'not' in 'if not most/all' phrases
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. 15 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 lights up on the word 'not' in the phrase 'if not', as in 'most, if not all' or 'if not the most important'. It is the spot where a writer is about to add a stronger or alternative claim. Turning it up pushed the model into looping, repetitive writing.
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.
-
) version of this image.⏎Description Iris germanica is believed to be the ancestor of many, if not most, of the modern bearded irises that are so popular with gardeners
-
three cities: Oklahoma City, El Reno or Enid.⏎At its statehood, Oklahoma was also home to second group generally⏎thought of as German, if not by officers of the United States Bureau⏎of
-
is one of the most important, if not the most important, metrics in valuing the bond.⏎Now, let's look
-
tanks, large flags, chest thumping marches, discipline-for-itself and martyr walls help inspire nationalism and patriotism? This becomes a particularly good question, if not a surprising one, because, historically speaking, ideas of nationalism and patriotism have never
-
eating plan.⏎Beyond promoting every nutrition expert’s mantra of “eat more vegetables and fruits”, the standard raw food diet is true to its name – most, if not all, foods need to be “raw”. This means dieters are prohibited from
-
ification which is in faith (1 Tim. 1:4; cf. Titus 3:9). In the modern church, however, many—if not most—believers who are reading Scripture are prone to skip over the genealog
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 think it is called to be called. I think it is called to be called. The term, which is known as to be called. The term, which is known as to be called. The term, which is known as to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the most common types of tomatoes available. They can be meaty like a beefsteak, but always juicy.⏎Vesuvio Piennolo<|endoftext|>Thank you to DadsWorksheets.com for providing a VARIETY of math practice sheets
-
of Frank Churchill’s secret engagement to Jane.⏎“Good God!” cried Emma, “Jane actually on the point of going as governess! What
-
in Europe and western Asia, but it was introduced and naturalized to Australia, New Zealand, and North America over time.⏎Want to save this post for later? Click Here to Pin It On Pinterest!⏎Are you interested in learning more about this
-
will present new data concerning a stereo photography studio that Dalmau opened in 1854.⏎Download data is not yet available.⏎How to Cite⏎Cuenca, C. (2019). The The “Animated
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
- L8:2127 excites
- L13:2739 weakly excites
- L13:76 weakly suppresses
- L14:1436 weakly suppresses 'and' linking a role to a name
- L7:2138 weakly suppresses
- L12:63 weakly suppresses
- L4:1646 weakly suppresses 'fixed' as a finance or cost term
- L6:1259 weakly suppresses
- L12:1504 weakly suppresses
- L1:2950 weakly excites
- L10:272 weakly suppresses
- L12:1055 weakly suppresses
- L5:563 weakly excites
- L14:2959 weakly suppresses
- L11:1085 weakly excites
- L7:823 weakly excites
- L9:884 weakly excites
- L10:660 weakly suppresses
- L6:2120 weakly excites
- L8:1235 weakly suppresses
- L12:2424 weakly suppresses
- L6:2078 weakly suppresses
- L5:1322 weakly excites
- L5:869 weakly suppresses
Feeds later layers
- L19:2863 weakly excites
- L21:761 weakly suppresses
- L16:815 weakly suppresses 'the' before a noun mid-sentence
- L18:855 weakly excites
- L21:1445 weakly suppresses
- L16:162 weakly suppresses
- L20:1560 weakly suppresses
- L17:2677 weakly suppresses
- L19:2401 weakly excites
- L26:2701 weakly excites
- L21:2456 weakly excites
- L16:271 weakly excites
- L22:1773 weakly suppresses
- L18:1797 weakly suppresses
- L18:1228 weakly suppresses
- L20:1051 weakly suppresses
- L26:1275 weakly suppresses
- L25:501 weakly suppresses
- L20:2229 weakly suppresses
- L16:801 weakly suppresses
- L17:1530 weakly suppresses
- L24:2590 weakly suppresses
- L18:1679 weakly suppresses
- L21:224 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
- BeenCalled
- _YES
- >true
- GetY
- oyer
- :Boolean
- onLoad
- /gpl
- +=⏎
- emple
Suppresses
- ials
- ино
- ports
- .nlm
- -san
- jedoch
- imming
- imagin
- berman
- Reviewed