'what' or 'how' after see or understand
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. 9 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 'what' (or 'how') right after verbs like 'see', 'understand', or 'work out', where a question-like clause is about to follow. Turning it up did not bring out this theme in the sample text, which went off topic.
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.
-
and has not in fact used measurements of the Antarctic "hole". There's many a slip twixt cup and lip there. It will be interesting to see what is in the relevant journal article when -- and if -- it has passed peer review
-
; let’s say we want to see what has changed between Commit 1 and Commit 3.⏎The diff will compare the
-
sight or at the very least supplementing what they see with extra-sensory information. For example, whales or bats,
-
lie completely within the designated 15-mile diameter count circle, and the site has opened its land up to birders since 1996 to see what interesting and unique birds call the site home or have stopped by on their way south
-
Level 5.0 - 6..5⏎Students will truly understand what they read with the help of these easytouse reading comprehension cards! Your students
-
stopped them. In evidencing dictionary items, OED editors are limited by and large to what’s available in print. This makes finding early examples of colloquial language very
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 don" 1. 1.1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
help. The sooner you receive care for your anxiety symptoms, the better. Feel free to reach out
-
majority of Company B, to whose leadership Little Wound was promoted as first sergeant. Major chiefs Red Cloud, Young
-
a heart-healthy lifestyle. Read more in these topics about reaching and staying at a healthy weight:⏎Seeing your doctor regularly and getting screening tests is important. The sooner your doctor diagnoses a disease, the more likely
-
"Maintaining and improving the chemical, biological, and physical integrity of the Great Lakes is the central commitment between the United States and Canada. The report documents how far we have come and how much further we need to go," said Dereth Gl
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
- L4:927 excites
- L6:502 suppresses
- L4:470 suppresses
- L4:227 suppresses
- L7:2058 excites
- L2:1957 suppresses
- L4:2197 suppresses
- L4:981 suppresses
- L5:458 excites
- L8:132 suppresses
- L8:369 suppresses 'of' after 'learned' or 'because'
- L3:2122 excites
- L8:1997 excites
- L1:2889 suppresses
- L1:1457 weakly suppresses
- L8:2146 weakly excites "as" in "it seems as if" phrases
- L4:2719 weakly suppresses
- L3:1538 weakly excites 'what' as a free relative
- L4:1582 weakly excites
- L5:570 weakly excites
- L4:534 weakly suppresses
- L4:956 weakly excites 'that' after a noun, introducing a clause
- L1:1616 weakly suppresses
- L8:7 weakly excites
Feeds later layers
- L13:344 excites
- L16:2975 excites
- L17:1282 excites
- L17:1407 excites "and" joining questions (wh-question context)
- L16:817 excites
- L17:2368 excites
- L12:434 excites
- L16:948 excites
- L18:1617 excites
- L11:2591 excites
- L12:1667 excites
- L19:1703 excites
- L15:1473 excites
- L19:2737 excites
- L17:1884 excites
- L13:2728 excites
- L17:2967 excites
- L15:2893 excites
- L16:1078 excites
- L17:2280 excites
- L15:467 excites
- L21:639 weakly excites
- L16:528 weakly excites
- L17:2718 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
- arsi
- eve
- 记者了解
- arp
- alam
- arra
- ever
- 我说
- briefed
- STDCALL
Suppresses
- ộng
- زي
- hài
- 过
- StateException
- Including
- 植物
- licer
- inconsistent
- 砚