'di' syllable in diabetes and diet 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. 25 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
This neuron reacts most strongly to the 'di' piece at the start of medical and diet words, such as diuretics, diabetes, antidiabetic and dietician. It also fires on a few similar word pieces. Turning it up or down did not clearly change the topic of the writing.
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.
-
throw up the food just eaten.⏎- People with anorexia could be conscious of eating in the presence of others too.⏎- Bank heavily on pills such as diuretics (increased urination to
-
soft drinks at all.⏎Though they haven’t yet identified the science behind the statistics, they say the link persists regardless of other risk factors like exercise, smoking ,diabetes, weight, and alcohol consumption, but they haven’t discovered the science behind
-
for centuries, and modern research has found that they may have a range of health benefits. For example, some studies have suggested that fenugreek may have anti-diabetic
-
considered more precious than gold.⏎Currently, this aromatic spice has been the subject of numerous studies where its compounds are analyzed, they have antioxidant, anti-inflammatory, antidiabetic, and antimicrobial properties that could offer protection against multiple diseases. Among the
-
vegetables are good sources of vitamins and minerals so it’s good to include a portion at breakfast, try a banana or fruit juice.⏎Breakfast can be good for dieters too, research has shown those who eat breakfast are less likely to be overweight
-
and cold symptoms. Ginger tea also contains anti-histamine properties that prevent airway contraction and helps treat respiratory allergies.⏎4}
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: if a city is connected with a once-cursored four-cursored city once again, then a city once again connected once again with a once-cursored city once again, and so on, then a city once again connected once
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
: 45–60 days⏎Native: Widespread, probably Asia⏎Hardiness: Frost-tolerant annual⏎Plant Dimensions: 16"–22" tall⏎Variety Information: Large, dark green oval
-
a carnelian cherry dogwood (Cornus mas) cultivar, tolerates a range of moist soils, including well-drained clay. Standing 15 to 25 feet high, with a 12- to 1
-
. The flowering stalk may stand up to 3 ft (1 m) tall. The sword shaped leaves are mostly basal, in two ranks, and about 2 ft (60 cm) long. German iris spreads by rhizomes;
-
will tolerate part shade.⏎To get the highest yield of tender and flavorful leaves, pinch the tip of each branch, starting in early summer when the plants are 6" tall, to encourage bushiness. However, if you want to enjoy the
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
- L19:877 weakly excites
- L15:184 weakly excites
- L10:250 weakly excites
- L8:736 weakly excites
- L0:1925 weakly excites
- L8:2791 weakly suppresses
- L20:1251 weakly excites
- L10:1711 weakly suppresses
- L22:2891 weakly excites
- L11:2870 weakly suppresses
- L22:1689 weakly suppresses
- L18:567 weakly excites
- L18:607 weakly excites
- L13:2851 weakly excites
- L5:755 weakly suppresses
- L17:1469 weakly excites
- L19:1851 weakly suppresses
- L4:740 weakly excites
- L17:333 weakly excites
- L6:1327 weakly excites
- L10:2966 weakly excites
- L22:535 weakly suppresses
- L13:2683 weakly suppresses
- L16:1157 weakly suppresses
Feeds later layers
- L26:1968 weakly suppresses
- L27:1962 weakly excites
- L26:2906 weakly suppresses
- L26:2336 weakly excites
- L27:2711 weakly excites
- L27:178 weakly excites
- L27:912 weakly suppresses
- L26:2355 weakly suppresses
- L26:792 weakly suppresses
- L26:2202 weakly suppresses
- L26:2176 weakly suppresses
- L26:1481 weakly excites
- L26:1645 weakly suppresses
- L26:1721 weakly excites
- L27:1641 weakly excites
- L26:2792 weakly suppresses
- L26:1812 weakly excites
- L27:597 weakly suppresses
- L27:2025 weakly excites
- L27:752 weakly suppresses
- L27:1823 weakly suppresses
- L26:2627 weakly excites
- L27:1557 weakly suppresses
- L27:2037 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
- hum
- ocese
- abs
- ously
- textView
- ellig
- wheel
- Abs
- vio
- izontally
Suppresses
- uhe
- uler
- sigue
- urs
- ULER
- owości
- 愍
- uled
- uhan
- ��