unclear: mid-word and descriptive phrases
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. 14 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 scattered spots in everyday descriptive text, such as the start of a word split across tokens or a phrase like 'a wide variety [of]'. The pattern is hard to pin down. Turning it up produced off-topic and repetitive text, so it does not reveal a clear meaning.
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.
-
propylene foam or formed thermoplastic.⏎What is a bumper absorber?⏎The bumper absorber is located behind the decorative plastic bumper cover. As its name implies, the bumper absorber takes up the impact energy in a collision, to protect
-
.⏎Shovels, Spades, and Trowels⏎Shovels are among the most basic and essential gardening tools. Shovels have many different-shaped blades, depending on whether they are designed for digging holes or for moving around piles
-
; to break up hard, clumpy soil; and to remove weeds. Hoes come in many different sizes and weights; they also come in a wide variety of blade shapes, including flat, angled, stirrup and looped-style blades.
-
Sweden about every fourth person aged 45 years or older has doctor-diagnosed osteoarthritis in one or multiple joints, the knee being the most frequent location. As we have a steadily ageing and increasingly obese population in Sweden the prevalence
-
sword and slay dragons.⏎St George’s Requests⏎However, we can remember that after saving the king’s daughter, George made requests, one of which was never to forget the poor, which is Christian thinking but also transcultural
-
utras of Patañjali (200 B. C.). The word ‘ekabhavika’ has been introduced and discussed here. Literally, it means ‘producing one birth’.⏎In the residual karma
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 something I has been was, it was, was, was, was, was, was, was, was, was, was, was, was, was, was, was, was, was, was, was, was
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
than lifting water off the Earth” such that spaceships won’t have the tremendous expense of lifting them from Earth in order to bring them along. They could theoretically travel from Earth to the lunar waypoint and find resources waiting for them before undertaking a journey
-
most are bitmapped formats -- especially hard to collaborate on -- and, worse, many incorporate data compression. Most of the rest require proprietary tools or operating systems to function, which makes collaboration across organizational boundaries difficult. Of particular importance: most vector formats
-
in the 14th century, Flemish artisans rose up in a series of bloody revolts against the aristocratic cloth dealers who had long monopolized power. The Count of Flanders and the
-
, bushy clumps. Axes extend through entire length of plant, but main⏎branches absent; dichotomous branching throughout. Axils rounded, with sides⏎symmetrically developed, with side branches at times growing almost⏎horizontally
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
- L11:1576 strongly excites
- L12:1810 suppresses
- L13:1198 excites
- L13:2112 excites
- L13:1330 suppresses
- L13:1807 excites
- L12:817 excites
- L13:1504 excites
- L12:1070 excites
- L12:1436 suppresses
- L13:2422 excites
- L13:998 excites
- L12:1425 suppresses
- L13:602 excites
- L13:2768 suppresses
- L13:1402 excites
- L12:2103 suppresses
- L12:1445 suppresses
- L13:1720 suppresses
- L8:1609 excites
- L11:1442 excites
- L13:174 excites
- L12:2194 excites
- L13:110 excites
Feeds later layers
- L15:555 suppresses
- L15:1181 excites
- L15:1637 excites
- L15:1822 excites
- L15:1941 suppresses
- L15:1689 suppresses
- L16:1745 excites
- L15:2404 suppresses
- L18:451 suppresses
- L15:1602 suppresses
- L15:2136 excites
- L16:1108 suppresses
- L16:2638 excites
- L16:787 excites
- L15:149 weakly suppresses
- L15:534 weakly excites
- L17:1198 weakly suppresses
- L17:64 weakly excites
- L15:192 weakly excites
- L19:261 weakly suppresses
- L21:1540 weakly excites
- L15:1073 weakly suppresses
- L17:767 weakly suppresses
- L17:1402 weakly excites
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
- 主场
- 行った
- 的那种
- build
- .build
- 행
- 星球
- pios
- builder
- ::_
Suppresses
- amac
- atem
- �数
- stantiate
- 日起
- etration
- "><?=
- cog
- quez
- quet