'ic' inside longer 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. 25 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 syllable 'ic' when it sits in the middle of a word or a name, such as 'Avic|enna', 'Kiczenski' or 'longic|rus'. Steering it up made the model fall into repeated 'sum of the sum' loops, and steering it down produced a mix of unrelated 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.
-
carried out research into the kinds of scoliosis treatments available at the time Richard was alive.⏎They included looking at 11th century Persian polymath Avicenna's theories on using traction.⏎Avicenna also advocated the massage techniques pract
-
on how you can make Orion the “star” of your nightscape photography. And we’ve added many excellent photos of Orion in the landscape by MaryBeth Kiczenski, Ralf Rohner, and Arnaldo Zannoni. So
-
eastern North America becomes British territory except St-Pierre and Miquelon Islands (France). British colonial governments for Quebec, Newfoundland (with Île d'Anticosti and Îles de la Madeleine), Nova Scotia (including present-day N
-
tenella. Coralline algae are comparable in abundance to stony corals. Other macroalgae include Halimeda tuna, Dictyota divaricata, Loboph
-
imulation (rTMS): Brain activity is triggered through weak electric currents, changing the excitability of corticospinal and corticocortial pathways.<|endoftext|>Doing u-substitution twice
-
central to north western Canada. There are four subspecies described by Hall (1981) with two (M.v.interior and M.v. longicrus) occurring in California.⏎This is a common bat found in all the mountain
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 top of a spring spring is on a top weight. A spring is spring under weight. An abocess is an abocess. A spring is spring. A spring is spring. A spring is spring. A spring is spring
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
plan, and institutional) and key audiences (graduate students, faculty, postdoctoral scholars, administrators, organizations) served.⏎John Locke ‘s work Some Thoughts Regarding Education was written in 1693 and still reflects conventional education priorities in
-
. A range of new features have been added to the new edition including: new structure to better meet the needs of teaching qualitative research; a new chapter on Grounded Theory plus updated coverage on the full range of<|endoftext|>Stairs - Did You Have
-
as distinguished from the lower cognitive faculties, sense, imagination, and memory, and in contrast to the feelings and desires. Reason comprises conception, judgment, reasoning, and the intuitional faculty. Specifically, it is the intuitional faculty,
-
Overton, Reasoning, Logic, and Thought Disorders: Deductive Reasoning and Developmental Psychopathology. D.P. Keating, Structuralism, Deconstruction, Reconstruction: The Limits of Reasoning.⏎".....Of interest to serious
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
- L3:698 excites
- L4:2456 weakly excites
- L24:3048 weakly suppresses
- L21:2513 weakly excites
- L20:2163 weakly excites
- L11:1148 weakly excites
- L19:1533 weakly excites "and" inside fixed institution names
- L22:2433 weakly suppresses
- L20:340 weakly excites
- L21:752 weakly suppresses
- L9:2482 weakly excites
- L17:156 weakly excites
- L16:2586 weakly excites
- L21:1084 weakly excites
- L10:1872 weakly excites
- L1:1246 weakly excites "word" after a language origin phrase
- L6:1145 weakly suppresses
- L22:1702 weakly excites
- L7:2317 weakly excites
- L4:1103 weakly suppresses
- L23:2281 weakly excites
- L18:1848 weakly excites
- L8:88 weakly excites line breaks inside wrapped sentences
- L4:92 weakly excites
Feeds later layers
- L26:324 weakly excites
- L26:1274 weakly suppresses
- L27:1304 weakly excites
- L27:1725 weakly excites
- L26:1444 weakly suppresses
- L27:1778 weakly excites
- L26:2702 weakly excites
- L26:2296 weakly excites
- L27:1286 weakly excites
- L27:2990 weakly excites
- L26:410 weakly excites
- L26:2424 weakly suppresses
- L27:2976 weakly excites
- L26:472 weakly suppresses
- L27:1800 weakly excites
- L27:2289 weakly excites
- L26:406 weakly excites
- L27:1327 weakly excites
- L27:2115 weakly excites
- L27:219 weakly suppresses
- L27:1075 weakly suppresses
- L27:814 weakly excites
- L27:979 weakly excites
- L27:1449 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
- ouce
- Yue
- 思想
- quares
- dressed
- uming
- >*/⏎
- xfd
- uju
- kor
Suppresses
- =top
- 阄
- 亿元以上
- ab
- ович
- 公
- cen
- GetName
- she
- →ns