unclear: brackets and parentheses in text
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. 11 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
The neuron is hard to pin down. It fires most on the word right after an empty pair of square brackets, which often marks a spot where text was removed, and it goes strongly negative on opening parentheses in citations. When turned up, the model produced strings of parentheses and numbers, so it seems tied to bracket and punctuation patterns.
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.
-
to great violence, with swelling, soreness to the touch, and sometimes redness of the skin…. This is a very formidable and extremely painful disease, [and] generally continues for many weeks…. The consequences of this disorder are often painfully
-
, stupidity and anger,” said Mangku Puseh Juga, one of 36 priests on hand to sanctify the temple.⏎“The main goal is (to) balance the macro world of nature and the micro world within each human.”⏎
-
[and] having a forest products industry that begins to move toward oaks. If these come true, that's going to change the nature of the industry and [it] will [have to] adapt. [Awareness of climate change]
-
. with a commission visiting the battlefield on September 28, 1926.:'27 North Carolina appropriated $50,000[when?] to purchase and landscape the site and to commission Gutzon Borgl
-
said that the image of Marx has been imprinted in his heart since he was a child, adding that “Marx molded by a Chinese artist contains emotion and [warmth].”<|endoftext|>Citadel of Herat⏎This citadel, which
-
which is really useful for us, because if you have the animals behave in the same way all the time, there’s no room for flexibility in the face of [climate] change.”⏎Even with a colony containing hundreds of
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 I were to put this into a system, I can't think of any way to think about it. Can you think of a way to put it in a system that could be something like this. For example, this is similar to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Biblical Theology "Biblical theology is the study of divine revelation as File Size: KB.⏎Get this from a library! Principles of biblical interpretation: (sacred hermeneutics). [Louis Berkhof] --
-
6.⏎Silver, Larry. “Nature and Nature’s God: Landscape and Cosmos of Albrecht Altdorfer.” Art Bulletin 81.2 (June 1999): 194–214.⏎
-
Historical, Doctrinal, and Practical Theology (3rd ed.) (edited by Philip Schaff) Vol. II at pages 684–
-
ETA⏎No references found.⏎Citations of this work BETA⏎No citations found.⏎Similar books and articles⏎Piero Antuono & Jan Beyer (19
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
- L10:66 suppresses opening parenthesis after a citation or term
- L4:64 excites
- L10:2562 excites
- L5:1499 excites opening bracket in math and definitions
- L7:2152 suppresses
- L10:893 excites
- L3:263 suppresses
- L5:2795 weakly excites
- L8:354 weakly excites opening parenthesis before a citation year
- L6:144 weakly excites
- L1:155 weakly suppresses
- L10:1883 weakly excites closing parenthesis after a number or unit
- L10:2376 weakly excites
- L1:919 weakly excites
- L3:2199 weakly suppresses
- L0:1951 weakly suppresses
- L0:2963 weakly excites
- L4:45 weakly suppresses
- L1:102 weakly excites
- L5:75 weakly excites
- L5:1890 weakly suppresses
- L7:668 weakly suppresses
- L3:57 weakly excites
- L8:18 weakly suppresses
Feeds later layers
- L12:2933 strongly excites
- L12:1287 strongly excites
- L16:1967 strongly excites
- L19:2339 excites
- L15:608 excites
- L18:980 excites opening parenthesis after a noun phrase
- L12:1502 excites
- L14:1516 excites
- L13:1698 excites
- L18:2928 excites
- L26:431 excites
- L16:1123 weakly excites
- L17:1509 weakly excites
- L16:1513 weakly excites opening parenthesis before a numbered list
- L13:2571 weakly excites commas before a continuation clause
- L18:682 weakly excites
- L20:1451 weakly excites
- L16:869 weakly excites
- L22:2877 weakly excites
- L14:1318 weakly excites "both" and "either" before alternatives
- L20:2173 weakly excites
- L27:109 weakly excites
- L14:3046 weakly excites
- L19:2354 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
- 서
- abbrev
- h
- outer
- ogg
- 所说的
- собственно
- out
- A
- /downloads
Suppresses
- ega
- .PO
- WOOD
- MX
- dex
- .dx
- Dx
- .lambda
- пуск
- LEX