semicolons joining two clauses
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. 17 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 a semicolon that joins two full sentences into one, usually where the second part explains or results from the first. Its top examples are very consistent, but turning it up or down only produced unrelated text.
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.
-
that winners of plurality primaries may be weaker nominees than ones who ran in primaries with a majority requirement; and 2) a relatively high number of open seats in districts that are safe
-
to the Indian Forest Act, 1927. The amendments have been laid down with the primary objective of abolishing the 1927 Act; as it was a reproduction of the colonial era and a sign of the oppression perpetrated
-
method first eliminates any anatomical features not associated with the lung parenchyma; the tissues involved in gas exchange serve as the main sites of COVID-19
-
their operation. Also, the hood will have better airflow across its work surface if there are minimal numbers of bottles, beakers, and laboratory apparatus inside the hood; therefore, it is prudent to keep unnecessary equipment and glassware outside of the hood
-
early whites and gums healthy.⏎3. Having one or more oral piercings⏎You may think keeping up with latest style trends like piercings is chic; however, it can wreak havoc
-
can lead to stroke. When the arteries narrow, this condition is called “narrowing.”⏎Arterial pressure may lead to a weakening of the blood vessels; this is what causes bloodshed inside the brain. Bad habits, such as smoking
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: 2004; 2004-1; 2004-1-1 (first-order) ?; 2004-1-1-1-1... ?; and so on
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
also exploring a potential new therapy that could target H3K27me3 and reverse the lowered levels. This work is still<|endoftext|>This coin represent King Eukratides I of Bactria, a land in central Asia (Afghan
-
prison or a $250,000 fine.⏎And the system has not been updated since women were allowed to serve in combat roles.⏎The Selective Service System has disbanded
-
and this is one of the most beautiful coins of the late Hellenistic period. Known for their strongly realistic portraits, this tetradrachm of Eukratides is no exception. He is depicted as a formidable warrior, wearing a
-
But except for the state of Texas, women cannot be drafted.⏎Within 30 days of turning 18, all male citizens must register for the Selective Service System,
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
- L5:1522 suppresses
- L3:452 excites
- L2:975 excites
- L6:7 weakly suppresses
- L16:1040 weakly excites
- L3:1163 weakly excites
- L3:115 weakly suppresses
- L8:7 weakly excites
- L16:262 weakly suppresses
- L16:28 weakly suppresses
- L2:387 weakly excites
- L1:331 weakly suppresses
- L0:1192 weakly suppresses
- L5:75 weakly excites
- L1:1262 weakly suppresses
- L16:96 weakly suppresses
- L4:854 weakly excites
- L1:957 weakly suppresses
- L0:904 weakly excites
- L8:16 weakly suppresses
- L16:170 weakly suppresses
- L16:145 weakly excites
- L4:51 weakly excites
- L3:104 weakly excites
Feeds later layers
- L18:1465 strongly excites semicolon joining clauses
- L20:2064 strongly excites
- L21:2311 suppresses
- L21:351 suppresses
- L18:2832 excites
- L27:347 suppresses
- L19:1198 suppresses
- L24:668 weakly excites
- L27:1672 weakly excites
- L18:822 weakly suppresses
- L26:2886 weakly suppresses
- L26:421 weakly suppresses
- L24:378 weakly suppresses
- L23:171 weakly excites
- L22:666 weakly suppresses
- L26:416 weakly suppresses
- L26:1447 weakly suppresses
- L19:1147 weakly suppresses
- L27:1042 weakly suppresses
- L20:435 weakly suppresses
- L20:336 weakly excites
- L21:233 weakly excites
- L19:1815 weakly suppresses
- L19:893 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
- however
- `}⏎
- 本报讯
- atically
- }/
- 年代
- angel
- //⏎
- }//
- },⏎
Suppresses
- ;!
- ;</
- ();)
- lsen
- ;
- ;&#
- ;p
- ;<
- ”;
- ;'>