'as well as' and list connectors
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. 18 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 word 'as' after 'as well', and on 'and' in lists of things being compared or added together. It is a weak fit overall. Turning it up produced repeated phrases like 'a box of any other product'.
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.
-
similar to Depart.⏎English → Indonesian⏎English → English⏎v 1: move away from a place into another direction; “
-
is still practiced within the Vietnamese and French adapted systems, as well as through the teachings of Jeffrey Yuen. The channel systems are considered by many to
-
in our nation, health care system and medical research, which does not have enough representation from the Black community.”⏎The report details the disparities for African Americans as well as many other groups, highlights areas where some progress has been made, and lays out
-
012 hurricane season, the industry continues to build upon critical lessons learned from 2008's major hurricanes, Gustav and Ike, as well as other powerful storms, such as 2005's Katrina and Rita and
-
reading instructions during the test, right? Also, this acquired familiarity is going to help you keep calm during the test. For this writing and language section as well as the others.⏎So, without further ado, let’s go over the three broad
-
Zimmer (1985), and Kaufman (1990), and some highlights are in Roberson (1980, 1986, 1993). The Kaufman book — Field Guide 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: the number of people who have been in the number of people in the number of people in the number of people in the number of people in the number of people in the number of people in the number of people in the number of people in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and useable - the rest is just science.⏎Since its 2003 launch, Drift Catchers have been used in 27 projects in nine states by scores of trained volunteers and community leaders. Below are two case studies.⏎
-
-income mothers next week. They will be invited to join the study, which is called Baby’s First Years, shortly after giving birth at one of ten hospitals in four cities across the United States (to avoid influencing the experiment, the researchers asked The
-
service is advertised as providing a large computing capacity (potentially many servers) much faster and cheaper than building a physical server farm.⏎AWS is located in 10 geographical "regions": US East (Northern Virginia), where the majority of AWS servers
-
queries, or even monitor employees. To understand people’s trust and attitudes towards workplace AI, we surveyed over 17,000 people from 17 countries: Australia, Brazil, Canada, China, Estonia, Finland, France,
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
- L16:1430 suppresses
- L17:2052 suppresses
- L15:2067 excites
- L16:1782 suppresses
- L16:2962 suppresses unclear: weak mid-phrase prepositions
- L16:1361 suppresses
- L15:1738 excites
- L14:2756 suppresses
- L17:1470 excites
- L15:292 excites
- L16:1733 excites
- L13:2083 suppresses
- L12:2832 suppresses the word 'even' as emphasis
- L14:2360 excites
- L17:1009 excites
- L17:586 suppresses
- L14:980 suppresses
- L12:115 suppresses
- L15:1693 suppresses
- L17:1017 suppresses
- L17:1079 suppresses
- L16:1240 suppresses
- L17:1151 excites
- L17:2264 weakly excites
Feeds later layers
- L19:2484 suppresses
- L19:354 suppresses
- L19:666 suppresses
- L19:1640 excites
- L19:2756 weakly excites
- L20:1923 weakly suppresses
- L19:356 weakly suppresses
- L19:519 weakly suppresses
- L19:331 weakly suppresses
- L19:2897 weakly suppresses
- L22:870 weakly excites
- L19:2298 weakly suppresses
- L19:355 weakly suppresses
- L20:2629 weakly suppresses
- L20:1189 weakly excites
- L22:2904 weakly excites
- L19:2204 weakly excites
- L19:1861 weakly suppresses
- L19:2274 weakly excites
- L19:416 weakly suppresses
- L21:1931 weakly suppresses
- L20:708 weakly suppresses
- L21:1126 weakly excites
- L21:2287 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
- bserv
- TestMethod
- grey
- ason
- ọi
- edian
- 彩
- nev
- 任何
- 个工作
Suppresses
- ature
- 本身
- مباشر
- מונה
- отов
- HL
- ologically
- ogn
- NAT
- 牯