unclear: fragments inside split names
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. 3 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 middle pieces of long or unusual words that are chopped into several parts, such as the 'ust' in 'Procrustean' or the 'o' in 'Robo'. It is strongly turned off by the word 'is' after 'do', and the steering outputs did not show a clear theme.
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.
-
actually look at the broad measures of aggregate economic activity that define recessions and recoveries, they don’t look like those letters—at best they would be Procrustean fits. For example, the Great Depression of 1929-
-
not compelled, as were those whose homes lay north of the Alps, to distort facts before they could make the plants of their native country fit into the procrustean bed of classical descriptions.⏎One of the chief of the commentators and herbalists
-
A NASA original, this galaxy is appropriately named for it’s hat-like appearance yet makes us wonder what exactly is going on in it’s center.⏎R Jay Gabany took this shot of the Trific Nebula, which “… illustrates three different types
-
balanced breakfast, setting up your metabolism properly for the day. If you have time constraints, consider options such as protein smoothies, healthy muffins, hardboiled eggs or granola with fruit, nuts and seeds.⏎Eliminate or Min<|endoftext|>
-
18 species of bat residing in the UK, including the common pipistrelle and greater horseshoe bat.⏎Monitoring bats is vital not just to
-
treatment of tuberculosis.⏎Because of the therapeutic climate in Asheville, N.C., he moved there in 1888 to establish the Winyaw Sanitarium for the treatment of diseases of the lungs and throat. Mode
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: I have been a part of the public or private. I have been a part of your life. I have been a part of your life, or my life. I have been a part of your life. I am a part of your
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, but as of this writing modules are not given⏎a chance to take action when a new thread is created or destroyed.)⏎However, what extension modules primarily do is to intercede in the HTTP protocol in order to customize how Apache processes and responds
-
when he was a kid. All you have to do is find a piece of paper and use your imagination to draw any prehistoric scene you
-
learn how disposed brine impacts surface water quality in Pennsylvania⏎The first thing we will do is get a file from HydroShare and upload it into HydroClient. This will allow
-
⏎Will Japan's nuclear crisis end in catastrophe?⏎When the unexpected happens, even at a nuclear power plant, don't freak out. All you have to do is cool the cores. Just add water. If you can't deliver it through a
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
- L0:532 weakly suppresses
- L2:1576 weakly excites
- L0:1456 weakly suppresses
- L2:115 weakly excites
- L1:534 weakly excites
- L0:313 weakly suppresses "and" joining paired nouns
- L0:1423 weakly excites
- L2:419 weakly suppresses
- L1:1519 weakly suppresses
- L0:2228 weakly suppresses
- L1:2762 weakly excites
- L0:1816 weakly excites
- L1:1427 weakly excites 'that' after reporting verbs
- L1:1681 weakly excites
- L1:2616 weakly suppresses
- L2:2003 weakly excites 'of' after abstract nouns like 'sense' or 'view'
- L2:625 weakly excites
- L0:1365 weakly suppresses
- L1:1739 weakly excites
- L2:1327 weakly excites
- L2:2894 weakly excites
- L1:2920 weakly suppresses
- L2:110 weakly suppresses
- L2:2558 weakly suppresses
Feeds later layers
- L4:809 suppresses
- L14:1616 weakly excites
- L4:3003 weakly excites
- L10:1683 weakly excites
- L4:1012 weakly suppresses
- L16:2287 weakly suppresses
- L10:2919 weakly suppresses
- L24:2173 weakly excites
- L13:404 weakly suppresses
- L17:1965 weakly excites
- L20:872 weakly suppresses
- L4:3039 weakly suppresses
- L13:905 weakly suppresses
- L15:608 weakly excites
- L18:2722 weakly suppresses
- L12:410 weakly excites
- L6:613 weakly suppresses
- L22:411 weakly excites
- L16:2879 weakly suppresses
- L6:1689 weakly suppresses
- L14:2062 weakly excites
- L15:1084 weakly excites
- L11:1007 weakly suppresses
- L10:1991 weakly suppresses the word 'voluntary'
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
- 扑
- 碉
- OUS
- ounding
- ValidationResult
- 党总支
- 党委书记
- GO
- uckle
- 执业
Suppresses
- yourselves
- __ ⏎
- [#
- Lamar
- yourself
- openness
- oneself
- himself
- �택
- favor