'at' in 'at least' and 'most' or 'least' qualifiers
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. 24 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 like 'at least' or 'at most', and on similar 'most' and 'least' words. When turned up, the model got stuck repeating 'most' and 'least', which matches the kind of word it likes.
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.
-
this planet?⏎Or does it lend credence to the idea that life could spring from inanimate material on this planet by showing that it happened not once, but at least twice?⏎Either way, expect the religious fundies to shrug it off
-
in freely moving animals. It is also capable of long<|endoftext|>According to legend, the first Valentine greeting was sent by St. Valentine himself. Actually, there are at least three
-
Vacciniums I think I may have identified here goes by the common names of Black Highbush Blueberry and Downy Swamp-Huckleberry. There are at least 3 species of Vaccinium/Gaylussacia in the photo above
-
A similar outflow is probably present under the disk as well and is hinted at in this illustration. The total amount of material entrained in the broad flow is at least 400 times the mass of the Sun per year. Note that
-
expression consists of a single line containing a positive integer number, an operator (one of +, - and *) and the second positive integer number. Every number has at most 500 digits. There are no spaces on the line. If
-
is underway in Australia to isolate exactly the type of tick and the specific bacteria that is causing Lyme disease here (there are at least 14 different types of bacteria that cause Lyme disease identified world wide).⏎
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: the small 1666-1667. In the small 1666-1667, the 1666-1667 and the 1666-16
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
15 different kinds of cancer are linked to smoking and other tobacco use. Nearly one-third of all cancer deaths are due to smoking. Quitting smoking is one of the best things you can do for yourself and for those you love. Click
-
choices that can reduce risk, including:⏎- Being physically active – More and more research indicates that physical exercise is one of the most important lifestyle choices for reducing the risk of developing Alzheimer’s. Exercise
-
talented instructors at Veritas Prep teach students how to prepare for GRE Verbal Reasoning questions. There are several practical strategies available to students that can help them conquer even the most difficult questions in this section.⏎Check out some valuable tips that students may
-
most snakes?⏎Florida is considered to be the state with the most snake species in the US, but is it also<|endoftext|>Be a Dent
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
- L21:1741 excites
- L20:361 suppresses
- L21:614 suppresses
- L17:30 excites
- L18:1547 excites
- L19:427 excites
- L11:331 suppresses
- L18:1084 excites
- L16:419 suppresses
- L19:644 weakly suppresses
- L4:49 weakly excites
- L1:59 weakly suppresses 'the' after a verb or preposition
- L21:1072 weakly suppresses
- L20:219 weakly suppresses "the" after time or place prepositions
- L19:2115 weakly suppresses
- L9:59 weakly suppresses "a" meaning per (a day, a year)
- L6:328 weakly suppresses
- L18:861 weakly excites
- L15:2587 weakly excites
- L18:641 weakly excites
- L17:2129 weakly excites
- L11:2473 weakly suppresses
- L11:875 weakly excites
- L9:24 weakly suppresses "the" before a named organisation or title
Feeds later layers
- L27:1338 strongly excites
- L26:1525 strongly excites
- L27:1086 strongly excites
- L27:2450 excites
- L25:2905 excites
- L26:631 suppresses
- L25:2986 weakly excites '1' in 21st or 11th, before 'st'
- L27:1060 weakly excites
- L26:670 weakly suppresses
- L26:2537 weakly excites
- L27:1811 weakly suppresses
- L26:2398 weakly excites
- L26:2458 weakly suppresses
- L25:2844 weakly excites
- L26:1094 weakly excites
- L25:626 weakly suppresses
- L27:2871 weakly suppresses
- L25:2511 weakly suppresses
- L27:1296 weakly excites
- L25:2102 weakly suppresses
- L26:1941 weakly excites
- L25:2221 weakly excites
- L27:662 weakly excites
- L25:1407 weakly suppresses 'in' right after a place or time
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
- 最
- best
- least
- most
- 最好的
- least
- 最佳
- most
- 最有
- 最大的
Suppresses
- nton
- uden
- vt
- tat
- rán
- 俟
- ведь
- vation
- vertiser
- ��