unclear: commas after dates and clauses
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. 19 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 is weak and hard to pin down. Its strongest moments come at commas after a date or a short phrase in news-style text, such as "Thu, 3 Jun". Turning it up or down did not produce a clear change in the writing.
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.
-
17:55 EST⏎- ASU instrument on NASA rover helps identify outcrop of long-sought rare rock on MarsThu, 3 Jun 2010, 14:52:
-
Comments are closed.<|endoftext|>The bittersweet truth about how taste works⏎Scientists have discovered an ion channel that allows taste buds to communicate with the brain.⏎Wed, Mar 06, 2013 at 03:2
-
program via a distance learning course for the first time. You can even watch a promotional video from a North Korean web site. (It helps if you speak Korean, which I don
-
tumor dies. In the period of a few weeks to months, it forms back. In rare cases, the tumor swells after treatment. This is mostly temporary and is easy to control.⏎How often is a cyberknife treatment repeatable?⏎As
-
Philadelphia.⏎After graduating, she practiced in Germany where she also supported women doctors, helping to establish a women’s clinic and hospital.<|endoftext|>It's Hiking Season, But How Many Calories Are You Really Burning?⏎When the weather gets colder, one
-
that you are ready to share it. First a few important points:⏎Now you have the modification as a source file. Sending the source file will not be helpful, as probably someone else has done other modifications to the original file in the meantime.
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: If, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
keeps the form of Amor fati (Love of Fate) through the important element of Nietzsche's philosophy, the "will to power" (der Wille zur Macht), the basis of human behavior, influenced by the Will to Live of
-
, 168 note) who remarks: — "Um dies (i.e. Varro's description of the extent of Roma quadrata) mit dem unten erörterten Pomerium und der Auffassung des Dion
-
zeichnung der Linie des Pomeriums, zweitens der parallelen Linie der Befestigung der Arx."⏎
-
ing Spiritualism and Materialism’, Ludwig Andreas Feuerbach wrote- “Der Mensch ist, was er ißt.” (Man is what he eats) and by
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
- L17:6 strongly suppresses
- L18:1651 strongly suppresses
- L18:823 strongly excites
- L17:14 suppresses
- L18:247 excites
- L17:18 excites unclear: mixed mid-sentence words
- L18:1956 suppresses
- L18:1444 excites
- L17:714 excites
- L18:560 excites
- L18:624 excites
- L18:765 suppresses
- L5:86 suppresses
- L16:2749 weakly excites
- L18:1018 weakly excites
- L6:7 weakly suppresses
- L18:569 weakly excites
- L18:477 weakly suppresses
- L1:1049 weakly excites
- L18:320 weakly excites
- L12:15 weakly excites
- L2:126 weakly excites
- L5:28 weakly excites
- L3:464 weakly excites
Feeds later layers
- L20:661 strongly suppresses
- L20:343 strongly suppresses
- L20:351 strongly excites
- L24:43 strongly suppresses
- L23:253 strongly suppresses
- L25:198 suppresses
- L24:512 suppresses 'the' inside organisation and convention names
- L21:185 excites
- L20:1198 suppresses
- L23:278 suppresses
- L25:594 suppresses
- L25:67 suppresses
- L25:406 suppresses
- L27:535 suppresses
- L27:133 suppresses
- L27:12 suppresses
- L27:16 suppresses
- L26:277 suppresses
- L26:161 suppresses
- L27:90 suppresses
- L23:105 excites
- L20:940 suppresses
- L27:293 suppresses
- L27:406 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
- entic
- base
- borough
- back
- cap
- logue
- ina
- van
- mana
- rea
Suppresses
- Releases
- Vertical
- *out
- Experiment
- ernen
- 不解
- ёр
- ,ev
- 陌生人
- ез