spot where a number is due
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. 25 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 switches on just before a number is about to appear, such as a percentage, a price, a calorie count or a year-like quantity. When it was turned up, the model's text drifted into odd repeated fragments, so the steering tells us little about it.
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.
-
a whole. Unfortunately, one hundred years later, women still make up a majority of the world’s poor.⏎Women’s wages in the U.S. are only 76% of men’s, but the disparity of long-term earnings between men
-
12th Social Studies & History⏎Anticipating the State of the Union, January 2010⏎Designed to precede President Obama’s 2
-
helps maintain a healthy digestive system and promotes regular bowel movements. Eating a 100-gram serving of tamales introduces 6.28
-
and orchards are the best choices for optimal nutrients, energy and vitality. Fresh fruits, vegetables, nuts, seeds, whole grains and legumes should ideally constitute 60-70% of your diet. Although processed food has an undeniable
-
greater level of experience, and less time spent raising children. When President John F. Kennedy signed the Equal Pay Act into law, women earned, on average, 60 cents for every dollar earned by men. In the forty-eight years that
-
If you're using a calorie counter, remember that Fat, Carbs, and Protein calories are just close estimations based on the Atwater factors:⏎Fat: 9 cal/g Carb: 4 cal/g Protein: 4 cal/g⏎
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: Iasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasasas
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
blood vessels (pyelonephritis, glomerulonephritis, amyloidosis, heart failure, renal vascular anomalies).⏎Reduction of
-
care provider checks for the disorder by using techniques called amniocentesis (pronounced am-nee-oh-sen-TEE-sis) or chorionic (pronounced kawr-ee-ON-ik) villus sampling
-
alopathy-PML. Updated January 8, 2016. Accessed September 26, 2016.⏎Progressive multifocal leukoencephalopathy (PML). AETC National Resource Center website
-
The best way to get rid of H. pylori is to use triple-therapy for 10 to 14 days. Usually antibiotics such as clarithromycin, amoxicillin and
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
- L21:339 excites spot just before a year
- L19:581 suppresses blank slot before a number
- L23:77 excites the space right before a number
- L18:377 excites number right after 'about' or 'less than'
- L22:372 excites empty spot before a number or "3D"
- L22:824 excites blank before an expected number
- L4:53 excites blank slot before a number
- L17:170 excites
- L20:436 suppresses
- L8:111 suppresses the space before a year in dates
- L17:285 weakly suppresses
- L21:613 weakly suppresses
- L22:216 weakly excites
- L18:93 weakly suppresses
- L16:314 weakly excites
- L13:923 weakly suppresses word right before a number or year
- L0:2567 weakly suppresses
- L4:1635 weakly suppresses 'light' as an adjective (pale, not heavy)
- L14:616 weakly suppresses
- L13:1372 weakly excites number before an age or week count
- L18:91 weakly suppresses
- L14:196 weakly excites
- L1:2159 weakly suppresses
- L18:203 weakly suppresses
Feeds later layers
- L26:103 excites blank space before a number
- L27:170 excites unclear: spaces in mixed-language text
- L27:283 excites the spot right before a number
- L26:15 suppresses
- L27:220 excites end-of-text token between documents
- L26:121 excites
- L27:48 weakly suppresses
- L26:216 weakly excites blank slot before a year or number
- L26:423 weakly suppresses
- L26:145 weakly excites
- L27:183 weakly suppresses
- L27:2249 weakly suppresses
- L26:97 weakly suppresses
- L27:265 weakly excites
- L26:172 weakly excites
- L26:769 weakly suppresses
- L26:17 weakly suppresses
- L27:238 weakly excites
- L26:224 weakly suppresses
- L27:1025 weakly suppresses
- L26:168 weakly excites
- L27:2604 weakly suppresses
- L27:317 weakly excites
- L27:276 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
- iced
- ionic
- ersions
- icens
- unted
- erts
- erti
- umber
- ices
- uther
Suppresses
- NIL
- yl
- ⏎→→⏎
- rag
- miss
- eid
- ={()=>
- ⏎→⏎
- mise
- corev