'de' inside Spanish names
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. 19 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 fires on the word 'de' (meaning 'of') when it sits inside a Spanish or Portuguese name, such as 'Nuestra Señora de Atocha' or 'Hombres de maíz'. Turning it up made the model write unrelated text about numbers and queues, so the steering does not show this theme.
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.
-
spelling), the Mayan equivalent of the Bible, which states that humans were literally created from maize, to Miguel Ángel Asturias‘ novel Hombres de maíz (Men of Maize) which is one of the best novels
-
án Cortés, and fought in the initial Spanish conquest of Mexico, but it is not clear whether his wife went with him. Some modern sources indicate that María de Estrada only arrived on the mainland in April 1520, with
-
but as the following discoveries prove, the hidden secrets of the seabed are about much more than just money.⏎Billion-dollar wreck: Nuestra Señora de Atocha⏎In July 1985, after 17
-
the Peace Corps. He created the Foundation Heart Ventures-Grupo de Ap<|endoftext|>is an elementary subatomic particle with infinitesimal mass⏎(less than
-
1934, the Franks r…⏎E-pasta adrese, uz kuru nosūtīt darba saiti:⏎Saite uz darbu:<|endoftext|>Branch of biology that groups and names organisms based on studies of
-
depending on the formed arrangement on the board between their piece and the other players' pieces (conjunction, opposition, quarter, etc.)⏎Often in this Libro de los juegos, a Table game was associated to a Chess game, through the die
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: the real and a one-up of the a. s. a. f. in the s. a. h. of the 2021-11-9-1. the a. a. f. is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Museum of Alberta, Musée national des arts et traditions populaires, Département de l'organisation des syst
-
mourning, as in Victorian times? I noticed it during a conversation with Bernard Lestienne, a French Jesuit who works in Brazil at the Instituto Brasileiro de Desenvolvimento. He was in New York for a conference, and as
-
school's large roof generates more solar power than the school consumes, especially in the summer months and during weekends. The NPO APERe (Association pour la Promotion des Energies Renouvables) and Brussels grid operator Sibelga selected the
-
04)⏎(4) « La compression veineuse est le traitement de base des affections ve<|endoftext|>With concern over pesticide use on produce as well
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
- L18:72 strongly suppresses
- L18:550 suppresses
- L15:65 excites
- L18:2749 suppresses
- L8:15 suppresses unclear: mid-phrase words across many languages
- L18:560 excites
- L17:1021 excites
- L16:477 excites
- L10:39 suppresses
- L3:6 suppresses
- L15:59 excites
- L7:122 excites
- L9:453 suppresses
- L18:66 excites
- L17:35 suppresses
- L11:840 weakly excites
- L5:136 weakly suppresses
- L14:381 weakly suppresses
- L18:178 weakly excites
- L17:98 weakly excites
- L18:567 weakly excites
- L18:566 weakly excites
- L17:56 weakly excites
- L0:1691 weakly suppresses
Feeds later layers
- L20:1835 suppresses
- L20:2754 suppresses
- L20:1972 suppresses 'of' in 'of course'
- L20:1446 suppresses
- L23:1939 suppresses
- L25:1625 suppresses
- L22:1798 suppresses
- L21:68 suppresses
- L20:519 suppresses
- L23:1015 weakly excites
- L21:1420 weakly excites
- L21:670 weakly suppresses
- L20:986 weakly excites
- L21:1021 weakly suppresses
- L20:337 weakly suppresses
- L22:404 weakly excites
- L22:789 weakly excites
- L20:1846 weakly suppresses
- L24:163 weakly excites
- L25:994 weakly excites
- L20:2487 weakly suppresses
- L20:1991 weakly suppresses
- L21:2342 weakly suppresses
- L22:2889 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
- ⏎→→⏎
- uci
- ода
- те
- uc
- يا
- ids
- ⏎→⏎
- لع
- muff
Suppresses
- innacle
- 停放
- 买到
- emotion
- velt
- berman
- nika
- section
- resenter
- RowIndex