unclear: weak mid-phrase prepositions
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. 16 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
Hard to pin down. Its weak highs fall on small connecting words such as "the", "on" and "of" in the middle of a sentence, and its strongest negative hits are on words like "manage" and "to" in comparisons. Steering did not show a clear 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.
-
) "The gun fired"⏎<|endoftext|>NASA is trying to decide whether to continue preparing the space shuttle Atlantis for a launch attempt on Tuesday or to roll it back into the Vehicle Assembly Building (VAB) to take
-
ate Acids and Bases practice, you can also practice Conjugate Acids and Bases practice problems.⏎What professor is this problem relevant for?⏎Based on our data, we think this problem is relevant for Professor Finnegan's class at
-
or Center for Neuromodulation⏎What is Parkinson's disease?⏎Parkinson's disease (PD or, simply, Parkinson's) is the most common form of parkinsonism, a group of motor system disorders.⏎It is a slowly progressing
-
ynursingpaper.com/wp-content/uploads/2020/06/cropped-155972031988413967-1-768x146-1
-
. Lythronax stands out from its contemporaries in having a much wider skull at the eyes and a narrow short snout, similar to its relative T. rex, which lived 10-12 million years later. Dr.
-
homes, from the ground, through sump pumps, cracks in the foundation, etc., it is considered a health hazard.⏎Any of the beams that slope from the ridge of a roof to the eaves
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 other... ,... the other... . ,... the... ... the... ... the... ... the... ... the... ... the... ... the...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Trim The Roots Of A Bonsai Tree? (Step by Step)⏎Caring for bonsai trees is a unique art form, different from how you would manage other plants.⏎Looking after non-bonsai plants involves re-potting them
-
no one better. He wrote more letters to her than he wrote to anyone. All her life, she wrote back: letter after letter filled with news
-
⏎The debate over whether to use dashes or underscores to represent spaces in URLs is rather heated in the web development community, but not quite as extremely so as that of whether to use tabs or spaces when indenting code. So, what is better
-
is clicked.<|endoftext|>Americans with Disabilities Act (ADA)⏎The Americans with Disabilities Act (ADA) provides civil rights protections to individuals with disabilities similar to those provided to individuals
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
- L15:2981 strongly suppresses
- L15:2703 suppresses
- L14:2076 suppresses
- L13:2320 excites
- L13:1692 excites
- L14:1991 excites
- L15:2067 excites
- L13:1991 excites
- L14:733 suppresses
- L14:1741 excites
- L14:2727 excites
- L13:2483 suppresses
- L15:1738 excites
- L14:399 excites
- L14:2595 suppresses
- L15:2223 excites
- L15:2248 suppresses
- L15:2503 excites
- L14:1042 excites
- L15:476 excites
- L13:2312 excites
- L14:1537 excites
- L15:1950 excites
- L14:978 suppresses
Feeds later layers
- L17:239 suppresses
- L17:807 excites
- L18:825 suppresses 'as well as' and list connectors
- L18:1536 suppresses end of web page disclaimer
- L18:1850 suppresses
- L18:556 excites
- L17:1586 suppresses
- L18:206 excites
- L19:1567 suppresses
- L19:355 excites
- L17:174 excites
- L19:1673 suppresses
- L20:697 weakly suppresses
- L17:1109 weakly suppresses
- L22:73 weakly excites
- L20:204 weakly suppresses
- L20:1189 weakly suppresses
- L19:401 weakly excites
- L24:174 weakly suppresses
- L20:708 weakly excites
- L18:540 weakly suppresses
- L20:2848 weakly suppresses
- L25:935 weakly suppresses
- L17:1437 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
- rosse
- -public
- Ƣ
- ieurs
- oğlu
- imized
- }> ⏎
- 轸
- łow
- dney
Suppresses
- other
- 其他
- (other
- 其他的
- other
- main
- footnote
- 得多
- 其它
- Other