repeated words and ellipsis dots
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. 14 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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
It reacts to a word or punctuation mark that repeats or stutters, like 'very, very' or the dots in '. . .'. Turning it up did not produce a clear matching theme in the model's writing, and turning it down caused repetition.
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.
-
agreed to assume claims by citizens of the United States against Spain. In Article III, the treaty stated that: "The boundary line between the two countries. . . shall begin on the Gulph [sic] of Mexico, at the mouth of
-
, 2003, the National⏎Transportation Safety Board (NTSB) made its boldest statement to date.⏎They stated that “ . . . the safe operation of fifteen-passenger⏎vans requires a knowledge and skill level
-
some 44,000 generations, equivalent to some million years of generations of a human population (but the evolutionary opportunities for humans would be far, far less, due to the small population numbers limiting the number of mutational possibilities;
-
is defined as to:⏎“engage in a course of vexatious conduct or to make a demand or a sexual solicitation or advance that one knows or ought reasonably to know is unwelcome.”⏎Severe Act or Pattern of Actions⏎Let’s
-
! How changed she was! How very, very different she looked from the last time she had seen herself in a big mirror!
-
in another(),so this prints 20. Thus final answer is 10 0 20. I hope i have been clear enough. ------------------ Come on in !! Drinks are on the house in the Big Moose Saloon !
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 have been thinking about the need to be more engaged, and I have been thinking about the need to be more engaged, but I have been thinking about the, I have been, a, I, I, an, I, I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in business.⏎This week’s article by Jeremy Rosen on Rav Lichtenstein is available from http://jeremyrosen.blogspot.co.uk/2015/08/rav-lichtenstein-zl.html⏎Your comment will be posted
-
the issue of Southern autonomy.⏎In 1924, the 28th president of the United States, Woodrow Wilson, died in Washington, D.C., at age 67.⏎In 1930, the
-
NR in decibels. From Step 1, we have the decibel value LdB as LdB = 10 log10 (P1/P0). Now let P1 = MAX^2 and P0 = MSE.
-
Pesticides in Produce⏎Researchers Highlight Pesticides in Produce,⏎Baby Food, Tap Water⏎Washington, D.C. – Environmental Working Group has released the eighth edition of its Shopper’s
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
- L9:7 excites
- L13:1006 excites
- L10:33 suppresses
- L13:755 suppresses
- L13:351 weakly suppresses
- L11:173 weakly suppresses
- L6:4 weakly suppresses
- L10:527 weakly excites
- L13:1401 weakly suppresses
- L13:1428 weakly excites
- L13:494 weakly excites
- L6:0 weakly excites
- L13:2687 weakly excites
- L10:3035 weakly suppresses
- L0:700 weakly suppresses
- L10:394 weakly excites
- L3:3 weakly suppresses
- L9:746 weakly suppresses
- L10:1216 weakly suppresses
- L1:1 weakly excites
- L5:394 weakly suppresses
- L5:872 weakly suppresses 'or' introducing an alternate name
- L11:966 weakly excites
- L13:287 weakly excites
Feeds later layers
- L17:1590 suppresses
- L22:269 weakly suppresses
- L20:3 weakly excites
- L18:501 weakly suppresses
- L18:574 weakly excites
- L15:78 weakly excites
- L19:1747 weakly suppresses
- L19:1058 weakly excites
- L25:1668 weakly suppresses
- L17:560 weakly suppresses
- L26:274 weakly excites
- L16:1943 weakly excites
- L17:2893 weakly suppresses
- L18:2055 weakly excites
- L17:2977 weakly suppresses
- L17:408 weakly suppresses
- L21:2623 weakly suppresses
- L27:1748 weakly suppresses
- L15:122 weakly excites
- L21:2616 weakly suppresses
- L16:1639 weakly suppresses
- L20:1526 weakly suppresses
- L18:546 weakly suppresses
- L17:1622 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
- bler
- ям
- 两者
- 者
- .TryParse
- (kind
- 研究所
- 的是
- 两种
- лась
Suppresses
- ainers
- -ahead
- GridColumn
- akening
- Ahead
- .nextSibling
- 玒
- AppBundle
- 鲠
- Posts