'long' in 'how long' questions
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. 0 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 word 'long' in questions like 'how long will it take'. Its opposite side reacts to the word after 'How' in other questions, such as 'how is' or 'how are'. Turning it up did not visibly change the writing.
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.
-
how long they will be sick, exposed to dialysis or ventilators. The less time
-
line at the point Q. Prove AQ = AD.⏎At the time of writing the hour and minute hands of my clock are at⏎right angles. How long will it be before they
-
⏎Find out what he or she uses⏎the computer for⏎The time spent at the computer⏎Keep track of how long your child⏎is on the computer⏎How Visikid works?⏎This small
-
el per litre adult limit in parts of Ukraine until at least 2040.⏎“This situation should also act as a warning and a reminder of just how long the legacy of nuclear accidents can be,” said Labuns<|endoftext|>Gun assaults in school
-
ilers seek to push their gas tanks to the limit and achieve hypermileage, exceeding EPA ratings for miles per gallon.⏎So, there you have it. How long 'till hyper is replaced by wondermileage or excessivemileage
-
aquarium spokesperson, said alligator snapping turtles are believed to live between 80 to 120 years in the wild. It currently isn't known how long they can live under professional care, she said, because zoos and aquariums
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 been in the past, and I have been in the past, and I have been in the past. So I think of this, and I have been in the past, and I have been in the past. So I think
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
fracking pollute the water supply? Will fracking make the United States energy independent? Does fracking cause earthquakes? How is fracking regulated? Is fracking good for the economy? Coupling a deep understanding of
-
how are these numbers obtained? And wouldn’t everything feature UPF50+?⏎There
-
a great addition to Lesson 1 – Sustainable Food and Agriculture⏎What are GMOs and are they okay to eat? What does organic farming look like? How are animals like chickens or cattle raised?⏎snapAG is a series of resources that invite
-
just ignore those rules, any more⏎Continue reading →<|endoftext|>How are ayurvedic treatments effective?⏎Originating in the Indian subcontinent over 5
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
None: this is the first layer.
Feeds later layers
- L2:721 excites
- L1:124 weakly suppresses
- L15:509 weakly excites
- L1:903 weakly suppresses
- L13:2936 weakly excites
- L20:1759 weakly excites
- L11:824 weakly suppresses
- L16:328 weakly excites
- L9:264 weakly suppresses
- L15:1945 weakly suppresses
- L25:33 weakly excites spot where a number is due
- L19:1310 weakly suppresses
- L9:2580 weakly suppresses
- L18:2213 weakly suppresses
- L8:2075 weakly excites
- L3:351 weakly excites
- L1:189 weakly suppresses
- L4:3011 weakly suppresses
- L4:1729 weakly suppresses
- L22:945 weakly suppresses
- L26:318 weakly suppresses
- L25:3032 weakly suppresses 'therm' in science words
- L3:661 weakly suppresses
- L4:154 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
- .Cast
- antage
- $MESS
- 竹
- *)_
- inkel
- ivalence
- CAST
- erte
- be
Suppresses
- arma
- [color
- 启示
- 专区
- อาร
- 梯队
- squads
- ourg
- PJ
- inse