unclear: mixed, health lists and 'weight'
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. 9 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 a jumble of things, most often words in lists of symptoms or side effects, such as 'weight' in 'abdominal pain, weight loss'. It goes quiet, or dips well below zero, on words about pressure and conditions in science experiments. Turning it up made the model write about temperature and pressure, but the other outputs had no clear shared theme.
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.
-
common name like Devil’s Ivy, you might wonder are pothos poisonous?⏎Are Pothos poisonous? Yes, Pothos plants are poisonous if ingested and can be irritant to the skin if touched. They are not known to
-
, pale stools, dark urine, itchiness. Accompanying symptoms of jaundice resulting from low bilirubin levels include fatigue, abdominal pain, weight loss, vomiting, fever, pale stools, dark urine.⏎Complications:<|endoftext|>
-
eating disorder is a terrifying thing to go through, whether you are witnessing it or dealing with it yourself. Eating disorders are often characterized by obsessions over food, weight, or appearance or anxiety surrounding food. These obsessions and anxieties over food
-
expired meat products, smoked fish and ready-prepared salads can cause a potential health hazard. It is always better to throw the foods which are past their ‘use-by’ dates.⏎Now that you know how to maintain your fridge, let’s talk
-
triggers for hives. While it varies from individuals, these may include:⏎- Allergens (e.g., foods, pollens, animals, insect stings).⏎- Irritants (e.g., cosmetics)⏎- Autoimmune diseases⏎
-
lily like the Easter or Asiatic lily, but is still toxic to dogs, humans and other animals. If your dog chews on the plant or eats pieces of it, he can become seriously ill and should receive immediate veterinary treatment.⏎
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: the temperature inside a star at the moment of the explosion, and the pressure inside a typical atmosphere of a star. Calculate the temperature in Kelvin and pressure in Pascals if you get the temperature of 1.5e6 kPa
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
structure. It was thought that perovskite didn't change structure over the enormous range of pressures and temperatures spanning the lower mantle.⏎Recent experiments that simulate the conditions of the lower mantle using laser-heated diamond an
-
was recorded in the Large Hadron Collider; where scientists smashed lead ions, creating a quark-gluon plasma to simulate conditions at the start of the universe. Maintaining plasma is a fine art; it
-
and 2900 kilometers in depth, as defined by the so-called transition zone (top) and the core-mantle boundary (below). Pressures in the lower mantle start at 237,000
-
and are composed of pure electron, ion, positron, or antiproton plasmas.⏎Plasmas commonly form by heating a gas to searing temperatures
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
- L5:132 excites
- L6:1982 weakly suppresses
- L8:1227 weakly excites 'on' after pressure or strain verbs
- L3:1217 weakly excites
- L6:2067 weakly excites
- L8:1638 weakly suppresses
- L0:1427 weakly excites
- L6:273 weakly excites
- L6:358 weakly suppresses
- L6:2639 weakly excites
- L8:3 weakly excites
- L6:2176 weakly excites
- L7:709 weakly excites
- L8:1119 weakly excites
- L8:1089 weakly suppresses
- L6:1078 weakly suppresses
- L2:2214 weakly excites
- L7:2679 weakly excites
- L6:2491 weakly suppresses
- L1:2574 weakly suppresses
- L8:2063 weakly excites
- L4:1945 weakly excites
- L8:1132 weakly excites
- L5:3050 weakly excites
Feeds later layers
- L18:2168 suppresses
- L20:2079 suppresses
- L18:3020 suppresses
- L11:1204 suppresses
- L21:2059 suppresses
- L15:2718 weakly suppresses
- L17:1142 weakly suppresses
- L11:1106 weakly suppresses
- L11:883 weakly suppresses
- L10:1350 weakly suppresses
- L24:2264 weakly suppresses
- L23:3033 weakly suppresses
- L23:2768 weakly suppresses
- L17:1844 weakly suppresses
- L22:2286 weakly suppresses
- L20:307 weakly suppresses
- L19:2447 weakly suppresses
- L15:1883 weakly suppresses
- L19:1104 weakly suppresses
- L22:835 weakly suppresses words before water, lava or magnetic fields
- L21:994 weakly suppresses
- L11:2335 weakly suppresses
- L19:759 weakly suppresses
- L20:1227 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
- S
- A
- 篷
- ver
- Black
- P
- 部落
- 遗漏
- ","",
- list
Suppresses
- 高温
- pressures
- -pressure
- 高昂
- chaud
- 炅
- caliente
- 혜
- perature
- 氪