'in' as a word-start 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. 25 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 responds to the letters 'in' when they start a longer word that has been split into pieces, as in 'in|coherent', 'in|rush' or 'in|humane'. Turning it down made the model get stuck repeating 'in the in the', which shows it is tied to that little 'in' piece.
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.
-
fighters to deter pursuit).⏎Several U.S. aircraft were shot down as the first wave approached land; one at least radioed a somewhat incoherent warning. Other warnings from ships off the harbor entrance were still being processed
-
concomitant users of warfarin and a skeletal muscle relaxant and, as pre-specified negative control comparators, among concomitant users of an in
-
n.) Hence: To carry on a process of deduction or of induction, in order to convince or to confute; to formulate and set forth propositions and the inferences from them; to argue.⏎7. (n.) To converse; to
-
20 V. Using additional external components, higher voltages can be achieved. The ADP1613 has an adjustable soft start function to prevent inrush current when the device is enabled. The pin-selectable switching frequency and PWM
-
............... . This enables the nephrons to maintain a stable GFR even when arterial blood pressure changes.⏎Which mechanism of renal autoregulation responds to an INCREASE in BP which causes an increased STRETCH in the afferent arter
-
2) (Learn how and when to remove this template message)|⏎Sendust is a magnetic metal powder that was invented by Hakaru Masumoto at Tohoku Imperial University in Sendai, Japan, about 1936 as
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: In in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
skills pay dividends beyond academic writing. By practicing persuasive writing, students also build a solid foundation in the kind of critical thinking and reading necessary for them to be well-informed citizens.<|endoftext|>"You're a strong skater, Piet, and you have
-
forest protection by buying tracts of land, but which often fail to address the needs or rights of local Indigenous peoples. Uninhabited forest areas often go unprotected, even if purchased through a buy-an-acre
-
Fauj’ in 1943 which was initially formed in 1942 by Rash Behari Bose. He had visited London during pre-independence period to discuss the
-
harvesting and selective harvesting at high and extra-high intensities<|endoftext|>I know you are used to reading about those famous and well-known freedom fighters who existed in the pre-independence regime; the likes of Chandra Bose, Abrindra Kumar, Kh
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
- L22:535 excites
- L24:1546 suppresses
- L9:2415 excites
- L23:1359 suppresses
- L23:1960 suppresses
- L21:2137 excites
- L19:1626 weakly suppresses
- L22:2526 weakly suppresses
- L17:1766 weakly suppresses
- L23:369 weakly excites
- L20:3029 weakly suppresses
- L21:2256 weakly suppresses
- L24:2514 weakly suppresses
- L6:2649 weakly excites
- L15:2931 weakly excites
- L17:706 weakly excites
- L15:3030 weakly excites
- L22:1112 weakly excites
- L24:2806 weakly suppresses
- L19:966 weakly excites
- L22:2162 weakly suppresses
- L18:939 weakly excites
- L8:1862 weakly excites
- L23:126 weakly excites
Feeds later layers
- L27:998 strongly suppresses
- L27:873 suppresses
- L27:918 suppresses
- L26:786 suppresses
- L27:1229 suppresses
- L27:1031 weakly suppresses
- L26:2107 weakly suppresses
- L26:2314 weakly excites
- L26:2078 weakly excites
- L27:1715 weakly excites
- L26:12 weakly excites
- L27:2701 weakly excites
- L26:79 weakly suppresses
- L27:2577 weakly excites
- L26:1478 weakly suppresses
- L26:2133 weakly excites
- L27:210 weakly suppresses
- L26:164 weakly suppresses
- L26:3009 weakly excites
- L26:1620 weakly excites
- L26:1189 weakly excites
- L27:1600 weakly suppresses
- L27:1523 weakly excites
- L27:882 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
- .lwjgl
- acement
- assist
- 的认可
- 用量
- 助け
- �认
- roduction
- pNet
- 潢
Suppresses
- 包括
- include
- including
- includes
- Including
- Includes
- including
- Include
- 还包括
- Including