'as' before 'well' or 'as much'
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. 22 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. 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 'as' in phrases like 'as well', 'as much as' and 'as well as', right before the model is likely to write 'well'. Its negative side is very weak and does not mean much. When it was turned up, the model wrote long runs of 'well well well', which shows the neuron is pushing toward that word.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
(both images from Fri Aug 9 2019)⏎“Similar particles have been detected near Saturn, suggesting⏎that the process is at work there as well,” says Liu. “Indeed,” adds⏎Angelopoulos, “this type
-
are all girls.⏎- Children that had died were named. So if one son called Jan Hendrik died, the next one born would be called Jan Hendrik as well. Usually, if you see two children with the same name, the oldest
-
your community soup kitchen, giving to others is a selfless act that benefits everyone involved. Did you know that generosity and giving to others can benefit your own wellbeing as much as it benefits those who receive your act of kindness?⏎It’s true.
-
.⏎- As human population and consumption increase, so do negative impacts on the planet.⏎- Human activities can significantly alter the natural environment, but impacts can be positive as well as negative.⏎- All human activity draws
-
is very effective in treating diabetes and heart disease in mice at the same time, and we believe it may turn out to be protective against asthma and other metabolic disorders as well.”⏎The laboratory animals showed no harmful effects from the treatment, though Hotam
-
new wonderful interactive SAMR visual for teachers. I just came across it while I was going through my Google Plus feeds and thought you may want to have a look as well. For those of you not yet familiar with what SAMR model is all
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 first few of the terms in the terms of the current $n$-ary powers of $e$. That is, instead of $e = \{1, 3, 5, 6, 9, 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
problems downloading a file, check if you have the proper application to view it first. In case of further problems read the IDEAS help page. Note that these files are not on the IDEAS site. Please be patient as the files may be
-
”⏎The U.S. progressives in this era were a diverse group of reformers including socialists like the writer Upton Sinclair and the frequent Socialist presidential candidate Eugene Debs, but also Theodore Roosevelt, who after serving as a Republican president earlier,
-
web site to request a printed copy of this publication.⏎- Our on-line publications are scanned and captured using Adobe Acrobat. During the capture process some typographical errors may occur. Please contact Sharon Hobrla, email@example.com if you notice any
-
for-mation of the dissimilar metal joint and the measurement of joint tensile strength⏎Welding Method 6, Paragraph 188.8.131.52 (Paraphrased, still in public comment
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
- L20:651 strongly excites
- L19:1570 strongly suppresses
- L19:2110 excites
- L19:2003 suppresses
- L17:1395 suppresses
- L20:1009 excites
- L18:1328 suppresses
- L19:164 suppresses
- L21:1973 suppresses
- L17:2765 suppresses
- L21:2038 suppresses
- L19:2502 excites
- L17:2081 excites
- L16:899 excites
- L18:2037 excites
- L18:409 excites
- L15:2010 suppresses
- L10:155 excites
- L21:1848 excites
- L18:2341 excites comma before an added clause
- L19:214 excites
- L18:2921 weakly suppresses period ending a sentence that announces a list
- L17:2623 weakly suppresses
- L19:2458 weakly excites
Feeds later layers
- L27:1410 strongly excites "as well" and "as much as"
- L27:1593 excites
- L27:172 weakly suppresses
- L27:457 weakly excites
- L25:1411 weakly excites
- L23:283 weakly excites
- L23:390 weakly excites
- L25:1498 weakly excites
- L27:490 weakly suppresses
- L27:655 weakly suppresses
- L25:995 weakly excites
- L24:2063 weakly excites
- L25:2940 weakly suppresses
- L27:1650 weakly excites
- L26:1612 weakly excites
- L25:2695 weakly suppresses
- L25:801 weakly excites
- L27:1420 weakly suppresses
- L25:2801 weakly suppresses
- L27:2971 weakly suppresses
- L25:1492 weakly suppresses
- L27:2285 weakly suppresses
- L27:680 weakly suppresses
- L27:1354 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
- well
- well
- Well
- Well
- WELL
- bien
- wells
- 及
- 以及
- "Well
Suppresses
- 懒得
- ()")⏎
- ьер
- ↩
- 干嘛
- лон
- ees
- 帱
- NOM
- ");//