unclear: mixed mid-sentence words
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 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 a grab bag of single words in the middle of sentences, such as 'pronounced', 'saw', 'observed' and 'TITLE', with no clear shared meaning. When it was turned up, the model wrote rambling, repetitive text, so it does not seem to control one clear topic.
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.
-
entire universe, US-based Space Foundation CEO Tom Zelibor had said in April.⏎On April 12, 1961, Gagarin pronounced his famous “Poyekhali!” (Let’s Go!) as the
-
data on on sicklefins, tagged them, and collected sperm and eggs for captive rearing. For more photos, visit: https://www.flickr.com/photos/usfwssoutheast/albums/72157667
-
civil liberties in an age of terrorism.⏎5. Justify their stances on this topic in letters to state legislators.⏎Resources / Materials:⏎-five index cards with instructions, as
-
an escape character, an escape is a single backward slash ( \ ) in Linux, programming, and regular expressions that perform a special function or to
-
War I. Yet the son declined to disappear from the scene and played roles both dramatic and useful over the eight decades of his maturity.⏎He worried Hitler, who saw him as a potential threat to the Anschluss uniting Austria with Germany
-
religious ceremony, in South Dakota this past summer.⏎Complete information about survivors could not be confirmed<|endoftext|>A simple example of HTML is:⏎<HTML> <HEAD> <TITLE>This is the Title</TITLE> </HEAD> <BODY
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: Let me look at this, I will make an argument, assume that I will assume the following, Let me assume that: "Suppose the following, let me make the assumption: " Let me make the argument
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
-reviewed articles.⏎Write a 7–8-page paper in Word format. Apply APA standards to citation of sources. Use the following file naming convention: LastnameFirstInitial_M5_A1.doc.⏎By Saturday, August 16,
-
Kansas’s support for the Federals “should be broken,” he would not “destroy that which is absolutely necessary for the subsistence–I may almost say the existence–of the surrounding inhabitants.” Price proposed destroying the Hannibal & St. Joseph
-
gaps in order to provide students with instruction. The problem is not that schools do not wish to hire certified teachers but that a suitable pool of certified teachers qualified to teach in these fields and willing to work in these districts does not exist. In addition
-
in Jewish communities around the world today.⏎|Extinct||Mishnaic Hebrew extinct as a spoken language by the 5th century CE, surviving as a liturgical language along with Biblical Hebrew for Judaism|⏎Is Arabic easier than Hebrew?⏎
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
- L13:687 suppresses
- L12:766 excites
- L12:2559 suppresses
- L14:2207 excites
- L14:2126 excites
- L14:974 excites
- L14:2973 excites
- L11:823 suppresses
- L14:714 weakly excites
- L13:2198 weakly excites
- L14:2358 weakly excites
- L14:889 weakly excites
- L12:603 weakly excites
- L14:1619 weakly excites
- L12:917 weakly excites
- L12:785 weakly excites
- L0:2696 weakly suppresses
- L14:2352 weakly suppresses
- L13:2384 weakly excites
- L14:902 weakly excites
- L6:2939 weakly suppresses
- L1:1681 weakly suppresses
- L13:2140 weakly suppresses
- L14:1789 weakly suppresses
Feeds later layers
- L16:1871 excites
- L16:984 suppresses
- L16:335 suppresses
- L16:2757 suppresses
- L20:697 weakly excites
- L16:2029 weakly suppresses
- L16:1622 weakly suppresses
- L16:252 weakly suppresses
- L16:291 weakly excites
- L16:716 weakly excites
- L17:1912 weakly suppresses
- L16:2762 weakly excites
- L18:2307 weakly excites
- L17:1788 weakly suppresses
- L16:593 weakly excites
- L21:1439 weakly suppresses
- L18:683 weakly excites 'is' after a long subject phrase
- L17:490 weakly excites
- L19:886 weakly excites the space before a rounded number
- L16:884 weakly suppresses
- L20:604 weakly excites
- L16:909 weakly excites
- L16:1663 weakly suppresses
- L16:2014 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
- BeNull
- ContentSize
- apphire
- .CommandType
- ifest
- _features
- kaar
- remainder
- anything
- elementType
Suppresses
- ',{
- _SERIAL
- MX
- ANTE
- ROLL
- MC
- sucht
- ective
- ditor
- 调配