'each' in 'each other'
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. 9 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 word 'each' when it is followed by 'other', as in 'with each other' or 'against each other'. Turning it up did not produce a clear theme in the written text.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
too thin, it will not dissipate heat correctly and be more prone to warping.⏎The brake rotor consists of two braking contact surfaces that must be parallel with each other, or else the brakes will pulsate when they are applied. In order
-
drop will equal the sum of the voltage drops measured across each resistor separately. This tells you that both the resistors' voltage drop waveforms are in-phase with each other, since they add simply and directly.⏎Measure voltage dropped across both capacitors
-
eventually stopped but only because Romeo and Juliet died. This was mainly because of the Friar.⏎The Capulets and Montagues were not on good terms with each other. Unfortunately though, Romeo and Juliet fell in love with each other regardless if
-
ix more than a few pages of paper to your fridge.⏎To see this for yourself, stick two fridge magnets together, back to back. They will slide against each other in one direction. If you slide them at right angles to that, they
-
.⏎Find the area of the shaded region created by the two overlapping⏎triangles in terms of a and b?⏎Position the lines so that they are perpendicular to each other.⏎What can you say about the equations of perpendicular lines?⏎ABCD is
-
⏎meant for such use. Another risk with NFS is that you must⏎devise a way to ensure that two or more servers do not interfere⏎with each other. Usually NFS file locking is handled by the⏎lockd daemon, but
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 blank Friday 19th, 2023 January 1, 2016 This is an error type of the unknown type of the unknown type of the unknown type of the "general" type of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
631–642, 2013. My thoughts precipitated a brief conversation with Erella Hovers in which, for the umpteenth time, she expounded the theory that a fossil
-
0 us, https://commons.wikimedia.org/w/index.php?curid=12708190<|endoftext|>
-
structure = function.⏎(Image from: https://commons.wikimedia.org/w/index.php?curid=1267900)⏎Want
-
08 — but declined to say which types of arms — several lower courts have had their say on whether Americans can own machine guns.⏎The U.S. courts of appeals for the Third,
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
- L6:1299 strongly suppresses
- L8:2074 strongly excites
- L7:663 strongly suppresses
- L6:2016 strongly excites
- L6:1463 strongly excites
- L6:2146 excites
- L6:809 suppresses
- L8:852 suppresses comma inside large numbers
- L8:448 excites
- L7:2026 excites
- L8:2858 excites
- L6:1695 excites
- L8:2388 suppresses
- L3:1661 suppresses
- L6:2069 excites
- L4:1891 suppresses
- L7:3055 suppresses
- L6:1122 suppresses
- L7:2370 suppresses
- L5:2136 excites
- L8:2375 excites
- L5:1387 excites
- L8:2078 excites
- L7:2878 excites
Feeds later layers
- L16:592 strongly excites
- L10:346 strongly excites "each" in "each other" phrases
- L10:152 strongly excites
- L11:173 excites
- L18:708 excites
- L17:259 excites
- L16:415 excites
- L15:1616 suppresses word fragments mid-word
- L18:322 suppresses
- L11:300 excites
- L18:828 excites
- L15:1900 suppresses
- L15:646 excites
- L19:395 excites
- L14:713 excites
- L14:611 excites
- L11:1384 excites
- L18:564 excites
- L20:3066 excites
- L10:42 excites
- L19:1204 excites
- L19:107 suppresses
- L20:486 excites
- L17:446 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
- 치
- ственно
- imately
- ับ
- ้า
- 得
- 冲
- ầy
- и
- ンド
Suppresses
- 他们都
- .raise
- nees
- atsby
- noop
- leshoot
- ()?;⏎
- "})
- iddles
- icerca