'per' in rate expressions
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. 10 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
The neuron fires on the word 'per' (or 'each', 'a') when it introduces a rate, like 'people per square mile', '3 inches per year' or '12,000 per family'. It stays mostly quiet elsewhere. Turning it up made the model write unrelated word problems and math-style text, which does not match the trigger.
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.
-
not want to move to Curry Island with his/her...⏎This section contains 570 words|⏎(approx. 2 pages at 300 words per page)<|endoftext|>Of all the technology featured at CES this week, the most exciting
-
there were 112 people, 46 households, and 32 families residing in the town. The population density was 1.2 people per square mile (0.4/km²). There were 106
-
7 and located in the area of Asia, with a land area of 796096 km² and population density of 278 people per km². Territory
-
875,000 barrels total) with a reasonable profit at the present price of crude oil, and 2,500 barrels per acre for an area of 2,248 acres (5,6
-
to one-third in 1 year.⏎Added to these principal items is a reduction in volume per plant. Thus, total manufacturing costs have been increased about a cent a pound, compared with the same months a year ago.⏎Using typical cost figures supplied
-
the socioeconomic benefits accrued to translocated farmers," they write. "At an average start-up resettlement cost of at least US$12,000 per family, this is an expensive development program."⏎From a macroeconomic standpoint, this
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: how to prevent a fire at all? I have never been helped by any rescue breathing apparatus. A. It is possible to escape. It is possible to escape. It is possible to escape. All Animals are safe from being harmed by a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
11, electricians earned an average of $52,910 per year, according to the Bureau of Labor Statistics. The average hourly pay for an electrician was estimated to be $25.44. The median annual
-
Bureau of Labor Statistics, earnings for mediators range from $28,090 to $102,202 per year with median annual wages estimated at $49,490. Hourly
-
ians. The lowest average salary for electricians, $39,640 per year, was reported in North Carolina.⏎The Bureau of Labor Statistics estimates that job growth
-
@example.org for more information. Updated June 25, 2015. (taken from http://genealogy.about.com/od/surname_meaning/p/boyle.htm ) BOYLE Last Name Meaning & Origin: A variant
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
- L9:1106 suppresses
- L4:141 excites
- L8:1850 suppresses
- L8:1249 excites
- L9:2105 suppresses
- L5:2215 excites
- L7:1350 suppresses
- L8:276 excites
- L7:2787 excites
- L5:1145 suppresses
- L8:1858 suppresses
- L9:515 suppresses
- L7:590 suppresses
- L5:2095 excites
- L6:337 excites
- L7:322 weakly excites
- L8:1537 weakly suppresses
- L8:1143 weakly excites
- L6:263 weakly excites
- L8:2997 weakly excites
- L0:2646 weakly excites
- L9:1177 weakly suppresses
- L4:2468 weakly excites
- L7:2287 weakly suppresses cutting and dividing in instructions
Feeds later layers
- L14:2097 excites
- L15:2029 excites
- L12:2621 excites
- L11:1407 excites
- L12:3035 excites
- L12:1931 excites
- L12:1980 excites
- L17:1454 excites
- L14:2704 excites
- L13:2389 excites
- L11:1066 weakly suppresses
- L12:1800 weakly excites
- L12:1471 weakly excites
- L17:2617 weakly excites
- L14:2830 weakly excites
- L12:842 weakly excites
- L19:2500 weakly excites
- L19:2592 weakly suppresses
- L13:2211 weakly excites
- L11:1654 weakly suppresses
- L20:2076 weakly excites
- L12:2169 weakly excites
- L11:267 weakly excites
- L11:655 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
- 每次
- latable
- 速度
- 每一次
- lse
- FINE
- 每
- الزمن
- każd
- 次
Suppresses
- NewLabel
- amous
- üz
- 无知
- circum
- reek
- Boeh
- tec
- Sachs
- aine