'st' starting a word fragment
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. 27 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 letters 'st' when they start the second half of a word that gets split up, like 'sto' in stoma or 'stav' in staves. When it was turned up, the model's replies drifted into odd, repetitive phrasing rather than anything clearly about those words.
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.
-
banding is specially designed to limit food consumption in severely obese patients. When placed around the patient's stomach, the saline-filled band forms a small gastric pouch and stoma. Inflation of the band can be adjusted to control the amount of food
-
first step should be to become familiar with the music notes.⏎In the Western system of music notes, written music is arranged on two sets of five lines, called staves. The upper stave is called the treble stave; the lower
-
take your chances in the no-holds-barred class, or on the roof.⏎Passengers were still pushing their way on board and trying to find somewhere to stow their endless boxes and bags of luggage when a bell clanged and we were
-
chunks of dirt and remove rocks. The axe-style blade on the head of a cutter mattock is ideal tool for chopping through root-laden soil near tree stumps.⏎Forks are particularly useful for turning soil, digging up root crops,
-
may rise by 3,000 percent by 2029 with better regulations.⏎The ever increasing demand for carbon credits could be the straw that stirs the milkshake for climate innovation in Mongolia and possibly resolves the air pollution problem
-
for domestic rubbish. This is because the biomass in domestic rubbish can only be gasified at low temperatures, but this in turn creates tars and waxes that stifle the process.⏎However, it turns out that these tars and waxes
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: maybe even digress even digress digress digdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdig
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a thin, flimsy plastic but you could, and still can, buy proper-lensed 3D blue and red glasses) use what is called anaglyph technology. This has since been superseded by Passive and Active 3D
-
wall next to the art, a statement from the artist about their intent for the piece.⏎When Marcel Duchamp submitted an upside-down porcelain urinal signed “R.Mutt”<|endoftext|>The Galileo Museum in Florence is an international reference on Galileo
-
Re became Tefnut's father. According to one myth, Tefnut became angry with her father Re and ran off to Nubia. The god Thoth went to find her and conjoled her into returning to Egypt.⏎
-
00-205. See general information about how to correct material in RePEc.⏎For technical questions regarding this item, or to correct its authors, title, abstract, bibliographic or download information, contact: (Michael McNulty
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
- L21:237 suppresses
- L23:206 suppresses
- L20:669 suppresses
- L24:105 weakly suppresses
- L18:151 weakly suppresses
- L1:3019 weakly suppresses
- L16:1129 weakly excites
- L18:1077 weakly excites
- L15:434 weakly excites
- L20:52 weakly suppresses
- L18:775 weakly suppresses
- L20:2642 weakly excites
- L21:475 weakly excites
- L18:2453 weakly suppresses
- L2:289 weakly suppresses
- L5:2370 weakly suppresses "from" after change or transfer verbs
- L21:2904 weakly excites
- L20:151 weakly excites
- L22:589 weakly excites
- L9:1544 weakly excites
- L22:3027 weakly excites
- L17:1948 weakly excites
- L4:2778 weakly suppresses
- L12:2943 weakly suppresses
Feeds later layers
None: this is the last layer.
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
- üler
- lients
- ows
- owed
- 男士
- ected
- ipse
- -win
- ieri
- Best
Suppresses
- dig
- уд
- よう
- cepts
- 二代
- code
- rens
- //**⏎
- 侗
- odia