closing brackets and parentheses
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. 22 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 the closing bracket, parenthesis, or full stop that ends a short aside or symbol, such as the ')' after '(Cádiz' or the '(%)' after a symbol. The pattern is common but loose, and turning it up or down did not give a clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
It stands for the number of expressions to follow. Each expression consists of a single line containing a positive integer number, an operator (one of +, - and *) and the second positive integer number. Every number has at most 500
-
places, in Sicily and Corsica, in Spain at Tartessus, and even beyond the pillars of Hercules at Gadeira (Cádiz)".⏎In the time of King
-
sidetrack⏎English → English⏎, v. i. [imp. & p. p. Departed⏎; p. pr. & vb.⏎.] [OE. departen to divide, part, depart, F.⏎d['
-
ued by||Banco Nacional de Cuba|⏎|Size||157 x 66 mm|⏎|Printer||American Bank Note Company (U.S.A.) – ABNC|⏎|Signatures||Ernesto (Che) Gue
-
makes sense for a function to return multiple things. In such cases, it's common to just return them in a tuple: "return (count, new_obj)".⏎Sometimes you might want to return objects of different types based on whether an operation
-
to rhyme with the word 'eye'. But, for whatever reason, it is usually pronounced as '/iy/' or “ee”.⏎How do you pronounce daemon in Golden Compass?⏎http://
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 way people are supposed to think about the way people are supposed to think about themselves. All of the way across all of history, all of the way across all of the lengths of the length of the length of all of the lengths.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
or other meat) is submerged in a salty liquid (“brine”) and left to marinate to add flavour, tenderise AND add moisture so the chicken stays juicier when roasted.⏎Is raw chicken OK in the fridge for 5 days
-
is of even greater concern when considering the changing demographics.⏎The intention to perform a behavior is the most important factor to actually engaging in future behavior, according to Fishbein and Ajzens (Fishbein & Ajzen 1
-
troops. Officers inspect an howitzer. The troops march. German troops and officers eat. Soldiers listen to the radio. Troops put on uniforms. German JU-52 aircraft taxi and take off. A fleet of aircraft in flight.
-
ard apple, carissa, dialium, flacourtia (Indian plum), marula, vangueria, tamarind, vitex and ‘jujube’ are picked in the wild.⏎They are used as accompaniments
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
- L0:110 excites
- L3:459 suppresses
- L19:39 suppresses
- L20:111 suppresses
- L16:48 excites
- L16:14 excites
- L16:0 excites
- L18:541 excites
- L16:27 suppresses
- L20:441 excites
- L21:195 excites
- L18:249 suppresses
- L18:569 excites
- L2:1083 suppresses
- L6:13 excites
- L19:231 suppresses
- L16:143 excites
- L20:198 excites
- L8:7 excites
- L21:1781 suppresses
- L20:1120 weakly excites
- L16:107 weakly excites
- L16:188 weakly excites
- L9:1377 weakly excites
Feeds later layers
- L24:2478 excites
- L27:1450 excites
- L27:2206 excites
- L27:325 excites
- L26:319 excites
- L26:744 weakly suppresses
- L27:1687 weakly excites
- L27:1494 weakly suppresses
- L27:128 weakly excites unclear: opening quotes and closing brackets
- L27:2442 weakly suppresses
- L26:2930 weakly excites
- L23:1506 weakly excites
- L24:2026 weakly excites
- L27:2504 weakly suppresses
- L27:766 weakly excites word endings split mid-word, like 'shower' in 'showered'
- L27:775 weakly suppresses
- L24:696 weakly suppresses
- L24:1988 weakly suppresses
- L25:2123 weakly excites
- L24:2403 weakly suppresses
- L23:1740 weakly excites
- L24:2699 weakly suppresses
- L24:2338 weakly excites
- L23:918 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
- ing
- ING
- ğun
- eft
- 지도
- ちは
- abd
- ドル
- 过的
- ments
Suppresses
- obel
- élect
- итет
- elas
- ramer
- 不限
- ework
- 楫
- alog
- lob