'of' inside organisation names
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. 15 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 'of' when it sits inside the name of an institution or event, such as 'Royal Society of Literature' or 'Fetal Medicine Foundation of London'. Turning it up produced off-topic, rambling text, so the steering does not show a clear effect.
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.
-
the plan doesn’t answer is how the city will cope with the climate change that has already been set in motion.⏎Enter a study prepared by Oakland’s Pacific Institute for the California Energy Commission, published in July but not widely circulated until this month.
-
Orionid meteor shower. Unlike the displays of previous years, however, the Orionids of 2012 have not put on a very good show. Sky watchers
-
. They have been reintroduced to National Reserves of Mongolia and China⏎CONSERVATION STATUS: Classified as Extinct in the
-
cooperate with each other is usually used to distinguish humans from other species.⏎“The dogs actually behave benevolently towards other dogs,” said lead researcher Friederike Range of the Messerli
-
private doctor⏎(The National Organization for Health Care Services)⏎Advanced Imaging Technology⏎- Ultrasound nuchal translucency (Certified by Fetal Medicine Foundation of London)⏎In humans, there are 23 types of chromosomes and most people
-
San Francisco⏎On November 10th, Frida and Diego arrive in San Francisco where Diego has been commissioned to paint murals in the Luncheon Club of the San Francisco Stock Exchange and the California School of Fine Arts [now the San
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 if I and others like my own ideas, my ideas, and my ideas as much as what and what about what it matters, and what it matters, what it is something to be and what it is, something to be and what
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ichia coli (E. coli), and against Streptococcus canis, which can be passed from companion animals to humans.⏎What diseases can you get from dog saliva?⏎This review focused on the most important viral and bacterial zoonotic
-
does not deserve all the blame. States such as California, Kentucky and New York began implementing zero tolerance policies for drugs, fighting and gang activity as early as 1989. By
-
reading, math, and writing. There are many resources available for parents who want to help their children brush up on their academic skills, including work
-
Tina explains how we can study fossilized feces and we reminisce about Jurassic Park.⏎Then, we answer fan-submitted Cosmic Queries. Can we learn anything from early humans that might help us adapt to climate change in the future? We explore
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
- L12:234 excites
- L4:4 suppresses
- L10:132 excites
- L13:132 excites "of" after a noun
- L10:469 suppresses 'of' after a noun phrase
- L11:153 suppresses
- L14:917 excites
- L14:1519 excites
- L6:29 weakly excites
- L9:1358 weakly excites
- L11:577 weakly excites
- L14:747 weakly excites
- L8:11 weakly excites
- L6:2907 weakly suppresses
- L11:558 weakly excites
- L14:453 weakly suppresses
- L5:763 weakly suppresses
- L9:1025 weakly excites
- L12:2499 weakly excites
- L6:388 weakly excites
- L10:2104 weakly excites opening parenthesis in formulas and citations
- L6:2111 weakly excites "that" starting a relative clause
- L9:463 weakly suppresses
- L5:259 weakly excites
Feeds later layers
- L16:2834 suppresses
- L16:1695 weakly excites
- L16:474 weakly excites
- L17:1111 weakly excites
- L17:1717 weakly suppresses
- L22:852 weakly suppresses
- L25:334 weakly excites
- L23:122 weakly excites
- L17:1161 weakly excites
- L17:2179 weakly excites
- L25:1135 weakly suppresses
- L20:756 weakly excites
- L20:1672 weakly suppresses
- L16:662 weakly excites
- L20:277 weakly excites
- L16:314 weakly suppresses
- L17:76 weakly suppresses
- L20:242 weakly excites
- L20:1303 weakly excites
- L16:2005 weakly suppresses
- L21:243 weakly excites
- L26:1167 weakly excites
- L19:294 weakly excites
- L18:2045 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
- ernen
- McKay
- ,err
- ��
- wyjątk
- ünd
- ansen
- setMessage
- <data
- ern
Suppresses
- agents
- prov
- cot
- 其所
- 工作者
- 릴
- pro
- tas
- riers
- sil