"way" in "find their way"
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. 2 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 lights up on the word "way" in phrases like "found his way home" or "find its way into". Turning it up made the model produce unrelated text, so the steering did not reveal the 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.
-
product of today’s waste. Most of these are then burned, which means they release toxic chemicals in the air. The one’s that are recycled, find their way back to retailers, who re-seal them, and claim the water is “
-
coolant moderator combination had to be one that did not absorb neutrons to any great extent. To make natural uranium “burn” every available neutron has to find its way into fuel material. Using a neutron absorbing coolant/moderator – like the ordinary water
-
made to put in a crop for his captor, who conveniently neglected to tell him when the war had ended. Fortunately, Walker escaped again and eventually found his way back home, traveling from Columbus, TX, to Galveston, to New
-
. They exist in parts of Arkansas, Georgia, Mississippi, Oklahoma, North Carolina, South Carolina, Tennessee, Texas and Puerto Rico. They have even made their way to California, Missouri and across the Pacific to Hawaii.⏎The insidious fire ants
-
river she notices a white rabbit and follows it into its rabbit hole, emerging again in Wonderland, a fantastical place where meeting mad characters is easy but finding your way home all the while keeping your sanity is less so.⏎Carroll, Lewis.
-
experience. We had to pioneer suspensions of liberty. We had to march through strange swamps of total mobilization of civilian effort. We had to find our way in unknown and ambushed forests of peace-making and through the unrevealed and
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: I have been thinking about the issue in the company. I have been thinking about the issue in the company. I have been thinking about the issue in the company. I have been thinking about the issue in the company. **The issue in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
" which is used to correct skewness in decoder application.⏎And version 7 or higher has "version information" which has information of version itself.⏎next page : How to encode - Capacity. Encoding to data code words.⏎ next page<|endoftext|>
-
, let's understand the abstraction in Java first. A Class can be defined as a template / blueprint for creating objects which defines its state and behavior. variable x: Remember from the Java Syntax chapter that a class should always start with an uppercase first
-
5 cm⏎Transmission or reproduction of materials protected by copyright beyond that allowed by fair use requires the written permission⏎of the copyright owners. Works not in the public domain cannot be commercially exploited without permission of the copyright⏎owner. Responsibility for any use
-
wind 雨 : rain 声 : sound⏎Meaning for Line 4 花
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
- L0:1224 suppresses
- L0:178 suppresses
- L0:1301 weakly suppresses
- L1:1853 weakly suppresses
- L1:810 weakly suppresses
- L1:2509 weakly suppresses
- L1:2922 weakly excites
- L1:915 weakly suppresses
- L1:2444 weakly suppresses
- L1:1978 weakly suppresses
- L1:332 weakly excites
- L1:628 weakly suppresses
- L1:2343 weakly excites
- L0:157 weakly suppresses
- L0:944 weakly excites
- L1:501 weakly suppresses
- L1:1340 weakly suppresses
- L0:2066 weakly suppresses
- L1:2264 weakly excites
- L1:2801 weakly excites
- L1:2208 weakly suppresses
- L1:1013 weakly excites
- L0:3050 weakly excites
- L1:603 weakly excites
Feeds later layers
- L21:521 suppresses
- L16:1023 suppresses
- L18:363 suppresses
- L17:1224 suppresses
- L4:557 weakly excites
- L20:2746 weakly excites
- L9:13 weakly excites
- L4:1484 weakly suppresses
- L17:1078 weakly excites
- L12:1259 weakly suppresses
- L25:817 weakly suppresses
- L17:678 weakly excites
- L16:2804 weakly excites
- L3:46 weakly suppresses
- L6:2675 weakly suppresses
- L19:2308 weakly excites
- L5:1009 weakly excites
- L23:1548 weakly suppresses
- L18:1477 weakly suppresses
- L18:670 weakly excites
- L27:670 weakly excites
- L9:413 weakly suppresses
- L19:853 weakly excites
- L13:3058 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
- 拭
- OrDefault
- inality
- requently
- Wis
- fists
- arson
- icity
- 铛
- 猩
Suppresses
- .poly
- lot
- од
- GLOBALS
- ------------
- dehyde
- Egyptian
- den
- <dd
- undert