unclear: mid-phrase words, mixed
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 mixed: they peak on different tokens. 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 grab bag of words that sit in the middle of a phrase, such as 'Placement', 'behind', 'succeed' and 'after'. No single theme stands out. When it was turned up, the model's writing drifted into unrelated rambling about math and reasoning.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
websites and started applications. It is also measuring time in front of the monitor.<|endoftext|>DRDO Placement Paper (EMT & Microprocessor)⏎Paper: DRDO Placement Paper (EMT & Microprocessor)⏎EMT and microprocessor:⏎Make sure
-
in “Internet Explorer” into the bar at the bottom left of the screen and click on it. It should open up after clicking.⏎· Type HydroShare.org into the browser and click on it – it should open up.⏎·
-
central and southern Andes for its edible root. In Peru and Bolivia, for example, this tuberous root vegetable is the second most widely grown root crop behind the potato. In these countries, the
-
agreed-on method for creating citations. It provides formats for citing differnet information resources within the body of the paper, for listing different information types on the Works Cited page, and even ways to set up headings and margins for the paper. There
-
confederation by distributing the major cities among his sons; the eldest was to be grand prince of Kiev, and the brothers were to succeed each other, moving up the hierarchy of cities toward Kiev, filling vacancies left by
-
a way to land the probe safely on the celestial body.⏎When Philae reaches the surface of the comet, the lander will harpoon itself into place
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 following is a summary of the first and last paragraph, but it's not a math problem. (The first and last paragraph. The first and last. I need not understand the meaning. But that is. The first and last.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
helps to boost fertility.⏎- Avocado plays an important role in women’s health. It helps in estrogen production and reproductive hormonal balance.⏎- Nutrients like folic acid, potassium, fiber and other minerals make it super food for good health.⏎
-
on family size and future increase in the overall population. In addition, a mother’s age, whether younger or older has an effect on outcomes of births including birth weight, multiple births, and birth
-
, level surface. From one or more of these vantage points, you should draw an imaginary plumb line from the top of the leg to the ground. If you cannot draw a
-
begin, both mosquito and microbial small and large subunit rRNA fragments are amplified from a metagenomic community DNA sample. Then, the community specific biotinylated antisense ribosomal RNA probes are synthesized in vitro using T7 RNA
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
- L8:7 suppresses
- L6:13 suppresses
- L8:16 excites
- L8:10 weakly suppresses
- L11:163 weakly excites
- L8:14 weakly suppresses
- L8:1319 weakly suppresses
- L6:8 weakly excites
- L6:12 weakly suppresses
- L8:12 weakly suppresses
- L10:535 weakly excites
- L7:303 weakly excites
- L9:1157 weakly suppresses
- L6:16 weakly suppresses
- L5:158 weakly excites
- L9:1766 weakly suppresses
- L3:1662 weakly excites
- L5:133 weakly suppresses
- L7:49 weakly excites
- L9:665 weakly suppresses
- L8:59 weakly excites
- L6:17 weakly excites
- L8:4 weakly suppresses
- L10:1373 weakly suppresses
Feeds later layers
- L14:585 excites comma opening an inserted clause
- L13:2153 excites
- L14:1951 weakly excites
- L13:803 weakly excites
- L17:2082 weakly suppresses
- L14:2192 weakly excites
- L15:3027 weakly excites
- L15:1381 weakly suppresses
- L21:248 weakly suppresses
- L16:2742 weakly excites end-of-document token after a finished text
- L16:1251 weakly suppresses
- L17:2673 weakly excites
- L15:270 weakly excites
- L14:2119 weakly excites
- L14:1618 weakly excites
- L15:1787 weakly excites
- L14:2203 weakly excites
- L16:1436 weakly suppresses
- L17:2567 weakly suppresses
- L23:189 weakly suppresses
- L20:1201 weakly excites
- L16:1449 weakly suppresses
- L18:1501 weakly suppresses
- L16:3027 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
- erus
- shedding
- isure
- ett
- ỵ
- "profile
- itches
- :block
- chod
- udes
Suppresses
- antic
- 麦克
- assic
- нем
- енд
- 分
- Indo
- ней
- 绿城
- ает