lowercase 'm' inside split words
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. 27 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 letter m when it appears inside a word that has been broken into pieces, like the m in mastiche or mystery. Turning it up or down gave repeated filler words such as one or none, so its effect on the text is unclear.
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.
-
time to complete the ultimate objective which is the discovery of the cures.⏎Pandemic board game, Z-MAN Games, Retrieved from https://www.zmangames.com/en/games/pandemic/⏎A game
-
, we have to start with ancient Greece, where citizens routinely chewed on the resin (or sap) from the Mastic tree. This substance was dubbed “mastiche
-
century that America finally implememented the promise of its ideals to all its citizens and in the process saved
-
23:34:592021-10-20 23:34:59assignment 2⏎mynursingpaper.com is a unique service that provides guidance with different types of
-
with a bleach solution and to take down your feeder for a week if you notice sick or dead birds.⏎Officials still haven't figured out the cause of the 'mystery' illness that has killed thousands of songbirds. However, the state says
-
berg says.⏎Think fat just hangs around and does nothing? It doesn’t⏎By Nanci Hellmich, USA TODAY⏎PHOENIX — Most people think of
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:onononononononononononononononononononononononononononononononononononononononononononononononon
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
otes (sweet potatoes), jicama (a tuber also known as sengkwang, yam bean, singkamas, Mexican turnip), amaranth (also called pigweeds) and mucuma (a tropical leg
-
notice him leaving the party early?► see thesaurus at seeNotice is not used in the progressive. You say: I notice you’re wearing glasses. ✗Don’t say: I’m noticing you’re wearing glasses.2 → be
-
the environmental and social risks of increased oil and gas production. Despite the heated debate, neither side has a monopoly on the facts. In this book, Daniel Raimi gives a balanced and accessible view of oil and gas development, clearly and thoroughly
-
.⏎Free homework help from online tutors⏎Every Day 1-10 p.m.<|endoftext|>Contact: Kara Villamil, or Mona S. Rowe⏎FOR IMMEDIATE RELEASE⏎DECEMBER 3, 1997⏎Images
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
- L23:270 strongly excites
- L19:334 suppresses
- L20:2203 suppresses
- L24:1190 excites
- L21:1679 excites
- L22:1138 suppresses
- L17:230 excites
- L18:553 excites
- L1:16 excites
- L2:1563 excites
- L16:73 excites
- L0:73 excites
- L1:35 suppresses
- L26:550 suppresses
- L26:2511 suppresses
- L8:250 weakly suppresses
- L8:14 weakly excites
- L4:1372 weakly suppresses
- L0:741 weakly excites
- L16:66 weakly suppresses
- L14:95 weakly excites
- L7:922 weakly suppresses
- L7:291 weakly excites
- L10:311 weakly suppresses
Feeds later layers
None: this is the last layer.
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
- ��
- ever
- ee
- (one
- 一头
- Matter
- eh
- 一团
- matter
- est
Suppresses
- inis
- inus
- inem
- ent
- on
- iss
- in
- anos
- an
- utation