the last zero in large numbers
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. 7 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 reacts to a zero that finishes a big number, such as the last 0 in 4,000 or 8000. It is a number-shape detector rather than a topic detector. Turning it up gave text with no clear link to numbers.
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.
-
number must be lower than the safe working load of the wheel you intend to use.⏎For example, a 4,000-lb cart with 4 wheels required to carry a 6,00
-
tars and waxes that stifle the process.⏎However, it turns out that these tars and waxes are no match for an 8000K (7,727°C) plasma arc which, through a combination
-
, Laundry, hostel and canteen effluent waste. The above waste is treated through Pond setup in the institute with a capacity of 24000 KLD (Kilo Liters per Day). The entire treated water is used
-
5000 copies and never went out of print.⏎This PDF is a full-text searchable reproduction
-
are 610 mg of Sodium in this food.⏎Vitamins - This food has both Vitamin A and Vitamin C. One serving contains 500 International Units of Vitamin A and 1.2 mg of Vitamin C.<|endoftext|>
-
sugar and fatty snacks.⏎People’s energy needs vary depending on age and activity levels but usually men require more energy than women. Men need about 2500 kcals and women about 2000 kcals per day.⏎Break
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: what is the use of the word "fear" in the sentence? What do you think? I see that it is important, and I know that the word "fear" is an example of the use of the word "f
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the growth of two human embryonic stem cell lines that lack PIG-A, and therefore do not contain any proteins known as glycosylphosphatidylinositol (GPI) anchor proteins on the cell’s surface. GPI anchor
-
We found that, across the western Palaearctic, urban blackbird populations contain less genetic diversity than rural ones, urban populations
-
⏎Semiarid and desert climates occur in regions with little precipitation. Desert climates are drier than semiarid climates. Semiarid climates, also called steppe climates, usually border
-
whose electrochemistry makes carbon capture appear nearly effortless. Their invention is a kind of battery: conductive electrodes coated with a compound called polyanthraquinone, which has a natural chemical attraction<|endoftext|>From: Michael Bar
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
- L3:1169 suppresses
- L3:494 suppresses
- L5:807 excites
- L5:138 suppresses
- L4:2830 excites
- L4:1158 weakly suppresses
- L6:7 weakly suppresses
- L5:158 weakly suppresses
- L4:2996 weakly suppresses
- L2:1607 weakly suppresses
- L6:513 weakly excites
- L6:13 weakly excites
- L1:2683 weakly excites
- L4:1088 weakly suppresses
- L4:1351 weakly suppresses
- L5:110 weakly suppresses
- L5:1336 weakly suppresses
- L5:2257 weakly suppresses
- L1:189 weakly suppresses
- L1:1365 weakly suppresses
- L5:1047 weakly excites
- L6:611 weakly excites
- L6:58 weakly excites
- L6:2300 weakly suppresses
Feeds later layers
- L17:106 strongly suppresses
- L16:75 strongly suppresses
- L19:2417 suppresses
- L20:121 suppresses
- L17:278 suppresses 'as' in 'as well as' and 'as much as'
- L20:174 suppresses
- L16:628 suppresses
- L18:102 suppresses
- L19:569 suppresses
- L21:381 suppresses zeros inside long number codes
- L18:205 suppresses
- L21:1072 weakly suppresses
- L20:413 weakly suppresses
- L26:875 weakly excites
- L17:594 weakly suppresses
- L19:311 weakly excites
- L19:906 weakly suppresses
- L18:139 weakly suppresses
- L24:1723 weakly suppresses
- L23:1756 weakly suppresses
- L17:886 weakly excites
- L26:1902 weakly suppresses
- L11:2856 weakly excites
- L27:1991 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
- ------------
- ever
- 醮
- ева
- edException
- ulaire
- turb
- ------------⏎
- ings
- kat
Suppresses
- gore
- SetBranch
- Search
- Jag
- 庆
- 锯
- Jong
- _SEARCH
- แนะ
- -guid