comma after 'still' or 'in addition'
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. 4 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 that follows a short linking word at the start of a sentence, such as 'Still,', 'In addition,' or 'most importantly,'. Turning it up did not make the written text show any clear pattern.
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.
-
time frame to take complete responsibility for getting themselves ready, ask questions to help prompt them if you think they aren’t paying attention to something crucial, and most importantly, when things aren’t crucial (e.g. do they have the right uniform packed
-
treated several other parishioners in the same way and with the same result, and the treatment was widely adopted by 18th-century medics.⏎Still, Cartwright might have settled for rural obscurity in his rectory had it not
-
the spread of germs,⏎- Keeping common surfaces clean, and⏎- Staying home when you are sick.⏎Currently there are no advisories restricting
-
generalization.)⏎Your reader will be thankful if you start with “This question…” or “This wig…” or, most likely, “This intrusion…” If<|endoftext|>A newly built apartment building consists of several floors with
-
services and evaluate video quality in six categories of digital defects on a scale of one to five (one being unwatchable and five being spectacular).”⏎“In addition, Comcast is working to establish local golden-eye programs in at least two markets.”⏎“My
-
great psychological benefits, which may spur you on to other healthy behaviors such as making regular exercise part of your life, too.”⏎Can anyone try intermittent fasting?⏎Still, most experts agree that some people should steer clear of IF altogether. That includes kids
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: 12, 20, 14, 16, 24, 32, 40, 48, 50, 60, 64, 72,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of this humongous ship, but many believe it had something to do with low grade rivets.⏎Wealthier passengers were offered luxurious dining tables with the best meals, but none of this
-
.Jack and the Beanstalk.The Seven Ravens.The Goose Girl.Brother and Sister A brother drinks from a stream and turns into a deer just when the
-
sprang. Amidst all the animosity in the play, it's easy to forget that the w
-
-set' of a clip with its own characteristics. Amongst other advantages, object-oriented editing can simplify the automation process compared to drawing track envelopes
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
- L3:155 strongly excites
- L2:63 strongly suppresses
- L3:1081 strongly excites
- L1:454 suppresses
- L3:195 suppresses
- L3:1071 excites
- L3:45 suppresses
- L2:501 suppresses
- L3:226 excites
- L3:1093 suppresses
- L2:307 excites
- L3:472 suppresses
- L3:81 suppresses
- L3:194 excites
- L2:329 excites
- L1:1759 excites
- L3:246 excites
- L3:208 excites
- L2:25 suppresses
- L0:2921 suppresses
- L3:21 excites
- L1:1192 suppresses
- L3:15 excites
- L1:456 suppresses
Feeds later layers
- L14:817 excites
- L10:856 suppresses
- L5:219 excites
- L6:173 excites
- L5:227 excites
- L5:33 suppresses
- L6:2492 excites
- L10:319 excites
- L7:828 excites
- L11:453 excites
- L6:189 suppresses
- L5:56 excites
- L10:390 suppresses
- L6:1579 excites
- L9:132 excites
- L16:2983 excites
- L10:679 suppresses
- L6:150 excites
- L9:630 suppresses
- L5:1614 excites
- L16:847 suppresses
- L6:1841 excites
- L9:94 excites
- L14:1920 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
- ####
- MBED
- odi
- oris
- ###
- вар
- ount
- mientras
- correctamente
- hog
Suppresses
- #{@
- eren
- )?$
- RoutingModule
- ClickListener
- 了好多
- */ ⏎ ⏎ ⏎
- ' ⏎
- _ ⏎
- ⇆