comma after 'in this blog post'
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
The neuron fires on the comma that ends an opening phrase like 'In this blog post,' or 'Below,'. It is pushed the other way by the words 'a' and 'an' before a noun. Turning it up made the model write about unrelated math and turning it down gave fragments about movies and reports.
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.
-
. These innovations hold the promise of transforming industries, revolutionizing the way we live and work, and addressing some of society’s most pressing challenges. In this blog, we’ll take a deep dive
-
heritage resources (including scientifically significant paleontological resources).⏎We are currently undergoing a renaissance in the discovery, research, and interpretation of fossils from Utah.⏎Below, click on the various districts (West Desert, Green River, Color Country, Canyon
-
your hair? The intricate anatomy of our hair strands comprises several distinct components, each playing a unique role in the appearance and health of our hair. In this article, we will explore the essential elements of hair anatomy: the cuticle, cortex,
-
Quality⏎Common Resources — Feb 7, 2018⏎The Facts on Fracking: A Podcast with Daniel Raimi⏎In this podcast, RFF’s Daniel Raimi gives a balanced and accessible view of oil and
-
range of scholarship about Caddo history and culture from anthropologists, historians, and other researchers.⏎In this extensive 2021 bibliography, you will find 386 pages of resources to learn more about Cadd
-
to its former glory.Built in 1902, the hotel boasts a stunning atrium, luxurious amenities, and a fascinating history.In this blog post, we’ll take a closer look at the West Baden Hotel, its history,
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: there are a *there* *a* *a* *a* *a* *a* *a *a *a *a *a *a *a *a *a *a *a *a *a *
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Junction. (A historical marker was placed at that spot in 1955.) News of Vilas’s work quickly spread; American Forestry magazine had an article about it
-
This is called parallax, and I’ve explained this in an earlier post for a solar eclipse that had very similar circumstances as this Swift observation. I’ve written a general parallax description as well.⏎So the wiggly path of the asteroid
-
aping Chesapeake Bay Watershed - U.S. Fish & Wildlife Service.⏎Here are some resources to help gardeners select native plants. The PA DCNR site has a listing of nurseries.⏎There are 133 species and subspecies
-
see some military time starting all the way back in 1873.⏎45-70 VS 444 Marlin⏎We have dedicated a post to the comparison of 45-70 vs 444
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
- L17:805 suppresses
- L17:2970 weakly excites
- L7:935 weakly excites
- L17:1139 weakly excites
- L18:2293 weakly suppresses
- L10:609 weakly excites
- L16:2714 weakly suppresses
- L15:2387 weakly excites
- L6:2099 weakly excites
- L16:3002 weakly excites
- L18:1999 weakly suppresses
- L10:2818 weakly suppresses
- L0:1253 weakly excites surname pieces split across tokens
- L11:2626 weakly suppresses
- L14:1847 weakly excites
- L15:1889 weakly excites
- L18:2631 weakly suppresses
- L16:2824 weakly excites 'in' after 'published' or 'appears'
- L17:2607 weakly excites
- L11:976 weakly excites
- L17:679 weakly suppresses
- L18:262 weakly excites
- L13:1970 weakly suppresses
- L15:2484 weakly excites
Feeds later layers
- L22:778 suppresses
- L24:1263 suppresses
- L25:45 excites
- L25:1586 suppresses
- L27:797 weakly suppresses
- L22:103 weakly excites
- L20:1707 weakly suppresses
- L20:2682 weakly suppresses
- L25:407 weakly excites
- L24:823 weakly suppresses
- L26:2817 weakly suppresses
- L24:970 weakly suppresses
- L24:413 weakly suppresses
- L22:811 weakly suppresses
- L20:2696 weakly suppresses
- L20:1727 weakly suppresses
- L25:544 weakly suppresses
- L24:2538 weakly suppresses
- L27:716 weakly excites
- L25:2475 weakly excites
- L20:2000 weakly suppresses "a" or "the" opening an object noun phrase
- L21:1505 weakly excites
- L21:1100 weakly suppresses
- L27:756 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
- -fontawesome
- добро
- ENDING
- iales
- 想不到
- lưới
- fec
- ending
- winnings
- etermined
Suppresses
- 有一个
- 有个
- 有
- an
- 有一
- a
- 也有
- 专门
- had
- 都有