unclear: mid-phrase function 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. 3 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 small words in the middle of a sentence, like 'a', 'more', 'in' or a comma, where the sentence is still unfinished. No single clear topic stands out, and turning it up did not produce a clear change in what the model wrote.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
P and Q are the feet⏎of the perpendiculars from X to the sides of a right angled⏎triangle. What position of X makes the length of PQ a minimum?⏎The circumcentres of four triangles are joined to form a⏎qu
-
all of the organelles and their functions. One fun—and tasty!—way to make the parts of the cell more concrete for students is to create edible cells. Using materials such as Jell-O
-
health consequences of teen sex, using measures of self-esteem and depression. The results of her work, published last May in The American Journal of Sociology, put her in the media spotlight. The New York Times showcased her study in a June article.⏎
-
. Throughout the years, the panels were also painted to cover discrepancies; two of them, including Romeo and Juliet, had faux finishes applied. Conservators concluded that all layers
-
1-2 hours depending on the size of your jar. A full melt pool will reach a depth of 1/2 inch. This will allow the fragrance to disperse more completely into the room. (Remember burn your candles for a maximum
-
, mind maps can make the job easier and less time-consuming.⏎Moreover, mind maps simplify problem-solving and planning. They
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 need to be more specific, and, I need to be more practical, and, I need to be more clear. Let me think. It's a bit a bit a bit a bit a bit a bit a bit a bit a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
25 to specifically target drivers in this age group.⏎* Since 1995, more than 400,000 young adults have learned life-saving defensive driving skills through DDC-Alive at 25.⏎
-
Nearly 7,000 people have been displaced and more than 5,000 homes destroyed, damaged or
-
e’s team demonstrated its principle with three lasers, but such crystals could store holograms created by up to a thousand lasers, which will allow more elaborate structures to be built. The crystals can also hold many holograms at the same time, each
-
weight loss by making the body absorb fewer calories from other foods. Studies show that doubling your daily fiber intake can result in up to 130 fewer calories being absorbed. You can add fiber to your food by making little changes to your diet
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
- L1:454 excites
- L0:150 suppresses
- L2:837 excites
- L2:179 excites comma after a clause subject
- L2:307 suppresses
- L2:63 excites
- L2:1205 excites
- L1:604 excites
- L2:339 suppresses
- L2:104 weakly excites
- L1:940 weakly suppresses
- L2:325 weakly excites
- L0:896 weakly suppresses
- L1:969 weakly suppresses
- L2:93 weakly suppresses
- L2:501 weakly excites
- L2:1243 weakly excites
- L2:80 weakly excites
- L1:136 weakly suppresses
- L0:632 weakly excites
- L2:185 weakly excites
- L1:959 weakly excites
- L2:219 weakly excites
- L2:747 weakly excites
Feeds later layers
- L4:211 excites
- L4:44 excites
- L4:161 suppresses
- L5:460 suppresses
- L6:335 suppresses
- L4:462 excites
- L5:1425 suppresses
- L10:724 excites
- L4:371 suppresses
- L6:1579 suppresses
- L6:2058 suppresses
- L8:197 suppresses
- L18:729 excites
- L4:374 suppresses
- L5:1371 excites
- L14:591 suppresses
- L8:444 excites
- L8:183 excites
- L5:56 weakly suppresses
- L7:386 weakly suppresses
- L5:131 weakly excites
- L4:80 weakly excites
- L6:342 weakly suppresses
- L6:2704 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
- 一大
- cabin
- aturas
- Cabin
- ieurs
- swith
- actors
- ()].
- porto
- apos
Suppresses
- Solution
- сай
- AW
- vê
- AYOUT
- b
- B
- لق
- 嘬
- ATA