the word 'same' as a referring word
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. 14 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 lights up on the word "same" when it points back to something already mentioned, as in "the same principle" or "at the same time". When it was turned up, the model's writing drifted into loose, repetitive phrasing, which doesn't clearly show what it does.
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.
-
an "insider." Some of the papers discuss the issue of objectivity, with the implication that objectivity is best obtained by standing outside the religion. This same subtext continues throughout the papers: that there is something amiss with being a
-
agency within the U.S. Department of Labor, workplace fatalities have been cut in half and occupational injury and illness rates have declined 40 percent. At the same time, U.S. employment has nearly doubled.⏎Under the OSH Act,
-
2009, 18- to 54-year-olds accounted for 95 percent of all Gulf War-era II veterans, whereas that same age group made up about 7 in 10 nonveterans.
-
water has a higher density than oil, it will sink at the bottom of the bottle, and the oil will stay above it because it’s less dense. The same principle applies to food coloring which will also sink at the bottom of the bottle.⏎
-
uent is called __?__ .<|endoftext|>Many people believe that every word in the Bible -- including the creation account presented in Genesis -- is literally true.⏎But those same people -- especially the children among them -- often face a tough challenge when they run
-
parallels between computational and biological viruses, and the means of battling against them. Just as we humans can minimize the risks of ill health through prevention and curing, the same is true of protecting the health of our vulnerable PCs.⏎In terms of
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: this. So, then, the next thing is to think of the next thing. Now, the point is to notice that the current state. Now, the next thing to notice, the next thing to notice, the next thing to notice
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
get the correct value from the table. For example, you select “Canada” and “VAT”, you want to get 5% in return.⏎Below this table, we want to be able to define the country in cell C11
-
of both numbers (and operators) directly below and above it. That means it begins in the same<|endoftext|>The books here address specifically how the mind learns and how this information can be applied both in academics and in life.⏎Bain, Ken.
-
use.⏎The Good News⏎Vaperma’s membrane-based Siftek, on the other hand, is expected to reduce emissions significantly. And in addition to this, it’s also supposed to be more economical. Thanks to the fact that S
-
cast ballots, one of the highest levels of participation in Venezuelan history. Incumbent President Hugo Chávez was elected for a third six-year term.Top of page⏎Venezuela had one of the highest murder rates in Latin America due,
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
- L13:547 excites
- L7:1253 suppresses
- L13:1666 suppresses
- L11:1907 suppresses
- L13:2979 suppresses
- L13:137 suppresses
- L8:1322 suppresses
- L7:620 excites
- L13:1966 excites
- L13:1040 suppresses
- L7:1794 suppresses
- L8:915 excites
- L12:1583 suppresses
- L10:2439 excites
- L12:2701 excites
- L13:2191 excites
- L13:1333 suppresses
- L7:975 excites
- L12:59 excites
- L10:1046 suppresses
- L13:3053 suppresses
- L12:732 suppresses
- L13:1818 suppresses
- L12:2864 suppresses
Feeds later layers
- L15:1209 suppresses
- L15:1570 suppresses
- L15:1306 suppresses
- L20:2429 suppresses
- L17:987 suppresses
- L16:2247 weakly suppresses
- L22:2538 weakly excites
- L16:2523 weakly suppresses
- L15:2991 weakly suppresses
- L17:2848 weakly suppresses
- L27:1006 weakly suppresses
- L15:109 weakly suppresses
- L15:204 weakly suppresses
- L16:1572 weakly suppresses
- L15:654 weakly excites
- L17:1450 weakly suppresses
- L15:124 weakly suppresses
- L17:2011 weakly suppresses
- L21:337 weakly suppresses
- L15:2879 weakly suppresses
- L18:2816 weakly suppresses
- L15:2237 weakly excites
- L15:3069 weakly suppresses
- L25:1415 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
- uiltin
- inality
- enaries
- KeyId
- Bullet
- Node
- guilty
- Line
- -line
- unseren
Suppresses
- ensuite
- però
- その後
- then
- thereafter
- 的同时
- 副市长
- ますが
- subsequently
- ;br