'which' or 'whom' after 'of'
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. 8 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 "which" or "whom" (and "them") when it follows a "some of" or "most of" phrase, like "a large chunk of which". Turning it up or down gave unrelated text, so it's not clear what the output does.
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.
-
, tar removal and methanation of the syngas.⏎Despite recycling efforts, the UK produces millions of tons of rubbish each year, a large chunk of which goes to methane-releasing landfill. Gasification and methanation are well understood
-
Dwarfs: The Creation of a Classic,” the exhibit celebrates Disney’s vision and the artistry of his staff—over 1,200 of whom labored for almost four years to make the film.⏎Diane Disney Miller,
-
the tau, and their respective neutrinos.⏎All known leptons have a negative or neutral charge. There are 6 types of lepton, 3 of which are negative, the other 3 neutral.<|endoftext|>From: Nicholas Holton,
-
of God (Anderson) saw potential in his work, and began mass-marketing them. He created several versions over the years, the last and most popular of which was this one created in 1941.⏎Christ at Heart's Door
-
as lithium carbonate and lithium hydroxide.⏎Lithium reserves worldwide are estimated at 39.5 million tons, about 5.5 million of which are in the United States, including at the Salton Sea, according to the
-
on the U.S. Team in the 1936 Olympic Games held in Berlin, Germany. They won 14 medals — eight of them gold: a quarter of the 56 medals won by the entire U.S
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 one that is about the heart. And I don't know what it is. Let me think. So, I'm thinking, I'm thinking, I'm thinking, I'm thinking, I'm thinking, I'm thinking, I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. This one, called iBART, was a thing of beauty. Free, too. It was written by two former high-school buddies—Ian Leighton, a sophomore at UC Berkeley, and David Hodge, a sophomore at the University
-
they drain faster than clay soils. If you live in an area with sandy soils, water two or three times a week, 20-30 gallons total during the first two years after planting, followed by half as much for the
-
, when her blood work came back everything was reported as normal. Understandably, she was both frustrated and confused about what was going on in her body. Sound familiar?⏎If you’re convinced your thyroid isn’t functioning properly but all your lab work
-
say you pose a multiple-choice question for which there is no single correct answer but for which there are perhaps more justifiable answers and less justifiable answers. Something like “Which of the following motivations best explains so-and-so’s actions in such-and
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
- L7:975 excites
- L7:323 excites
- L7:1253 suppresses
- L2:2319 weakly excites
- L2:520 weakly excites
- L7:620 weakly excites
- L5:1619 weakly suppresses
- L4:307 weakly suppresses
- L7:298 weakly suppresses
- L7:931 weakly suppresses
- L7:94 weakly excites
- L3:156 weakly excites
- L7:2264 weakly excites
- L7:1714 weakly excites
- L7:808 weakly excites
- L7:1089 weakly excites
- L4:632 weakly excites
- L7:1170 weakly suppresses
- L3:191 weakly suppresses
- L6:581 weakly excites
- L7:1886 weakly excites
- L7:1264 weakly excites
- L6:2493 weakly excites
- L7:567 weakly excites
Feeds later layers
- L9:454 suppresses
- L9:2181 weakly suppresses
- L22:505 weakly excites
- L11:2219 weakly suppresses
- L15:503 weakly suppresses
- L9:854 weakly suppresses
- L20:812 weakly suppresses
- L10:2324 weakly suppresses pronouns referring back to a noun
- L12:2896 weakly excites
- L15:2548 weakly excites
- L11:434 weakly suppresses
- L10:976 weakly suppresses
- L10:2259 weakly excites
- L17:1439 weakly excites
- L13:2316 weakly excites
- L13:2031 weakly suppresses
- L10:2579 weakly excites
- L16:2489 weakly excites
- L13:1126 weakly suppresses
- L9:562 weakly excites
- L13:2540 weakly suppresses
- L13:2365 weakly excites
- L11:2863 weakly suppresses
- L17:400 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
- "_"
- ��
- "|"
- ":""
- riority
- са
- Idx
- WXYZ
- 塍
- ętr
Suppresses
- reate
- 也不例外
- ves
- ewe
- isten
- 一路
- finally
- 积极配合
- 汲
- ::_('