middle of technical terms in science text
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. 10 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 the piece of a scientific or technical name that comes after the start, such as 'one' in 'pheromone', 'ome' in 'proteasome', or 'Survey' in 'Geological Survey'. It is mostly about word completion in science writing. Turning it up made the model repeat simple sentences, and turning it down gave unrelated story text.
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.
-
only the males of the species got attracted towards the light source, mistaking it for female moths. Female moths secrete a chemical called pheromone, which has a scent and acts as sex appeal for moths. The infra
-
either can’t make enough insulin to keep blood sugar levels down (type 1), or can’t effectively use the little insulin that it does produce (type 2).⏎First, to understand what happens in the bodies of people who have diabetes,
-
of the gardenia’s foliage. They thrive when the weather is warm and dry. Whiteflies feed on the gardenia’s sap and secrete a sticky substance known as honeydew. This, in turn, attracts sooty mold,
-
products, have been detected in about one-third of the wells sampled by the National Water-Quality Assessment (NAWQA) Program of the U.S. Geological Survey (USGS).1 Chloroform and other trihalomethanes,
-
endonuclease from the 2009 pandemic influenza strain using the high intensity X-ray beams at the European Synchrotron Radiation Facility (ESRF).⏎The researchers were able to determine the 3D atomic structure of the enzyme
-
itin side chains onto whatever protein is in close proximity, and when a protein is thus ubiquitinated, it is marked for destruction by the cell’s proteasome.⏎This means that you can,
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 have been, so, and I will be, and I will be, and I will be, and I will be, and I will be, and I will be, and I will be, and I will be, and I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
&D and Opus 12 Present: “Utilization of CO2 Emissions to Make Renewable Fuels and Chemicals"⏎Wednesday, July 29, 2020⏎10:00 AM PDT.⏎Opus
-
; hence, the popular phrase "Proverbs 31 Woman," while ignoring the advice for men. However
-
SS⏎Using an SPSS data file, you'll test hypotheses with t-tests and regression analysis to predict risks and trends.⏎Tuesday, Oct. 14, 6-8 p.m.⏎If you are registered for a quantitative course at
-
the Abnormal Silverline (Cigaritis abnormis), the Giant Red Eye (Gangara thyrsis) and the Large
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
- L9:994 excites
- L9:571 excites
- L5:2491 suppresses
- L8:634 excites
- L8:621 excites
- L8:1219 weakly suppresses
- L6:1180 weakly suppresses
- L7:1019 weakly suppresses
- L7:1924 weakly suppresses
- L8:2142 weakly suppresses
- L9:742 weakly suppresses "of" after an agency or department name
- L8:658 weakly suppresses
- L5:320 weakly suppresses
- L9:247 weakly excites
- L7:1821 weakly suppresses
- L7:931 weakly suppresses
- L9:1045 weakly suppresses
- L8:877 weakly excites
- L5:666 weakly suppresses
- L1:779 weakly excites
- L9:694 weakly suppresses
- L4:10 weakly excites
- L8:699 weakly excites
- L9:993 weakly excites
Feeds later layers
- L12:590 suppresses
- L15:2570 suppresses
- L12:2477 excites
- L11:399 suppresses
- L13:761 weakly excites
- L11:237 weakly suppresses
- L11:1025 weakly excites
- L11:1239 weakly suppresses
- L11:767 weakly suppresses
- L20:1684 weakly suppresses
- L22:847 weakly excites
- L15:1478 weakly suppresses
- L12:778 weakly suppresses
- L16:1016 weakly excites
- L12:143 weakly excites
- L15:669 weakly suppresses
- L13:385 weakly excites
- L18:2049 weakly excites
- L12:1825 weakly suppresses
- L16:1223 weakly suppresses
- L20:333 weakly suppresses
- L12:1092 weakly excites
- L12:1700 weakly suppresses
- L14:78 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
- olds
- sons
- AYS
- annis
- actics
- IELDS
- nger
- yssey
- ****
- 丝
Suppresses
- Hogan
- vetica
- ToDevice
- VELO
- 的路上
- PointerException
- DOCUMENT
- 有一次
- iqueta
- ukkit