'at' in 'not at all' and 'at all'
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. 21 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 word 'at' in the phrase 'not at all' or 'at all', as in 'there aren't any trees at all'. Turning it up made the model write 'all at all' 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.
-
; it could rise, fall, and rise again, as in a W. (Or, in the direst situation, the recovery may not be a recovery at all, but just an L-shaped continuation of economic misery.)⏎But not everyone subscri
-
the water and waste in the system to flow smoothly out to the sewer or septic tank. If there were no vent, the system would drain slowly or not at all, not unlike a water-filled straw when you seal one end with your finger
-
s" the others, it masks them more or less: in other words, the other sound are perceived as much weaker than they actually are, or even not at all. This phenomenon is of course taken into account by all the audio systems of
-
percent reported eating any meat, which historically was a diet staple, and many of them had only had one meal in 24 hours. Some had not eaten at all.⏎The Central African Republic has the highest level of hunger in the world,
-
influence how substances are metabolized and experienced. Such biological differences can influence whether substance use continues and increases, remains “social” and occasional, or does not occur at all. In the early stages, however, those more genetically vulnerable to addiction are
-
had only 18 percent marketable ears. This wasn't much better than the 6 percent marketable ears produced in the plots that received no sprays at all."⏎The authors predict that growers could realize increased profits with Bt sweet corn
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 a 3D map, and I have a 2D grid where I can move, but I also have another, more open, and, more, than once, with the same number of lines and columns. Now,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
tragic existence, blazing a trail of death and destruction that prompted mobs of angry villagers to take up torches and pitchforks against him on the silver screen. Never once during his rampage, however, did the monster question the honesty of his ultimate
-
won’t arrive at the Jezero Crater until February 18, 2021. But if the helicopter soars, it’ll provide never-before-seen views of the Red Planet.⏎Popayán Colombia https://www
-
6th Field Artillery bands played at the monument's dedicationon July 3, 1929. By 1949,<|endoftext|>Never underestimate the power of a paper towel roll.⏎Just the other day my five-year
-
, thus increasing the resistence. The microcontroller senses this and turns off a few of the LEDs as a result. Make one of these guys and you’ll never again have to
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
- L18:569 excites
- L20:1922 suppresses
- L16:752 excites
- L10:111 suppresses
- L20:2853 suppresses
- L10:39 excites
- L19:2184 suppresses
- L8:7 excites
- L18:2553 suppresses
- L18:1933 excites
- L5:158 weakly suppresses
- L7:49 weakly suppresses
- L17:1778 weakly suppresses
- L6:13 weakly excites
- L18:1763 weakly excites comma in 'not X, but Y' contrasts
- L6:2339 weakly suppresses "as" completing an "as ... as" comparison
- L17:1613 weakly excites
- L17:1644 weakly suppresses
- L7:662 weakly excites
- L18:178 weakly suppresses
- L5:28 weakly excites
- L7:110 weakly suppresses
- L10:386 weakly excites
- L15:386 weakly suppresses
Feeds later layers
- L27:599 strongly excites
- L26:835 strongly excites
- L24:1252 strongly excites
- L24:856 excites
- L25:501 suppresses
- L26:840 suppresses
- L26:1210 excites
- L26:2537 suppresses
- L23:1458 excites
- L27:769 weakly excites
- L23:1305 weakly excites
- L26:1800 weakly excites
- L25:1241 weakly excites
- L23:2703 weakly suppresses
- L22:1524 weakly excites
- L23:2966 weakly suppresses
- L27:1119 weakly excites
- L23:2095 weakly suppresses
- L27:1315 weakly excites
- L25:582 weakly suppresses
- L24:380 weakly suppresses the word 'at' in place phrases
- L23:2671 weakly excites
- L22:2641 weakly suppresses
- L26:684 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
- at
- At
- _at
- At
- at
- -at
- →at
- .at
- (at
- /at
Suppresses
- .za
- INCT
- ЛА
- hra
- PHA
- PSA
- стал
- cratch
- põ
- mdir