'of' after 'that' in comparisons
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. 6 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 'of' when it follows 'that' or a similar word to compare two things, as in 'similar to that of a gas'. Steering did not clearly change the topic, so what it does beyond this comparison habit is not certain.
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 the United States and, until the beginning of the 20th century, was used by nearly all American cities. Its structure is similar to that of the ___2___ and national governments, with an elected mayor as chief of the
-
or plants that has characteristics of both parents; hybrid.⏎- A mule is a cross between a horse and
-
, but in one’s chest, one’s hands, and one’s feet, all together.”⏎Long describes Homer’s overall style to be that of ‘sensuous immediacy’, writing: ‘Homer delights in the vibr
-
Black Brant used in this case was a 4-stage rocket. As it separated the radar pattern matched that of a ballistic missile with multiple re-entry warheads coming down, further convincing the Russians
-
not provide mechanical properties that match that of healthy bone.⏎Nanobiocom developed an intelligent material made of bioactive components
-
The presence of charged ions means that a plasma is highly electrically conductive and responds strongly to magnetic and electric fields. Its behavior is most comparable to that of a gas, as the plasma has no defined volume but instead assumes the volume of
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: the main idea of this problem is to find the area of the region of the set of points in the complex plane such that $z = e^{i k} $, where $k=2$ and the equation $z^2
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
remember, and there is where a curriculum might come in handy. You can organize your lessons by setting up a program for your studies yourself, or acquiring all the necessary information from websites like https://growcurriculum.org/ and customizing it to
-
responsible adult is present in the home when the homebound teacher is working with the student. The parents must also make every effort to keep homebound instruction appointments as scheduled.<|endoftext|>“Anyone
-
disabilities. With almost 85% of adults and 96% of children visiting a medical professional each year, healthcare providers should be better equipped with the necessary knowledge and skills to provide high-quality healthcare to all
-
who would betray Him. He had already sealed the deal in verses 14-16. Judas asked Christ<|endoftext|>Treating Developmentally Disabled With Dual Diagnosis of Dementia or Alzheimer’s Disease⏎Alzheimer’s Disease is a leading cause
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
- L5:2517 weakly suppresses
- L5:1732 weakly excites
- L3:1741 weakly excites
- L4:149 weakly suppresses
- L1:2143 weakly suppresses
- L4:2987 weakly suppresses
- L4:1375 weakly excites
- L3:1992 weakly excites
- L4:2180 weakly excites
- L3:2147 weakly excites
- L1:1092 weakly excites
- L5:2714 weakly suppresses
- L2:2229 weakly suppresses
- L2:2307 weakly suppresses
- L3:189 weakly excites
- L4:1034 weakly suppresses
- L2:2587 weakly suppresses
- L4:1330 weakly suppresses
- L1:1788 weakly excites
- L4:2922 weakly excites
- L5:2719 weakly excites
- L4:1210 weakly excites
- L5:1601 weakly excites
- L4:1484 weakly excites
Feeds later layers
- L15:1496 excites
- L12:1915 excites
- L13:2812 excites
- L14:2163 excites
- L9:1995 excites
- L21:2666 weakly excites
- L14:1795 weakly suppresses
- L14:2243 weakly excites
- L11:2109 weakly excites
- L9:264 weakly excites
- L13:1954 weakly excites
- L11:2336 weakly excites
- L12:1200 weakly excites
- L15:2730 weakly excites
- L12:1518 weakly excites
- L15:570 weakly excites
- L9:1558 weakly excites
- L15:3053 weakly suppresses
- L14:1013 weakly excites
- L11:1040 weakly suppresses
- L12:2692 weakly excites
- L9:2332 weakly excites
- L15:2219 weakly suppresses
- L12:1395 weakly excites 'as' in 'as ... as' comparisons
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
- tridge
- Bail
- attribute
- 需要用到
- recognizable
- attributes
- require
- bail
- characteristic
- ируется
Suppresses
- anje
- менно
- 不然
- WAY
- راف
- andom
- 增
- affles
- kiệm
- jeta