'the' after 'lack' or 'have'
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 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 'the' when it comes right after words like 'lack', 'have' or 'possess', where a noun such as knowledge or resources is expected next. Turning it up made the model write more ordinary, general text with no 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.
-
knowledge necessary to make informed decisions and take responsibility for their own financial well-being. Faced with an array of complex financial services and sophisticated products, many graduates lack the knowledge and skills to make rational, informed decisions on the use of their money and
-
the most-common challenges in security planning⏎- Some state policies require, without providing adequate implementation funding, that LEAs adopt specific measures, and many LEAs lack the resources and expertise to prepare competitive applications.⏎- The lack of conclusive research evidence about
-
pest, the bronze birch borer. River birches are less susceptible to the borers than are paper birch trees. And though river birches lack the classic white bark, their curly, cinnamon-colored bark has its own charm.⏎De
-
Jewish friends leshana tova (happy New Year), to send greeting cards, and perhaps plant trees in Israel in their honor.⏎Although the high holidays lack the merriment of Purim, the gift giving of Hanukkah, and
-
. There were and still are efforts to address the current challenges to space security but the instruments resulting from, or proposed by, those efforts lack the same legal force as the original, constitutive legal instruments; they are somewhat vague
-
Eventually the disease will cause plants to defoliate and die.⏎If a plant is infected by the pathogen, it cannot be cured. Homeowners who have the disease should remove and destroy all infected plants. Nearby uninfected impatiens
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: the most recent, and most recent, in the book of the story of the city of the world, I have been told about the recent and reclusive, in the book of the world, I have been to the city of the world
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
” and its guardian Izad (god), who was considered one of the greatest Iranian deities in the “Zarwani” beliefs, annually on Bahman 22 (February 11). Other names for this feast are “
-
out of date, and any more recent senses may be completely missing.⏎- The inedible parts of a grain-producing plant.⏎- To separate out the chaff, early cultures tossed baskets of grain into the air and let the wind blow away
-
Presence of Christ⏎By Chuck Collins⏎Center for Reformation Anglicanism⏎Transubstantiation became the official doctrine of the Roman Catholic Church at the Fourth Lateran
-
wine of Holy Communion change into the body, blood, and soul of our Lord when magic words are spoken over them by a fancy-dressed priest. Transubstantiation and its Anglican clone are excuses to worship Jesus in the bread and
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
- L1:1029 weakly excites
- L2:115 weakly excites
- L0:1301 weakly suppresses
- L1:323 weakly excites
- L2:2891 weakly excites
- L2:1166 weakly excites
- L1:388 weakly suppresses
- L1:1746 weakly excites
- L2:2347 weakly excites
- L2:2567 weakly suppresses
- L0:924 weakly suppresses
- L1:1477 weakly suppresses
- L2:2796 weakly suppresses
- L1:1464 weakly excites
- L1:702 weakly suppresses
- L2:111 weakly suppresses
- L2:293 weakly excites
- L2:2328 weakly suppresses
- L0:356 weakly suppresses
- L1:2115 weakly suppresses
- L1:2957 weakly excites
- L2:386 weakly excites
- L1:2683 weakly excites
- L0:178 weakly suppresses
Feeds later layers
- L11:2529 suppresses 'here' in 'read more here' links
- L9:1959 weakly excites
- L6:384 weakly excites
- L6:2556 weakly suppresses
- L7:171 weakly suppresses
- L6:2412 weakly excites
- L12:2805 weakly excites
- L6:2823 weakly excites
- L7:1758 weakly suppresses
- L11:1972 weakly suppresses
- L22:1980 weakly suppresses
- L23:1973 weakly excites
- L13:2803 weakly suppresses
- L23:2666 weakly excites
- L21:2557 weakly excites
- L17:3050 weakly suppresses
- L20:1060 weakly suppresses
- L15:1602 weakly excites
- L22:1337 weakly excites
- L16:1773 weakly suppresses
- L22:2136 weakly excites
- L8:2704 weakly suppresses
- L10:2754 weakly suppresses
- L6:568 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
- leep
- morning
- omon
- plex
- OfFile
- cost
- �
- headache
- cona
- contin
Suppresses
- ibase
- rawid
- entin
- ’y
- ícul
- !」
- ース
- .dest
- .det
- nds