'of' in 'end of' and 'middle of' phrases
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. 12 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
The neuron lights up on the word 'of' in time phrases like 'by the end of', 'middle of' or 'as of', where a period of time is about to be named. It was quiet or pushed the other way on phrases about 'times' and 'cases' or situations. Turning it up did not produce a clear theme in the model's replies.
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.
-
hundreds of elementary students across the toronto district school board have been Children who attend preschool enter kindergarten with improved skills but that advantage is nearly halved by the end of the year quot ensuring that young children enter kindergarten
-
ways she could, including hosting musical soirées that included guests such as Wolfgang Amadeus Mozart.⏎Her music, though largely considered out of fashion by the end of her life, took many
-
of Know-how discovered that college students reported a 42% improve in stress and nervousness and a 38.7% improve in depression throughout the 90 days following lively shooter drills, in contrast with the 90 days
-
ekenen en vergelijken" on the citizen science platform MijnTuinlab (My Garden lab).⏎The project was enthusiastically received, as by the end of September the perimeters of more than 10,000 gardens had
-
who owned small firearms were encouraged to register them, while new requests for licences to carry firearms were suspended for a year. At the end of 2012, a draft arms control law was before Congress.Top of
-
of academic failure.⏎Over 30,000 students in DC Public Schools (DCPS) and public charter schools are considered at-risk. As of the
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 am not the case where I have to be. I am not the case where I have to be. I am not the case where I have to be. I am not the case where I have to be. So, what is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
long as the allergy exists. Food challenge tests for safe re-introduction are performed under the supervision of a registered clinical expert, such as a diet<|endoftext|>There are times in history when prudent decisions have allowed France to maintain
-
vision very dark and fuzzy. When used in this way, a vignette will generally not be consciously detectable by the untrained eye. However, there are times when a more dramatic vignette is a useful effect, such as in a black
-
the lender will be assessing the borrower's ability to repay the loan based on a worst-case scenario interest rate. The calculation for individuals with irregular income may be different from those
-
cases, small samples of tissue are taken for further analysis. It may also be necessary to retain entire organs for a more detailed examination. This is strictly limited to cases where such retention<|endoftext|>Published August 30, 2004⏎
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
- L10:456 suppresses
- L10:262 suppresses
- L10:2346 suppresses
- L8:1760 weakly excites
- L8:1522 weakly excites
- L6:2090 weakly suppresses
- L4:779 weakly excites
- L8:2424 weakly excites
- L10:753 weakly excites
- L11:1893 weakly excites
- L9:2124 weakly excites
- L9:2819 weakly excites
- L7:3049 weakly suppresses
- L11:949 weakly excites
- L7:160 weakly suppresses
- L6:2362 weakly excites
- L7:2461 weakly excites
- L9:2425 weakly excites
- L10:2644 weakly suppresses
- L5:783 weakly suppresses
- L8:2520 weakly excites
- L1:3044 weakly excites
- L1:2550 weakly suppresses
- L4:219 weakly suppresses
Feeds later layers
- L16:1538 suppresses
- L13:2282 suppresses
- L23:1157 suppresses
- L20:2078 suppresses
- L17:1625 suppresses
- L18:2172 suppresses
- L23:1612 suppresses
- L22:2667 suppresses
- L27:2569 suppresses
- L14:2514 suppresses
- L16:2929 suppresses
- L13:368 weakly suppresses
- L25:2688 weakly suppresses
- L18:483 weakly excites
- L13:2122 weakly suppresses
- L20:924 weakly suppresses
- L22:2430 weakly suppresses
- L25:2200 weakly suppresses
- L21:1309 weakly suppresses
- L14:770 weakly suppresses
- L20:1355 weakly suppresses
- L14:2564 weakly excites unclear: word before a technical term
- L19:2680 weakly excites
- L16:2593 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
- 敛
- iers
- mites
- ister
- lian
- Calder
- son
- calor
- kov
- rimon
Suppresses
- 情形
- 等情况
- 的情况
- situations
- 的情形
- cases
- 情况
- 情况下
- cases
- ситуации