the word 'more' before a comparison
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. 18 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 responds to the word 'more', especially in phrases like 'more than', 'more serious' or 'more recent'. The negative side is fragments of place names and words split in the middle. Steering the neuron did not produce a clear change in the 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.
-
plastic ― into one general area. However, it does not converge into a floating island of trash, like some people have come to believe.⏎Instead, it’s more like a cloud of pollution, and much of it is made up of particles called
-
volume laying the groundwork for his thesis that mathematics relies on a foundation constructed on assumptions. Along the way, he references certain principles of statistics and calculus as well as more esoteric branches of mathematics. He also touches on the work of several famous mathematic
-
its imaging infrastructure was seen as a benchmark for European-wide cooperation in imaging, leading to the decision to place the administrative centre in Turku.⏎The consortium builds on more than a decade of cooperation between dozens of
-
nose are fragile, so it is no wonder that they can easily bleed. However, sometimes nose bleeds can become frequent. This may be an indication of a more serious problem.⏎One of the most common causes of frequent nose bleeds is dry
-
learning anywhere, anytime; and responsive to the 21st century context of an interconnected global economy and rapidly emerging global society?⏎That was an organizing question for more than a thousand people who gathered in Doha, Qatar, for the inaugural World
-
a Minute ” was interesting and had this fascinating graphic comparing price changes from Sears, Amazon and Best Buy. In addition, this academic article explains yield management while this more recent study looks at internet dynamic pricing.<|endoftext|>A diplomatic crisis is engulfing part
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: "This is a story... it's a story... this is... this... this... this... this... this..." but... this... this... this... this... this..." and... this... this... this..." from and and and
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
like church and tax books being destroyed by the flood, like the land taxes in the year 1731 from the villages Meedhuizen and Slochteren and the church books before 1700.⏎Graphics by Lynn Water
-
Luther's preaching and writing sparked the Protestant Reformation, which spread through much of Europe, including France. Inspiration pour des voyages à Londres, en Angleterre, en Ecosse, au Pays de Galles et en Irland
-
of York, artisans, £10 million grant, constructing a chapel, craft apprenticeships, Danish invaders, Decorated Gothic, Decorated Gothic nave, doleritic sedimentary rock, Early English, Early English north and south transepts
-
, of which only three have survived: the⏎Auberge de France, the Auberge d'Angleterre and the Auberge de Provence and⏎After the foundation of Vallet
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
- L5:858 strongly excites
- L1:47 strongly excites
- L16:139 strongly suppresses
- L0:41 suppresses
- L16:463 suppresses 'for' introducing a purpose or target
- L17:253 suppresses
- L16:66 suppresses
- L16:113 excites
- L0:2211 suppresses
- L2:1866 suppresses
- L15:96 suppresses
- L3:1055 suppresses
- L16:60 suppresses
- L1:32 excites
- L0:2765 excites
- L14:95 excites
- L0:2724 suppresses
- L10:1699 weakly excites
- L17:1637 weakly excites
- L4:189 weakly excites
- L17:725 weakly suppresses
- L4:344 weakly suppresses
- L16:118 weakly suppresses
- L10:1473 weakly suppresses
Feeds later layers
- L21:812 strongly suppresses
- L20:601 strongly suppresses
- L19:1237 strongly suppresses
- L27:2224 suppresses
- L19:1639 excites
- L19:277 excites
- L24:2266 suppresses
- L19:1728 suppresses
- L23:1368 suppresses
- L19:572 suppresses
- L22:2727 suppresses
- L19:848 excites
- L20:2203 excites
- L19:593 suppresses
- L20:1710 excites
- L19:1070 suppresses
- L19:322 excites
- L21:325 excites
- L20:2652 excites
- L27:1818 suppresses
- L20:1040 excites
- L19:1853 excites
- L21:1679 excites
- L27:2962 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
- ifferences
- t
- enuity
- dığı
- orning
- sb
- ght
- dığımız
- entifier
- e
Suppresses
- Than
- -than
- _than
- 링
- Than
- board
- THAN
- 手套
- BOARD
- icals