article 'a' after a role or title
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. 23 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" or "an" that comes right after a phrase like "and", "of which he was", or "not", where a job title or role is about to follow. Turning it up gave text that did not clearly relate to this pattern.
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.
-
at IBM Research – India for ten years. His work has led to twelve patents from the USPTO. A Senior Member of the IEEE and the ACM and a Fellow of the UK Higher Education Authority, he is the recipient of the INAE
-
“The goal of the research was to better understand the economic impact of North Carolina’s seafood industry,” says Jane Harrison, North Carolina Sea Grant’s coastal economics specialist and a project lead. “The shellfish results indicate the strength of the growing aquac
-
—still half the rate of the global average.1⏎Gayle Tzemach Lemmon, a senior fellow at the Council on Foreign Relations as well as a journalist and the best-selling author of The Dressmaker of Khair Khana,
-
Festivals, and Associate Producer/Writer for Westinghouse TV. He currently serves on the Executive Board of the Stage Managers’ Association, of which he was a founding member and former chairperson. He has taught at both Emerson College and Stock
-
has been sworn in as the new Director of the Central Intelligence Agency. As a career diplomat, he is the first person to serve in this role who is not a career intelligence officer. We believe that his experience and expertise will be a valuable asset
-
.<|endoftext|>Dr Ana Cruz is a Research Fellow with Modernising Medical Microbiology, and a key part of the new CRyPTIC Project, where we will aim to
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: what is the value of the least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least least
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
activities of the Mennonites of Russia can be divided into three periods: 1789-1860, 1860-191
-
is a cousin of the former first lady, is the chief rabbi of the black Jewish community. He also leads Beth Shalom B’nai Zaken Ethiopian Hebrew Congregation in Chicago. Uniquely among Hebrew Israelite rabbis,
-
from which the priests wash their hands and feet before engaging in Divine service; the compounding of the incense and of the anointing oil; and the Shabbat. I shall focus here upon the washing of the hands.⏎Hand-w
-
to purchase a Torah scroll and ram’s horn for the upcoming High Holy Day services. Originally called the Little Rock Congregation, the group formally incorporated as B’nai Israel in 1867 and elected Morris Navra as its first president
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
- L20:702 excites
- L20:2148 excites
- L19:1200 weakly excites
- L1:517 weakly suppresses
- L20:1020 weakly suppresses
- L4:2935 weakly suppresses
- L21:1148 weakly suppresses
- L11:2438 weakly suppresses 'from' after origin verbs like imported or brought
- L6:2871 weakly suppresses
- L3:3042 weakly suppresses
- L18:2451 weakly suppresses
- L22:1639 weakly suppresses
- L3:349 weakly excites
- L21:2041 weakly suppresses
- L19:1127 weakly excites
- L12:667 weakly suppresses
- L1:1826 weakly excites
- L21:1697 weakly suppresses
- L17:1343 weakly excites
- L15:666 weakly excites
- L9:747 weakly excites
- L5:2768 weakly excites 'to' after 'happen' (happen to)
- L20:2659 weakly suppresses
- L22:1341 weakly suppresses
Feeds later layers
- L27:1920 weakly excites
- L27:1125 weakly suppresses
- L27:842 weakly excites
- L25:1448 weakly suppresses
- L27:2227 weakly suppresses
- L27:2544 weakly excites
- L24:2913 weakly excites
- L24:485 weakly suppresses
- L24:1353 weakly suppresses
- L26:568 weakly suppresses
- L27:1719 weakly excites
- L25:1481 weakly excites
- L24:316 weakly excites
- L24:1631 weakly suppresses
- L25:497 weakly suppresses
- L27:1573 weakly excites
- L24:1441 weakly excites
- L26:1874 weakly excites
- L25:1825 weakly excites
- L25:2793 weakly suppresses
- L24:130 weakly excites line break at the end of a line
- L26:1936 weakly excites
- L25:1802 weakly excites
- L25:1361 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
- essian
- ci
- 自然
- educt
- 外籍
- (IConfiguration
- �
- 冮
- 과정
- natural
Suppresses
- istol
- engers
- letic
- mania
- ista
- eah
- arious
- .setAuto
- uales
- obbies