unclear: chinese academic text, 'when' loops
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. 20 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 is mostly quiet and fires weakly on Chinese academic text and a few scattered words. Turning it down made the model repeat the word 'when' over and over.
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.
-
法、分析法来进行研究,探讨泰国学生在学习汉语助动词中的困难及偏误情况,分析他们产生偏误的原因。我们发现经过基于对比研究设计的汉语助动词课堂教学后
-
bs for Thai Learners Based on the Comparative Analysis⏎- 人文学院-已发表论文 ⏎“助动词”在汉语中具有较大的作用,使用率高,但是泰国学生经常在学习和使用汉语助动词中
-
が含まれ、1人の参加者がいます。2020年7月6日 at 9:45 AMに Megan Smith さんが最後の更新を行いました。⏎2020年7月6日 9:4
-
densities."⏎The study, "Multi-State Trials of Bt Sweet Corn Varieties for Control of the Corn Earworm (Lepidoptera: Noctuidae)," analyzed the performance of Bt sweet corn, comparing its rate of inf
-
Hawaii's Oahu flattens from an island into a low-lying seamount. Researchers writing in the upcoming Feb. 15 issue of the journal Geochimica et Cosmochimica Acta estimate that the volcanic island will
-
plants, among them poison hemlock and fool's parsley ... Cow parsley can be confused with giant cow parsley/giant hogweed (Heracleum mantegazzianum), the sap of which can cause severe burns after coming in contact with 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: when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when when
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in the concentration of species in solution.⏎It’s frustrating that the article makes it seem like the applications of a really awesome discovery in pure science are mere moments away, rather than explaining why a discovery like this is awesome and encouraging peoples’ interest in
-
人文学院-已发表论文 ⏎“助动词”在汉语中具有较大的作用,使用率高,但是泰国学生经常在学习和使用汉语助动词中产生各种的偏误。本文通过调查法、统计
-
tendency, in some quarters, to jump to a solution like bariatric surgery, and rhapsodize about what an acceptable safe effective treatment alternative it is, without ever examining the underlying reasons for the obesity.⏎Reasons for obesity⏎Probably
-
Augmented Reality links knowledge/content with relevant visual imagery.⏎3 Steps to Creating an Awesome Virtual Museum in Class You're spending an afternoon browsing the exhibits at an art museum. If you're anything like me, you'd probably appreciate the art a lot
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
- L18:1910 excites
- L19:219 suppresses
- L16:2782 excites
- L15:2140 excites
- L15:830 excites
- L15:1774 suppresses 'only' in 'not only ... but also'
- L19:1084 suppresses "and" joining clauses after a condition
- L16:1223 excites
- L18:845 weakly excites
- L18:409 weakly excites
- L18:1864 weakly excites
- L18:1249 weakly excites
- L18:1003 weakly excites
- L19:400 weakly excites
- L16:1556 weakly suppresses
- L18:1582 weakly suppresses
- L17:1081 weakly suppresses
- L18:251 weakly excites
- L18:2248 weakly excites
- L19:547 weakly excites
- L19:1966 weakly excites
- L0:193 weakly excites
- L12:758 weakly suppresses
- L15:2201 weakly excites
Feeds later layers
- L23:1645 strongly suppresses
- L27:659 strongly suppresses
- L25:633 strongly suppresses
- L23:449 strongly suppresses
- L21:1194 strongly suppresses
- L24:148 suppresses
- L22:676 suppresses
- L24:1542 excites
- L21:2960 suppresses
- L25:565 suppresses
- L25:1593 suppresses
- L22:1286 suppresses
- L23:2814 suppresses
- L24:2018 suppresses "first" in "for the first time"
- L21:1023 suppresses
- L26:1403 excites
- L22:1056 suppresses
- L26:2065 suppresses
- L26:814 suppresses
- L25:626 weakly suppresses
- L21:2942 weakly suppresses comma before a clause that contrasts a claim
- L21:1797 weakly suppresses
- L25:925 weakly suppresses
- L23:353 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
- 那么简单
- awei
- ENDIF
- rein
- 槿
- ayscale
- 衽
- ắng
- idences
- ложение
Suppresses
- 时
- 的时候
- when
- 时候
- when
- 時
- When
- 的時候
- 时表示
- When