"at" before end, bottom or start
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 23 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron fires on the word "at" in phrases about a position in a text or place, like "placed at the end", "at the bottom" or "at the beginning". Turning it up made the model write "at the end of the day" over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
surname or family name.⏎Is surname your last name?⏎In the English-speaking world, a surname is commonly referred to as a last name because it is usually placed at the end of a person’s full name, after any given names.⏎What exactly
-
about it! Check out the brainteasers below to see the power of the brain in action, and answer the questions at the bottom of the page.⏎Your brain can see things that aren’t there.
-
. If you want the list double-spaced, highlight all the text and set the paragraph to be double-spaced. Do not simply hit the Enter button twice at the end of each line, it will become an editing nightmare and when the sorting
-
. Students continue solving problems and finding answers as they move through the Fact Land. For some extra fun, students collect special tokens in certain boxes and tally them up at the bottom. Once students have made their way through the “land,” they count
-
.”)⏎He likes to eat popcorn when he goes to see a movie.⏎To play in the snow is one of life’s greatest pleasures.⏎(The infinitive appears at the beginning of the sentences. This sounds very formal.)⏎He went downstairs to get
-
postures) and movements designed to increase strength and flexibility. Breathing is actively encouraged throughout the class and a relaxation exercise (savasana) is practised at the beginning and/or end of class. You will leave class feeling energised and
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: what if we can create a system that can just immediately and immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately immediately
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
any paper they like. Here’s the easiest way I know.⏎- Choose any piece of paper.⏎- Fold the paper in half.⏎- Starting and ending at the fold, make the first stroke of the number 2.⏎- Cut along the
-
than just black blobs that mimic the shapes of things casting them. At the center, known as the umbra, the shadow is very dark indeed. But at the edges is a lighter, softer-edged area called the penumbra. When
-
focus at all points in a scene. We see the center of a scene with ideal light sensitivity and perfect focus and then focus and brightness will fall off increasingly toward the periphery. Therefore,
-
eater or steam through conduction. When the air circulates around the radiator using convention, the room is heated.⏎The distance from the center of a circle to the circumference. One-half of the diameter of a circle.⏎A tool used for checking
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L19:1989 suppresses
- L22:1467 excites
- L21:2680 excites
- L22:717 excites
- L21:893 suppresses
- L20:2201 excites 'the' in location phrases like 'at the end'
- L21:2835 excites
- L21:2851 suppresses
- L20:2458 suppresses
- L20:2571 suppresses
- L20:1604 suppresses
- L20:2257 suppresses
- L19:1967 excites
- L21:970 excites
- L21:2498 excites
- L18:2946 excites
- L22:1205 excites
- L22:1470 excites
- L19:1920 suppresses
- L19:2448 suppresses
- L16:623 excites
- L18:2149 excites
- L22:1450 suppresses
- L22:2371 weakly suppresses
Feeds later layers
- L27:599 strongly excites
- L26:835 strongly excites
- L24:1252 strongly excites
- L25:501 suppresses
- L24:856 excites
- L26:1210 excites
- L27:1119 excites
- L26:840 suppresses
- L26:2537 suppresses
- L26:1800 weakly excites
- L24:2468 weakly excites
- L24:2353 weakly excites
- L27:3040 weakly suppresses
- L27:2816 weakly excites
- L24:1282 weakly excites
- L25:2274 weakly suppresses
- L26:507 weakly excites
- L24:380 weakly suppresses the word 'at' in place phrases
- L24:2573 weakly excites
- L26:971 weakly excites
- L25:1268 weakly excites
- L25:724 weakly excites
- L27:2999 weakly excites
- L27:2323 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- at
- At
- At
- _at
- at
- →at
- -at
- .at
- (at
- "At
Suppresses
- hone
- antry
- ddl
- udson
- ypress
- 埠
- umont
- isle
- 呼
- orda