation to look at the snippets yourself. The mark comes from the same Claude call that wrote the explanation, from the same evidence.

random control

A random control is a push of the same size as the steering push, but along an arbitrary direction in the same layer instead of along the neuron. Pushing hard on anything changes the text, so a change on its own proves little. The control shows what a meaningless push does, which lets you tell a neuron’s real effect from the noise of being pushed at all. The steering card has a toggle that shows the two side by side. Each layer has 50 random directions precomputed at the same strengths and prompts; the card shows the first.

connection

A connection is a strong direct link between two neurons in different layers. When the earlier neuron fires, it writes into the running summary of the text, and the later neuron reads from it: if what one writes lines up with what the other listens for, they are connected. On the map, amber links mean the earlier neuron excites the later one and blue links mean it suppresses it. The links only cover direct reads, so two neurons can work together without being linked here. Technically, the link strength is the cosine between the earlier neuron’s output direction and the later neuron’s input direction, and only the strongest links are kept.

explanation

An explanation is the short account of what a neuron seems to do, shown at the top of a featured neuron’s page. It was written by Claude, not by a person, from the neuron’s strongest snippets, its logit lens and its steering text. It is a guess from limited evidence, not ground truth, and the evidence it was written from is on the same page so you can judge it. Each explanation has a simple version and a technical version, which the mode toggle switches between. The writer was Claude Haiku 5.5, asked for a label, both versions and a confidence mark in one call per neuron.

polysemantic

A polysemantic neuron responds to several unrelated things at once: the same neuron might fire on chemistry words, on a kind of punctuation and on a few names, with no single story that covers them. This is the normal case, not the exception, because the model has far more things to track than it has neurons and packs several into each one. It is the main reason neurons are hard to read and labels are often incomplete. Featured neurons were picked for having a cleaner pattern than most, so the rest of the model is messier than the sliders suggest. The usual explanation is superposition: features stored as overlapping directions rather than one per neuron.