mid-word fragments of long 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. 26 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 tends to fire on a piece of a longer word that has been split up, such as 'Ful' in 'Fuligulinae' or 'ker' in 'handkerchief'. Turning it up did not produce an obvious theme; turning it down made the model repeat the word 'moments'.
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.
-
ustrades were ornately molded in pre-cast faux stone and lighting was generously provided from electric globes on eleborate cast-iron poles… large plinths marked the foot of the structure at both ends. The four rusticated pl
-
Dictionary and Cyclopedia⏎- n. A sea-duck of the subfamily Fuligulinœ and genus Clangula; a garrot. The common golden
-
54th Union Army regiment, the first all-volunteer black regiment. The sculpture, dedicated to black soldiers, is located in Boston
-
ant that contains at least 70% alcohol.⏎- Other recommendations⏎- Avoid keeping your smartphone and handkerchief in the same pocket. Frequently wash your hands, especially when you are in
-
early as February-April. Crops that need to be raised in warmth either indoors, in a greenhouse or under cloches e.g. tomatoes, peppers, courgettes are planted out after being ‘hardened-off’ (kept out during
-
Platonis et Aristotelis / Comparison of Plato and Aristotle (1458). This book accused the humanist circle patronized by the Greek-born Platonist Cardinal Johannes Bessar
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: moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments moments
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
angles formed by the cords on the disk—the sum of the forces and the equilibrium conditions. Provenance: Lorraine collections.<|endoftext|>White sauce,
-
and limitations cannot be understood only through electrical specifications and nameplate data, because redundancy is included in the design. In order to obtain a thorough understanding of electromagnetic force/torque generation, one must consider thermal characteristics
-
of oil and sweat in the skin. It becomes scaly and even wounds and scars take more time to restore to health.⏎Repetitive facial expressions cause muscle contractions. As the elasticity of the skin diminishes,
-
submaximal voluntary isometric contractions with the arm in a pronated position for ten healthy and young male subjects
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
- L23:2530 suppresses
- L20:2641 suppresses
- L16:39 suppresses
- L22:1754 weakly excites
- L20:2922 weakly excites
- L10:1486 weakly excites
- L24:105 weakly suppresses
- L3:3020 weakly excites
- L11:2029 weakly excites
- L0:2757 weakly excites
- L5:2165 weakly suppresses
- L20:295 weakly excites
- L14:717 weakly excites
- L23:115 weakly excites
- L10:876 weakly suppresses
- L24:1182 weakly suppresses
- L5:593 weakly suppresses
- L9:2087 weakly suppresses
- L6:1400 weakly excites
- L22:273 weakly excites
- L16:110 weakly suppresses
- L18:633 weakly excites
- L3:388 weakly suppresses
- L0:2578 weakly suppresses
Feeds later layers
- L27:422 weakly suppresses
- L27:1073 weakly suppresses
- L27:1361 weakly excites
- L27:1246 weakly excites
- L27:2136 weakly suppresses capital M inside abbreviations
- L27:1816 weakly suppresses
- L27:1326 weakly excites
- L27:3036 weakly suppresses
- L27:1915 weakly suppresses
- L27:303 weakly suppresses
- L27:586 weakly suppresses
- L27:2012 weakly suppresses
- L27:489 weakly suppresses
- L27:1201 weakly suppresses
- L27:2786 weakly suppresses lowercase 'm' inside split words
- L27:44 weakly suppresses
- L27:2413 weakly excites
- L27:945 weakly suppresses
- L27:2138 weakly excites
- L27:2114 weakly suppresses
- L27:2113 weakly suppresses
- L27:1117 weakly suppresses
- L27:2070 weakly excites
- L27:2544 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
- Caller
- .newaxis
- edly
- vari
- 弦
- 正宗
- дей
- ;$
- alan
- directory
Suppresses
- moments
- moment
- mom
- tor
- Moments
- Mom
- moment
- Moment
- 矩
- mom