'only' before a verb or adjective
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. 18 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 'only' when it sits in the middle of a phrase like 'is only effective' or 'are only just beginning', right after a verb. It is a hedging word before the main point. Turning it up gave lists of counting or repeated descriptive phrases, which does not clearly match the trigger.
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.
-
cal infection, which causes grave illnesses, including pneumonia, meningitis, and middle-ear infections. Children and the elderly are especially at risk. Vaccines are only effective against a few of the pneumococcal types and increasing resistance to antibiotics
-
of blah magazine’, cardboard red and blue ’3D’ glasses?! These classics marked the beginning of mainstream 3D for film and TV audiences – in fact only recently there were a spate of classic 3D movies played on Channel
-
the canal became an issue. So in 1905 the city completed the bridge over the canal. The first bridge was a transporter bridge, on of only a few ever built, and had a gondola which was raised and transported
-
“Overall, young people’s ability to reason about the information on the Internet can be summed up in one word: bleak.”⏎The question of how it works is only slightly less obvious. The short answer is, “Mud sticks”. If you
-
the aptitude and the inclination should have the opportunity to receive entrepreneurial training and access to resources like low-interest loans to become self-employed.⏎Of course, training is only the first step. Once trained, they should be connected job opportunities. Life skill
-
its native ecosystems means a return to a more natural state, a point we made in a post last week. Second, ecosystem dynamics are incredibly complex and we are only just beginning to understand the role that even a single species can play.⏎Wolves
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 best of the best (or the best of the best of the best) is therefore (A) A (A) A (A) A (A) A (A) A (A) A (A) A (A
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
19 percent).⏎Among Gulf War-era II veterans in 2009, 31 percent of veterans aged 18 to 24 years were enrolled in school. As for older
-
and 10% with colorectal cancer.⏎Those three cancers accounted for slightly more than half of all cancer diagnoses.⏎Nearly 60% of survivors were 65 or older, the report found. The largest number of survivors with prostate
-
in history,” citing the lack of students being exempted from the draft as the reasoning.⏎But nothing is said about the fact that less than half of the population ages 18-26 are actually considered. If the military wanted to have
-
rivers, and in small streams, lakes, marshes, and even roadside ditches containing adequate year-round water flow. In areas where deep, calm water is not available, beavers that have enough building material available will create ponds by building dams
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
- L15:2956 excites
- L16:1867 suppresses
- L13:2011 excites
- L17:1746 suppresses
- L14:1201 excites
- L14:1150 suppresses
- L16:2328 suppresses
- L14:835 suppresses
- L15:2592 suppresses
- L16:390 suppresses
- L14:2639 weakly excites
- L15:1963 weakly excites
- L14:2689 weakly suppresses
- L14:1459 weakly excites
- L17:3004 weakly suppresses
- L12:2655 weakly excites
- L15:826 weakly suppresses
- L6:1149 weakly suppresses
- L12:2754 weakly suppresses
- L8:343 weakly suppresses
- L13:1365 weakly excites
- L13:2879 weakly suppresses
- L14:770 weakly suppresses
- L12:1917 weakly excites
Feeds later layers
- L19:2188 excites
- L19:2926 excites
- L19:1179 suppresses
- L21:2415 weakly suppresses
- L21:1226 weakly suppresses
- L24:2068 weakly suppresses
- L19:1151 weakly excites
- L22:2290 weakly excites
- L27:620 weakly excites
- L21:956 weakly suppresses
- L22:758 weakly excites
- L19:1131 weakly excites
- L19:3025 weakly suppresses
- L19:2909 weakly suppresses
- L21:2742 weakly excites
- L25:437 weakly suppresses
- L25:478 weakly suppresses
- L21:1589 weakly excites
- L19:1717 weakly suppresses
- L19:1019 weakly excites
- L25:1813 weakly suppresses
- L23:377 weakly suppresses
- L20:1923 weakly excites
- L19:2006 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
- oningen
- 缓缓
- mons
- powers
- ineff
- 精心
- 随着
- tiny
- getDoctrine
- tap
Suppresses
- closest
- None
- Neither
- [X
- only
- Neither
- overall
- erland
- [Y
- еть