smell and taste words in health text
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. 5 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 lights up on words for smells, tastes and colours of things in health and medical writing, like 'odour', 'smelling', 'metallic taste' and 'opacities'. Turning it up made the model write about feeling unwell and symptoms, which fits that 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.
-
ur-producing bacteria that live on the surface of the gums, tongue and throat. When these bacteria in the mouth break down protein, they release a sulphurous odour – causing bad breath. Improperly-digested food, known as ama
-
cold intolerance⏎- Disordered appetite, increase or decrease in weight⏎- Increased thirst or urination⏎- Loss of sex drive⏎- Bloody or foul-smelling discharge⏎- Acne or abnormal hair growth⏎- Visual difficulties⏎- Imp
-
era) and the maripa palm (Attalea maripa). Palm mesocarp oil which is 49% saturated, is generally reddish in colour because of high beta
-
West, but across the continent.⏎We know that statements<|endoftext|>Cataracts are, unfortunately, one of the most common eye problems in pets. They are opacities in the lens of the eye and can affect any age, species, or
-
to a doctor to get the treatment you need.⏎Does your urine have a sweet-smelling stench? As Amy Krambeck, a specialist at the Mayo Urinary
-
notice a strong urine smell. First, think about the amount of water you’ve drunk in the last couple
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 don't think that the idea of a place where the sun and the moon are separate. But I don't accept that the sun and the moon have their own day. So, 1:1:1:1:1:
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
?⏎: a characteristically anonymous, timeless, and placeless tale circulated orally among a people.⏎Is Cinderella a folktale or fairy tale?⏎Cinderella is considered a fairy tale, not a folk tale. Folk tales have a basis
-
, does it?"⏎Ryan started to wonder about how the universe began, asking Lisle, "What about the Big Bang?"⏎Lisle answered, "The Big Bang teaches<|endoftext|>Fig.1 above refer to a point moving along a circular path.
-
implies that the universe is...?⏎- Decreasing in Entropy⏎4. The reason galaxies move apart is because of...?⏎- The force of the big bang which caused the galaxies to move away at a velocity related to their size.⏎-
-
Whitewright in 1891⏎Like many other Texas cities, Whitewright owes its founding to the arrival of the railroad and the westward movement of
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
- L4:980 weakly excites
- L3:2452 weakly suppresses
- L4:2344 weakly excites
- L3:2032 weakly suppresses
- L0:2609 weakly suppresses
- L4:1352 weakly excites
- L4:783 weakly suppresses
- L4:2115 weakly excites
- L2:728 weakly suppresses
- L3:2391 weakly suppresses
- L1:394 weakly excites
- L1:2296 weakly excites
- L3:2663 weakly suppresses
- L2:695 weakly excites
- L4:1410 weakly excites
- L2:2771 weakly suppresses
- L4:1172 weakly excites
- L3:1235 weakly suppresses
- L4:2797 weakly excites
- L4:2565 weakly excites
- L4:1405 weakly excites
- L4:2893 weakly suppresses the word 'means' after a subject
- L4:2767 weakly suppresses
- L4:2856 weakly excites
Feeds later layers
- L6:2161 excites
- L9:3021 weakly suppresses
- L8:3022 weakly excites
- L6:2725 weakly excites
- L16:2644 weakly excites
- L8:2082 weakly excites
- L11:1675 weakly suppresses
- L9:466 weakly suppresses
- L7:1679 weakly excites
- L10:1084 weakly excites
- L12:2388 weakly suppresses
- L7:1087 weakly suppresses
- L10:546 weakly excites
- L15:357 weakly excites
- L21:2796 weakly suppresses
- L6:2449 weakly excites
- L9:244 weakly excites
- L7:2185 weakly suppresses
- L19:1982 weakly excites
- L7:1646 weakly excites
- L9:921 weakly excites
- L12:1296 weakly suppresses
- L10:1231 weakly suppresses
- L9:783 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
- ตาร
- ignant
- ripp
- ')[
- еть
- aup
- $',
- ,’
- .'.
- pery
Suppresses
- sciences
- compos
- rost
- 一次次
- astronom
- ⎝
- 天赋
- 人生
- 医科
- prolong