the third digit of years like 201x or 202x
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. 22 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 third digit of a year, such as the "2" in 2021 or 2023, especially in dates and citations. Turned up, the model produced repeated phrases and questions, and turned down it gave generic everyday 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.
-
identifiers, Wikipedia articles with PLWABN identifiers, Wikipedia articles with RKDartists identifiers, Wikipedia articles with<|endoftext|>Date: 4th Oct 2021 @ 2:48pm⏎In our Fridays Are lesson, we
-
.gov/opub/ted/2010/ted_20100805.htm (visited November 27, 2021).<|endoftext|>- Managerial Economics Tutorial⏎- Managerial Economics - Home⏎
-
.C.) known as well as the Saite Dynasty.<|endoftext|>Baby Mantis Shrimp Perceive UV Light Like Their Parents⏎- Published6 Sep 2022⏎- Author Ravindra Palavalli Nettimi⏎- Source
-
? Scholarship application essay example this column presents those 10 problems and provides solutions to each. how do you know? Apr 29, 2020 · problem-solving strategies
-
for acyclic graphs of pointers.<|endoftext|>The Death And Life Of The Great Lakes Pdf⏎- and pdf⏎- Tuesday, June 1, 2021 2:11:00 AM⏎- 1 comment⏎
-
⏎to access all content.<|endoftext|>Great Slaty becomes World's Largest Woodpecker as Ivory Billed Goes Extinct⏎22 November 2021 | Vibhav peri | Crises⏎The Great Slaty Wood
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: when you're doing an practice or trying to get back to the correct or successful action when you're not doing or trying is trying, or when you're on the same or trying to get back to the same or successful action when you're
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
much abridged summary of the relevant facts with a⏎Asturian Spanish pronominals (Corbett, 2000:124-146, García González, 1979)⏎3rd sing
-
made atonement for the dead, so that they might be set free from their sin.” II Maccabees 12:42-45⏎This belief in resurrection will become known as the Resurrection of Saints…”
-
October 3<|endoftext|>Fusion Science and Technology / Volume 43 / Number 1 / January 2003 / Pages 128-133
-
of diarrheal disease (incidence rate ratio = 1.39, 95% confidence interval: 1.03, 1.87), but only after a dry period. After wet periods, the risk
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
- L19:1260 excites
- L1:1404 weakly suppresses
- L6:1938 weakly excites
- L16:3018 weakly excites
- L21:2965 weakly suppresses
- L10:2435 weakly excites
- L10:2924 weakly excites
- L21:2468 weakly suppresses gemstones and jewelry words
- L6:1912 weakly excites
- L20:1853 weakly suppresses
- L0:88 weakly excites
- L17:1205 weakly suppresses
- L21:1470 weakly suppresses
- L5:2990 weakly suppresses
- L8:2188 weakly suppresses
- L4:2561 weakly suppresses
- L14:994 weakly excites
- L0:2531 weakly suppresses
- L9:1895 weakly excites
- L1:2342 weakly excites
- L4:1135 weakly excites
- L5:11 weakly excites
- L9:2206 weakly excites
- L17:257 weakly suppresses
Feeds later layers
- L25:845 weakly suppresses
- L27:2431 weakly excites
- L25:1766 weakly excites
- L25:458 weakly excites
- L23:1977 weakly suppresses
- L25:1010 weakly excites
- L27:572 weakly suppresses
- L23:1026 weakly excites
- L26:1348 weakly excites
- L24:529 weakly excites
- L24:1191 weakly suppresses
- L24:1198 weakly excites
- L25:58 weakly excites colon in clock times
- L25:3070 weakly suppresses
- L24:2000 weakly excites "United" in US and UN names
- L25:1862 weakly excites
- L25:1613 weakly suppresses
- L24:2270 weakly excites
- L27:1743 weakly excites
- L23:2240 weakly excites
- L25:2113 weakly excites
- L24:1647 weakly excites
- L23:2016 weakly excites
- L26:310 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
- asename
- enstein
- oda
- 但是我
- 槛
- riba
- ovan
- ynam
- 질
- -spacing
Suppresses
- feito
- CKET
- ORY
- 对她
- ético
- sixth
- führung
- ppers
- venth
- pré