'in' after high or low
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. 2 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 fires on the word 'in' when it follows 'high' or 'low', as in 'high in fiber' or 'low in fat'. It tends to go quiet on question words like 'when' followed by a verb. Turning it up did not produce any clear change in 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.
-
eat celery! Celery contains a lot of vitamins and is high in fiber, therefore it is safe for consumption and can benefit your cat. Celery
-
broth, ginger ale, and saltine crackers, also soothe the stomach.⏎- Avoid eating foods that are high in fat and salt.⏎- Try to eat before you get hungry and before nausea occurs
-
blood pressure [*].⏎Focus on foods high in potassium too. Potassium is an electrolyte needed to balance out too much sodium
-
fiber to create the proper stool consistency. If your stools are too hard, add fluid; if too loose, add fiber.⏎There are many foods that are high in fiber per serving, such as beets, beans, peas, sweet potatoes,
-
but because you can never collect too many.<|endoftext|>Scientists have discovered that a diet which is high in salt, can have major effects on the brain and in mice, can lead to
-
of plant protein and risk of death (from any cause) - the more plant protein in the diet, the lower the risk of death.⏎What foods are high in plant protein? Legumes (beans,lentils,
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 this for a long time" (so long) as many times: (so many times: so long: so long: so long: so long: so long: so long: so long: so long
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
trends. Second, discuss how this data may be used by hospital systems, medical practitioners, and government agencies.<|endoftext|>Cyberspace Computers and Cyberspace When does reality end and cyberspace begin? Can crimes committed in cyberspace be
-
What is your name?: saha wasta na anjeun?|⏎|Where do you live?: timana anjeun kantun?⏎|When can we meet?: timana tiasa anjeuna sadaya papangg
-
celebrations responds to the first questions⏎posed by the faithful regarding this subject:⏎- Who celebrates the liturgy?⏎- How is the liturgy celebrated?⏎- When is the liturgy celebrated?⏎- Where is the liturgy celebrated?<|endoftext|>A particle
-
?⏎(b) In what quantity?⏎(c) At what price?⏎(d) When do they want (it)?⏎(e) What kind of advertisement do they like?⏎(f
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
- L1:148 suppresses
- L1:1695 weakly suppresses
- L0:2301 weakly excites
- L0:960 weakly suppresses
- L1:2063 weakly excites
- L1:1388 weakly excites
- L1:505 weakly suppresses
- L1:170 weakly excites
- L1:1919 weakly suppresses
- L0:2406 weakly excites
- L1:2128 weakly suppresses
- L1:1941 weakly excites
- L1:2689 weakly excites
- L1:92 weakly excites
- L1:1108 weakly excites
- L1:1756 weakly excites
- L1:1371 weakly suppresses
- L1:347 weakly suppresses
- L0:781 weakly suppresses
- L0:2625 weakly excites
- L1:2660 weakly excites
- L0:1879 weakly suppresses
- L0:1566 weakly suppresses
- L1:2533 weakly excites
Feeds later layers
- L16:583 suppresses
- L5:55 suppresses
- L20:1387 suppresses
- L17:923 suppresses
- L21:1557 suppresses
- L4:2725 suppresses
- L19:1929 suppresses timestamps and dates in text
- L17:1662 suppresses
- L18:1505 suppresses
- L4:1834 suppresses unclear: time spans and 'once you've'
- L7:451 suppresses
- L19:2799 suppresses
- L9:619 suppresses
- L16:1504 suppresses
- L6:213 suppresses
- L18:2750 suppresses
- L10:1212 suppresses
- L4:2784 suppresses
- L18:1933 weakly suppresses
- L10:591 weakly suppresses
- L9:383 weakly suppresses
- L8:1749 weakly suppresses
- L8:1916 weakly suppresses
- L4:1093 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
- erve
- tá
- itori
- position
- степ
- strncmp
- 夷
- 阶级
- Sizer
- EFI
Suppresses
- icals
- repetition
- uin
- /action
- ,ID
- owo
- ewis
- ;p
- <IEnumerable
- curring