'to' in 'X to Y' phrases
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. 5 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 responds to the word 'to' in pairs like 'year to year' or 'person to person', where something is compared or passed across. Its opposite side reacts to 'which' after 'from', as in 'from which'. Turning it up did not bring out these phrases in the output.
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.
-
000. That year, Canadian women had given birth to 479,000 children.⏎- With Canada's population steadily growing from year to year, the number of deaths has understandably increased over time, and it will continue
-
however, when James Kirk took command, and ordered a pursuit of the Narada toward Earth. (Star Trek)<|endoftext|>It's just fluctuating wildly from year to year as it always has done but by making some complex "allowances", climate
-
either top to bottom or bottom to top). If the trimmer is placed parallel to the nail (cutting from side to side), the nail is crushed and may splinter.⏎The handles of the trim
-
comply may result in legal action.⏎WHAT YOU NEED TO KNOW:⏎Diphtheria is a disease caused by a bacterial infection. The infection spreads quickly from person to person through sneezing or coughing. It
-
in South Korea. Human-to-human transmission in this outbreak as in all others has been limited, indicating that this MERS-CoV does not easily spread from person to person. The exact source of MERS-CoV is unknown at this time.
-
below the 48°N parallel is 474/year with April and May generally being the most active months. The numbers vary quite widely from year to year. For example, in 2011 only three icebergs
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 need to be able to find a way to get a job. I need to be able to get a job. I need to be able to get a job. I need to be able to get a job. I need to be
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
say the prints date to the middle of the Jurassic Period, a time from which few dinosaur fossils have been recovered.⏎Researchers reported this month on their study of about
-
end they don't look at the writer by what he writes, but rather evaluate him by where he lives, or by the location of "the room" from which he writes, as the Peruvian writer Vargas Llosa noted in his
-
of Ireland upon the horizon." In Iona accordingly he resolved to make his home. The spot from whence St. Columba made this sorrowful survey is still called by the isles
-
from basic waxing and waning of the moon to the arrival of comets and eclipses.⏎The word "space" was specifically mentioned as the place from where the meteorite that made Toph's bracelet and Sokka's sword came.
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
- L4:1593 excites
- L3:1119 excites
- L4:192 excites
- L1:2195 excites
- L4:1002 weakly suppresses
- L2:1015 weakly suppresses
- L4:2146 weakly suppresses
- L1:2387 weakly excites
- L1:2984 weakly suppresses
- L2:2450 weakly excites
- L3:1079 weakly suppresses
- L4:3068 weakly excites
- L4:1726 weakly excites
- L3:2937 weakly suppresses
- L3:807 weakly suppresses comma inside dates and years
- L2:1350 weakly excites
- L0:2442 weakly suppresses
- L3:2372 weakly excites
- L3:1736 weakly suppresses
- L1:1414 weakly excites
- L3:2229 weakly suppresses
- L3:2166 weakly suppresses
- L0:896 weakly excites
- L1:2350 weakly excites 'reasons' or 'permitting' after a qualifier
Feeds later layers
- L11:658 strongly suppresses
- L9:1382 suppresses
- L18:2122 suppresses
- L7:1563 suppresses
- L16:1283 suppresses
- L19:1997 suppresses
- L7:1393 suppresses
- L10:1054 suppresses 'from' after movement or transfer verbs
- L20:1980 weakly suppresses
- L7:1533 weakly suppresses
- L7:1399 weakly suppresses "to" between a starting and ending value
- L6:951 weakly suppresses
- L10:1603 weakly suppresses
- L13:1329 weakly suppresses
- L19:1871 weakly suppresses
- L11:2214 weakly suppresses
- L17:1651 weakly suppresses
- L10:42 weakly suppresses
- L13:2955 weakly suppresses
- L20:499 weakly suppresses
- L7:626 weakly excites
- L16:2231 weakly excites
- L9:1961 weakly suppresses 'of' in titles and place names
- L19:2374 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
- lest
- xAD
- addPreferredGap
- letic
- ModelProperty
- 场
- GROUP
- itudes
- xes
- ♭
Suppresses
- oire
- 往
- Carthy
- 年开始
- iente
- 로부터
- ersen
- ič
- ultan
- getInt