unclear: pieces of medical and Latin-style 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. 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 weakly on the middle pieces of longer, technical words, such as 'submucosa' and 'oligodendrocytes'. The pattern is faint and the examples do not share a clear theme. Turning it up made the writing drift into unrelated descriptions.
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.
-
biting. Tail lesions are painful.⏎- Ear lesions: Ear lesions are a collective term for lesions in/at the ears of pigs, but also includes otohematomas, also called blood-ears. Ear
-
). Following the emperor Nero’s death in June 68, Titus was energetic in promoting his father’s candidacy for the imperial crown. Licinius Mucianus, legate of Syria, whom he reconciled with Vespas
-
illness and generally does not infect the nervous system. It reactivates later in life in people who have suppressed immune systems. The virus damages oligodendrocytes, the cells in the central nervous system (CNS) which produce my
-
wall consists of 4 layers of tissue. From deep (external) to superficial (internal) these are the serosa, muscularis externa, submucosa and mucosa. This layered arrangement follows the same general structure in all regions
-
, chief, enteroendocrine cells⏎Lamina propria: connective tissue⏎Muscularis mucosa: two smooth muscle layers⏎|Submucosa||Connective tissue, submucosal (Meissner’s)
-
smaller particles using chemical and mechanical digestion.⏎The stomach can perform these roles due to the layers of the stomach wall. These are the gastric mucosa, submucosa, muscularis externa and serosa. All parts of the GI tract
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: 2000000000000000000. I am... I am 20000000000000000000.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
D. de Caprona, The Role and Meaning of Structures in Genetic Epistemology. R
-
The day after the nationwide general strike was called on October 4, 1934, Lluís Companys, now president of the Generalitat, proclaimed the "Republic of Catalunya within the Federal Republic of Spain"—an
-
to cross the bridge just before the long tunnel. Source: Facebook⏎The efforts of those who managed the huge task of building the railway through the Sierra de Alfàbia – part of the substantial Sierra de Tramuntana mountains on the north
-
studies in more clinically relevant models. Proceedings of the National Academy of Sciences USA (PNAS) has published the study online this week.⏎Argentinian scientist Andrés Dekanty, a Juan de la Cierva researcher in the Mil
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
- L1:2677 weakly suppresses
- L2:2915 weakly suppresses
- L2:24 weakly excites
- L1:651 weakly suppresses
- L0:2537 weakly suppresses
- L1:2022 weakly excites
- L2:2601 weakly suppresses
- L1:2576 weakly suppresses
- L1:1688 weakly excites
- L2:2855 weakly suppresses
- L1:835 weakly suppresses
- L2:2737 weakly suppresses
- L2:2432 weakly excites
- L1:198 weakly suppresses
- L0:222 weakly excites
- L1:2937 weakly suppresses
- L2:833 weakly suppresses
- L0:1149 weakly suppresses
- L1:2378 weakly suppresses
- L0:253 weakly suppresses
- L1:1130 weakly suppresses
- L0:2664 weakly suppresses
- L1:488 weakly suppresses 'of' after group or combination nouns
- L1:685 weakly excites
Feeds later layers
- L5:30 suppresses
- L4:1473 weakly suppresses
- L7:2446 weakly suppresses
- L5:10 weakly suppresses
- L5:2110 weakly suppresses
- L19:2180 weakly excites
- L26:2656 weakly suppresses
- L6:2952 weakly suppresses
- L12:1416 weakly suppresses
- L9:735 weakly suppresses
- L17:1854 weakly excites
- L14:728 weakly excites
- L22:2328 weakly excites
- L6:2027 weakly excites
- L18:1001 weakly excites
- L6:1723 weakly suppresses
- L20:103 weakly suppresses
- L4:1888 weakly suppresses
- L6:69 weakly excites
- L5:1142 weakly suppresses
- L26:318 weakly excites
- L4:1543 weakly suppresses
- L4:1265 weakly suppresses
- L20:658 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
- urma
- держива
- abay
- alom
- ecause
- alpha
- suggest
- лей
- eden
- raud
Suppresses
- etical
- 化进程
- reira
- gien
- ernity
- 庆典
- étique
- aunch
- öff
- Launching