decimal point or digit inside a 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. 15 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 reacts to the dot in numbered sections like '3.2' or '5.9 g', and it is pushed the other way by digits inside numbers, such as the 0 in '10'. The effect is weak, and turning it up or down made the writing change in unrelated ways.
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.
-
two end systems⏎B. It specifies the quality of the data & connection⏎C. It specify the size of data⏎D. All of the above⏎6-4. The class-based addressing is also known
-
3.2), the sash should be positioned to produce the recommended face velocity, which
-
3. And the map of the world, which colours a country blue if it's leaning
-
3. Demonstration on pest management technologies like pheromone traps, fruit bagging
-
3.5 million African-American slaves, and now the Union Army began to recruit a second
-
3.16. The consolidation and merging of data sources provides multiple advantages for compilers.
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 working on the problem of the time period in a math class, but I'm stuck, but I've been stuck for a long time. I'm thinking of, and in the hour of the class, but I'm stuck
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
different, which makes remembering them a difficult task. The following are some creative ways that can help you create new passwords and actually remember them all.⏎Remember 100 different passwords with 1 rule⏎One way to generate unique passwords that you
-
/10.5479/si.GVP.BGVN200701-333060⏎28.98°S, 140.25°W; summit elev
-
mm;⏎9. External material⏎1.5mm thick cold-rolled steel plate with paint treatment;⏎10. Box door
-
definition for<|endoftext|>I have recently been re-reading two books I purchased a few years ago, when I was still teaching languages traditionally. The books are:⏎100+ Fun Ideas for Practising Modern Foreign Languages in the Primary Classroom and⏎
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
- L14:1 excites
- L14:375 suppresses
- L14:7 suppresses
- L14:402 excites
- L14:9 suppresses
- L4:647 weakly excites
- L14:817 weakly excites
- L3:485 weakly excites
- L7:576 weakly excites
- L14:784 weakly suppresses
- L7:1060 weakly excites
- L14:33 weakly suppresses
- L13:595 weakly suppresses
- L13:695 weakly excites
- L3:2542 weakly suppresses
- L7:310 weakly excites
- L13:933 weakly suppresses
- L7:728 weakly excites
- L14:921 weakly excites
- L14:338 weakly suppresses
- L13:925 weakly suppresses
- L13:534 weakly suppresses unclear: sentence-ending periods
- L10:636 weakly excites unclear: 'your' in contact boilerplate
- L9:630 weakly excites
Feeds later layers
- L16:2152 suppresses
- L16:1695 weakly suppresses
- L16:1258 weakly excites
- L20:685 weakly excites
- L17:2406 weakly excites
- L19:1839 weakly excites
- L18:707 weakly excites
- L17:2402 weakly suppresses
- L27:1341 weakly suppresses
- L18:660 weakly suppresses
- L16:521 weakly suppresses
- L17:570 weakly suppresses
- L19:1792 weakly suppresses word just before a 'reason is' explanation
- L20:1917 weakly excites
- L22:443 weakly suppresses
- L20:1036 weakly excites
- L17:295 weakly suppresses
- L20:2950 weakly excites
- L17:1155 weakly excites opening parenthesis after a sentence
- L18:2162 weakly excites
- L23:2205 weakly suppresses
- L16:2440 weakly suppresses
- L25:1675 weakly excites
- L19:2562 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
- aps
- f
- igos
- 改进
- F
- SEP
- apa
- (F
- expanded
- Sept
Suppresses
- 维奇
- ��
- 记者在
- liğinde
- 苟
- configFile
- nevertheless
- 😉⏎⏎
- linger
- pressions