unclear: digits inside amounts and figures
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. 11 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 a digit or word in the middle of a number, price or quantity, such as the zeros in a large sum of money. It turns negative on numbers written as percentages or fractions. Turning it up made the model drift into arithmetic word problems, though the link is loose.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
maturity.⏎Let's take Bond A, issued by Company Alpha, as an example. It has the following data:⏎- Face value: $1,000⏎- Annual coupon rate: 5%⏎- Coupon Frequency: Annual⏎-
-
outcome of $400,000 compared to an average teacher.⏎Effect of Good Teachers on the Economy⏎What
-
decrease the body's ability to use carbs for key functions. The USDA's Dietary Reference Intakes recommend getting from 45 to 65 percent of daily calories - from 900 to 1,300 calories in a
-
, over a second pulley wheel in the fixed block, and back to the free block, the mechanical advantage is 3. A 300-lb load can be raised by a pull of 100 lb on the free
-
idea been as popular with collectors and taxonomists of these groups in the past, the current number could well be in excess of 100,000 species. Recent estimates suggest that there may be more than 500
-
, and not stopped climbing.⏎According to U.S. Centers for Disease Control and Prevention data, opioids accounted for 75% of the 100,000 overdose deaths in the last 12 months,
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: - something about a place, a person, a country, a region, of a specific name of a country or existing in a local language of a specific city of a country or existing in a region of a specific
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
that represents a fraction of the number 100.⏎In other words, it is a part of a general number that should be considered divided by 100 units. The symbol with which this fraction is represented is ” % “.⏎
-
tolerance⏎To use the formula, first figure out what you want your error of tolerance to be. For example, you may be happy with a confidence level of 95 percent (giving a margin error of 0.05),
-
% is the same as 25/100 and equivalent to 0.25 . One of the most common uses of
-
converted to a percentage such as 0.10 percent, which indicates that one-tenth of one percent of a person’s blood is alcohol. The amount of alcohol it takes to reach
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
- L1:925 weakly suppresses
- L6:1287 weakly excites
- L8:2738 weakly suppresses
- L7:1428 weakly suppresses
- L7:1945 weakly excites
- L6:1573 weakly excites
- L10:1587 weakly excites
- L10:2318 weakly suppresses
- L1:2654 weakly excites
- L7:2595 weakly suppresses
- L3:2310 weakly excites
- L1:1591 weakly excites
- L8:5 weakly suppresses
- L0:951 weakly excites
- L8:110 weakly excites
- L6:2316 weakly suppresses
- L4:2659 weakly excites
- L10:1485 weakly suppresses
- L2:1998 weakly suppresses
- L5:1895 weakly suppresses
- L5:1694 weakly suppresses
- L7:1888 weakly suppresses
- L6:1547 weakly suppresses
- L1:2996 weakly suppresses
Feeds later layers
- L12:2728 suppresses
- L12:2823 excites
- L12:2236 excites
- L12:322 excites
- L13:1130 suppresses
- L24:316 suppresses
- L15:1724 weakly suppresses
- L12:2195 weakly suppresses
- L13:1629 weakly suppresses
- L14:1863 weakly excites
- L12:1800 weakly excites
- L12:2116 weakly suppresses
- L12:625 weakly excites
- L22:2133 weakly suppresses
- L13:3000 weakly excites
- L12:2002 weakly excites
- L27:1325 weakly suppresses
- L15:2534 weakly excites
- L23:1224 weakly suppresses
- L23:1101 weakly suppresses
- L15:708 weakly suppresses
- L18:1977 weakly suppresses
- L13:2810 weakly excites
- L12:2664 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
- .base
- base
- base
- (base
- -base
- _base
- atre
- →base
- etre
- uminum
Suppresses
- че
- ://{
- FormGroup
- acked
- record
- ⏎
- FromArray
- ngrx
- FlatButton
- illez