"Al-Aqsa" and "Lunar Module" name parts
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. 5 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 responds to the second half of certain names, such as the "sa" in Al-Aqsa or "Module" in Lunar Module. It mostly stays quiet, and it strongly switches off on words like "ring" and "of" in "Kingdom of". Turning it up or down gave text that did not clearly connect to this.
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.
-
recalled that “248 took refuge there. They included 1 Islamic Jihad, 28 Hamas, 50 to 60 Al-Aqsa Martyrs. The remainder were ordinary townsfolk and included 100 uniform
-
the Israeli Supreme Court over the eviction of six Palestinian families from Sheikh Jarrah in occupied East Jerusalem.⏎- The next day, Israeli Police stormed the Al Aqsa mosque.⏎- A few days later, Hamas and other Palestinian groups started firing rockets
-
over the Pacific from an observation plane.⏎The European Space Agency's first cargo mission to the international space station ended in a spectacular fireworks show today, with the fiery re-entry of the
-
limestone and cullet⏎Information provided by the Glass Packaging Institute http://www.gpi.org.⏎Answers: C, A, B, B, C<|endoftext|>This article
-
third-stage boosters that were used for seismic studies and six Lunar Module Ascent Stages that were crashed at the end of their missions, as was
-
Pacific Ocean was colder than normal, but it was warmer than usual in the Atlantic. That combination caused a drought in Colorado that was just as bad as the Dust Bowl of the 1930s.⏎The ocean currents slipped back into their
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 set of all integers. The set of all integers is closed under addition and multiplication. Let's say, $ \mathbb{Z} $, and $ \mathbb{Z} $. Let's say $ \mathbb{
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and the operation of convolution define a commutative ring. Any ring without divisors of zero can be extended to a quotient field or
-
ai and Ayutthaya, and was incorporated into the Kingdom of Siam in 1786. Patani was divided into the three
-
chlorine/chloramine also reduces TTHM’s, insecticides, pesticides, synthetic fertilizers<|endoftext|>The Red Sea is a prime environment and resource for the Kingdom of Saudi Arabia. Centuries of shipping and fishing activities supported local communities. In recent
-
classically Diophantine geometry), a subfield of algebraic number theory.⏎- The study of the real points of an algebra
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
- L4:2174 suppresses
- L4:2211 weakly excites
- L3:2845 weakly suppresses
- L4:829 weakly suppresses
- L3:1989 weakly suppresses
- L4:993 weakly excites
- L3:1042 weakly excites
- L4:1100 weakly suppresses
- L3:1858 weakly suppresses
- L4:385 weakly excites multiplication and scaling in math
- L3:1664 weakly suppresses acronym endings like HSUS and SDS
- L4:970 weakly excites
- L0:2296 weakly suppresses
- L1:1369 weakly suppresses
- L4:2281 weakly excites
- L0:1350 weakly suppresses
- L3:503 weakly excites
- L3:1715 weakly suppresses
- L4:2388 weakly suppresses
- L4:2598 weakly suppresses
- L4:2953 weakly excites
- L3:1727 weakly excites
- L2:336 weakly suppresses
- L2:2015 weakly suppresses
Feeds later layers
- L6:511 suppresses
- L13:2594 suppresses
- L23:2666 weakly excites
- L7:1078 weakly excites
- L18:628 weakly suppresses
- L7:373 weakly suppresses
- L22:1562 weakly excites
- L25:1468 weakly excites
- L11:797 weakly suppresses
- L9:2299 weakly excites
- L12:2307 weakly excites
- L9:2569 weakly excites
- L24:2778 weakly excites
- L25:2303 weakly excites
- L6:3061 weakly excites
- L16:1915 weakly excites
- L16:2970 weakly excites
- L8:2647 weakly excites
- L6:2862 weakly suppresses
- L9:2545 weakly excites
- L6:54 weakly excites
- L9:1455 weakly excites
- L23:2451 weakly excites
- L9:847 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
- lée
- @student
- MITTED
- peri
- utschen
- atedRoute
- TERNAL
- виз
- meno
- utation
Suppresses
- olar
- ouser
- contains
- constraints
- ��
- トー
- جيب
- ideals
- -quarter
- zaw