"terms" in "in terms of"
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. 10 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 reacts very strongly to the word "terms" in the phrase "in terms of", as in "advantages in terms of speed". Turning it up made the model's text wander into repetitive loops instead of anything about that phrase.
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.
-
approval – especially in matters concerning finance and law. As the world changes, most of our financial matters have moved seamlessly online. Our digital identity is lagging in terms of adoption but it is only due to the sensitivity of the matter at hand.⏎
-
from five healthy volunteers harbored a large and<|endoftext|>Although a Last Will and Testament remains the most common estate planning tool, a living trust is not far behind in terms of popularity. If you find yourself to be the beneficiary of a living trust,
-
a site of memory, including the ‘ESMA Memory Site Museum,’ dedicated to telling the stories about captivity in the building.⏎“Argentina had a great advance in terms of truth trials, but there’s a lot left to do for women,” said
-
If equity isn’t at the center of the decisions that are made to open up outdoor spaces, “we risk perpetuating existing inequalities that have a real cost in terms of the health of communities of color as well as others who have historically lacked equitable
-
and multilingual society. “It was after having been in the USA, where I studied Library Science, that I saw that Malaysia has a lot to offer in terms of its diversity. You grow up aware of differences but being comfortable with it.”
-
the highest quality in terms of natural flavor and taste, and which are free from additives and preservatives.
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 am trying to create a system that will help users build a virtual reality environment. What is the first thing I think about doing? I need to create a system that will allow me to build a virtual environment, what I need to do
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
batch of 41 country pages, with more to follow. They are:⏎Afghanistan, Angola, Argentina, Bahrain, Belarus, Belgium, Burkina Faso, Burundi, Cambodia, Cameroon, Colombia, Costa Rica, Czech Republic
-
.⏎The passages in classical literature which are to be considered, are the following: —⏎(1) Dionysius II 65: οὔτε γὰπ τὸ χωρίον τοῦ
-
States and Mexico; The opposite occurs with Cubans and Puerto Ricans living in the United States.⏎This is reflected in a survey carried out between
-
the presence of Paleozoic rocks among the field of cainosoic rocks. Twenty years ago B.Pilya found some strange glasses here and P.V.Florensky began to study them. With a point of view of meteoritical
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
- L6:293 strongly suppresses
- L8:62 strongly suppresses
- L4:1572 strongly excites
- L7:9 strongly suppresses
- L5:2525 strongly suppresses
- L6:296 strongly excites
- L5:418 excites
- L8:79 suppresses
- L5:667 suppresses
- L4:1045 excites
- L6:1031 excites
- L6:1058 suppresses
- L6:2400 suppresses
- L4:2029 excites
- L4:1781 suppresses
- L5:1127 excites
- L4:2382 suppresses
- L8:1546 suppresses
- L6:108 suppresses
- L2:87 suppresses
- L5:34 excites
- L5:58 excites
- L7:701 excites
- L4:1854 excites
Feeds later layers
- L17:70 strongly excites
- L12:50 strongly excites
- L19:204 strongly excites
- L20:131 strongly excites
- L17:152 excites mid-word fragments, often 'intric'
- L18:513 excites
- L14:118 excites
- L14:248 excites
- L14:254 excites
- L16:414 excites
- L14:126 excites
- L16:269 excites
- L16:283 excites
- L14:280 excites
- L15:953 excites
- L21:189 excites document break token mid-sentence
- L19:541 excites
- L24:39 excites
- L14:830 suppresses
- L22:164 excites
- L18:417 excites
- L14:421 suppresses
- L25:17 excites
- L17:259 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
- itary
- udio
- 我又
- lider
- 目的
- -flex
- imately
- OfWork
- entially
- aised
Suppresses
- ningar
- jag
- }{⏎
- j
- rational
- <j
- pert
- \Db
- gra
- vido