comma inside a question
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. 20 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 a comma in the middle of a question, usually right before an alternative such as "or" (for example "Are statements descriptive, prescriptive or emotive,"). Its opposite is the word just before a question mark, like "later?" or "right?". Turning it up or down made the model write strings of question marks or odd symbols.
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.
-
is the sequence of events in the conversion of the chemical energy derived from ATP hydrolysis to a mechanical stroke? How are these altered in a processive motor, where each of the catalytic heads is subject to mechanical forces from intramolecular
-
-based learning can create developmental imbalances we never anticipated.⏎Are you really going to read this blog word for word? Or are you much more likely to skim it, looking for ideas that, for whatever reason, resonate with you?⏎Will you resist
-
use of ethical language.⏎What is the meaning and nature of a moral judgement? Are statements descriptive, prescriptive or emotive, or do they correspond to an objective, universal goodness which we attempt to access through
-
they have refused to host a public hearing. Are they afraid that they would again have a standing-room-only crowd of concerned citizens rejecting the HCP, as they did in 2016? Are they trying to avoid having
-
one interesting class of question is aimed at finding out the variety of situations those laws allow. For example, is there only a single kind of rotating black hole, or do the laws of general relativity admit an infinite variety of such objects?
-
’t carefully considered before. Is it someone with an advanced degree? Is it a Nobel Prize winner? Is it someone who has memorized a great deal of information, or someone who knows how to find and evaluate information? A biological researcher? A
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: Yes? No?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
originated 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
-
place.⏎Ready to start? Let’s dive in!⏎A little history of the SSL certificate⏎They say “to know your future you must know your past”, right? Put into a closer, business perspective, I’d rephrase that with “
-
’s Heartland Theory is, why this theory is important and what the strengths and limitations of this theory are.⏎Ready to learn more about Mackinder’s Heartland Theory? Read on…⏎- What is Mackinder’s Heartland Theory?⏎- Why
-
.”⏎SALLY FIELD, acceptance speech at 2007 Emmy Awards⏎What differences do you see in countries who have endured war compared to those who have not?⏎Featured image: children during Pol Pot provided by http://www.ridec
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
- L16:1078 strongly excites
- L18:2812 strongly excites
- L17:1069 strongly excites
- L16:2817 strongly suppresses
- L15:2412 strongly suppresses
- L11:296 suppresses
- L17:2579 suppresses
- L15:1473 excites
- L5:1800 suppresses
- L16:1123 excites
- L17:1407 suppresses "and" joining questions (wh-question context)
- L13:2080 suppresses
- L13:2939 excites
- L16:669 excites
- L6:1476 excites
- L5:458 excites
- L14:2652 weakly excites
- L13:943 weakly excites
- L2:1640 weakly excites
- L1:2000 weakly excites
- L15:725 weakly suppresses
- L14:1800 weakly suppresses
- L13:2701 weakly excites
- L17:142 weakly suppresses
Feeds later layers
- L24:402 strongly suppresses
- L25:541 strongly suppresses
- L27:233 strongly suppresses
- L26:613 strongly suppresses
- L23:164 strongly suppresses
- L21:1499 strongly suppresses question marks ending a question
- L26:953 strongly suppresses
- L27:838 suppresses
- L27:975 suppresses
- L24:2628 suppresses
- L27:1576 excites
- L27:1080 excites
- L21:1538 suppresses
- L24:2138 suppresses comma before 'right?' tag
- L26:1767 suppresses
- L25:1025 suppresses
- L22:321 weakly excites
- L23:2365 weakly excites
- L23:100 weakly excites comma or full stop ending a clause
- L26:775 weakly suppresses
- L27:1551 weakly excites
- L23:1526 weakly suppresses
- L24:1922 weakly excites
- L24:902 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
- ':{'
- jerne
- //--------------------------------------------------------------⏎
- ':['
- >()->
- )::
- ":-
- regation
- ',[
- ⾦
Suppresses
- ?
- ?
- ?⏎⏎
- ?⏎⏎
- ?⏎
- ?⏎
- ?"
- ? ⏎
- ?”
- ?”