word before a 'for' or 'with' object
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. 13 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 the word right before a phrase that adds detail, such as 'dentist with', 'would know', 'study for', or 'has stories'. Turning it up made the model produce number-heavy math text, and turning it down made it lose its thread and write stuttering fragments.
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.
-
age 7, a child will have started to get adult molars in place, and the position of these and the primary teeth provide the child's general dentist with a good perspective of the development of
-
too small that no one can see, or making the text and background colours clash that make it impossible to read.⏎If you are typography nerd, then you would know something about the “Classic Scale”. This makes the readability of your PowerPoint smoother and
-
and Oxfam in Vietnam have collaborated on a research report on the effects of disasters and climate change on multidimensional inequality in Vietnam. Please check the complete study for more information on the data and analysis.<|endoftext|>Parus niger (Southern black
-
I never did find out what was the matter. Maybe a deer? This is just one example of a horse spooking. Anyone who has been around horses has stories of horses spooking, sometimes for no reason at all.⏎Quick Re
-
down in the photo, is carved and ground into an unmistakable human female form, while the back remains mostly⏎unworked gray cortex. In the photos above note the clear morphological similarity of the figure to that of the well known Neolithic
-
the types of broken bones, including:⏎TransverseA bone break is considered transverse when the fracture extends across the bone.⏎SpiralAs you may have guessed, a spiral fracture spirals around the bone.⏎ObliqueAn oblique fracture
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 point that, when we are in the space of the points that are on the surface of the space that are defined by the set of points that are the points on the curve in the space space that are the curve on the curve in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
loose with our terms, but that just makes it more confusing.⏎You ought to be able to think these through - take a shot and we'll tell you where you're on and off-target.⏎Ok, thats a good idea.⏎1. Unfortunately
-
natural predators. These predators include other insects and parasites. Insects that feed on whiteflies include the lady beetle and lacewings. Parasites that use whiteflies as a host species are the tiny wasps in the Encarsia genus.
-
both Greek and English, offers the visitor a fascinating tour in a simple and understandable way, utilizing modern audiovisual media. Even small fragments which seem insignificant at first glance, create an opportunity to reconstruct the facts,
-
of Lagos. She has been trying to get a loan to support her maize business to no avail; a number of microfinance banks that Akande approached declined to give her a loan. Among the challenges she also faces is the high price of fertil
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
- L11:361 suppresses
- L12:2279 suppresses
- L9:2390 suppresses
- L12:969 suppresses
- L9:1322 excites unclear: medical symptom lists and materials
- L12:328 excites
- L12:372 suppresses
- L4:463 excites
- L11:1327 excites
- L8:408 excites
- L12:2145 excites
- L12:2899 weakly excites
- L11:579 weakly suppresses
- L11:294 weakly excites
- L12:2722 weakly excites
- L9:361 weakly excites
- L10:616 weakly excites
- L12:1207 weakly suppresses
- L6:1231 weakly suppresses
- L5:2236 weakly excites
- L11:708 weakly suppresses
- L1:753 weakly suppresses
- L12:1065 weakly suppresses
- L9:341 weakly excites
Feeds later layers
- L14:1060 excites
- L15:576 excites
- L14:778 excites
- L15:1943 suppresses
- L16:749 excites
- L19:2665 excites
- L16:1768 excites
- L14:551 excites
- L15:1513 weakly excites
- L15:1569 weakly excites
- L14:252 weakly excites
- L14:1124 weakly excites
- L14:453 weakly excites
- L17:2515 weakly excites
- L14:674 weakly excites
- L14:1252 weakly suppresses
- L16:98 weakly suppresses
- L16:1315 weakly excites word or phrase being named or defined
- L15:448 weakly suppresses
- L15:182 weakly excites
- L16:304 weakly excites
- L14:2042 weakly excites comma after 'For X,' introducing a viewpoint
- L14:2669 weakly excites
- L15:360 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
- xn
- fully
- culture
- /off
- 内的
- führer
- _average
- _or
- umble
- խ
Suppresses
- 】【
- 年后
- getResult
- getTable
- نز
- 棵树
- 蛎
- setResult
- redi
- estone