'the' after a verb or preposition
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. 1 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 "the" when it comes right after a verb or preposition, such as "follow the", "within the" or "maintain the". Turning it up gave text that did not stay on one topic.
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.
-
Romanisation of Chinese⏎Romanisation (or latinication) refers to the written representation of a Chinese word by using the Latin (Roman) alphabet.⏎To appreciate the reasons for variances in the romanised spelling of Chinese words one has to
-
splints due to excessive muscle tightness or other factors. To avoid this problem, follow the “10% rule”: increase mileage no more than a 10%
-
of 2016, and it's very satisfying to have come so far in the process that the plans will become reality. It's interesting to follow the customer in this exciting project, and we will do everything in our power to earn
-
Transmission, Distribution, & Refining⏎Multispectral and Hyperspectral Remote Sensing Techniques for Natural Gas Transmission Infrastructure Systems⏎The goal is to help maintain the nation's natural gas transmission infrastructure through the timely and effective detection of natural gas leaks
-
concept of cell organization in response to these morphogen gradients is well documented, little is known about how these subtle concentration changes get established the first place, particularly within the dynamic environment of a real tissue. This research provides evidence that tiny biophysical forces
-
soil slides and/or gullying which primarily affect roads themselves, wiping them out and blocking vehicle circulation, and secondly, destroying or burying agriculture lands. Along the 138 km surveyed, we calculated an amount of soil material released of
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: I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of the world's oldest, and largest monotheistic religions. These religions share a variety of customs, beliefs, and even practices
-
their forms in the monotheism.⏎The UNITY and ONENESS is a general topic of conversation everywhere and
-
six centuries before Christ. The Greeks and the Romans, meanwhile, gave the power of flight to their gods and to the mischief-making followers of those gods.⏎Polytheistic or non-literate cultures very often contain a belief 1n winged
-
the kind created to soothe the terrors of the monotheists were less necessary. The Romanian gypsy, half Christian, half pagan prayed
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
- L0:774 suppresses
- L0:26 weakly suppresses
- L0:1966 weakly excites
- L0:2507 weakly suppresses
- L0:178 weakly suppresses
- L0:1224 weakly suppresses
- L0:1301 weakly suppresses
- L0:3005 weakly suppresses
- L0:1020 weakly suppresses
- L0:2495 weakly excites
- L0:1860 weakly suppresses
- L0:333 weakly excites
- L0:2746 weakly suppresses
- L0:30 weakly suppresses
- L0:2209 weakly excites
- L0:2740 weakly excites
- L0:713 weakly suppresses
- L0:95 weakly suppresses
- L0:2175 weakly suppresses
- L0:890 weakly suppresses
- L0:517 weakly suppresses
- L0:2699 weakly excites
- L0:535 weakly excites
- L0:1913 weakly excites
Feeds later layers
- L16:792 excites "the" inside a prepositional phrase
- L17:30 excites
- L9:24 excites "the" before a named organisation or title
- L26:1941 excites
- L16:1776 excites 'the' after a verb or preposition
- L21:614 excites
- L18:641 excites
- L16:2266 excites 'the' before study or journal names
- L25:1414 excites
- L3:87 excites 'the' after 'member of'
- L17:1481 excites 'the' after 'some of' or 'most of'
- L19:1376 excites
- L20:361 excites
- L15:716 excites
- L20:219 excites "the" after time or place prepositions
- L24:364 weakly suppresses 'at' in 'at least' and 'most' or 'least' qualifiers
- L18:2380 weakly excites "the" after certain verbs and prepositions
- L4:52 weakly excites
- L17:359 weakly excites
- L16:419 weakly excites
- L17:2190 weakly excites 'the' starting an object phrase
- L18:799 weakly excites
- L19:427 weakly excites
- L21:202 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
- over
- next
- first
- overhead
- best
- chemical
- above
- output
- area
- answer
Suppresses
- rade
- 的战略
- 的理想
- aille
- 的地位
- veled
- صة
- 的身体
- 的位置
- 各省