"to" after "it's important"
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. 1 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 lights up on the word "to" when it comes right after phrases like "it's important" or "remember", where a tip or advice is about to follow. When it was turned up, the model's replies did not show a clear theme.
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.
-
anged wheel⏎When evaluating wheel products for a specific application, it's important to take into consideration all possible factors that will affect performance. We discuss a few potential
-
tight, you’re wrong.”⏎While this isn’t wrong advice, it’s important to understand where this tourniquet procedure came from and why we’re constantly being
-
out any forms, or give out money or credit card details.⏎- Don't accept any offers that seem too good to be true - they probably are.⏎Remember to always be
-
wheel. It's also important to include a safety variable for any unforeseen circumstances. We recommend a minimum variable of
-
compacted or poorly drained. Instead of treating the symptoms using pesticides, it’s important to discover and treat the reason the<|endoftext|>As the summer comes, families need to cool
-
nationwide. But it’s also important to remember one of the most vital purposes of offering school meals: fighting hunger so kids
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 going to the moon, and I have been going to the moon again. I have been going to the moon twice, and I have been going to the moon three. I have been going to the moon four. I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
calories, which helps you lose weight. If calories consumed in food are less than calories spent during exercise and other daily activities, you will lose weight. It is impossible not to! Many people have learned how to lose belly fat by starting a running
-
more than one edge. So when I draw this graph on my 2D screen, it can happen that edges will be crossing each other.⏎It is impossible to avoid crossing edges, but I guess there is some algorithm which is able to
-
Kennedy Shelley⏎Governors in several states have closed down beaches in order to promote social distancing during the 2020 Coronavirus scare.⏎It is impossible to find a tanning bed place that is still open while everyone is COVID-
-
familiar with the condition and don’t know what to look for. You may want to consider getting a second opinion if your symptoms do not respond.<|endoftext|>It is impossible to think of the Guatemalan, Mexican and Mesoamerican diet without maize
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
- L0:1225 suppresses
- L0:1513 weakly suppresses
- L0:2216 weakly excites
- L0:3024 weakly suppresses
- L0:884 weakly excites
- L0:2237 weakly excites
- L0:2006 weakly suppresses
- L0:2923 weakly suppresses
- L0:1301 weakly suppresses
- L0:1832 weakly excites
- L0:1202 weakly suppresses
- L0:2812 weakly suppresses
- L0:1763 weakly excites
- L0:178 weakly suppresses
- L0:2408 weakly excites
- L0:1239 weakly excites
- L0:1889 weakly suppresses
- L0:2808 weakly suppresses
- L0:977 weakly suppresses "like" after "you'd" or "something"
- L0:1907 weakly excites
- L0:886 weakly suppresses
- L0:1918 weakly suppresses
- L0:1198 weakly excites
- L0:1985 weakly suppresses
Feeds later layers
- L7:1056 weakly excites
- L13:959 weakly suppresses
- L10:1753 weakly suppresses
- L11:1860 weakly excites
- L3:509 weakly excites
- L14:1878 weakly suppresses
- L13:1711 weakly suppresses
- L4:1239 weakly suppresses
- L20:942 weakly excites
- L2:2747 weakly suppresses
- L25:2307 weakly suppresses
- L16:2397 weakly suppresses
- L16:744 weakly excites
- L15:2385 weakly suppresses
- L5:2893 weakly suppresses
- L12:452 weakly excites
- L17:1822 weakly excites 'should' in advice and claims
- L21:2823 weakly suppresses
- L2:494 weakly suppresses
- L13:2663 weakly excites
- L5:1883 weakly excites
- L8:1465 weakly excites
- L3:1030 weakly excites
- L9:3021 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
- eyed
- quên
- 烊
- усп
- geom
- geometric
- äm
- canf
- precaution
- stand
Suppresses
- """,⏎
- _MEDIUM
- ilians
- rack
- geries
- iationException
- hält
- _trampoline
- alie
- illo