"of" inside "American Journal of"
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. 8 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 "of" in journal names, especially "American Journal of", which is the usual next word in a scientific journal title. Turned up, the model's text got unrelated and wandering, so the steering does not show this pattern.
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.
-
's Oahu flattens from an island into a low-lying seamount. Researchers writing in the upcoming Feb. 15 issue of the journal Geochimica et Cosmochimica Acta estimate that the volcanic island will continue
-
Dispersal of Aedes Aegypti in an Urban Area After Blood Feeding as Demonstrated by Rubidium-Marked Eggs," American Journal of Tropical Medicine Hygiene⏎52 (2):177-17
-
a National Science Foundation (NSF) grant, Kaur and her team have published an account of their ground-breaking research in the August edition of the American Journal of Primatology. The issue features a special section on ''Disease Transmission, E
-
study was published in Physical Review letters on 6. November 2017 and on the cover of
-
round," he said.⏎The results are published in the American Journal of Hypertension.<|endoftext|>Nicolaus Copernicus’ 540
-
and aids. The study, published in the May issue of the American Journal of Preventive Medicine, also finds men and those without a usual source of medical care
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: "I'm not going to get a check. Just a check. Check 11110001110001100000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
arn). It is estimated neutrino cross-section for interaction increases linearly with energy of incident neutrino.⏎Reference: Griffiths, David, Introduction to Elementary Particles, Wiley, 1987.⏎See also: Antineutrino
-
and terrace deposits along the Cimarron River from Freedom to Guthrie in northwestern Oklahoma⏎U.S. Geological Survey, Water Resources Division ;⏎Available from the Earth Science Information Center, Open-file Reports Section,<|endoftext|>Source: California Invasive
-
fractals, dynamical systems, patterns of nature created through simple rules, and aesthetic graphics drawn from the universe of mathematics and art.⏎Chaos and Fractals is divided into six parts: Geometry and Nature; Attractors; Cellular Autom
-
activity disorder (ADHD) in North American children compared to placebo. A 2015 study by M.A. Stein and colleagues in the journal CNS Drugs extended this research, determining that guanfacine also improved academic and social functioning
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
- L4:983 suppresses
- L3:1758 suppresses
- L5:415 weakly suppresses
- L7:162 weakly suppresses
- L6:13 weakly excites
- L2:1334 weakly suppresses
- L2:70 weakly excites
- L2:589 weakly suppresses
- L0:149 weakly suppresses
- L1:883 weakly suppresses
- L3:2526 weakly suppresses unclear: mid-word pieces of long words
- L6:2858 weakly suppresses
- L5:960 weakly suppresses
- L7:414 weakly excites
- L6:12 weakly excites
- L6:7 weakly suppresses
- L2:313 weakly excites
- L2:508 weakly suppresses
- L1:2877 weakly excites
- L7:320 weakly suppresses
- L1:2642 weakly suppresses
- L3:2932 weakly suppresses
- L4:389 weakly suppresses
- L6:64 weakly suppresses
Feeds later layers
- L13:1513 excites
- L10:952 suppresses
- L11:393 excites
- L9:1026 excites
- L16:2151 excites
- L11:122 suppresses
- L10:457 weakly suppresses
- L10:1316 weakly excites
- L9:33 weakly suppresses
- L16:498 weakly suppresses
- L15:646 weakly excites
- L12:2305 weakly suppresses
- L18:819 weakly excites
- L11:1753 weakly excites
- L13:1001 weakly suppresses
- L12:576 weakly excites 'of' in fixed phrases like table of contents
- L9:2846 weakly suppresses
- L10:227 weakly excites middle of fixed names and phrases
- L9:716 weakly excites
- L11:966 weakly suppresses
- L9:845 weakly excites
- L10:2026 weakly suppresses
- L15:1607 weakly excites
- L20:244 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
- مبر
- remain
- �
- remains
- итет
- met
- isch
- \xb
- beh
- ству
Suppresses
- зол
- indle
- 涩
- ered
- cession
- Seeds
- 妞
- ene
- росс
- goodbye