the word 'would' before a verb
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. 17 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 "would" when it comes before the verb that follows, as in "it would be better" or "what would you do." When it was turned up, the model's text wandered to unrelated topics, so the effect on the writing is not clear.
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.
-
/ Guide because after using our Pdf solutions you will never find need of them. You will get all the Pdf files at the end of this post, so we would like to request you to read full article.⏎|Maharashtra state Board||
-
done this way because it’s easy for normal Drupal users to control on a shared host. That’s not to say it isn’t a valid solution just that it would typically be faster to cache to a file.⏎Anyway, perhaps one of the most
-
using illeism: “One thing I didn’t want to do was make an emotional decision … I wanted to do what was best for LeBron James … what would make him happy.”⏎James was lampooned for speaking this way and accused of
-
derivation of Kirchhoff's voltage law.⏎Batteries do make a good example for this simply because they are current sources with completely isolated grounds. This example would be
-
Plan⏎Make A Plan⏎Disaster can strike quickly and without warning. It can force you to evacuate your neighborhood or confine you to your home. What would you do if basic services were cut off? Local officials and relief workers will be
-
still in the top ten. Most of such calls are followed by specific attribute access.⏎Why use anecdotal evidence? I don't doubt the numbers, but it would be⏎better to use the standard library
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 to to do something to the end of a to a to a to... I am to a to a to...... (I am to a to a to...) I have to to the end of the end of the end
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, he began travelling to the USA and Europe where he received more commendations for his artistic gift.⏎His grandest work is the Symphony of Light, which was supposed to
-
. It is deeply influenced by the country's culture, religion, and historical events, and has evolved over time to reflect changing artistic styles and techniques. In this blog post, we'll explore the fascinating world of
-
obvious; we hope some of the technical information is useful. For hearing interviewers, consider who is on your interview team. Collaboration with Deaf people–cinematographers, editors, Certified Deaf Interpreters (CDIs), production assistants
-
use them with your low beams during inclement weather.⏎<|endoftext|>The equipment of the standing army is thoroughly modern, the infantry being provided with Mauser rifles and the artillery with Kr
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
- L2:1436 strongly excites
- L16:120 strongly excites
- L6:299 strongly excites
- L0:1872 suppresses
- L16:100 excites
- L15:89 excites
- L15:114 excites
- L15:792 suppresses
- L16:296 suppresses
- L15:78 excites
- L15:115 excites
- L6:840 weakly suppresses
- L15:1587 weakly excites
- L3:1697 weakly excites
- L15:82 weakly suppresses
- L15:1404 weakly suppresses
- L15:96 weakly suppresses
- L16:28 weakly suppresses
- L15:129 weakly suppresses
- L15:201 weakly excites
- L0:983 weakly suppresses "similar" after "are"
- L15:451 weakly excites
- L15:1616 weakly excites word fragments mid-word
- L16:79 weakly suppresses
Feeds later layers
- L18:719 strongly excites
- L26:466 excites
- L21:108 suppresses
- L22:1848 excites
- L18:924 suppresses
- L22:2361 excites
- L27:2265 excites
- L18:1442 suppresses unclear: word fragments and digits inside numbers
- L26:1989 suppresses
- L22:2970 suppresses
- L19:1258 suppresses
- L22:2274 suppresses
- L21:2617 suppresses
- L20:2885 suppresses
- L26:1760 suppresses
- L26:2006 excites
- L20:2681 excites
- L18:3058 excites
- L20:626 suppresses
- L18:872 excites the digit 8 in lists and numbers
- L26:1595 suppresses
- L21:2715 weakly excites
- L26:2364 weakly excites
- L19:3002 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
- 불구하고
- ट
- forces
- gne
- 峭
- ailable
- logen
- harbour
- →ev
- driv
Suppresses
- ".$
- arity
- credited
- ></
- overtime
- esimal
- ".$_
- ><
- fact
- �认