'well' inside a compound phrase
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. 24 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 reacts to the word 'well' when it joins another word into a phrase, like 'well-funded', 'well known' or 'well-being'. Turning it up didn't produce anything clearly tied to that idea, so the steering text is hard to read as a 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.
-
with minefields and traps. Following Prime Minister Narendra Modi’s decision to strike at terrorists across the Line of Control in 2016, Pakistan responded with well-tested covert tools—ratcheting up the levels of violence against Indian forces in
-
” of the justice system, including new technologies used in the courtroom, as well as court facilities and security, and watch a demonstration by Pocket, the dog whose well-trained nose aids Capitol Police in protecting the courthouse.⏎Thereafter, they get an
-
Clearinghouse for Comprehensive School Reform presents a suggested progression for the reform process. Whether you are just beginning to think about change or your school's reform efforts are well underway, Step-by-Step provides tools to assist you during this process.⏎State
-
this year (Nature, 2009). In 2015, Stephen Hawking and Russian billionaire Yuri Milner announced the start of a well-funded program, ‘Breakthrough Initiatives’, which aims to find extra-ter
-
, Fake News is a deliberate spread of misinformation, with the intent to mislead in order to gain financially or politically.⏎Fake News vs Satire⏎If a well known comedy programme featured a parody of a news presenter saying that Donald Trump’s hair
-
<|endoftext|>When it comes to workplace safety, few topics get as much attention as that of fire safety. Between fire drills and how to contain fires, most people are well aware of the fact that this is a significant risk. For many facilities, however
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: when all right to right all right to right all right to right all right all right right right right right right right right right right right right right right right right right right right right right right right right right right right right right right right right
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
19th century, has added many decorative elements to the castle although retaining much of the military appearance of what was then the summer residence of the Marquis de Armijo.<|endoftext|>Thread: FNG methods of attack question⏎05-0
-
& Co”, and Alfred Deakin (St.⏎Kilda Cemetery), thrice Prime Minister of Australia. Residing at 7⏎Barnalo Grove, Armad<|endoftext|>🇮🇳 Further to
-
, so these children often grow up to be productive workers and rule-followers. The downside of this strategy is that they could grow up to be sheep-like and uncreative, requiring instructions from others to function, rather than feeling free to think for
-
195 square kilometres and enters Lake Huron at Point Clark. The Pine River sub-watershed is located in Bruce County, Township of Huron-Kinloss and includes the Village of Ripley.⏎The Bruce Peninsula Biosphere Association (
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
- L19:134 weakly excites
- L18:178 weakly suppresses
- L17:1403 weakly suppresses
- L20:342 weakly suppresses
- L21:38 weakly excites
- L3:1066 weakly suppresses
- L17:2984 weakly excites blank before a number in prevalence claims
- L4:1587 weakly excites
- L9:2511 weakly excites
- L6:2675 weakly excites
- L5:3054 weakly excites 'to' after ease or safety adjectives
- L1:1133 weakly suppresses
- L23:619 weakly suppresses
- L11:1916 weakly suppresses
- L9:1408 weakly excites
- L12:2408 weakly suppresses
- L7:2412 weakly suppresses
- L19:2221 weakly excites
- L21:928 weakly suppresses
- L20:413 weakly suppresses
- L12:172 weakly excites
- L20:669 weakly suppresses
- L21:304 weakly excites
- L13:23 weakly excites
Feeds later layers
- L27:1484 weakly excites
- L27:572 weakly excites
- L26:880 weakly excites
- L27:2138 weakly suppresses
- L25:1749 weakly excites
- L25:670 weakly excites
- L25:1866 weakly excites
- L25:1714 weakly suppresses
- L25:2340 weakly suppresses 'to' after a verb, before 'avoid'
- L27:757 weakly excites
- L26:789 weakly excites
- L25:569 weakly excites
- L26:1879 weakly suppresses
- L26:2437 weakly suppresses
- L27:1296 weakly suppresses
- L25:1582 weakly suppresses
- L25:2353 weakly excites
- L25:1692 weakly excites
- L27:1836 weakly suppresses
- L27:672 weakly suppresses
- L25:1530 weakly suppresses
- L25:2593 weakly suppresses
- L25:753 weakly suppresses
- L25:1527 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
- -offs
- ows
- -off
- Herz
- .off
- acon
- off
- ULATION
- bac
- ometown
Suppresses
- eti
- interest
- eno
- Interest
- loth
- *e
- setObject
- interest
- 箸
- Nor