'of' after a person noun
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. 15 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 'of' when it follows a word for a person, like 'a friend of', 'a comrade of' or 'cousin of'. Turning it up made the model repeat 'of mine', 'of yours' and 'of hers', so it seems to help the model use possessive 'of' phrases.
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.
-
. Whenever he noticed a peasant beating his horse, Bergh would order the peasant to stop beating it at once.⏎On his way home from Russia, this friend of animals stopped in England, where he observed the work of the Royal Society for the
-
river and done trading until that Regiment was relieved by a South Carolina Regiment. The next morning as some of our men went to wash as usual, a comrade of Co. G was shot in the leg which was a signal for hostilities when
-
arring exists on this skin of the leg. Generally, weaker chemicals are used for the legs as they tend to
-
friend of Jesus who raised him from the dead (John, Chapters 11 &
-
friend of Allah, and one who does not work is considered by Allah, to be His
-
, France was plunged into a succession crisis over which distant cousin of the king would inherit the throne. Militarily and diplomatically, Francis'
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: It's not a bad idea to get a new idea in a new way to help help help...?... It... It... it... it... it... it... it... it... it... it... it... it... it...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document⏎Unformatted text preview: itself, and if ( ) ( ) lim lim x c x c h x
-
2. Sign up to view the full content.⏎This preview has intentionally blurred sections. Sign up to view the full version.View Full Document⏎Unformatted text preview: 60 | | | | CHAPTER 1 FUNCTIONS AND MODELS Not all
-
concrete permeability⏎* Increases concrete strength⏎* Improves resistance to corrosion<|endoftext|>Prometheus shines brightly in this image, taken as part of the ongoing campaign to
-
dots” in a real-life political, often international scenario to create a plan of action.⏎You may be familiar with the word מוֹדִיעִין, also referring to intelligence (as well as to a city situated
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
- L14:1807 excites
- L10:39 weakly excites
- L4:619 weakly excites
- L8:1100 weakly excites
- L11:199 weakly excites
- L5:2217 weakly suppresses
- L3:1110 weakly suppresses
- L4:2319 weakly excites
- L9:17 weakly excites
- L0:707 weakly excites
- L2:370 weakly excites
- L7:130 weakly suppresses
- L13:122 weakly excites
- L9:2202 weakly suppresses
- L14:943 weakly suppresses
- L4:705 weakly suppresses
- L8:10 weakly excites
- L2:2022 weakly excites
- L14:2034 weakly excites
- L0:2533 weakly excites
- L0:1754 weakly excites
- L4:2068 weakly suppresses
- L8:14 weakly excites
- L3:1912 weakly suppresses
Feeds later layers
- L17:2185 excites
- L16:961 excites
- L17:2736 excites
- L17:1771 excites
- L26:711 excites 'years' after an age number
- L19:2304 excites
- L27:447 excites
- L20:450 excites
- L25:2020 excites
- L22:142 suppresses 'of' after a quantity or noun phrase
- L23:567 excites
- L27:658 weakly excites
- L18:953 weakly excites
- L23:258 weakly excites
- L22:1993 weakly excites
- L17:488 weakly excites
- L18:2510 weakly excites
- L18:2693 weakly excites
- L24:2771 weakly excites
- L18:636 weakly suppresses
- L24:1650 weakly suppresses
- L19:3054 weakly excites
- L22:2457 weakly excites
- L22:2697 weakly suppresses
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
- mine
- hers
- ours
- of
- theirs
- yours
- mine
- của
- Mine
- _of
Suppresses
- FileType
- 不限
- 作为一个
- 是一个
- глав
- Wrapped
- фи
- 落到
- ancode
- entials