'which' 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. 16 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 'which' when it follows a small word like 'to' or 'from', as in 'a common purpose to which' or 'opium poppies, from which'. Turning it up gave text that did not show this pattern, so the link is only in the examples.
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.
-
teams are highly structured and are sometimes confused with the work group. Teamwork can be difficult to establish within a traditional team without the identification of a common purpose to which all workers can agree and commit.⏎The self-directed
-
to a garden, a theme to which all other parts relate. This may be a walk or a summer-house or a
-
and parsley. When crushed, wild chervil smells sweet and carrot-like.⏎Wild chervil has a more bitter flavor than garden chervil, to which it is related. As with garden chervil, cooks primarily use wild ch
-
medical course at the University of Tübingen, but the Franco-Prussian war interrupted his studies. After the war he returned to the university, from which he received the M.D. degree with honors in 1877.⏎
-
Italian botanists of the Renaissance devoted them-selves chiefly to interpreting the works of the classical writers on Natural History, and to the identification of the plants to which they referred. This came about quite naturally, from the fact that the Mediterranean flora
-
with fibromyalgia. Decongestants are stimulants which the doctor says can make the symptoms of co-morbid conditions (a category to which fibromyalgia belongs) worse.⏎There's an additional risk of an unintention
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 is the right way to do for you? for you to know how to help you with your to you for you to help with your for you to the way of for you to the for you to the (what is the right way
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
three types of configuration files: machine, application, and security.⏎Configuration files contain elements, which are logical data structures that set configuration information. Within a configuration file, you use tags to mark the beginning and end of an element. For example,
-
41 patent models.⏎This sampling of patent models from the Textile Collection describes the two major groupings, textile machinery and sewing machines. In both groups, the examination of the models begins with the earliest of the inventions. In this early
-
ation: To ensure that an RTG infrastructure will be in place to meet requirements after the 1997 Cassini<|endoftext|>When building cascading deck stairs, it is important to make sure that the stairs are properly supported and that each stair
-
sections:⏎In addition, the Wiki contains state-by-state information on ratifications of international human rights treaties, communications/ complaints mechanisms and inquiry procedures.⏎On the Wiki, read:⏎Our eventual aim is to make all information about children's rights available in
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
- L15:853 strongly suppresses
- L8:1052 suppresses
- L11:1781 suppresses
- L11:949 excites
- L14:520 suppresses
- L15:1468 excites
- L15:2521 excites
- L13:2780 excites reciprocal 'each other' or 'one another'
- L10:371 suppresses
- L13:2516 suppresses
- L9:361 excites
- L14:2854 suppresses
- L3:311 suppresses
- L4:1593 suppresses
- L11:2663 excites
- L4:23 weakly excites
- L4:1330 weakly suppresses
- L15:379 weakly excites
- L5:729 weakly suppresses
- L5:2424 weakly excites
- L5:224 weakly excites
- L10:1603 weakly excites
- L5:1734 weakly excites
- L10:289 weakly excites
Feeds later layers
- L17:2238 strongly excites
- L20:1912 excites
- L17:2878 excites
- L18:162 excites
- L17:2643 suppresses
- L20:242 suppresses
- L20:2146 excites
- L17:2162 weakly excites
- L18:740 weakly excites
- L20:386 weakly excites
- L17:2431 weakly suppresses
- L18:2476 weakly suppresses
- L17:2524 weakly suppresses
- L19:689 weakly excites
- L19:366 weakly excites
- L17:2690 weakly excites
- L25:987 weakly excites opening bracket before an acronym
- L19:1364 weakly excites
- L20:685 weakly suppresses
- L25:666 weakly suppresses
- L18:1007 weakly suppresses
- L21:1872 weakly suppresses
- L19:2455 weakly suppresses
- L25:281 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
- bú
- zę
- prostitution
- São
- 供电
- ходим
- brazil
- bildung
- rosso
- 阄
Suppresses
- 对她
- 给我们
- notin
- 对企业
- olia
- 共和
- 对孩子
- stone
- 出炉
- fur