opening parenthesis after a sentence
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. 17 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 reacts to an opening bracket that starts a side note right after a finished sentence, such as a source credit or an explanation. When it is pushed negative, it reacts to the word 'And' that starts a new sentence instead. Steering changed the story in ways that do not clearly follow this pattern.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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.
-
state’s current coastal Master Plan, which incorporated worst-case scenarios for relative sea-level rise calculated two years ago— which the new figures now make out-of-date. (The state’s estimates of sea-level rise and subsidence are listed on page
-
(I say “apparently” because there are differing accounts of what exactly happened). Coleman, riding in the rear seat, was not wearing a seat belt. (She may have been unable to see over the cockpit when strapped in.) Her mechanic
-
from traditional South African folk songs, and protest songs from famous musicians from Miriam Makeba and Hugh Masekela to U2 and Stevie Wonder. (Source: PBS)⏎The conclusion of "Have You Heard From Johannesburg?, Free at
-
a big one with what looks like a fuzzy beard of brown leaves underneath its green fronds. It’s naturally found around desert oases in the Colorado Desert. (The Colorado Desert is not in Colorado, but is named for the river. Joshua
-
, Dr. Patterson has maintained a database now listing 370 genetic disorders in dogs, with 5 to 10 new ones added each year. (By comparison, there are some 5
-
Another case is where we need to authenticate some data, i.e. produce a kind of integrity check that only a party with a given private key could produce. (In this case, the general solution is to combine a hash code with encryption.)⏎
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: if, for, say, if, for, if, for, if, if, if, if, if, if, if, if, if, if, if, if, if, if, if, if, if,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
health (not to mention a whole host of other health issues).⏎Read on to learn more about some of the drinks that may contribute to aging your brain faster. And for more, check out The Worst Foods Increasing Your Risk of Dementia.⏎S
-
real impact on their future. They will look around at their environment to determine if they want to stay where they are or move somewhere else when they get older. And they may get their first job at 16 or 17, which
-
a reader and a textbook. First, it provides extensive readings from the central writings in ethics along with writings from literary figures who offer concrete examples of lived ethics. And second
-
and, later, to ensure a multiplicity of routes such that the system could survive a nuclear attack. It was not to keep the system’s data safe. And as ARPANET gave way to the purely academic world of the early internet,
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
- L16:489 suppresses
- L16:432 suppresses period after month or initial
- L14:155 excites sentence-ending periods in prose
- L14:128 suppresses
- L16:240 suppresses comma after 'sometimes' or 'the truth'
- L9:2447 excites
- L10:478 suppresses
- L14:762 suppresses
- L13:53 excites period ending a claim before elaboration
- L4:102 excites
- L15:313 excites decimal point inside numbers
- L11:1194 excites
- L5:168 suppresses
- L4:7 excites
- L16:482 excites
- L14:422 excites periods after initials and abbreviations
- L13:813 suppresses
- L16:203 weakly suppresses
- L9:2302 weakly excites
- L10:612 weakly excites full stop ending boilerplate or a title
- L16:869 weakly excites
- L10:1004 weakly excites
- L16:285 weakly excites
- L3:854 weakly suppresses
Feeds later layers
- L19:979 excites
- L18:1841 excites
- L18:2045 excites
- L26:22 excites the 'res' fragment inside words
- L18:2217 excites
- L18:1403 excites
- L18:1791 suppresses
- L26:236 weakly suppresses
- L24:216 weakly suppresses
- L19:2493 weakly excites
- L25:1113 weakly suppresses
- L24:411 weakly suppresses
- L19:1262 weakly suppresses
- L27:286 weakly suppresses
- L18:621 weakly suppresses
- L25:373 weakly suppresses
- L18:892 weakly suppresses
- L18:1749 weakly suppresses
- L18:1216 weakly excites
- L24:1853 weakly excites
- L20:68 weakly excites
- L18:1671 weakly excites
- L22:368 weakly suppresses
- L19:2416 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
- cased
- pled
- ponent
- оказ
- cle
- unction
- LOAT
- uum
- zá
- uits
Suppresses
- ,json
- ,r
- ,options
- rão
- ?,?,
- ,np
- ,R
- ,’
- ertools
- though