mid-word fragments in 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. 12 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 middle piece of a longer word, often a scientific or technical name split into parts, such as "ribofl[avin]", "subf[amilies]" or "list[eria]". The examples are varied, so the pattern is only loosely defined.
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.
-
Files?⏎We can zip and unzip files in Windows 10 easily. To zip files:⏎Step
-
and recombineering in E. coli". PLoS ONE 8 (2): e57075. . . http://www.plosone.org/article/info%3Adoi%2F10.137
-
500,000 species in existence.⏎The superfamily is presently divided into 19 separate families with more than 90 subfamilies recognised. Currently
-
or raised to move the cutting blade to or away from the material to be cut.⏎A drill press with a moveable arm that can be swung to various positions on the work
-
cheese⏎Cottage cheese has a low fat and calorie content. It is full of calcium, phosphorus, selenium, vitamin B12, riboflavin (vitamin B2
-
full potential by passively listening to the teacher and filling in worksheets they must engage in This is called a mixed number which is a mixture of a whole bbc bitesize to add the two fractions together print out the worksheet below from twinkl or copy
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 question is about the same day, and the same day is the same day." The same day. The same day. The same day. The same day. The same day. The same day. The same day. The the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
immediately. Just as I say, “ Imagine the tip of your left pinky finger… where the nail meets the skin…” Do you think of it? Did
-
, because the impairments in shifting gaze and attention that we found in 7-month-olds may be a fundamental problem in autism,” said Robert T. Schultz, PhD, director of the Center for Autism Research at CHOP and a co-author on
-
grounds. They will make the return dry season journey in June, to take advantage of the permanent water of the Tarangire River.⏎Numbers of wildebeest have fallen from 50,000 to 6,00
-
and young adults,” said Jennifer Duke, PhD, senior research public health analyst and co-author of the study. “And unfortunately
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
- L8:1135 strongly excites
- L10:1219 excites
- L11:1385 excites
- L8:391 excites
- L9:804 excites
- L11:1336 suppresses
- L9:1840 excites
- L10:1231 excites
- L7:845 excites
- L11:640 excites
- L9:814 excites
- L11:849 suppresses
- L10:2456 excites
- L9:1048 suppresses
- L11:844 suppresses
- L11:1119 suppresses
- L9:581 weakly suppresses
- L9:968 weakly suppresses
- L11:60 weakly excites
- L11:1500 weakly excites
- L5:468 weakly excites
- L10:1674 weakly excites
- L10:882 weakly excites
- L8:2691 weakly excites
Feeds later layers
- L13:1567 excites
- L17:764 excites
- L17:2851 excites
- L13:1661 suppresses
- L20:1426 excites
- L13:2538 excites
- L13:1892 suppresses
- L14:1724 excites
- L13:2227 weakly excites
- L16:1461 weakly excites
- L13:2528 weakly excites
- L18:2449 weakly excites
- L18:1009 weakly suppresses
- L13:2026 weakly excites
- L24:2664 weakly suppresses
- L14:2069 weakly suppresses
- L17:1647 weakly excites
- L15:677 weakly suppresses
- L14:2377 weakly excites
- L15:2709 weakly excites
- L13:1285 weakly excites
- L13:1536 weakly excites
- L14:1091 weakly suppresses
- L16:988 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
- 诸如
- OTA
- Wolfe
- 相对
- rst
- aversal
- azed
- ivity
- 不论是
- ATURE
Suppresses
- /W
- /T
- /P
- 毛孔
- /S
- /R
- /E
- jo
- /AFP
- /V