unclear: mid-word subword pieces
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 mixed: they peak on different tokens. 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 fires on a random-looking piece of a longer word, like the 'stein' in a surname or the 'gone' in 'bygone'. When it was turned up, the model's writing drifted toward plain everyday sentences with no clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
link between blood-vessel health and Alzheimer’s, but this study is “the first one that shows that treating mild cognitive impairment can have results,” says Anton Porsteinsson, MD, director of the Memory Disorders Clinic at the University of Rochester School
-
apses, they are merged with the past and become a part of history.⏎Hence, history, in this sense, is the knowledge of the bygone incidents, events, conditions and circumstances of the people in the past. Biographies
-
the feel of a bygone era. Time has effectively stood still along these winding country roads, in the small
-
parts of the lymphatic system.|⏎|Germ Cell Tumor||GTC. A neoplasm derived from germ cells., normally presenting inside the gonads (ovary and testis).|⏎|Blastoma||More common in
-
998).⏎The South Africa Education system, which is grounded on the country’s Constitution, views learners as citizens who can secure a new democracy (Joubert, 2009). Learners’ voices should therefore be heard as
-
enteritis. A bacteria called clostridia was isolated as the organism responsible.⏎Necrotizing enteritis is associated with poor hygiene and an inadequate diet, and it was
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: A. The numbers 1, 2, 3, 4, 5 are all related to me, and I need to think about the following question: A. The numbers 1,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2:15|⏎|Last Modified:||11 May 2011 04:
-
the largest family is the Eulophidae with nearly 4,500 included species followed closely by the Encyrtidae and Pteromalidae. The phylogenetic relationships between the
-
“yarrow” is an Anglo-Saxon word that means “to prepare” because it was
-
decrease in pressure, thus a net downward force on the paper.<|endoftext|>On May 28, 1828, began the official life of Kaspar Hauser, a young man who appeared mysteriously in the streets of Nure
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
- L2:49 suppresses
- L0:1224 suppresses
- L0:178 suppresses
- L1:87 suppresses
- L0:1301 suppresses
- L2:166 weakly suppresses
- L2:527 weakly excites
- L2:2722 weakly suppresses
- L1:853 weakly excites
- L2:471 weakly suppresses
- L0:2565 weakly suppresses
- L0:1060 weakly suppresses
- L1:1664 weakly suppresses
- L0:894 weakly excites
- L2:70 weakly excites
- L2:1321 weakly suppresses
- L1:1029 weakly excites
- L2:41 weakly suppresses
- L2:2093 weakly suppresses
- L1:69 weakly excites
- L1:951 weakly excites
- L2:1852 weakly excites
- L2:641 weakly suppresses
- L2:1334 weakly suppresses
Feeds later layers
- L7:982 suppresses
- L5:1677 suppresses
- L10:2623 excites
- L4:834 weakly excites
- L18:3009 weakly suppresses
- L21:764 weakly excites
- L18:581 weakly suppresses
- L9:2611 weakly excites 'is' right after 'there'
- L10:1032 weakly suppresses
- L21:1823 weakly suppresses
- L13:1879 weakly excites
- L6:3013 weakly excites
- L11:119 weakly excites
- L10:1413 weakly suppresses
- L22:695 weakly suppresses
- L25:2374 weakly suppresses
- L18:2959 weakly excites
- L27:91 weakly excites
- L23:1005 weakly suppresses
- L6:2429 weakly suppresses
- L5:504 weakly excites
- L22:1039 weakly excites
- L16:1297 weakly suppresses unclear: document ends and answer colons
- L22:1534 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
- house
- 薨
- ply
- 享
- house
- 贵
- 下车
- ier
- 堌
- สวย
Suppresses
- omorphic
- redits
- tober
- itore
- })(
- charts
- 。",⏎
- .refs
- IMPORTANT
- 舌头