unclear: blank before a year or number
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. 11 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 weakly on a blank spot that sits just before a year or a number, often in a citation or date. The signal is faint, so it is hard to say what it really does. Turning it up made the text drift into lists of numbers and years.
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.
-
introduced. As an added⏎bonus, an introduction to JShell, Java's interactive programming⏎tool, is included.⏎|Java: A Beginner's Guide, 9
-
aanews.noaa.gov⏎– Erich Luening⏎Print this page<|endoftext|>Today is Sunday, Feb. 3, the 34th day of 2013. There are 331 days left in the year
-
modern conflict, writes Audrey Borowski.⏎A Theory of the Drone. Grégoire Chamayou. Trans. by Janet Lloyd. The New Press. 2015.⏎In the past decade, the deployment of hunter-killer
-
es wrought by the present technological age.⏎Michael Freeman, Bathurst Heights Secondary School, North York, ON.⏎1971-1979 | 1980-1985 | 1986-
-
is not (Shectman & Horowitz, 2003). Studies find that people do not treat computers as people outside
-
, 19. Martin Beatty, 20. Patt Carr.⏎Sources: NAI, (National Archives of Ireland
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: " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
's life.⏎Methane has been present in much larger quantities in past geological ages. And large quantities of it exist in combination on the ocean beds.⏎Its greenhouse gas effect is felt to be around twenty-five times that or carbon dioxide, and
-
泰国学生在学习汉语助动词中的困难及偏误情况,分析他们产生偏误的原因。我们发现经过基于对比研究设计的汉语助动词课堂教学后,“后测(Post-test)”的选择正确率高于
-
— that the Earth stood still in outer space while the sun and other celestial bodies rotated around it. Such findings outraged the Roman Catholic Church.⏎One problem with Copernicus’ observations
-
, and preferably 1.5 degrees Celsius, from pre-industrial levels. The aim is to use the best available scientific data to reduce the global amount of greenhouse gases, mainly carbon dioxide and methane emissions, in the Earth’s atmosphere.⏎Leading
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
- L8:7 excites
- L8:16 suppresses
- L8:3 excites
- L2:36 excites
- L6:12 excites
- L8:10 excites
- L8:4 excites
- L6:7 suppresses
- L5:46 excites
- L3:59 excites
- L8:14 excites
- L5:503 excites
- L9:5 excites
- L5:117 excites
- L8:13 suppresses
- L2:70 excites
- L5:315 suppresses
- L7:66 excites
- L0:1301 weakly suppresses
- L4:19 weakly suppresses
- L7:21 weakly excites
- L1:1340 weakly suppresses
- L5:198 weakly excites unclear: weak "and" in lists
- L2:1710 weakly suppresses
Feeds later layers
- L12:130 excites
- L13:41 excites the word 'the' after a lead-in phrase
- L14:31 excites
- L12:4 excites document-end markers, mixed
- L14:167 excites
- L17:0 excites end-of-document token boundary
- L27:47 excites
- L27:81 excites
- L16:828 excites
- L27:386 excites
- L14:12 excites
- L14:934 excites
- L14:22 excites
- L27:39 excites
- L14:17 excites
- L14:10 excites
- L14:25 excites
- L27:55 excites
- L12:403 excites
- L14:14 excites
- L14:1 excites
- L27:314 excites
- L25:11 excites
- L14:21 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
- ies
- (
- f
- 1
- cap
- op
- ational
- ou
- 2
- 3
Suppresses
- ATUS
- Crimea
- 射手
- SHOT
- 自然是
- sembled
- flowed
- _Statics
- BOSE
- COURT