'think' after 'I' or 'I don't'
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. 5 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 the word 'think' in first-person speech, such as 'I think' or 'I don't think', usually in quoted remarks. Turning it up gave unrelated text.
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.
-
if you serve one to them on a plate."⏎Shaunfield said what kids learn in the garden they're likely to bring back home to their parents.⏎"I think
-
xic patients that anything is wrong with them, Dwyer told LiveScience.⏎“Unless we conceive of it as more of a metabolic function, I don’t think we’ll get past the
-
The studio probably had little choice but to create a form letter.⏎I don’t think the form letter nature of the rejection is necessarily impersonal. It sticks to objective
-
in our children:⏎|Image: San José Library on Flickr|⏎2. DRAW! My favorite gift to give at birthday parties is art supplies. I don’t think kids can every have enough drawing paper, pencils, crayons, markers, or
-
“In fact you can’t even differentiate, you can’t even separate the tech from the social. They are irretrievably interwoven together. And I think this is the biggest problem with these kinds of projects, and with technology more generally
-
Read more about some trail-blazing female dentists, as well as the forward-thinking men who supported women in the profession:⏎Breaking the Barrier⏎I don’t think many of my professional brethren like it much that
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: a cake and a cake of a cake. Let's say the number of the cake is 3, and the number of the other cake. Let's say the number of the other cake. Let's say the number of the other cake
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
been developed over many years, thus demonstrating an influence of past and future cases. Similarly, the doctrine of precedent ensures that similar cases have similar results. Procedural grounds for objections protect witnesses from harassment and potential confusion. Therefore, all of these
-
can cause agitation, anxiety, insomnia, increased heart rate and blood pressure, and even psychosis.⏎Adderall is a highly-addictive amphetamine whose potential side effects include hallucinations,
-
Jim Sublett, chief of allergy immunology at the University of Louisville School of Medicine in Kentucky has repeatedly said in literature that over-the-counter remedies can have adverse side effects.⏎The side effects
-
- Faster passenger boarding at the PTI can aid station operators in running a safer and quicker service.⏎- (Side effect) improved passenger movement will help to reduce larger crowds and subsequently improving safety.<|endoftext|>
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
- L4:1694 weakly excites
- L3:1866 weakly excites
- L0:3006 weakly suppresses
- L1:3057 weakly suppresses
- L4:2695 weakly excites
- L2:2050 weakly excites
- L4:2188 weakly excites
- L1:1027 weakly suppresses
- L2:2522 weakly excites
- L3:2624 weakly excites
- L0:1006 weakly suppresses
- L1:1504 weakly suppresses
- L4:2193 weakly suppresses
- L1:2812 weakly excites
- L0:531 weakly excites
- L4:730 weakly excites
- L0:234 weakly excites
- L2:1598 weakly suppresses
- L3:1460 weakly excites
- L2:864 weakly excites
- L2:2101 weakly excites
- L3:1048 weakly excites
- L4:2515 weakly excites
- L3:540 weakly excites
Feeds later layers
- L6:2740 suppresses
- L9:401 weakly suppresses
- L12:1732 weakly suppresses
- L10:1324 weakly suppresses
- L13:2825 weakly excites
- L19:1545 weakly suppresses
- L13:1883 weakly suppresses
- L8:660 weakly excites
- L25:1274 weakly excites
- L10:2094 weakly suppresses
- L13:210 weakly suppresses
- L17:71 weakly excites
- L12:1854 weakly suppresses
- L17:68 weakly suppresses
- L9:917 weakly suppresses
- L8:423 weakly suppresses
- L6:785 weakly suppresses
- L13:899 weakly suppresses
- L11:300 weakly suppresses
- L7:2248 weakly excites
- L14:562 weakly suppresses
- L13:738 weakly suppresses
- L27:1818 weakly suppresses
- L11:2158 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
- rall
- ooke
- leitung
- erox
- ière
- `\
- elman
- .weixin
- uiltin
- scient
Suppresses
- Her
- еств
- /her
- idential
- her
- 都市报
- 叆
- unks
- 垓
- her