'says' after a quote
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. 14 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 the word 'says' when it comes right after a quotation and introduces who said it, as in news stories. Turning it up gave unrelated text, so the steering does not confirm the pattern.
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.
-
funded by the N.C. Commercial Fishing Resource Fund Grant Program.⏎“The goal of the research was to better understand the economic impact of North Carolina’s seafood industry,” says Jane Harrison, North Carolina Sea Grant’s coastal economics specialist and a project lead.
-
tardigrades are vulnerable to high temperatures, though it seems that these critters would be able to acclimatize to increasing temperatures in their natural habitat,” says Ricardo Nieves, an author of<|endoftext|>January 10, 20
-
was first noted, a number of misconceptions, myths, and dogmas have arisen regarding fundamental aspects of boll weevil ecology in the subtropics,” says Showler. “These include its food sources, overwintering, and di
-
DCIS.⏎“The transition from DCIS to invasive ductal cancer is a key event in breast cancer progression, but it remains poorly understood,” says principal investigator Dr. Carlo M. Croce, director of Ohio State´s Human
-
to the holy grail of foot-and-mouth vaccines. Unlike the traditional vaccines, there is no chance that the empty shell vaccine could revert to an infectious form,' says Professor Stuart.⏎Dr Charleston adds: 'The ability to produce a vaccine outside of
-
Earth's surface. But when researchers looked at the western foothills southeast of Fresno, "we were no longer able to identify where the Moho was present," says Hersh Gilbert, a co-author
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: what if this? well, what is the value of a different approach to the problem of the question, and what is called this? this. for what, and, and so, and, and, and, and, and, and
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
fulness in children has been associated with obstinacy — a refusal to bend and comply with the wishes of a superior. We assume that our children
-
eltis polypodioides. It gets its name from its effective, defense mechanism against droughts. Being able to survive decades without water and still be able to return to a healthy state after some watering. This has made a lot of
-
are SDG 11 and SDG 16. SDG 11 purposes to “make cities inclusive, safe, resilient and sustainable” and SDG 16 seeks to “promote just, peaceful and inclusive societies”
-
firsthand experience of paying bribes or using a contact to get a job done in public office.⏎In 2012 India has ranked 94th out of 176 countries in Transparency International's Corruption Perceptions Index, tied
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
- L7:7 suppresses
- L7:8 suppresses
- L13:3 suppresses
- L6:325 suppresses
- L6:60 weakly suppresses
- L7:135 weakly suppresses 'of' after 'issue' in citations
- L10:43 weakly excites "terms" in "in terms of"
- L13:1000 weakly suppresses
- L6:443 weakly excites
- L6:823 weakly suppresses
- L9:143 weakly suppresses
- L8:319 weakly excites
- L9:35 weakly suppresses
- L9:330 weakly excites gap where a number was removed
- L13:338 weakly suppresses
- L13:866 weakly excites
- L8:187 weakly suppresses
- L9:933 weakly suppresses
- L10:458 weakly excites
- L7:1723 weakly excites
- L1:2604 weakly suppresses
- L13:1372 weakly suppresses number before an age or week count
- L6:562 weakly suppresses
- L5:248 weakly excites
Feeds later layers
- L18:1014 excites
- L15:1601 weakly excites
- L15:248 weakly excites
- L20:48 weakly excites
- L16:2600 weakly excites
- L16:326 weakly excites
- L19:152 weakly suppresses
- L16:2521 weakly excites
- L18:2448 weakly suppresses
- L16:1135 weakly excites
- L16:292 weakly excites
- L23:2043 weakly suppresses
- L24:511 weakly excites
- L23:2522 weakly suppresses
- L16:662 weakly excites
- L15:576 weakly suppresses
- L24:1395 weakly excites
- L17:2057 weakly excites
- L22:832 weakly suppresses
- L15:1524 weakly excites
- L18:1837 weakly suppresses
- L16:2182 weakly excites
- L18:2529 weakly excites
- L23:1922 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
- onis
- ingly
- ronics
- elon
- ombok
- 不代表
- URL
- 还有很多
- uess
- oring
Suppresses
- 剽
- sinking
- ожно
- riot
- WHEN
- 每
- BAM
- utow
- тра
- 体内