'to' after 'come' or 'jump'
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 lights up on the word 'to' in phrases like 'come to a conclusion' or 'jump to conclusions', where the sentence is about reaching a conclusion. When it was turned up, the model's text drifted toward phrases like 'come up with' and 'come to'. When turned down, the text fell into repeated apologies.
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.
-
, to the time of the creation of the world. Provisionally, based on the studies published in Biblical Chronology over the past two years, I have come to the year 3932 B.C. as the year of creation,
-
importance of connections between the biological diversity of the Sky Islands and that of the Sierra Madre Occidental of Mexico. Many separate scientific analyses of the region have come to the same conclusion: the biological diversity of the individual mountain ranges is in peril if
-
933 at<|endoftext|>Bedross Der Matossian⏎One of the outstanding issues in Armenian Genocide historiography has been the inability of historians to come to a consensus regarding the causes, the aim of the perpetrators, and the process of
-
fairly far down on a tree that relies on a pyramidal shape to really look nice, then it too should probably be taken out.”⏎But don’t jump to hasty conclusions, Trinklein said. Just like you can’t unring
-
of the reason they won…other than sheer suicidal bravery and determination…was another belief that underlies the two involving women discussed here: Resurrection.⏎I came to identify this concept as Maccabean. Book Two of the Books of the
-
University Police will assess the situation for danger of imminent harm.⏎2. A team led by the Dean of Students and the Chief Diversity Officer will review the report and reach out to you to provide care, offer support, and learn more about the incident
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: it's not. It's not. I don't know what I mean. But it's not. It's not. Maybe. Maybe. Maybe. (c) -) I need to know what I mean by the "d"
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and need to know where is the Ganges river. I know it is in India but I need more details.⏎I would really appreciate your help with this question :-)⏎Hi Rebecca, it is a good question but difficult to answer because of the
-
completely useless :P I just felt like correcting the other code (too tired to check who it was that posted it, and I can't see just now anyway :P)⏎void genFactorial(int x, int t)⏎int total = t
-
the salt out of their bodies.⏎Again, the Green Sea Turtles are a protected species and it is VERY important to NEVER TOUCH THEM!⏎Happy Diving! :)<|endoftext|>Few deny there are major global challenges ahead — from climate change and loss of
-
here about those 'Box-Shaped Watermelons'!? Or how about a yellow center instead of pinkish-red! Seedless?⏎See you next time! :-)<|endoftext|>A Safe Hospital, Reduce Risk, Save Lives⏎The Mission of Safe Hospital
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
- L6:2009 excites
- L10:2309 weakly suppresses
- L2:1205 weakly suppresses
- L9:846 weakly excites
- L13:906 weakly excites
- L13:2684 weakly suppresses
- L5:1243 weakly suppresses
- L2:2784 weakly suppresses
- L2:1346 weakly suppresses
- L13:2864 weakly excites
- L7:135 weakly excites 'of' after 'issue' in citations
- L0:2897 weakly excites
- L0:793 weakly suppresses
- L10:2688 weakly suppresses
- L14:2446 weakly excites
- L11:1593 weakly suppresses
- L12:947 weakly excites
- L10:1485 weakly excites
- L6:1693 weakly excites
- L11:1786 weakly suppresses
- L1:887 weakly suppresses 'be' after 'might', 'also' or 'only'
- L2:275 weakly suppresses 'of' after effectiveness or instead
- L13:967 weakly excites
- L8:2829 weakly suppresses
Feeds later layers
- L16:592 suppresses
- L18:746 weakly suppresses
- L16:1371 weakly excites
- L25:2167 weakly suppresses
- L16:2743 weakly excites
- L17:1171 weakly suppresses
- L18:828 weakly suppresses
- L24:682 weakly excites
- L18:2068 weakly suppresses
- L17:2969 weakly suppresses
- L18:1882 weakly suppresses
- L17:259 weakly suppresses
- L25:1600 weakly excites
- L19:327 weakly suppresses
- L22:1027 weakly excites
- L19:1610 weakly excites
- L20:1453 weakly suppresses
- L21:1524 weakly suppresses
- L19:973 weakly suppresses
- L18:850 weakly suppresses
- L21:1703 weakly suppresses
- L25:147 weakly excites
- L26:2240 weakly excites
- L17:1286 weakly 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
- 为首的
- \/
- олн
- 为首
- "]=="
- iode
- _|
- stime
- Bail
- =db
Suppresses
- 라이
- OLDER
- ingo
- 嘿嘿
- ARIO
- pop
- Osw
- earned
- ENDER
- AGO