empty spot before a number or "3D"
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. 22 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 a blank spot in the text that is directly followed by a number, like "3D software", "3D printer" or "360-degree". Turned down, the model wrote lists of numbers, and turned up it wandered into unrelated sentences.
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.
-
questions in complete sentences and paragraphs based on 7 days, worldwide (use zoom feature to zoom in or out). Your responses
-
get to the part where I show you literally how to design a complete house using a popular free and open source 3D software, we must first learn some designing techniques. Using a very easy
-
way of making things involved taking lots of parts and screwing or welding them together. Now a product can be designed on a computer and “printed” on a 3D printer, which creates a solid object by building up successive layers of material
-
’s subjects, and the public congregations and usual places of divine worship; be<|endoftext|>New anti-flu drugs could be on the market within years after scientists create 3D map of part of the virus
-
—the popular online file repository for 3-D objects—and expect to change the world. “I think one of the
-
which all interactions are reduced to defining the determination of the Matter or Economy in a society: social<|endoftext|>Touring the Milky Way’s a blast with this brand new 360-degree interactive panorama. More than 2 million infrared photos taken by
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: 1. 2. 3. 4. 5. 6.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
04, 2010 · LEDC flooding case study Bangladesh 1998 1. LEDC Flooding Case Study: Bangladesh 1998 2. Between July-September 1
-
can add the cardinal numbers and the numerators. First the cardinal numbers. That's only one 3. This will remain. And then the numerators. 1 + 3 = 4. We then get the answer 34
-
1, data 0, rloc 0, and pref 0. The GLib 0 resource in a shared library corresponds exactly to the code 1 resource in an application.⏎Before going any further, you need to make sure
-
cart with 4 wheels required to carry a 6,000-lb load would use the following calculation:⏎[(4,000 lbs + 6,000 lbs) × 1.25] ÷
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
- L18:106 strongly suppresses
- L19:581 suppresses blank slot before a number
- L14:191 excites blank spot before a run-on number
- L4:53 excites blank slot before a number
- L21:173 excites
- L21:339 excites spot just before a year
- L16:471 suppresses blank spot before a year
- L16:9 suppresses
- L18:377 excites number right after 'about' or 'less than'
- L5:334 suppresses
- L7:229 weakly excites first digit of a split number
- L6:0 weakly excites
- L16:66 weakly excites
- L20:344 weakly excites hyphen inside compound words
- L8:30 weakly excites blank slot right before a number
- L21:189 weakly suppresses document break token mid-sentence
- L20:433 weakly excites
- L6:1279 weakly suppresses
- L15:12 weakly excites
- L17:816 weakly suppresses
- L18:48 weakly suppresses
- L16:55 weakly excites
- L18:1534 weakly suppresses space between month and day number
- L11:1601 weakly suppresses
Feeds later layers
- L27:170 strongly excites unclear: spaces in mixed-language text
- L23:77 excites the space right before a number
- L26:423 suppresses
- L27:448 excites
- L27:237 suppresses
- L27:113 excites
- L26:172 excites
- L26:611 suppresses
- L25:33 excites spot where a number is due
- L27:306 suppresses
- L25:95 suppresses
- L27:283 excites the spot right before a number
- L27:1017 suppresses
- L24:675 suppresses
- L26:97 suppresses
- L25:736 excites
- L26:542 suppresses
- L27:287 suppresses
- L26:2873 suppresses
- L26:57 excites
- L25:222 suppresses
- L25:91 excites
- L27:183 suppresses
- L27:1652 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
- '/')
- ibel
- �藏
- ionic
- 喟
- prof
- .hxx
- ****************************
- владе
- orarily
Suppresses
- ⏎
- ⏎
- ⏎⏎
- ⏎
- ⏎
- ⏎⏎
- ⏎ ⏎