'a' after a preposition
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. 9 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 'a' or 'an' when it comes after a preposition like 'for', 'from' or 'of', as in 'look out for a definition'. Its opposite is 'the' after 'one of'. Turning it up did not produce a clear theme in the generated text.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
in general, but if they get in this habit early, they will be building a routine that can last a lifetime. Ask what your child wants to do for a
-
the panelists could come to a common understanding. This raised a few questions in my head. Disappointed with the expert advice, I decided to look out for a definition of my own.⏎As what’s needed in such a situation, I looked
-
ill-equipped men in his command. Fremont, whose Federals were also ill-equipped and trained, had nearly 40,000.⏎Price left a small garrison at Lexington, which was overrun by Major Frank J. White’s
-
.⏎We tend to dwell on the deaths, but often overlook the survivors of a life altering collision. Crashes don’t discriminate! Everyone is at risk, from a newborn properly restrained in a car seat, to a bicyclist or pedestrian in the
-
Hitler on both occasions, thus putting himself firmly on the Gestapo’s list of enemies; in 1938, as the Nazi vice was closing on an independent Austria, the archduke, at obvious risk to his life, volunteered
-
System that contains information on reported live births in the U.S.⏎The researchers were interested in pinning down the average age of first-time mothers since the age of a woman’s first birth has a bearing on the total number of births she may have
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 know about a date, but I have to go on the record." So, I'm not sure about anything, but I have to go on the record. [Inception] 04.12.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
number of cancers and scientists around the globe are busy trying to harness the power of the immune system in the fight against cancer. The HPV vaccine is just one of the first; expect to hear more and more about the topic in the future.⏎The
-
. As you develop your<|endoftext|>Catherine Rampell had a post last week declaring that “College is (Still) Worth It“. The piece is the latest in a series that she and her Economix blog colleague
-
, the United States has lost over 50 percent of its managed honeybee colonies over the past 10 years.⏎Pollinator gardens are just one of the many ways to provide both shelter and food to insects teetering on the brink
-
This video is part two of a seven-part National Academies series, Climate Change: Lines of Evidence. The video
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
- L7:312 excites
- L6:61 suppresses
- L8:636 suppresses
- L6:29 suppresses
- L5:125 excites "of" after all three, both, ten
- L3:528 excites
- L8:475 excites
- L7:396 excites
- L3:47 weakly excites
- L4:217 weakly suppresses
- L4:16 weakly suppresses
- L3:489 weakly suppresses
- L8:1232 weakly excites
- L1:235 weakly suppresses
- L6:819 weakly suppresses
- L8:508 weakly excites
- L1:2074 weakly suppresses
- L3:545 weakly suppresses
- L7:2164 weakly suppresses period inside country abbreviations
- L4:731 weakly excites
- L8:333 weakly excites
- L7:401 weakly suppresses
- L4:48 weakly excites
- L3:472 weakly suppresses
Feeds later layers
- L11:1280 excites
- L14:655 excites
- L14:169 excites
- L18:381 excites
- L15:716 excites
- L11:55 suppresses
- L18:172 suppresses
- L14:600 excites
- L12:652 excites
- L14:871 excites
- L15:1205 excites
- L17:362 excites
- L18:715 weakly suppresses
- L13:976 weakly suppresses
- L12:2478 weakly excites
- L17:146 weakly suppresses period after abbreviations and dates
- L15:37 weakly excites
- L20:251 weakly suppresses
- L16:66 weakly suppresses
- L22:194 weakly suppresses
- L16:528 weakly excites
- L14:1876 weakly excites
- L21:637 weakly excites
- L13:831 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
- GED
- dreaded
- <decltype
- gaping
- 方方面面
- ozilla
- 等内容
- このような
- figcaption
- こんな
Suppresses
- increment
- قيد
- ion
- ión
- l
- 束
- Increment
- rement
- IONS
- -ne