'a' after 'in' or 'such'
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. 2 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 'a' in set phrases like 'in a world', 'in a way', or 'in such a way'. It is strongest when the 'a' starts a new phrase after a preposition. Turning it up made the text wander without a clear theme.
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.
-
’s tricky (tricky) tricky (tricky)”⏎–Run-D.M.C., “It’s Tricky”⏎Long-term planning is hard, especially in a democracy. America’s elected regimes have short residence times, as Presidents come and go
-
ildfermentation.com/). Teaching others how to do the same is one great way of keeping traditional (and ‘living’) foods, well, alive! In a world inundated by highly processed or Genetically Modified foods that use chemicals as pres
-
ee and a, two ee and a, three ee and a, four ee and a’. I actually never learnt this way and I’ll get into that in a second.⏎By doing the above you have vocalised the 16 sixteenth
-
make life so difficult for them. Its unconventional aim is to teach students that it is O.K. to ''act autistic'' and also how to get by in a world where it is not.⏎Trained in self-advocacy, students
-
first thing that we should keep in mind while applying growth methods is that it should be properly and methodically pruned or trimmed. It should be trimmed in such a way that it encourages further growth of the plant and does not make<|endoftext|>Amaz
-
reality. A tathagata is one that has achieved awakening as to the nature of the Universe, as did Siddhartha Gotama. Being in such a state of condition is coming to know one’s own nature as is expressed in our
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 a job to a project, and I have a budget. So I am trying to create a budget for this, I have a job to a project, and I have a budget. I am trying to create a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, selecting a location, glazing, environmental controls<|endoftext|>Advent Reflections: Week 2, Day 5⏎“Moses said to his father-in-law, “Because the people come to me to inquire of God. When they have
-
nevertheless had no tolerance for bad acting. Edwin bought an interest in the Winter Garden Theatre at 667 Broadway in New York City together with his brother-in-law John Sleeper Clarke. The brothers John Wilkes, Edwin, and Junius
-
faith. Deposed in the "Glorious Revolution" of 1688-89, he fled to France. His daughter and son-in-law succeeded him as Queen Mary II and King William III. James II died in
-
22. Most wanted a group of Islamic leaders to decide while a minority felt that the Prophet's brother-in-law, Ali, was the rightful successor as it should be someone from the prophet's
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
- L0:1301 weakly suppresses
- L0:178 weakly suppresses
- L0:1224 weakly suppresses
- L1:1735 weakly excites
- L1:526 weakly suppresses
- L1:2173 weakly excites
- L1:2278 weakly excites
- L1:2801 weakly excites
- L1:1973 weakly suppresses
- L0:27 weakly excites
- L0:1563 weakly suppresses
- L0:1671 weakly suppresses
- L0:895 weakly excites
- L1:1785 weakly suppresses
- L1:2363 weakly excites
- L0:1176 weakly suppresses
- L1:2340 weakly suppresses
- L0:1644 weakly excites
- L1:28 weakly excites
- L0:2561 weakly excites
- L1:319 weakly suppresses "on" after effect or impact nouns
- L0:60 weakly excites
- L0:508 weakly suppresses
- L0:1722 weakly suppresses
Feeds later layers
- L11:881 excites "with" after combine, mix or react verbs
- L4:3008 weakly suppresses
- L16:1550 weakly excites
- L6:2562 weakly suppresses
- L16:1492 weakly suppresses
- L3:1945 weakly excites 'to' in 'person to person' phrases
- L11:1993 weakly excites
- L10:1726 weakly excites
- L5:2276 weakly excites
- L21:2551 weakly excites
- L14:1691 weakly suppresses
- L8:2017 weakly excites
- L22:2096 weakly excites
- L25:2798 weakly suppresses
- L18:1509 weakly suppresses
- L16:1978 weakly excites
- L5:1105 weakly suppresses
- L18:2425 weakly excites
- L18:2954 weakly suppresses
- L27:2514 weakly suppresses
- L22:839 weakly excites
- L18:96 weakly excites
- L4:2358 weakly suppresses
- L12:1050 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
- scious
- 工业大学
- correlated
- relat
- tagged
- Against
- تعاون
- har
- over
- änner
Suppresses
- ikip
- |M
- okers
- |m
- тяж
- PageRoute
- oker
- |"
- ulfilled
- ��