unclear: hyphens and size comparisons
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. 6 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 is hard to pin down. It fires most on hyphens inside compound words and on phrases that compare sizes, such as 'twice the size' or 'times the height'. The steered text did not show a clear pattern.
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.
-
the speeds of two satellites placed in stable circular orbits of radius R and 2R respectively?⏎C) 1 / 2⏎D) 1 / �
-
refer to their objects⏎> through a Map. Then changing a leaf object requires only that the value⏎> associated with key represented by the leaf be (re-)inserted into the⏎> Map. The Tree itself need not change at all
-
for example, in the controversy concerning the effect of electromagnetic fields on health, the field strengths have almost always been reported in milligauss.⏎The mechanical units are the same in all cgs systems. Length,
-
designed to support orbits for the satellites, to perform orbit control tasks for small-, mini- and micro-satellites.<|endoftext|>Until 1864, Georgia had
-
technologies, databases on the fundamental properties of “nano-structured” materials, and other essential supporting tools and capabilities. U.S. industry
-
North Pacific Gyre.⏎Twice the size of Texas – though some scientists have argued that its mass has extended to twice
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: "What is the purpose of the first part of the sentence, 'To the best of my knowledge' I am able to achieve' (A. P. S. I.)" I will do it, the purpose of the first part
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the expressions of V(t), I(t), the trigonometric formula sin(X)sin(Y)=1/2(cos(X-Y)-cos(X+Y)), and the
-
power is then written as⏎Here the average of sin2 ωt over a cycle is 1/2 and that of sin ω t cos ωt is zero. Substituting these values, we obtain average power
-
9 Copyright Math Open Reference. All rights reserved<|endoftext|>Use chain rule⏎F'(x) = dF(x)/dx = (dF(x)/dx2)(dx2/dx) = 2x (dF(x)/dx2
-
the input-output pair. It is called supervised learning because the process of an algorithm learning from the
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
- L5:33 excites
- L5:227 weakly suppresses
- L5:27 weakly excites
- L5:219 weakly suppresses
- L3:286 weakly excites
- L5:75 weakly excites
- L4:493 weakly suppresses
- L4:592 weakly excites
- L2:2796 weakly suppresses
- L3:2140 weakly suppresses
- L2:359 weakly suppresses
- L4:522 weakly suppresses
- L4:97 weakly suppresses
- L1:146 weakly excites
- L5:1833 weakly suppresses
- L4:896 weakly excites
- L3:338 weakly suppresses
- L4:458 weakly excites
- L5:2234 weakly suppresses
- L2:1684 weakly excites
- L4:25 weakly excites
- L4:544 weakly excites
- L4:1908 weakly excites
- L0:758 weakly suppresses
Feeds later layers
- L16:1275 suppresses
- L19:789 weakly suppresses
- L9:335 weakly excites 'of' after nouns like anniversary or onset
- L9:1854 weakly suppresses
- L17:101 weakly suppresses
- L18:2385 weakly suppresses
- L10:1018 weakly excites
- L17:2764 weakly excites
- L16:1216 weakly suppresses
- L7:2611 weakly excites
- L23:1542 weakly suppresses
- L14:636 weakly excites
- L20:3031 weakly suppresses
- L9:2462 weakly excites
- L15:1162 weakly suppresses
- L12:1896 weakly excites
- L11:1431 weakly suppresses commas and 'and' in noun lists
- L10:1568 weakly excites
- L7:2164 weakly excites period inside country abbreviations
- L12:277 weakly suppresses
- L14:319 weakly suppresses
- L18:1492 weakly suppresses
- L12:2959 weakly excites
- L7:1031 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
- loid
- cep
- legit
- aran
- objectId
- avel
- arkan
- eland
- nexus
- +f
Suppresses
- itivity
- devote
- ;⏎
- 隳
- ;")⏎
- .";⏎
- ("-");⏎
- 同事们
- --*/⏎
- ⏎ ⏎