preposition right after a passive verb
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 a small word like 'to', 'on' or 'off' that comes right after a verb in the passive form, such as 'paid to', 'shot on' or 'ripped off'. When it was turned up, the model's text drifted into ordinary prose without a clear theme, so the steering output does not show much about it.
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.
-
for works that are sold for money. That is because registration helps to prove that the copyright of a work belongs to a certain author.⏎If an author gets paid to make a work for someone else, the person who pays for making the work (
-
local schools across the country engaged in efforts to improve science curriculums.⏎Under the grants announced today, the colleges and universities will each receive $1 million to $2 million over four years to improve science education in several areas, including modern
-
or less safe. Yet they were also unwanted and were prodded to leave as soon as possible. They had every incentive to do so, but getting
-
that, they are likely to be in a hostile environment.⏎Danger is a constant companion on the battlefield, and because of that, there are things that are done to save lives that aren’t as applicable in the civilian world.⏎Why Are We Ta
-
caliber Iver Johnson⏎It took two shots from a .32-caliber Iver-Johnson revolver to assassinate William McKinley. He was shot on September 6, 1901 in the late afternoon by Leon C
-
means is useless, I don’t believe that they can be properly roasted at home. Folks, there are educated professionals who acquire specialty equipment and are hipped on this subject. Do you think you’ll be able to properly roast coffee beans using
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: (x+1) where yz and zy. The problem is to find for the variables. Wait, no. I need to find the right part. Wait, no. I need to find the right part. Hmm,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Sunday night at the age of 88.⏎Rubin, who did much of her work at the Carnegie Institution, has been called a “national treasure” by the organization’s president. During the ’60s and ’70s
-
lawyers and teachers. This undisciplined lot went on to brave bullets, were jailed, lathi-charged and suffered untold humiliations at the hands of the colonial rulers. In other words, nationalism and patriotism in India was constructed as
-
create a culture where sexual violence is prevented and reporting is encouraged.⏎This issue is current and demands attention. Vanessa Guillen, an Army Specialist, was recently killed by a fellow Specialist in Fort Hood, Texas. After she went missing, her family
-
garment that Bella Hadid wore, which was a spray-on dress. Hadid walked slowly onto a platform in the middle of the runway. Here she was sprayed by two technicians with a white material. With a little cutting and folding, the material
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
- L7:345 suppresses
- L7:1813 excites
- L6:2485 weakly excites
- L6:12 weakly excites
- L4:37 weakly suppresses
- L3:115 weakly suppresses
- L5:2896 weakly suppresses
- L4:1494 weakly excites
- L7:174 weakly suppresses
- L6:412 weakly suppresses
- L4:1394 weakly excites
- L6:7 weakly suppresses
- L2:956 weakly suppresses
- L7:1629 weakly excites
- L0:1704 weakly suppresses
- L7:2618 weakly excites
- L1:2801 weakly excites
- L6:2790 weakly suppresses
- L5:2890 weakly suppresses
- L6:2131 weakly excites
- L1:2519 weakly suppresses
- L5:2114 weakly excites unclear: calculus notation, mixed topics
- L2:1899 weakly excites
- L4:1318 weakly suppresses
Feeds later layers
- L15:2606 suppresses
- L9:1562 suppresses 'by' before a byline or agent
- L9:1627 suppresses
- L11:361 suppresses
- L13:1515 suppresses "your" before doctor or health provider
- L19:2845 suppresses
- L9:1842 suppresses
- L19:2853 suppresses
- L19:2508 suppresses
- L16:2208 suppresses
- L15:799 suppresses
- L12:2887 suppresses
- L15:1053 excites
- L17:2197 suppresses
- L12:2014 suppresses
- L12:2145 excites
- L15:1455 suppresses
- L20:1794 suppresses
- L17:978 weakly suppresses
- L15:3044 weakly suppresses
- L13:2832 weakly suppresses "by" after a passive verb
- L14:2508 weakly suppresses
- L9:1830 weakly suppresses
- L12:1286 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
- quences
- Украины
- Richie
- 和个人
- 的身体
- обучения
- 实事
- печат
- BUSINESS
- ัว
Suppresses
- [&](
- }};⏎
- directors
- iners
- '};⏎
- iting
- igator
- '=',
- 监管部门
- '));⏎