'the' in 'on the other hand' phrases
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 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' in set phrases like 'on the other hand', 'on the one hand' and 'on the right'. When it was turned up, the model got stuck repeating the word 'one', which is the word the neuron seems to push toward.
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.
-
25 times more potent than carbon dioxide.⏎In the Communication, the Commission recommends realizing bio-waste potential as a renewable source of energy and recycled materials, on the basis of enhanced implementation of existing legislation at the national level with support at the EU
-
on hands-free mobiles while driving⏎Brake has renewed its call for the government to look again at the law surrounding hands-free mobile phones and driving, on the back of new research which concluded that talking on a hands-free phone can be just
-
child's development and, on the 'Words for Life' website, there are some useful tips about how you can
-
to be made. A pop icon or populist president might discover advantages in gainsaying the grammar of standard English in favour of a local dialect or patois. On the other hand, scrupulous attention to the rules of prescriptive grammar might be
-
weekly earnings by worker characteristics, including educational attainment on the right-hand side:⏎Source: New York Times, Bureau of Labor Statistics⏎For
-
genotypes may represent an important source of genes, with which in the future breeding programs can create varieties with high adaptability to adverse impacts of climate change. On the other hand, local collection of these genotypes will contribute to the identification, protection
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 fact that, while most of the world is focused in the process of our own process of our own own own own of our own of our own of our own of our own of our own of our own of our own of our own
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
non-torpor Mars mission approaches and established the medical plausibility of torpor. We have expanded our team and put together a development plan that we are in the process of executing.⏎SpaceWorks makes clear that "enabling access to Mars is
-
regions, and local governments have started thinking about their own Green New Deals (GNDs) to address this crisis now. Municipalities and localities are in the unique position to enact truly transformative climate justice policies due to their ability to be much
-
a cobot is?⏎You might have come across this term “cobots” while searching for a solution to improve your production processes but weren’t quite sure what the term referred to. As you’ve probably noticed yourself, cobots are trending more
-
by the War Department, and continues to the present day, despite military public health efforts to reduce tobacco use. Military recruits, often young and away from home for the first time,
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
- L17:30 suppresses
- L16:792 excites "the" inside a prepositional phrase
- L9:24 excites "the" before a named organisation or title
- L19:427 suppresses
- L4:52 excites
- L10:120 suppresses
- L16:419 excites
- L18:641 suppresses
- L2:205 weakly excites
- L7:533 weakly suppresses
- L13:975 weakly excites
- L18:811 weakly excites
- L15:1848 weakly suppresses
- L11:1168 weakly excites
- L18:171 weakly excites
- L3:291 weakly excites
- L8:636 weakly suppresses
- L9:662 weakly suppresses
- L15:2320 weakly suppresses
- L16:1293 weakly suppresses
- L10:1377 weakly suppresses
- L19:2283 weakly suppresses
- L12:435 weakly suppresses
- L16:1949 weakly suppresses
Feeds later layers
- L21:1237 excites
- L26:1028 excites
- L27:556 excites
- L24:979 weakly excites
- L21:2976 weakly excites
- L24:823 weakly excites
- L23:1004 weakly excites
- L25:1926 weakly excites
- L25:2369 weakly excites
- L25:572 weakly excites
- L24:2713 weakly excites 'a' before few or variety
- L24:1552 weakly excites
- L25:1990 weakly excites
- L25:2165 weakly excites
- L23:742 weakly suppresses
- L25:549 weakly suppresses
- L26:2620 weakly excites
- L23:2525 weakly excites
- L23:1361 weakly excites
- L24:2538 weakly excites
- L21:1489 weakly excites
- L27:2548 weakly suppresses
- L23:1633 weakly excites
- L26:1432 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
- 一页
- one
- one
- 一步
- 一位
- 一个人
- 一种
- 一站
- 一个问题
- satu
Suppresses
- -final
- isateur
- LEMENT
- tract
- msm
- heck
- lement
- .bn
- ulant
- anye