comma after a 'that' or 'who' clause
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. 6 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 a comma that comes right after a word like 'that' or 'who', where the sentence pauses to add an extra clause. Turning it up did not produce a clear, matching change in the written 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.
-
000 acres of pastures for cattle and sheep that, starting in the late 19th century, gave way to development of the
-
include:⏎- Limit or abstain from alcohol.⏎- Avoid overeating and, if overweight, implement healthy strategies to lose weight.⏎- Cut back on refined carbs
-
have considerably larger suppression costs and a much greater chance of escaping...⏎The ionospheric Alfvén resonator (IAR) is an area of space that, due to sharp Alfvén speed gradients, has the ability to trap electromagnetic plasma
-
the powerless, and the individual responsibility and autonomy of each person under God. In the New Testament, four different Greek words are used in reference to those people who, today, we call "Elders." A study of these four words -
-
Every data-driven trend comes with vulnerabilities that, if left unchecked and unprotected, can cause great damage.⏎In such cases, no
-
cards over three rounds, each new round begins after seven cards have been won. This is accomplished by translating cards in both languages. The winner is the player who, in the final round, translates all of the played cards error-free.⏎Novices
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: the number of the month of the number. I am not sure about the details of the plan. I need to think about something. I am not sure about the details of the plan. I need to think about something. I am not
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
There are three possible levels at which this problem can be⏎tackled, and how it is used depends on the levels of the students.⏎As an exercise in confidence intervals it is relatively simple from⏎a calculational point of view; the
-
web-based systems.⏎So, despite spending millions of dollars on these ERP systems, organizations are still being relegated to the use of manual, paper-oriented processes and forms to fulfill various core
-
stamens 6, included, inserted at perianth base, anthers attached ± near middle; ovary ± sessile, style 1, entire or 3-branched⏎Fruit: capsule, loculicidal
-
The scientists could directly relate the lorikeets' deaths to inadequate artificial diets, and to the unsanitary conditions of trays and other containers from which they were being fed<|endoftext|>Published on March 1, 2013 by Carol⏎
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
- L5:1217 weakly excites
- L1:1573 weakly suppresses
- L0:340 weakly suppresses
- L3:2118 weakly suppresses
- L2:2065 weakly excites
- L2:2388 weakly suppresses
- L2:2235 weakly excites
- L5:1506 weakly excites
- L3:305 weakly excites
- L0:2194 weakly suppresses
- L2:2897 weakly excites
- L2:156 weakly suppresses
- L4:1238 weakly excites
- L5:2657 weakly suppresses
- L5:713 weakly suppresses the spot after a comma in a number list
- L5:1858 weakly excites
- L0:807 weakly excites
- L3:1214 weakly excites
- L0:136 weakly suppresses space after a dash in number ranges
- L2:1870 weakly suppresses
- L1:1972 weakly suppresses
- L5:611 weakly excites
- L4:1539 weakly excites
- L4:631 weakly suppresses
Feeds later layers
- L9:459 suppresses
- L15:212 weakly suppresses
- L9:2267 weakly suppresses 'with' after exchange nouns
- L7:14 weakly suppresses
- L10:2279 weakly excites
- L22:201 weakly excites
- L22:761 weakly excites
- L17:769 weakly excites
- L12:2140 weakly excites 'the' after place prepositions
- L24:2543 weakly suppresses
- L23:1151 weakly suppresses
- L13:95 weakly suppresses
- L11:2199 weakly excites
- L18:1134 weakly suppresses
- L11:1392 weakly suppresses 'of' in fixed technical phrases
- L26:2134 weakly excites
- L14:126 weakly excites
- L21:2947 weakly suppresses
- L9:1562 weakly suppresses 'by' before a byline or agent
- L20:101 weakly suppresses
- L16:965 weakly excites
- L24:457 weakly excites
- L26:2424 weakly suppresses
- L15:622 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
- atra
- among
- 兮
- ập
- here
- /current
- -pop
- .Pop
- 说我
- periodically
Suppresses
- stro
- \Resource
- riding
- oice
- 力还是
- чная
- zte
- 裱
- 床位
- 很大的