'same' in 'at the same time'
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. 17 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 'same' in set phrases like 'at the same time' and 'in the same way'. When it was turned up, the model's text got a little more repetitive, but it did not show a clear theme tied to this phrase.
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.
-
Ars. "It started as an effort to prove we could create a commercial quality game using NASA content that is fun. NASA was looking for a project at the same time the Army Game Studio was looking to branch out from America’s Army. It
-
28-29 October 2015 in Luxembourg.⏎Today’s digital world offers children across Europe countless opportunities for personal growth and learning. At the same time, this digital world potentially exposes them to risks. We live in a time
-
throughout life. Brain trauma, vascular lesions, the effects of toxic substances such as alcohol, all of which accumulate over a lifetime and impact brain function. In the same way that some people see their hair just grow into adulthood while others keep a mop
-
950s and through scientific drilling in the oceans in the 1960s, the foundation for plate tectonics was laid at the same time, however, Wegener's groundbreaking theory was turned upside down.⏎Earthqu
-
pinpointing and eliminating habits that interfere with the productive use of time.⏎At the same time, many methods provide insight into the development of new and better habits, which
-
of the scavenge air amount, and thereby an increase of the exhaust gas temperature level, which leads to reduced efficiency of the diesel engine itself, at the same time creates a remarkable potential for increased power output on both power turbines and steam turbines
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 the second, third, and, 4th, 5th, and, 5th, 6th, 6th, 6th, 7,7,7,6,6,6,6
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
on these cloud fortresses, and will do so even more in years to come. Remarkable as the cloud is, there is no reason to think it is anywhere near completion. The generation of data and the need for its analysis continue to grow
-
⏎The first gas masks were made in the late 1700s, and used exclusively by miners. These were very primitive, though. They were nowhere near as efficient as filtering out toxins as today's gas masks.⏎Gas masks were
-
er oscillator producing a sine wave of about 37 kHz, which is then boosted to about 2400 volts. That’s pretty good, but nowhere near the 15 kilovolts needed for a self
-
at http://www.fastmr.com/catalog/browse.aspx?category=16&page=1. The online catalog offers a description for each report, which can cost anywhere from $75.00 to $11,400.
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
- L4:2695 suppresses
- L16:282 weakly excites
- L4:1505 weakly suppresses
- L7:1508 weakly excites
- L4:1330 weakly suppresses
- L16:80 weakly suppresses
- L6:2553 weakly suppresses 'for' after praise words like 'known for'
- L15:1145 weakly excites
- L3:323 weakly excites
- L3:1490 weakly suppresses
- L10:411 weakly suppresses
- L5:2924 weakly excites
- L5:888 weakly suppresses zero after a decimal point
- L1:1673 weakly suppresses
- L4:902 weakly excites
- L9:750 weakly excites
- L4:2453 weakly suppresses
- L16:1156 weakly excites
- L15:2846 weakly excites
- L16:2682 weakly suppresses
- L12:2855 weakly excites
- L11:1215 weakly suppresses
- L6:1351 weakly excites
- L7:1947 weakly excites
Feeds later layers
- L23:875 suppresses
- L26:551 suppresses
- L26:1060 weakly suppresses
- L18:1925 weakly excites
- L22:3027 weakly suppresses
- L19:2080 weakly excites ampersand joining two items
- L19:2201 weakly suppresses
- L19:933 weakly excites
- L18:2037 weakly suppresses
- L24:847 weakly suppresses
- L19:1183 weakly suppresses
- L19:643 weakly excites
- L23:742 weakly excites
- L23:1414 weakly suppresses
- L22:2966 weakly suppresses
- L21:2973 weakly suppresses
- L19:1070 weakly suppresses
- L25:2850 weakly suppresses
- L23:621 weakly suppresses
- L21:1242 weakly suppresses
- L18:726 weakly suppresses
- L18:1829 weakly excites
- L27:1896 weakly suppresses split names and comparative stems
- L25:2232 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
- principle
- underlying
- -sex
- ../../
- adesh
- iating
- هة
- ести
- 专业
- 原理
Suppresses
- rels
- asti
- differences
- atures
- Differences
- closer
- ACTER
- 更强
- cess
- esModule