unclear: words before an explanation
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. 20 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 words that set up an explanation or a list, such as 'due' in 'was due to', or an opening bracket before a note. Its negative side fires on the word 'a' before a role name, like 'a lexicographer'. Turning it down made the model produce empty blank-filled templates, but turning it up had no clear effect.
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.
-
frequent intervals becomes.⏎When initial temperature is measured in those products which incorporate frozen ingredients extra care may need to be taken to insure that the temperature reading is of the (average) temperature of the contents of the container and not just the temperature of the
-
with fuel. For centuries, it was firewood, but by the early 1900s the use of firewood was increasingly inconvenient.⏎This was due to two major problems. First of all, population growth meant that most mountains around
-
not everyone has had their own experience with SAD, many have watched someone else go through it. When talking about seeing someone else go through this, (Male 282) says, “they’re different, distant, standoff
-
1___ in the United States and, until the beginning of the 20th century, was used by nearly all American cities. Its structure is similar to that of the ___2___ and national governments, with an elected mayor as chief of
-
. For such individuals, concentrations in information systems, information technology, or computer studies generally would be more appropriate.⏎These recommendations for formulating concentrations in computer science are based on "Computing Curricula 2001: Computer Science" by
-
by an average of 5.6% per year from 2009-13 with a stable exchange rate and low inflation. This growth was due partly to high international prices for Peru's metals and minerals exports, which account for
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: (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____) (____)
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
best access the data incorporated in printed and electronic dictionaries. This is sometimes referred to as 'metalexicography'.⏎A person devoted to lexicography is called a lexicographer.⏎General lexicography focuses on the design, compilation, use
-
us, Species⏎List the taxa in order from largest to smallest.⏎A heterotrophic eukaryote that absorbs nutrients from organic materials in the environment is a.⏎The system
-
memory, which does not have a ROM, to bring the loader to the memory once at start up.⏎A microcomputer system with a ROM and a RAM has a monitoring program in its ROM in any case, so the computer can be completely operated
-
most commonly affect women and occur in adolescence and young adulthood, though they can occur in anyone, of any gender, at any age.⏎Eating disorders most commonly affect women and occur in adolescence and young adulthood, though they can occur in anyone,
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
- L17:2664 strongly excites
- L19:2077 strongly excites
- L18:2255 strongly suppresses
- L15:1538 excites
- L16:685 excites
- L17:2866 suppresses
- L18:2845 excites
- L19:2164 suppresses
- L16:1199 excites
- L19:1193 suppresses
- L18:2185 suppresses
- L18:394 excites
- L17:644 excites
- L12:2186 excites
- L13:2012 excites
- L18:2819 suppresses
- L17:623 suppresses
- L16:1965 excites
- L19:1246 excites
- L18:419 weakly suppresses
- L14:441 weakly excites
- L19:2591 weakly suppresses
- L15:2158 weakly suppresses
- L17:371 weakly suppresses
Feeds later layers
- L25:360 strongly suppresses
- L26:234 strongly suppresses
- L23:224 strongly suppresses
- L26:13 strongly excites
- L22:351 suppresses
- L27:218 suppresses
- L21:416 suppresses
- L27:921 suppresses
- L26:1589 suppresses
- L27:475 suppresses
- L27:904 suppresses
- L25:300 suppresses
- L26:256 suppresses
- L25:474 suppresses
- L27:1373 suppresses
- L26:378 suppresses
- L22:161 suppresses
- L22:317 excites
- L26:1059 suppresses
- L24:146 suppresses
- L25:51 excites
- L25:161 suppresses
- L24:260 suppresses
- L22:795 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
- exampleInput
- &apos
- odox
- 듯
- &T
- eful
- -FIRST
- xhttp
- JsonValue
- eyJ
Suppresses
- **
- ____
- __
- **
- ___
- ____
- ______
- (__
- (**
- ________