'co' starting a word piece
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. 26 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 syllable 'co' when it starts a word that has been split into pieces, as in 'coexist', 'coining' and 'coauthored'. The steering outputs did not clearly reflect this, so the pattern rests on the text snippets alone.
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.
-
Arabia’s Information Minister put his profile on Facebook and got more than 5,000 “friends.” And when the king fired a senior cleric for criticizing co-education at a new Saudi university, the cleric’s supporters struck back in cy
-
be members traveled near and far to get some of the mysterious liquid. In 1627 the Seneca people, brought French missionary, Joseph De La Roche d’Aillon, to the sacred spot, and he blessed the spring
-
ball⏎Here to Help Average Joes & Janes Play Better Basketball⏎A Basketball term meaning to embarrass some one usually while slamming the ball over them.⏎James Naismith wrote the original 13 rules of basketball.<|endoftext|>a sheet with
-
the father of chaos theory and discovered the Lorenz attractor that often occurs in chaotic systems.⏎Lorenz is perhaps best known for coining the term butterfly effect. Whilst a professor at the Massachusetts Institute of Technology,
-
Tunis where several African communities coexist, made news headlines.⏎Ksiksi remembers when Tunisia became the first
-
source philosophy in order to produce the premier software of their markets and many nonprofit organizations have coalesced in support of open source software. Even governments have heralded their support
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: a lot of a lot of a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot a lot
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Concrete poured and formed on the construction site by concrete contractors is obviously at the mercy of weather conditions at the time of curing. Precast concrete panel construction, therefore, results in a stronger, less porous, higher
-
It’s important to note core stability is a key component of pelvic and core health. You can find all of the core stability and pelvic floor exercises in the Prenatal and Postnatal Exercise Guide.⏎Perform these seated, sitting on a Swiss ball
-
similar to gray mold. The color of the decayed area is light to dark brown, and black in advanced stages. A second symptom of infection starts at lenticels. Symptoms are brown to black specks with a white to light tan center
-
Institution: Lutheran Children and Family Service, Philadelphia, PA.<|endoftext|>Updated: May 19, 2021⏎A great video to help explain Concussion Management by our friend, Dr. Mike Evans before we get into this:⏎
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
- L1:460 suppresses
- L17:123 weakly excites
- L16:91 weakly suppresses
- L1:2342 weakly suppresses
- L20:59 weakly suppresses
- L1:110 weakly suppresses
- L0:7 weakly excites
- L5:1458 weakly suppresses
- L21:296 weakly suppresses
- L16:34 weakly suppresses
- L22:14 weakly excites
- L1:2159 weakly excites
- L0:2780 weakly excites
- L4:343 weakly suppresses
- L16:1659 weakly suppresses
- L1:81 weakly suppresses
- L6:19 weakly excites second half of proper names
- L4:122 weakly suppresses
- L20:452 weakly excites
- L0:2107 weakly excites
- L6:942 weakly suppresses
- L20:38 weakly suppresses
- L17:218 weakly suppresses
- L7:2696 weakly excites
Feeds later layers
- L27:131 weakly suppresses
- L27:802 weakly excites
- L27:2983 weakly excites
- L27:1132 weakly suppresses
- L27:2242 weakly excites
- L27:159 weakly excites
- L27:682 weakly excites
- L27:1962 weakly excites
- L27:1067 weakly suppresses
- L27:601 weakly suppresses
- L27:2143 weakly suppresses
- L27:1228 weakly suppresses
- L27:890 weakly suppresses
- L27:276 weakly excites
- L27:3061 weakly excites
- L27:2164 weakly suppresses
- L27:549 weakly suppresses
- L27:1452 weakly excites
- L27:1508 weakly excites
- L27:536 weakly excites
- L27:2714 weakly suppresses
- L27:2199 weakly excites
- L27:959 weakly suppresses
- L27:2861 weakly excites the letter f starting fume or fountain
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
- +-
- lets
- 年底前
- stream
- ৎ
- MARK
- ände
- --
- wide
- <>
Suppresses
- .didReceiveMemoryWarning
- üzere
- antan
- ounder
- ilio
- earer
- igham
- rawler
- (PC
- onus