"from" after a change in amount
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. 6 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 on the word "from" when it names the starting value of a change, as in "doubled from 3.5 to" or "increased by 62%, from 47.3 years." Turning it up did not produce a clear theme in the replies.
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.
-
1980 and 2010, the average years of schooling for females aged 15–19 in MENA more than doubled from 3.5 to
-
damaging groves. But this year the industry is "making a comeback", according to Ocean Spray. The group forecasts the Florida grapefruit harvest will almost double from 19 million boxes a year ago to 28 million this year. The
-
rise over the past three and a half decades. The average age of first-time mothers has increased by 3.6 years over last 36 years, from the age of 21.4 years in 1970 to
-
grow from 6Gbps to 10Gbps. The net result of these simultaneous forces is a 25% increase in monthly transit fees, from $16,000 per month to $20,00
-
alcohol and marijuana use changed little from 2009 to 2012. However, lifetime RR for use of other illicit drugs increased substantially, from an RR of 1.8 (95% CI, 1.
-
artners are using the iPad to learn and study.⏎And the AP found that the percentage of preschoolers with a social media account increased by almost a third, from about one in ten to about one-half of one percent.⏎Here’s what you
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 number of 3-digit numbers with a second digit of 3 and first digits of 1 and 2 and 3, and 4, 5, 6, 7, 8, 9, 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to health. It wasn’t long, however, before new drugs were developed that rendered this treatment obsolete.⏎As such, the building’s role would shift several times over the next six decades before it was added to the National Register of Historic Places in
-
that the fractional red-shift d / is constant throughout a given spectrum. Each wave-length is increased, and, consequently, the energy in each quantum is reduced, by the constant factor 1 + d / . We might naturally conclude that the total
-
student blogs were nothing more than a class journal posted online. They were digital notebooks. However, over time, they became a place for student voice. They evolved as my thinking about blogging evolved. So, here is what I’ve learned along the
-
compromise responses to climate changes – both ‘normal’ variation and shocks as well as new variability and increased uncertainty.⏎Third, local support systems are changing, and so is people’s ability to make use of them. The report shows that changes in locally
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
- L3:754 suppresses
- L2:1832 weakly suppresses
- L3:1942 weakly suppresses
- L4:2080 weakly excites
- L4:2337 weakly excites
- L5:46 weakly excites
- L5:27 weakly excites
- L4:1356 weakly excites
- L5:57 weakly suppresses
- L4:1308 weakly suppresses
- L5:2530 weakly excites
- L1:1971 weakly suppresses
- L4:1992 weakly excites
- L4:1262 weakly suppresses
- L1:2801 weakly excites
- L1:2650 weakly suppresses
- L4:1238 weakly suppresses
- L2:78 weakly suppresses
- L4:2877 weakly suppresses
- L3:713 weakly excites
- L5:2216 weakly excites
- L5:2398 weakly excites
- L0:2760 weakly suppresses
- L3:342 weakly excites
Feeds later layers
- L11:2662 excites
- L12:1467 excites numbers after dates and figures
- L7:2592 excites
- L9:743 excites
- L8:2924 excites
- L17:2732 excites
- L18:1660 excites
- L8:229 excites
- L8:2018 excites
- L14:2475 excites
- L11:2970 weakly excites
- L13:2224 weakly excites
- L14:867 weakly excites
- L7:582 weakly excites
- L9:1257 weakly excites
- L12:1294 weakly excites
- L13:2502 weakly excites
- L7:2210 weakly excites
- L20:2968 weakly suppresses 'more' or 'not' before a likelihood claim
- L13:2958 weakly excites
- L11:1958 weakly suppresses
- L23:1518 weakly suppresses
- L7:213 weakly excites
- L8:1401 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
- 换成
- @update
- 改动
- 改为
- 改成
- 变为
- 替换
- 修订
- đổi
- 更改
Suppresses
- ết
- щен
- 넷
- _rng
- رب
- fähig
- nox
- uncan
- 甥
- unes