'that' starting a new clause
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 lights up on the word 'that' when it is used like 'this' or 'that' to refer back to something just said, as in 'and that percentage' or 'of course that isn't always the case'. When it was turned up, the model's text drifted into rambling and repetition rather than a clear theme.
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.
-
é’s population is 2.5 million. The urbanization rate is currently 3.3%; 58% of the country is urbanized and that percentage continues to grow annually.⏎Cameroon’s population is made up of its indigenous
-
<|endoftext|>Homeowners across a large swath of Pennsylvania are reporting sightings of spotted lanternflies hatching from egg masses.⏎“Egg-hatching season is here, and that has some people on edge,” said Emelie Swackhamer, a
-
. You can communicate a thought, but your thought may not be understood. In some cases, your thought may not even reach the proper target. Thats why it pays to ask questions to make certain that people understand what you are saying
-
good listener.⏎Over a period of 240 years there has been 44 presidents and the U.S has had both good times and bad times and that all depended on who was running the country and what choices they
-
stakes and adrenaline during a game. (Or a whimsical ‘why not’). But of course that isn’t always the case.⏎A total of 64% of Australian adults
-
00. But of course that’s not what you’ll actually receive.⏎Your employer is required to withhold federal and
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: What is the value of the parameter in the parameterization of the curve $ \frac{1}{0} $ when $ \text{0} $, $ \text{1} $, $ \text{1} $,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2. Every day the boss enters the office⏎at half past nine⏎3. My friend will ring me up at eight o’clock⏎4. Where was Kate going when you met her yesterday?⏎does your brother work?⏎6. What
-
God Himself standing over him in judgment. “Where were you hurt?” the general asked in his high-pitched voice (Patton sounded
-
expelled the only red state in the national organizaton" -- but<|endoftext|>Raising Our Children to Be Gatekeepers⏎And the LORD said unto Cain, Where is Abel thy brother? And he said, I know not: Am I my brother
-
)<|endoftext|>High: 77°F ~ Low: 59°F⏎Monday, Sep. 15, 2014⏎"Where's The Beef"Posted Friday, October 9, 2009,
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
- L3:700 excites
- L3:828 suppresses
- L3:487 weakly suppresses
- L3:854 weakly suppresses
- L1:74 weakly excites
- L1:1273 weakly excites
- L1:2689 weakly suppresses
- L2:2681 weakly suppresses
- L1:1053 weakly suppresses
- L3:246 weakly excites
- L0:593 weakly excites
- L2:2557 weakly suppresses
- L3:762 weakly suppresses
- L3:155 weakly excites
- L3:745 weakly excites
- L1:1169 weakly excites
- L3:189 weakly excites
- L0:2338 weakly suppresses
- L3:569 weakly excites unclear: mid-phrase function words
- L2:1527 weakly excites
- L3:1068 weakly excites "that" pointing back to a noun
- L1:29 weakly suppresses
- L2:2830 weakly excites
- L2:2137 weakly suppresses
Feeds later layers
- L10:965 excites the word 'that' introducing a clause
- L17:1821 weakly excites
- L15:2787 weakly excites
- L20:586 weakly excites
- L15:810 weakly suppresses
- L20:2799 weakly suppresses
- L14:1540 weakly excites
- L17:1851 weakly excites
- L24:54 weakly excites
- L11:1110 weakly excites
- L26:2486 weakly suppresses
- L15:470 weakly excites the word 'this' after a preposition
- L22:1365 weakly suppresses
- L17:2680 weakly excites
- L10:1770 weakly excites
- L24:1345 weakly excites
- L11:2788 weakly suppresses 'side' in 'side effects'
- L19:1935 weakly excites
- L9:917 weakly excites
- L14:84 weakly suppresses
- L9:2011 weakly excites
- L18:789 weakly excites
- L12:1812 weakly excites
- L19:1004 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
- PGA
- followed
- asc
- uada
- eway
- register
- alta
- 唬
- registro
- asthan
Suppresses
- kräfte
- ện
- וץ
- äl
- quartered
- -floor
- vọng
- 揆
- 韧
- ?q