'of' after nouns like 'variation' and 'section'
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. 2 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 'of' when it follows a noun like 'variation', 'section', 'variant' or 'author' (as in 'a variation of' or 'a section of'). Its negative side is a comma right after a person's role, such as 'a biologist,'. Turning it up did not produce a clear, consistent theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
's especially ironic as the Caroline minuscule was invented under king Charlemagne as a way to counter the Merovingian cursive (in fact a variation of the Roman cursive) that
-
Last Winter (Agence France-Presse)⏎Beekeepers Sue EPA over Failure to Protect Bees from Pesticides (by Noel Br<|endoftext|>The following variation of a Bulls and Cows problem was given at the Fall 2008
-
content region. A view replaces the node to display a list or some other small variation of content.⏎Next you are asked to make the view more complicated. You scratch your
-
, after Filipino rebels seized control of a remote section of Malaysia’s Sabah state as part of an unresolved territorial dispute that stretches back centuries
-
xylem cells by the Ophiostoma fungus. Attacked trees exhibit wilted foliage with reddish or purplish discoloration within a section of the crown or throughout the entire crown. The foliage eventually turns brown and tends to
-
Terminal?⏎If you are using open-source packages, then sooner or later, you will encounter the .tgz or .tar.gz files. This is a variation of zipped files; most open-source packages use this format to zip their files.
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: I have been working on this week, and I have to have been thinking about... so I have to have been thinking about... and I have to have been thinking about... so I have to have been thinking about... and I have to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
temperatures, though it seems that these critters would be able to acclimatize to increasing temperatures in their natural habitat,” says Ricardo Nieves, an author of<|endoftext|>January 10, 2002⏎Parshat Va
-
e.g., infectious bacteria, would be harmless to other organisms, e.g., humans. In 1928, Sir Alexander Fleming, a Scottish biologist, observed that Penicillium notatum, a common mold, had destroyed st
-
cientific, while others took the time to address the main arguments. Alex Gilbert, an energy analyst, took to Twitter to explain the problems with the study in detail. While the authors
-
campaign on energy policy. “We desperately need to find a new form of energy.”⏎Colin Campbell, a retired geologist, says governments are too caught up in short-term issues to focus on the long-term
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
- L1:878 suppresses
- L1:427 excites
- L1:192 weakly excites
- L1:295 weakly suppresses
- L1:653 weakly excites
- L1:2075 weakly suppresses
- L1:137 weakly suppresses
- L0:537 weakly excites the 'or' in 'and/or'
- L1:1861 weakly excites
- L1:898 weakly excites
- L1:385 weakly excites "as" after "known"
- L0:2009 weakly suppresses
- L0:588 weakly suppresses
- L0:2002 weakly excites
- L1:1007 weakly suppresses name pieces inside a multi-part word
- L0:1091 weakly excites
- L1:463 weakly excites
- L0:448 weakly excites
- L0:456 weakly suppresses
- L1:152 weakly suppresses
- L1:1385 weakly suppresses
- L0:2434 weakly suppresses
- L1:2408 weakly excites
- L1:947 weakly excites
Feeds later layers
- L6:54 suppresses
- L9:1674 suppresses
- L5:192 excites
- L5:692 suppresses
- L6:361 weakly excites
- L5:194 weakly excites 'a' or 'of' right after a verb
- L5:999 weakly excites
- L8:1677 weakly suppresses unclear: mixed, mostly math and instructions
- L17:258 weakly excites
- L6:695 weakly suppresses
- L5:1055 weakly excites
- L4:2010 weakly excites
- L6:426 weakly excites
- L4:1192 weakly excites
- L17:1974 weakly excites
- L5:2600 weakly excites
- L15:2146 weakly suppresses
- L20:2044 weakly suppresses
- L5:121 weakly suppresses
- L9:1139 weakly excites
- L4:139 weakly excites
- L8:498 weakly excites
- L10:1354 weakly excites
- L15:1961 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
- ially
- entially
- ologies
- esis
- ings
- aint
- IFO
- DbType
- iffs
- �
Suppresses
- keh
- 六个
- 五个
- 뉴
- herr
- yaptığı
- ję
- 找回
- DEN
- ousse