reference or starting point
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. 20 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 words like 'starting', 'the', or 'are' when the sentence is about a baseline, a zero point, or a place to begin, as in 'our starting point' or 'as the origin'. Pushing it up made the model talk about reference points and zero levels. The negative side is about timing words like 'shortly'.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
kid is, so long as we are going on-pace through the material and through the content, and reaching those educational standards and outcomes, because that’s our starting point.⏎“Our core question is, What’s the experience we want kids to have
-
by the distance from the center of the press where the forces act. Using the center of the press as the origin, distances to the left would have negative values while the distances to right would
-
glycemic index comes into the scene. Glycemic index (GI) classifies foods based on how quickly they elevate blood sugar levels, taking pure glucose as the comparison point.⏎Foods with a high GI, such as white bread and white
-
use to set goals, taking the time to plan what you want to achieve can make the difference to obtain what you want. Sort of if you know where you are going, you’ll get there. Otherwise, you may end up in a place
-
success welding galvanized tube using E70C-6 metal cored wire such as Hobart’s Galvacor; the parameters given above are a good starting point for metal cored wire. Others have found that self-shielded flux
-
of geography and the environment, who specializes in dendroclimatology.⏎“This also helps put climate change into context, as it is hard to know where you are going if you do not know where you’ve
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 a program that allows small programs (less than 1/4 of a quarter of a test case) before the program is actually running. We are running a small program that is under test with a program before the test case is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
candle long enough. If you light your candle and extinguish it shortly thereafter tunneling will occur. You need to burn your candle for 1 hour
-
in his body.⏎If you or anyone else has a stroke, pay special attention to the time when the symptoms started. Some treatment options are most effective when presented shortly after stroke starts.⏎Stroke signs and symptoms include:⏎Trouble talking and understanding what
-
: it will close at midnight on November 1st, shortly before the real election day on November 4. But already more than 4
-
he followed her when she returned to her home in California, and the couple were married at San Francisco in 1880. They travelled together to Europe shortly afterwards, and lived variously at London, Bournemouth, Davos and elsewhere
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
- L12:2003 excites
- L19:2255 excites
- L19:2794 excites
- L19:2005 suppresses
- L15:2410 weakly suppresses
- L10:2896 weakly suppresses
- L12:2192 weakly suppresses
- L19:3042 weakly suppresses
- L8:1778 weakly suppresses
- L18:877 weakly suppresses
- L8:1888 weakly suppresses
- L6:1676 weakly excites
- L10:962 weakly suppresses
- L19:610 weakly excites
- L7:246 weakly suppresses
- L14:2020 weakly suppresses
- L7:2751 weakly excites
- L10:2737 weakly suppresses
- L4:1081 weakly excites
- L17:1334 weakly suppresses
- L9:1866 weakly excites
- L8:2489 weakly excites
- L13:2564 weakly excites
- L2:1440 weakly excites
Feeds later layers
- L22:1177 suppresses
- L21:2129 suppresses
- L27:2562 excites
- L24:1099 excites
- L23:1295 weakly excites
- L24:1318 weakly excites
- L21:3029 weakly excites
- L22:638 weakly excites
- L21:1595 weakly excites
- L21:2115 weakly excites
- L23:1703 weakly suppresses
- L27:2815 weakly excites
- L21:2726 weakly suppresses
- L21:2086 weakly suppresses
- L22:2643 weakly suppresses
- L23:1552 weakly suppresses
- L23:2872 weakly suppresses
- L22:2989 weakly suppresses
- L22:2003 weakly suppresses
- L22:2466 weakly excites
- L27:2406 weakly excites
- L25:1422 weakly excites
- L21:1350 weakly suppresses
- L21:161 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
- zero
- Zero
- level
- reference
- 参照
- level
- 零
- zero
- 参考
- Level
Suppresses
- enting
- uchsia
- ABOVE
- ahan
- ущ
- 顿
- <Document
- 스트
- ocha
- getStyle