mid-word fragments before a completion
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 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
This neuron fires on short pieces of a longer word, such as "ad" in Benkhadra, "st" in diethylstilbestrol, or "val" in valentines, where the word is not yet finished. It goes negative on words like "brain" in puzzle-website text. Turning it up gave repetitive phrases about "the first time" and "the second half", which does not match the trigger.
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.
-
roles in establishing a green economy. From Secretary of State Hilary Clinton, who created a new office for women at the State Department, to Amina Benkhadra, Morocco’s Minister of Energy, Mines, Water and Environment, to Christian
-
sky changes through the seasons. We’ll hear some of the stories our ancestors told about the cosmos and how these relate to what we know today.⏎In association with STFC Rutherford Appleton Laboratory⏎If your child has additional needs, please get
-
ens experienced in the womb can lead⏎to enlarged and hypersensitive prostate in adulthood.⏎with the best known of the endocrine disrupters, particularly⏎diethylstilbestrol, or DES, led to great scientific confidence⏎about the predictive
-
network can behave chaotically during the learning phase due to large step sizes.⏎- Scott E. Fahlman: An Empirical Study of Learning Speed in Back-Propagation Networks, September 1988<|endoftext|>Cardio and weight
-
IL at George Street, a home where teen boys learn independent living skills; a community-based Diagnostic Evaluation service that helps people of all ages; SouthEast Treatment and Reentry Team (START), a program that works with teen girls in residential placement through
-
2.16 g of myristic acid, 93.96 g of palmitic acid, 9.288 g of stearic acid, 79.92 g of total monounsaturated
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: these things around public learning communities around the world classeylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylylyly
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
? ? spring ? ? ? hole⏎Answerwaist coat tail gate way side arm pit fall out put down cast off spring time line man hole⏎See another brain teaser just like this one...⏎Or, just get a random brain teaser⏎If
-
a matrix⏎Find nth maximum⏎Matlab Basics II - Create a vector with a repeated entry⏎Create sequnce
-
iated. Note that the spelling doesn't necessarily stay the same.⏎See another brain teaser just like this one...⏎Or, just get a random brain teaser⏎If
-
x * (x + x) * x.⏎Mile to Kilometer⏎Given A4 sizes find A3's long side⏎Polygon Interior Angle Sum⏎Bit to Gigabyte⏎Find the treasures in MATLAB Central and discover how the community can
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
- L20:429 excites
- L5:2 weakly suppresses
- L16:507 weakly excites
- L17:123 weakly excites
- L16:441 weakly excites
- L17:366 weakly excites
- L25:206 weakly excites
- L21:355 weakly suppresses
- L1:2442 weakly suppresses
- L22:156 weakly excites
- L11:2505 weakly excites
- L14:409 weakly suppresses
- L9:18 weakly suppresses text boundary after a document break
- L25:2407 weakly excites
- L10:905 weakly excites
- L5:2683 weakly suppresses
- L20:1779 weakly excites
- L19:229 weakly excites
- L19:314 weakly excites
- L14:892 weakly suppresses
- L2:1086 weakly excites
- L7:3 weakly suppresses
- L2:1700 weakly excites
- L2:9 weakly suppresses sentence-final periods vs dots in code
Feeds later layers
- L27:2775 weakly suppresses
- L27:2765 weakly excites
- L27:1021 weakly suppresses
- L27:294 weakly suppresses
- L27:2317 weakly excites
- L27:2951 weakly excites
- L27:2604 weakly suppresses
- L27:2661 weakly suppresses
- L27:893 weakly suppresses
- L27:3016 weakly suppresses
- L27:332 weakly suppresses
- L27:1968 weakly suppresses
- L27:2152 weakly excites
- L27:1721 weakly suppresses
- L27:417 weakly excites
- L27:2615 weakly excites
- L27:2882 weakly excites
- L27:3057 weakly suppresses
- L27:2680 weakly suppresses
- L27:1908 weakly excites the 'd' that starts 'die' or 'dead'
- L27:2730 weakly excites
- L27:2327 weakly suppresses
- L27:2201 weakly suppresses
- L27:934 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
- onenumber
- äd
- ตล
- าม
- 首发
- 군
- cycles
- 维奇
- COLUMN
- heets
Suppresses
- tool
- 工具
- Tool
- Tool
- ulaire
- mtx
- ey
- duel
- bout
- beiten