word right before a term's name
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. 4 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 that introduces a name, like 'called', 'known as' or 'named', right before the thing being named. When it was turned up, the text often drifted toward phrases like 'is called' or 'the system is called'.
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.
-
daily functioning. This condition, known as mild cognitive impairment, progresses to Alzheimer’s in roughly 10% to
-
its axis. In such movement particle has charge, produce electric waves ( waves property of particle) and its speed ( frequency) is : c>1. We call such particle - ? electron? and its energy is: E=h*f.⏎In
-
It’s called a phone booth.⏎Phone booths used to be everywhere, providing an office for agent
-
Peritoneal Cancer: Causes, Symptoms, Diagnosis, …⏎(2 days ago) The inside of your abdomen has a slick, transparent lining. It’s called the peritoneum. When
-
of you who don’t know what a genetic algorithm is, it is basically an optimisation method that mimics evolution. You have a set of candidate solutions (called a population). The best solutions get combined with each other (mating) and
-
, deer have food they avoid, and<|endoftext|>Seven miles to the south west of Sheffield there's a piece of wild moorland surrounded by roads. It's called Big Moor. I estimate that if you circumnavigated it on foot by
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: 20% of the month, 10% of the year, 30% of the year. This month, 10% of the year, 10% of the year, 10% of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
For backgrounds, remember yin and yang. If you use a light background, use dark or black colour for text. If you have a dark background, use white or light colour for text. The text needs to stand out from the background.⏎
-
.⏎In the examples shown, we used stickers (rooms and stars) from our In the house sticker pack. We drew the other symbols onto blank stickers.⏎#2 Outdoor I-Spy⏎Use your daily walk to play I- Spy and spot
-
1785-87 (circa)⏎Publication: Troisieme Voyage De Cook, ou Voyage a l'Ocean Pacifique…⏎Color: Black & White⏎The copper engravings from
-
. It produces dense clusters of hanging tubular flowers, the colour varies from white, pink, red, green to yellow. The corolla is a partially inflated
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
- L3:2662 weakly suppresses
- L2:75 weakly suppresses
- L3:272 weakly excites
- L3:476 weakly excites
- L3:1001 weakly suppresses
- L3:967 weakly excites
- L0:1734 weakly excites
- L2:2141 weakly excites
- L3:1773 weakly suppresses
- L0:2837 weakly suppresses
- L3:59 weakly excites
- L2:36 weakly excites
- L1:87 weakly suppresses
- L0:3029 weakly excites
- L0:2132 weakly excites
- L0:1679 weakly excites
- L1:691 weakly suppresses
- L2:2201 weakly suppresses
- L2:1275 weakly suppresses
- L3:1757 weakly suppresses
- L2:111 weakly suppresses
- L1:1689 weakly excites
- L3:2893 weakly suppresses
- L2:2630 weakly excites
Feeds later layers
- L5:944 excites
- L6:1197 excites
- L8:1224 excites
- L12:2252 excites
- L5:2715 suppresses
- L11:1231 weakly excites
- L5:1317 weakly excites 'a' right after 'being'
- L22:2344 weakly suppresses number after 'account for' or 'over'
- L7:384 weakly suppresses
- L5:647 weakly excites
- L12:2902 weakly excites
- L10:972 weakly suppresses
- L5:1183 weakly excites
- L15:1977 weakly excites
- L21:945 weakly excites
- L6:1258 weakly excites "mean by" in questions and definitions
- L18:1329 weakly excites
- L8:1348 weakly suppresses
- L17:2516 weakly excites
- L8:2122 weakly excites hand and body action verbs in instructions
- L6:662 weakly excites
- L12:74 weakly excites
- L7:1128 weakly suppresses
- L18:2718 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
- 叫做
- called
- named
- 叫
- 简称
- �
- 称号
- 名
- name
- 名字
Suppresses
- Wilkinson
- enville
- spilled
- ügen
- vore
- prefs
- ordan
- ologne
- opacity
- :normal