parentheses defining names or acronyms
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. 12 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 opening bracket that comes right after a full name, just before its short form, such as 'American Psychological Association (APA)'. It is strongly switched off on the word 'the' inside legal-style definitions like '(the Act)'. Turning it down made the model produce many repeated '(the NYT)' style definitions.
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.
-
such as in packet mode. The information transmission may or may not start immediately as requested by sender, the additional delay being caused by medium
-
il in any form as it contains compounds that can induce uterine contractions.2⏎Use
-
, and parody. Teachers may even make multiple copies for classroom use.⏎The legal standard for deciding what constitutes fair use of a copyrighted work involves examining and balancing the effects of four different relevant factors:⏎Retrieved and used with permission from ASCD http
-
constitutional anti-discrimination clause is set out in Article 23(2) that reads “no person shall be treated in a discriminatory manner by any person acting by virtue of any written law or in the performance of the functions of any public office
-
now at least 877 times the background extinction rate over geological time.⏎BioScience, published monthly, is the journal of the American Institute of Biological Sciences (AIBS; www.aibs.org). BioScience is a forum for integrating
-
one another,<|endoftext|>Down with Zero Tolerance⏎It might be time to ease up a bit when it comes to disruptive students and detention. The American Psychological Association (APA) recently reviewed 1
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 'sponsors' or the 'sponsors' (the 'sponsors' or the 'sponsors' or 'sponsors' or 'sponsors' or 'sponsors' or 'sponsors' or 'sponsors
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.⏎Under the Yukon Human Rights Act (the Act), harassment is defined as to:⏎“engage in a course of vexat
-
given by the trial judge was improper. The process will now be remitted back to the planning stage that was previous to the Peel Watershed Planning Commission’s (the “Commission”) Final Recommended Report (“Final Recommended Plan”), instead of being remitted
-
ak Dun v. Yukon, 2015 YKCA 18 (the “appeal decision”), the Yukon Court of Appeal unanimously allowed the Yukon government
-
workplace sexual harassment.⏎Harassment occurs in many forms, and is always wrong. However, legal protections have limits, and the Yukon Human Rights Act (the Act) only offers protection against harassment that occurs based on certain protected characteristics (called
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
- L10:1934 excites
- L3:85 excites
- L11:3013 weakly excites
- L6:16 weakly excites
- L11:1624 weakly suppresses
- L6:1112 weakly suppresses
- L1:1433 weakly excites
- L5:1550 weakly suppresses
- L8:85 weakly suppresses
- L10:3035 weakly suppresses
- L7:2478 weakly suppresses
- L0:1388 weakly excites
- L5:2656 weakly excites
- L6:76 weakly excites
- L10:1632 weakly suppresses
- L11:2810 weakly suppresses
- L11:671 weakly excites
- L4:2353 weakly excites
- L1:2005 weakly suppresses
- L1:2908 weakly excites
- L10:367 weakly excites
- L3:1050 weakly suppresses
- L9:2493 weakly excites
- L1:2496 weakly suppresses
Feeds later layers
- L13:2248 suppresses
- L25:360 suppresses
- L23:224 suppresses
- L18:1501 suppresses
- L13:1360 suppresses
- L16:2306 suppresses
- L17:738 suppresses
- L17:95 suppresses
- L16:1226 weakly suppresses
- L15:2658 weakly suppresses
- L20:1923 weakly excites
- L26:13 weakly excites
- L20:995 weakly excites
- L14:1576 weakly suppresses
- L21:782 weakly suppresses
- L20:521 weakly suppresses
- L14:2455 weakly excites
- L23:845 weakly excites
- L16:978 weakly suppresses
- L26:2854 weakly excites
- L26:234 weakly suppresses
- L21:1503 weakly suppresses
- L17:2715 weakly suppresses
- L21:416 weakly suppresses
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
- ropol
- -Level
- ighet
- ailability
- ipur
- isoft
- cular
- 点钟
- يلا
- ylum
Suppresses
- ("
- (
- (“
- ("<
- ("
- ("__
- (
- ف
- "](
- ("*