number right after 'top' or 'bottom'
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. 11 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
Fires on the number that follows the word 'top' or 'bottom' in a ranking, such as 'top 10' or 'bottom 15 percent'. When turned up, the model's writing moved toward rankings and 'highest' lists.
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.
-
low or no earnings, perhaps 40 percent of college-educated workers make less than the top 25 percent of high school graduates.⏎This alternative reading of the figure helps to
-
08 NBA season, there were a total of 245,811 points scored by 450 players (data here). The top 20% of scorers (led by #1 Kobe Bryant with 2
-
5 years, the main causes of death in the world have remained fairly consistent.⏎The top 10 leading causes of death in the world:⏎- Heart disease⏎- Cancer
-
ondall State School 's Year 3 Literacy Test results have increased significantly since the ELF program began. The percentage of students identified as being in the bottom 15 percent of the state has decreased from 21.6 percent to
-
than any other continent.⏎What continent has the largest animals?⏎List of largest land carnivorans⏎|Rank||Common name||Native range by continent|⏎|1||Polar bear||North America, Eurasia|⏎|2||Brown bear
-
. The figure also shows, for example, that something approaching half of college graduates (the figure doesn’t allow for too much precision) make less than the top 25 percent of high school graduates. And somewhere approaching half of high school graduates
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: "I don't mean to say, but I see. I mean that I mean." So far, I have been thinking about, and I have been using. I can see. I mean it is clear, so it is clear.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
site, additional surface material was collected and one 1 X 2 meter test pit was excavated to determine the site's approximate extent, cultural affinity, and stratigraphic position. From cultural material recovered on the surface and in the test excavation
-
(6-8) Show all Grade Level⏎Resource Type: Activities⏎Results 1 - 7 of 7 matches⏎From Grid to Home⏎Lisa
-
marriage behind everyone’s back, everything seems to be testing them; including a fight that broke out and ended in Tybalt’s murder and Romeo being<|endoftext|>Show Contents Previous Page Next Page⏎Chapter 3 - The Apache Module Architecture and API⏎
-
|⏎Widely distributed along the base of the Southern Alps lies a series of unfossiliferous sedimentary beds, consisting for the most part of well-stratified gravels, sands, and clays, with occasional lignite, whose
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
- L10:2995 excites
- L10:1207 suppresses
- L5:1902 excites
- L6:2088 suppresses
- L6:1934 suppresses
- L6:2402 weakly excites
- L7:2566 weakly suppresses
- L7:2337 weakly suppresses
- L10:2133 weakly excites
- L6:2045 weakly suppresses
- L9:5 weakly excites
- L9:647 weakly excites
- L6:12 weakly excites
- L2:1895 weakly excites
- L10:0 weakly excites
- L8:2872 weakly excites
- L9:3040 weakly suppresses
- L8:1523 weakly excites
- L9:2569 weakly excites
- L3:1832 weakly excites
- L7:1776 weakly suppresses
- L8:1800 weakly excites
- L10:2363 weakly suppresses
- L9:1232 weakly suppresses
Feeds later layers
- L17:2737 excites
- L15:2789 excites
- L27:2575 excites
- L14:2092 excites
- L18:2149 excites
- L22:969 excites
- L22:1605 excites
- L19:1263 excites
- L26:1525 excites
- L22:2560 excites
- L21:1581 excites
- L13:513 excites
- L12:2921 excites
- L19:2558 weakly excites
- L21:542 weakly excites
- L19:2312 weakly excites
- L20:1150 weakly excites
- L23:2127 weakly excites
- L13:571 weakly excites
- L17:1168 weakly suppresses
- L13:2110 weakly excites
- L18:2465 weakly excites
- L18:2934 weakly excites
- L17:2688 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
- 排序
- ranking
- highest
- highest
- 最快的
- 排名第
- 排名
- 最高的
- largest
- 排名第一
Suppresses
- arel
- aras
- asions
- .Done
- aton
- utions
- اث
- ardless
- nung
- วร