the word 'at' in place 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. 24 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 lights up on the small word 'at' when it starts a phrase about a place, time or level, such as 'at the end' or 'at all levels'. When it was turned down, the model got stuck repeating 'at least' over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
. If you want the list double-spaced, highlight all the text and set the paragraph to be double-spaced. Do not simply hit the Enter button twice at the end of each line, it will become an editing nightmare and when the sorting
-
to spread renewable energy technologies to enhance energy supply in urban areas. We particularly support conceptual and analytical awareness and understanding on gender, poverty and urban energy so that authorities at all levels, civil society and private sector partners are aware and understand the linkages
-
options so you don’t get bored with your workouts. When designing your fitness plan, consider adding some stretching and flexibility training. Simply add a few minutes of stretching at the end of your cardio and weight training workouts.⏎Cardio, or aerobic,
-
.”)⏎He likes to eat popcorn when he goes to see a movie.⏎To play in the snow is one of life’s greatest pleasures.⏎(The infinitive appears at the beginning of the sentences. This sounds very formal.)⏎He went downstairs to get
-
code, x would never be zero, that wasn't the issue. The fact is that it is pointless even having the x if you set total to 0 at the start because total will never increase, and that's even if you ignore that
-
exposes them to risks. We live in a time of rapid technological and social change, and a better, safer internet for children is high on the agenda for policymakers at both the national and European level.⏎In 2013, the European
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: at least alt alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts alts
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
functions are generaelly a subset of hash functions that fulfil at least two extra criteria:⏎In order to fulfil these criteria (or at least, as a by-product of needing to fulfil these criteria), secure hash functions
-
or disorder is the genotype with two recessive alleles. So I guess in order for the normal genotype, or what is considered normal, a person has to have at least
-
across disparate operating systems.⏎Kubernetes clusters are actually a set of nodes categorized into master nodes and worker nodes. Every cluster needs to have at least one of each. While the master node is meant to control the state of
-
of dog food, but all of them must meet certain requirements to be considered “real” dog food. For example, the food must have a meat content of at least 70% and it must not contain any artificial colors or flavors.
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
- L20:2348 excites "at least" and minimum-quantity phrases
- L19:1171 excites
- L22:160 suppresses
- L21:1339 excites
- L22:1904 excites ages right before a number
- L23:570 excites
- L19:2912 excites
- L20:1281 suppresses
- L23:253 excites
- L20:833 excites
- L23:2966 suppresses
- L23:476 excites
- L21:140 excites
- L17:2409 suppresses
- L22:825 suppresses 'for' before a span of time
- L21:2498 excites
- L22:2713 excites
- L21:2859 excites
- L20:2184 excites
- L21:1647 weakly excites
- L8:2858 weakly excites
- L22:2720 weakly excites
- L16:513 weakly suppresses
- L22:2255 weakly suppresses
Feeds later layers
- L27:599 strongly suppresses
- L26:835 strongly suppresses
- L25:501 strongly excites
- L26:840 excites
- L26:1210 suppresses
- L26:2537 excites
- L26:1800 suppresses
- L27:1119 weakly suppresses
- L27:675 weakly suppresses
- L26:2330 weakly suppresses
- L26:1876 weakly suppresses
- L27:1315 weakly suppresses
- L25:582 weakly excites
- L27:144 weakly suppresses
- L26:207 weakly suppresses
- L26:992 weakly excites
- L25:3014 weakly excites
- L25:1145 weakly suppresses
- L25:2592 weakly excites
- L27:2999 weakly suppresses
- L27:1918 weakly suppresses
- L26:789 weakly excites
- L25:509 weakly suppresses
- L25:562 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
- juana
- onian
- udad
- headed
- onden
- DialogContent
- onomies
- INCT
- CUS
- ąż
Suppresses
- at
- At
- At
- _at
- →at
- at
- -at
- .at
- (at
- [at