'not' or 'only' setting up a contrast
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. 7 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 words like "not" or "only" that come right after a verb such as "is" or "are" and set up a contrast, as in "they are not effective" or "though not exclusively". Its negative side fires on words like "typically" and "also" in similar spots. Turning it up made the model write rambling, repetitive text that did not clearly relate to the topic.
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.
-
source to generate energy. This is known as the Warburg Effect.⏎While metabolic changes are an important feature in the transformation of normal cells into cancer cells they are not now thought to be cancer’s primary cause. Despite this, metabolic changes remain an
-
899 and has been used with success in chemical, ceramic, carbon allotropy, steel/alloy, composite materials and plastic industries for decades, it was only in late 1980′s that its commercial benefits became available to
-
the novel which is also an alternative history–this one, The Grasshopper Lies Heavy, portraying a world in which the Allies won World War II, though not exactly along the lines of history as we know it. Despite the book being banned
-
authorities were suspicious, and added up the total of his deposits. They found that he had deposited $162,<|endoftext|>An intelligent and impassioned, albeit not
-
host of health benefits, including weight loss, improved lung function, lower risk of stroke, cancer and heart disease.⏎But just because the humble apple is ubiquitous doesn't mean it's without some mystery as well. Read on to learn more about the
-
Until now, semi-conductors have been the main tool for waste heat recovery, but they are not effective when temperatures drop below 100 degrees Celsius. Semi-conductors
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 am not a fan of the words. I am not a fan of the words. So, the words are: **The words are:** **"I am not a fan of the words"** I am not a fan of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. The Stabilizer level of swimming pool water must be maintained at 30-50 ppm. The Stabilizer level is important, but is typically only tested at the beginning of the swimming season when a pool is opened. The
-
-related hormones, but research shows it also improves stamina and relieves fatigue. Eleuthero, often called Siberian g
-
it's driven by "thermal and geophysical disturbances in the inner Earth", or links between tectonic plates and the climate.⏎But another is that the Earth passes through the plane of the Milky Way galaxy every 3
-
to 30 grams of protein at each meal to optimize muscle protein synthesis. Though most people eat a combination of protein-containing foods at each meal, this infographic shows what 25 grams of protein from different sources looks like.⏎Researchers in Ireland
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
- L0:914 weakly suppresses
- L5:1434 weakly suppresses
- L1:1018 weakly suppresses
- L5:2324 weakly excites
- L3:2698 weakly suppresses
- L1:1806 weakly excites
- L6:824 weakly excites
- L3:1410 weakly excites
- L6:1917 weakly excites
- L4:2090 weakly excites
- L2:1054 weakly excites
- L6:1054 weakly excites
- L2:1331 weakly excites
- L1:2294 weakly excites
- L6:2425 weakly suppresses
- L1:731 weakly suppresses
- L1:1947 weakly excites
- L5:1287 weakly suppresses
- L5:2029 weakly suppresses
- L2:2809 weakly excites
- L3:695 weakly suppresses
- L0:367 weakly suppresses
- L3:1808 weakly suppresses
- L4:659 weakly excites
Feeds later layers
- L11:2178 suppresses
- L8:2777 suppresses
- L9:118 excites
- L8:2298 suppresses
- L8:1944 suppresses
- L12:892 suppresses
- L9:1468 weakly suppresses "however" as a mid-sentence aside
- L8:142 weakly suppresses
- L13:1595 weakly excites
- L8:2707 weakly suppresses
- L8:2513 weakly suppresses
- L20:1821 weakly suppresses
- L14:396 weakly excites
- L22:2343 weakly excites
- L15:137 weakly suppresses
- L9:2714 weakly suppresses
- L19:2693 weakly suppresses
- L16:1923 weakly excites
- L11:2941 weakly suppresses
- L16:3000 weakly excites
- L13:742 weakly suppresses
- L9:2284 weakly suppresses
- L19:2359 weakly suppresses
- L16:2212 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
- typeid
- modo
- ervised
- ditor
- AdapterManager
- ��
- edition
- MessageType
- 亚
- ferences
Suppresses
- iz
- ht
- ak
- not
- OK
- 혁
- 北
- nut
- ACK
- fact