"per" in rate 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. 19 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 fires very clearly on the word "per" in rate phrases like "per passenger mile," "per kilometer" or "per person." It is one of the strongest and most reliable triggers among these five. Turning it up made the model drift into numbers and quantities, but the steering text is not a clean match.
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.
-
global warming pollution per person for the miles traveled (and may cost less, too).⏎On average, taking a trip by bus produces the least amount of greenhouse gas per passenger mile, followed by train travel, then air. Cars, light
-
This would combine use-phase with build-phase emissions.⏎Today, the European Union requires average carbon dioxide emissions for all new passenger cars to be 130 grams per kilometer driven. This figure is supposed to decrease by about 27
-
optimize water efficiency.⏎Greenhouse gas emissions:⏎U.S. cotton showed strong improvement in the 10-year average compared to the previous decade for GHG emissions per pound of lint produced.⏎The Global Cotton LCA (Cotton Incorporated,
-
s to deliver full LCA analyses of all models, to demonstrate their environmental footprint. This will probably include a sustainability cap imposed on the total amount of carbon emitted per unit from cradle to grave.⏎I also expect an increasingly demanding public to base
-
emissions and reduces landfill odors.⏎This year's LMOP winning projects will avoid the emissions of 546,000 tons of carbon dioxide equivalent per year, the equivalent of annual greenhouse gas emissions from nearly 100,
-
more, wheat up to $5 more. According to another computer model, organic farmers are expected to net an average of $50 to $60 more per acre than their conventional counterparts.<|endoftext|>It is sometimes felt that saving energy by natural
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: the way that a writer's response to a question is a response to the question itself. That is, the question is the writer's response to the question, and the response is a response to the question. The question is the writer's
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Talk to your doctor about the right medicines for you.⏎A common and safe remedy is to take supplemental vitamin B6 (pyridoxine), to a maximum of 100 mg/day. The American Congress of Obstetricians and
-
and E. coli or salmonella infection.⏎How to use⏎Use Ampicillin as directed by your doctor.⏎Take Ampicillin by mouth at least one-half hour before or 2 hours after eating.⏎Take Ampicillin with a full
-
times a month, while older crabs may take thirty to fifty days to grow large enough before needing to shed again. A newly shed crab will be about a third bigger than it was before.⏎Is my red claw crab dead or molting?⏎
-
2 inches apart.⏎- For summer turf establishment, a thin layer of mulch (no more than⏎1/4 inch thick) can be spread over the seeds with a peat spreader 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
- L18:1546 suppresses
- L16:1154 suppresses
- L8:2654 excites
- L15:1503 excites
- L17:1040 excites
- L10:3036 suppresses
- L10:2844 excites
- L9:2347 suppresses
- L10:141 excites
- L18:1726 suppresses
- L6:442 excites
- L15:1521 excites
- L15:1423 suppresses
- L9:43 excites
- L11:2874 excites
- L15:1926 suppresses
- L11:805 suppresses
- L12:1686 excites
- L13:1705 suppresses
- L12:408 suppresses
- L16:1279 suppresses
- L18:2183 suppresses
- L12:1655 suppresses
- L2:1018 excites
Feeds later layers
- L25:158 excites
- L20:1692 excites
- L21:2443 excites
- L21:2573 excites
- L22:991 excites
- L20:2616 excites
- L20:2483 excites
- L20:2624 excites
- L25:437 excites
- L20:1404 excites
- L22:941 excites
- L25:611 excites
- L23:2208 excites
- L22:1917 excites
- L24:806 excites
- L21:1348 excites
- L22:395 excites
- L21:1563 excites
- L23:1404 excites
- L21:1957 excites
- L20:232 suppresses
- L21:2852 weakly excites spot where a number or 'time' comes next
- L24:1201 weakly excites
- L21:1628 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
- amount
- volume
- capita
- unit
- hours
- 体积
- 金额
- 小时
- amount
- units
Suppresses
- dings
- masters
- 四大
- 축
- _CONTINUE
- master
- rice
- ibName
- rir
- esub