"with" after a noun, introducing contents
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 lights up on the word "with" when it follows a noun and describes what that noun includes or comes with, like "a complete history with facts" or "a timeline with points". It is quieter, and sometimes negative, on "with" after a verb that names a partner, like "curry favour with". Turning it up did not give a clear theme in the output.
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 different cultures, and how it could be used to improve human relations. Learn more:<|endoftext|>.45-70 Government Ballistics⏎A complete history with facts, uses, pros, cons, and ballistic information.⏎45-7
-
legislated by the Mexican Federal Government and ruled by the Federal Labor Laws under Article 74. Employees, public and private, are entitled to the day off with or without pay, a fact that I have had conflicting information for. Supposed
-
this activity, but this is a good opportunity to remind them that even an ordinary incident can be retold in an exciting way.⏎The activity involves making a timeline with points they choose from their lives, and then freew
-
Email Based, Online Homework Assignment Help in Classification Of Relays⏎Transtutors is the best place to get answers to all your doubts regarding classification of relays with examples. Transtutors has a vast panel of experienced in classification of relays
-
link to download ebooks to your computer. Orders paid by PayPal require up to 8 business hours to verify payment and release electronic media. For immediate downloads, payment with credit card is required.⏎This math unit presents four activity pages in which students read
-
SHOKÉ,⏎LESSONS IN ELEMENTARY PHY-⏎PHYSIOLOGY. By Sir MICHAEL FOSTER,⏎K.C.B., M.D. With Illustrations. Pott 8vo. 15.⏎Sir MICHA
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 value of the expression $2020^2 - 2020$? Let me think... I don't know what 2020 is, but it's not 2020,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
them throughout the period when other, similar regimes curtailed Jewish rights to curry favor with the Nazis. Still, Salazar didn’t want more Jews settling permanently in his
-
A Comparison Of High-Level And Low-Level Programming Languages⏎A way to categorize programming languages is through high-level or low-level. The low-level languages communicate directly with the central processing unit or the CPU. They can perform very basic commands but are
-
Require school police and related security officers to have training in adolescent development.⏎- Attempt to reconnect alienated youth or students who are at-risk for behavior problems or violence with the mainstream student body. Use threat-assessment procedures to identify those at risk before
-
inscribe their names in the book of life for yet another year. Religious Jewish people work diligently to perform mitzvot (good deeds) to find favor with God during the 10 days of repentance between Rosh Hashanah
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
- L4:882 strongly suppresses 'with' after provide or familiarize
- L2:863 suppresses
- L0:2342 suppresses
- L5:910 excites
- L5:763 suppresses
- L6:2407 suppresses
- L5:28 weakly excites
- L1:1452 weakly excites
- L8:2991 weakly suppresses
- L4:1462 weakly suppresses
- L6:1627 weakly suppresses
- L2:274 weakly excites
- L0:1716 weakly suppresses
- L6:2613 weakly excites
- L7:176 weakly excites
- L5:1694 weakly excites
- L1:2082 weakly excites
- L8:1575 weakly excites
- L6:1184 weakly suppresses
- L3:2840 weakly suppresses
- L5:748 weakly suppresses
- L7:476 weakly suppresses
- L7:1443 weakly suppresses
- L3:2999 weakly suppresses
Feeds later layers
- L16:1860 excites
- L21:1060 excites
- L17:2363 excites
- L14:1049 excites
- L16:2035 excites 'of' after a noun phrase
- L12:2058 excites
- L25:1883 excites
- L17:2463 weakly excites
- L17:2436 weakly excites
- L18:2711 weakly excites
- L10:1839 weakly excites
- L17:2009 weakly excites
- L20:2525 weakly excites
- L14:2290 weakly excites
- L10:886 weakly excites
- L17:2527 weakly suppresses
- L23:1875 weakly excites
- L18:1068 weakly excites
- L14:1478 weakly excites
- L19:2406 weakly excites
- L16:2988 weakly excites
- L17:1728 weakly suppresses
- L15:504 weakly excites
- L18:2123 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
- bearing
- ])**
- 高达
- uncomp
- pied
- partials
- ឥ
- outh
- fines
- 浓厚
Suppresses
- rice
- 哪里
- ично
- rire
- 干嘛
- thane
- cly
- 干什么
- 哪个
- 化进程