the letter r in the middle of words
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. 27 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 letter r in the middle of a word, such as the r in Avranches, Citrullinemia, midrashic, paraphrased and lacrimation. When turned up, the model drifted into odd loops like 'everything everything', which does not match the letter pattern.
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.
-
Norman cathedral. It is the seat of the Bishop of Coutances and Avranches, previously the Bishops of Coutances. Standing 80 metres (
-
button to use the lights. A specimen can be tested under multiple lights at the same time. The instrument offers highly accurate and reliable test results<|endoftext|>What is Citrullinemia Type 1?⏎Citrullinemia type I is a
-
Mary Potter Engel⏎Midrash refers to the rabbinic tradition of interpretation that explores scripture beyond its literal meanings. The verse by verse interpretations generated by midrashic commentary bind the Talmud,
-
"Just a moment," and turned to the subject and repeated the same request. If the subject asked for additional information, the experimenter repeated, or paraphrased the request. If the subject persisted, the experimenter said
-
.⏎tAsAM ruditashabdena gavAM haMbhAraveNa cha |⏎vrajasyotthApanaM chakrurghoShavR^iddhAH samAgatAH ||2
-
Llangynhafal, Llanychan, Llanbedr D.C. and Llanrhydd. Llannerch contained the parishes of Llanfair D
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:icicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicicic
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
The phase-difference can, therefore, be defined as ΔΦ=ΦV-ΦI.⏎We define the instantaneous power similarly to the DC power by P(t)=V(t)×I(t). When using
-
power frequency in the United States), calculate impedances for all components and determine all voltage drops using Ohm's Law (E=IZ ; I=E/Z ; Z=E/I). The polar magnitudes of the results should closely agree with your vol
-
for all the areas.⏎To address this problem, we just need to recreate categories for all the features. First, delete the single category from them.⏎v.category input=gadm36_d
-
="500"> <polygon points="200,10 250,190 160,210" style="fill:lime;stroke:yellow;stroke-width:1"/> </svg
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
- L21:3031 excites
- L25:459 suppresses
- L20:777 suppresses
- L25:252 excites
- L23:2074 suppresses
- L19:89 excites
- L2:202 excites "of" after "many" or "which"
- L0:117 suppresses
- L23:233 suppresses
- L23:310 excites
- L22:1753 suppresses
- L20:895 excites
- L22:1306 excites
- L8:3 excites
- L8:7 excites
- L20:747 excites
- L18:569 excites
- L24:938 suppresses
- L26:606 suppresses
- L26:1564 excites
- L16:41 suppresses
- L25:956 excites
- L8:16 weakly suppresses
- L8:14 weakly excites
Feeds later layers
None: this is the last layer.
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
- =(
- ={{
- =ax
- =F
- =\"
- ={
- ="
- =[],
- ]="
- =s
Suppresses
- ic
- icit
- am
- icos
- onne
- ons
- ame
- icot
- icol
- ов