'and' joining the last item in a list
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 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
It fires on the word 'and' when it comes just before the final item of a list, such as 'produce, export, and import'. It is weakly active there and turns strongly off at the start of a new document. Turning it up did not produce a clear visible change.
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.
-
from the knee) (Figure 2).⏎Angular Limb Deviations are usually easily diagnosed by looking at the limbs standing directly in front of, directly behind, and/or from the side of the foal. You can also observe the foal
-
their third annual "IBM Next Five in Five" includes a mention of solar power. The list includes five innovations that will change the way people work, live, and play over the next five years. Accordingly, IBM thinks energy saving solar technology will
-
in Proceedings of the National Academy of Sciences.⏎A major challenge for tissue engineering is to identify the essential environmental ingredients that cells need in order to communicate, migrate, and organize into living tissues. One of these ingredients is the presence, outside the cell
-
consequences for people and society.⏎In this report, the Ada Lovelace Institute explores the datafication of health: what it is, how it occurs, and its impacts on individual and social wellbeing. We draw on examples to synthesise existing
-
finding any old stick to bash natural gas with is that nations will burn coal instead. Or just ignore the zealots and move forward rapidly to produce, export, and import methane. While Canada’s elite stands there slack-jawed going but what
-
lot of talk over the last several years about the nutrition of school meals – where the ingredients come from, how they’re prepared, what the food tastes like, and how the meal is presented. These are all important conversations for elevating the quality
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 don't have the same experience as the people who are in the same situation as me. I don't have the same experience as the people who are in the same situation as me. I don't have the same experience as the people
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
<|endoftext|>Cancer & Tumorigenesis⏎Cancer describes the disease with uncontrolled
-
Li<|endoftext|>What was the role of residential segregation in enabling the racial wealth gap? Why does the
-
<|endoftext|>Charles Messier (France) had just completed an observation of Jupiter and was in the
-
<|endoftext|>Public holidays in Pakistan are different than bank holidays in the country. While all of 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
- L2:210 strongly excites
- L2:1021 excites
- L2:246 excites
- L2:36 excites
- L2:1899 excites
- L2:839 excites
- L2:602 suppresses
- L2:112 excites
- L2:175 excites
- L2:2242 suppresses
- L2:1710 suppresses
- L2:234 excites
- L2:503 suppresses
- L2:165 excites
- L2:86 suppresses
- L2:1578 excites
- L2:356 excites
- L2:111 suppresses
- L2:305 suppresses
- L2:637 suppresses
- L2:216 excites
- L2:989 suppresses
- L2:28 suppresses
- L2:854 excites
Feeds later layers
- L6:24 excites
- L5:187 suppresses
- L6:208 suppresses
- L15:306 suppresses
- L6:277 suppresses
- L6:374 excites
- L6:212 weakly excites
- L11:539 weakly suppresses
- L6:22 weakly excites
- L7:241 weakly excites
- L6:14 weakly excites
- L4:10 weakly excites
- L11:1353 weakly excites end-of-document token after a web snippet
- L10:859 weakly excites unclear: word fragments, weak
- L6:70 weakly excites
- L6:34 weakly excites
- L18:388 weakly suppresses
- L5:195 weakly excites
- L9:1524 weakly excites
- L19:1676 weakly suppresses
- L8:56 weakly excites
- L26:1564 weakly excites
- L12:114 weakly suppresses
- L6:879 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
- idades
- itous
- izo
- .microsoft
- ochond
- ilde
- angelog
- ypress
- ��
- ipients
Suppresses
- herself
- 她说
- 像是
- 一件事情
- 她是
- 那个人
- 某个
- \Id
- oneself
- 一艘