"more", "drop" and other words about amount of change
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. 3 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 fires on words that describe how much something goes up, down or differs, such as "more", "dropped" and "drop in blood pressure". The pattern is loose, and turning it up did not give a clear theme in the model's 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.
-
increases with the increase of precipitation. It also depend on the direction of dominant airflows at 850 hPa. In the mid-eastern region, more than 40% acid rains appear when the dominant wind directions are NW,
-
exercises, which can be changed more often while providing new stimuli to the muscles. A good rule of thumb is that a movement should be changed when the performance plateaus over several training sessions or the movement produce increasing pain that is not corrected by a
-
and other hormones linked to breast cancer.⏎One hormone the study examined was estrone, which is a form of estrogen. It was discovered that levels of this hormone dropped 9.6 percent in women who dieted, and 11.
-
8 from Section 3.1, we drew a histogram of the length of eruption of California's Old Faithful geyser and found that the distribution is symmetric. Draw a boxplot of these data. Use the boxplot and quartiles
-
9-0.965). Children were more vulnerable to rotavirus (p = 0.008, OR<|endoftext|>
-
cards held in a black hand sold for about 20 percent less than cards held in a white hand. What’s more, “the race effect was more pronounced in sales of minority player cards.”⏎The eBay study mimics a similar one from
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: I have been thinking about lately. The user is a set of knowledge. The user is a set of knowledge. The user is a set of knowledge. The user is a set of knowledge. The user is a set of knowledge. The
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Piedmont⏎By Michael Fine⏎N.C Cooperative Extension⏎Every year, my family embarks on a challenge to extend our winter vegetable season later into the winter than the previous year. In 2013, our first growing season,
-
automatic switching circuit plugs into the inverter and the wall outlet. Then the appliance plugs into the automatic switching circuit.⏎The solar panel, battery, and inverter may be purchased off-the-shelf from a variety of locations. The last two parts
-
the thief will have to make payment for what he has taken; if he has no money, he himself will have to be exchanged for money, so that payment may be made.⏎And if a Levite does not give money to get back his
-
unspoken which means): from Outdated French, or from Latin significantia, from significare ‘point out, portend’. Hence the actual application form is fairly minimal, and a majority of the applicants’ info is present of their CVs
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
- L0:1588 excites
- L1:317 weakly suppresses
- L2:1899 weakly excites
- L0:449 weakly suppresses
- L0:2568 weakly suppresses
- L2:417 weakly excites
- L1:533 weakly suppresses
- L1:501 weakly excites
- L0:728 weakly excites
- L0:953 weakly excites
- L2:1652 weakly suppresses
- L1:1251 weakly suppresses
- L0:2485 weakly suppresses
- L2:695 weakly excites
- L1:1013 weakly suppresses
- L2:786 weakly excites
- L2:2857 weakly suppresses
- L0:2029 weakly suppresses
- L2:1410 weakly suppresses
- L2:3042 weakly excites
- L1:840 weakly suppresses
- L1:662 weakly suppresses
- L1:955 weakly suppresses
- L2:940 weakly suppresses
Feeds later layers
- L5:189 excites
- L5:193 suppresses
- L7:1367 excites
- L6:557 suppresses
- L4:955 excites
- L6:1148 weakly excites
- L21:1544 weakly excites
- L10:720 weakly suppresses
- L17:1175 weakly excites
- L6:175 weakly excites
- L10:2940 weakly suppresses
- L7:783 weakly excites
- L5:2159 weakly excites
- L20:1313 weakly suppresses
- L8:276 weakly excites
- L8:724 weakly excites
- L4:244 weakly suppresses
- L7:1920 weakly excites
- L13:173 weakly suppresses
- L10:1941 weakly suppresses
- L22:2341 weakly excites
- L4:1635 weakly excites 'light' as an adjective (pale, not heavy)
- L17:1766 weakly suppresses
- L10:455 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
- rido
- 绮
- plotted
- национа
- 回暖
- orgas
- plot
- 停滞
- 丰满
- Consort
Suppresses
- KG
- KB
- 款
- $
- 筐
- 료
- eyes
- tamb
- n