'what' as a free relative
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. 3 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 'what' when it means 'the thing that', as in 'built what he called' or 'painted what he felt'. Turning it up did not bring out that theme in the model's replies.
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.
-
six stories and broke his back. He died the next day.⏎Astrodome Capsule Stunt⏎Karel Soucek was a Canadian stuntman who built what he called a cushioned capsule to make a successful plunge over Niagara Falls. He
-
gives rise to the question of what is the actual rate of price inflation. They don’t use the CPI ratio put out by the Census Bureau. They use what is called a “chained” price deflator from the GDP report which means
-
bre colours behind him. He still differed quite a bit from the Impressionists, though: Impressionists painted what they saw, but Van Gogh painted what he felt.⏎Van Gogh and Georges Seurat became very good friends, and
-
students often struggle just to keep up with the pace and the pressure. In order to complete what can be a daunting plate of assignments, students stay up late into the night and
-
The idea was that legislators would convene at a safe remove from voters and, thus insulated from the din of narrow interests and widespread but ephemeral passions, do what was in the long-term interest of their constituents and of the nation. Now information
-
the Drone, French philosopher Grégoire Chamayou seeks to comprehend how drones have revolutionised warfare and to deconstruct the various narratives at the heart of what has become a conceptual and legal landmine.⏎The drone’s philosophy can be summed
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: I have no data. The map is a, and the rest. The rest is a and open, but no longer than system, with some combination, and no clear map. The map is a, and the adjacent. The rest of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
-emerge. This phenomenon is described in leaflet 8: Direction of cure.⏎Other reasons for giving only one medicine⏎at
-
en, 52 to 48 percent. Clark's defeat was attributed to intense anti-Clark campaigning conducted by direct mail, mailgrams, and leaflet distribution during the final weeks of the campaign, attacking Clark for his positions on abortion
-
ever had an allergic reaction to a medicine.⏎How to take oxcarbazepine⏎- Before you start the treatment, read the manufacturer's printed information leaflet from inside the pack. It will give you more information about oxcarbazep
-
help our planet.⏎You will need:⏎- Scrap paper⏎- Old books/newspapers/leaflets etc⏎- Needle and thread⏎- Something pokey, like a screwdriver
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
- L2:111 weakly suppresses
- L0:2728 weakly suppresses
- L2:561 weakly excites numbers right after 'these' or 'first'
- L1:461 weakly suppresses
- L1:1155 weakly excites
- L1:1280 weakly excites
- L1:1469 weakly excites
- L1:2336 weakly excites
- L2:1030 weakly suppresses
- L2:549 weakly suppresses
- L2:49 weakly suppresses
- L1:1513 weakly suppresses
- L2:686 weakly suppresses
- L2:2395 weakly suppresses
- L1:1390 weakly suppresses
- L2:1086 weakly suppresses
- L2:722 weakly suppresses
- L2:2692 weakly excites
- L1:1913 weakly excites
- L0:31 weakly suppresses
- L1:1801 weakly excites
- L2:228 weakly excites
- L0:71 weakly excites
- L2:1481 weakly excites
Feeds later layers
- L4:2586 suppresses
- L9:75 weakly excites 'what' or 'how' after see or understand
- L4:3019 weakly suppresses
- L22:1146 weakly suppresses blank before a time or count
- L9:656 weakly excites
- L23:2694 weakly suppresses
- L6:502 weakly excites
- L10:2017 weakly suppresses
- L4:2380 weakly suppresses
- L15:3024 weakly excites
- L10:2893 weakly excites
- L17:625 weakly excites
- L12:2213 weakly suppresses
- L5:2609 weakly suppresses
- L4:1279 weakly suppresses
- L10:161 weakly excites
- L4:3061 weakly suppresses
- L16:1765 weakly suppresses
- L16:2993 weakly suppresses
- L6:1377 weakly suppresses
- L5:1251 weakly suppresses
- L8:1564 weakly excites
- L21:2054 weakly suppresses
- L11:421 weakly 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
- ordova
- uetype
- APON
- ..."); ⏎
- ibern
- itous
- 是一款
- звуч
- 您的孩子
- ."); ⏎
Suppresses
- 一刻
- duced
- acias
- nada
- ziehung
- siempre
- >[]
- DRAW
- Draw
- identally