"until" after a negated clause
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. 6 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 lights up on the word "until" in sentences like "it wasn't until 1775 that...", where something is said to have happened later than expected. Turning it up made the model's writing drift into repetitive or rambling text, which doesn't show a clear 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.
-
980s, it wasn’t until the LIGO detector came online in 2002 that the hunt for
-
of Elizabeth I). However, it wasn’t until 1775 that Alexander Cummings added the u-bend and then
-
Up until the 1950s, the Folger also housed all nine of Gregory’s plaster reliefs in storage. It wasn’t until Dr. Frederick Hard—Scripps College president and Shakespeare scholar—negot
-
in 1531. In 1621, English colonists brought the first wild turkeys back to Massachusetts. However, it was not until after 1800 that domesticated turkeys became common in North America
-
as, 3 October 1997⏎While the flag of the Arapaho was first considered during the second World War, it was not until June 15, 1956 that one was adopted by the
-
founding of the University of Glasgow providing religious and educational status. It was not until the 16th century that Glasgow became prominent in world affairs. The city
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: Use of a black background, my own equipment, and the knowledge that I have used the computer Answer this question in the same row of rows Question: "Where is this?" Answer: "Here is a row of rows
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
formality of the painting is indicative of the manifestation of abstraction and expressionism. Owning this liminal space of amalgamating tradition and contemporary, Akhlaq went on to influence future great artists like Shahzia Sikander, Rashid
-
of Pandit houses, asking them to convert, flee, or perish; as prominent Pandits were being murdered, the Chief Minister of the state Dr. Farooq Abdullah abandoned his people and ran away to London in order to let this slow
-
.org/10.1080/13527266.2011.620764⏎BELK, R. (201
-
This we have already explained in the introduction to Surah Al Muddaththir; and in E. N. 5 of Surah Al-Muzzammil also we have explained what great burden of the coming down of Revelation he had
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
- L5:2957 suppresses
- L4:1675 excites
- L5:2390 suppresses
- L4:2882 suppresses
- L4:2393 weakly suppresses
- L5:781 weakly suppresses
- L4:661 weakly excites
- L5:2097 weakly excites
- L5:64 weakly excites
- L4:1632 weakly excites
- L4:1906 weakly suppresses
- L2:1259 weakly excites
- L5:1462 weakly excites
- L3:2705 weakly suppresses
- L4:1372 weakly suppresses
- L5:2393 weakly suppresses
- L5:2761 weakly excites
- L3:2620 weakly excites
- L5:681 weakly suppresses
- L2:42 weakly suppresses
- L4:88 weakly suppresses
- L4:2460 weakly suppresses 'City' in 'New York City'
- L5:1103 weakly suppresses
- L5:332 weakly suppresses
Feeds later layers
- L8:1869 suppresses
- L15:1386 excites prepositions after passive verbs
- L7:2970 suppresses
- L14:2150 excites
- L15:958 excites
- L9:2681 weakly excites
- L13:831 weakly excites
- L13:1515 weakly excites "your" before doctor or health provider
- L23:1345 weakly suppresses
- L7:1852 weakly excites the phrase 'by the time'
- L20:1612 weakly excites
- L23:2676 weakly suppresses
- L12:652 weakly excites
- L11:812 weakly excites
- L10:394 weakly excites
- L13:1162 weakly excites
- L11:1327 weakly excites
- L17:1995 weakly excites
- L17:2902 weakly suppresses
- L8:1962 weakly suppresses
- L7:2904 weakly excites
- L8:2420 weakly excites
- L7:507 weakly suppresses
- L19:883 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
- ILE
- leston
- enses
- fully
- ды
- ile
- ustin
- icias
- Reuters
- ество
Suppresses
- 协办
- aches
- uwagę
- dõi
- atted
- 伤心
- ouce
- Mahmoud
- dislikes
- 威胁