'to' after 'in order for' phrases
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. 3 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 'to' when it comes after a named subject, as in 'in order for a region to be approved' or 'expect there to be'. Turning it up led to rambling, repetitive text rather than anything clearly connected to this 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.
-
.2: Variable Relationships⏎- For each pair of variables, do you expect there to be a relationship? That is, do you think a change in one variable is
-
it was not nearly as successful as it could have been. Or as successful as we, the outsiders looking in, would like for it to have been.⏎The Inventor's Mindset⏎One thing that should be stated
-
IR 1458+10B, in a paper in the current issue of The Astrophysical Journal.⏎"At such temperatures we expect the brown dwarf to have properties that are different from previously known brown dwarfs and much closer to those
-
icultural Area”. AVA’s are official wine grape growing regions of the United States designated by the Alcohol and Tobacco Tax and Trade Bureau. In order for a region to be approved as an AVA, it must be able to prove its identity as
-
a reference to the fact that they can form when no storms are present, and even though they usually occur under a cloud bank, it is possible for the weather to be very calm and even fairly sunny. It is rare for such a watersp
-
radiation. In order for the layers to adhere to each other, the light must pass through the layer being cured and into
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: the first line, the second line, the third line, the fourth line. So many things to do. So many things to do. So many things to do. So many things to do. So many things to do. So many
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. To the unprofessional eye the object might look good, but the truth lies in the details: use of the correct reversible materials; intrusive interventional methods kept to a minimum; careful documentation of the research and the work process before, during and
-
is a year round affair. Earlier this year we committed ourselves to becoming an anti racist borough and this work supports how we are working hard to keep to that promise.⏎We want to present this important part of British history as an essential
-
risk of breast cancer and prostate cancer and also of the spread of the cancer to other places. This is according to a very recent paper (3) from Tulane University in New Orleans. Studies of cultured cancer cells and human cancers grafted to
-
changed their minds about her. They considered her knowledgable and wise and they seeked her council. The Owl no longer gave her advice, she kept it to herself and was sad for the folly of her kind.⏎View cards
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
- L1:906 weakly suppresses
- L1:1506 weakly excites
- L0:2141 weakly excites
- L2:2949 weakly excites
- L0:613 weakly excites
- L2:1189 weakly suppresses
- L1:1380 weakly suppresses
- L1:2175 weakly suppresses
- L2:1899 weakly excites
- L2:3017 weakly excites
- L1:874 weakly suppresses
- L0:1068 weakly excites
- L2:1208 weakly suppresses
- L0:1551 weakly suppresses
- L1:124 weakly suppresses
- L0:686 weakly suppresses sentence-starting 'This' at text start
- L1:753 weakly excites
- L1:2682 weakly suppresses
- L2:111 weakly suppresses
- L0:678 weakly suppresses
- L1:988 weakly suppresses
- L1:2816 weakly suppresses
- L1:2792 weakly suppresses
- L0:1431 weakly excites
Feeds later layers
- L4:374 excites
- L14:2204 excites
- L5:53 excites
- L4:1089 weakly suppresses
- L18:2794 weakly suppresses
- L15:1412 weakly excites
- L15:2137 weakly excites
- L7:1217 weakly excites
- L8:256 weakly excites
- L4:212 weakly excites
- L10:372 weakly excites
- L11:723 weakly suppresses
- L14:198 weakly excites
- L15:2227 weakly suppresses
- L11:1311 weakly suppresses
- L14:822 weakly excites
- L14:1740 weakly excites
- L13:2015 weakly excites
- L13:1715 weakly suppresses
- L20:1074 weakly suppresses
- L12:60 weakly excites
- L11:2467 weakly excites
- L24:2637 weakly suppresses
- L16:1832 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
- happen
- 挛
- bü
- 构
- 来讲
- oci
- 生物学
- unj
- \Migrations
- ander
Suppresses
- eful
- thalm
- edith
- orrow
- łość
- ivor
- 쓰
- newList
- rollment
- idth