"it" as a subject pronoun
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. 8 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
Fires on the word "it" when it starts a clause and refers back to a thing or organisation named earlier, such as a program, a stamp scheme or a satellite. Turning it up did not produce any clear theme in the writing.
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.
-
history in this sense `scientific history'.⏎Although the object of research and the subject matter of scientific history are the events and episodes of the past, the laws which it deduces are not
-
Five miles south of here is Egan Canyon, where one of the oldest gold mines in Nevada was located. Indians mined gold there. A stage station was located by Major Howard Egan in 185
-
An element can be represented on several diagrams. You can quickly locate which diagrams it occurs in throughout the model, using the Element Usage dialog. Having identified a diagram
-
to purchase wetlands and wildlife habitat for inclusion into the National Wildlife Refuge System — the only federal lands where wildlife comes first.⏎Since the stamp program was created, it has generated over $700 million and protected over 5.2 million
-
aviolet background—that transform it into a mere dark-matter shadow of its former self.⏎Simultaneously, as the satellite galaxy moved closer to the massive system, it encountered the overwhelming gravitational force of the much larger mass. This force wrenched lumin
-
of color on the moon for the first time.⏎The vehicle the astronauts ride in will look different as well. Rather than a rocket and capsule built by NASA, it will be constructed by SpaceX. The private firm pitched its massive Starship rocket for
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 user might be interested in using a different approach to find the solution. Let me think about some of the possible approaches. First, I will try to use the given equation $ f(x) = \frac{f(x)}{x
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
modified file (here called source.cpp) and the non-modified version of the file (here called source.cpp.orig.)⏎diff -u -p source.cpp.orig source.cpp⏎That lists the difference between the two files in the unified diff
-
Mr Kamesh.Let's walk through the program once again. 1) ValHold v = new ValHold(); v.i=30; another(v,i);⏎This code creates a obj. reference v and passes that reference to the
-
can be the ideal vehicle for children to begin to develop some insight into the experiences of individuals escaping from conflict situations, and empathy for their plight. Despite the theme these books all contain threads of hope and indicate the power of kindness.⏎Azzi in
-
sample configuration file which resides under Drupal directory ‘sites/default/default.settings.php’.⏎# cd /var/www/html/drupal⏎# cp sites/default/default.settings.php sites/default/settings.php⏎Now make a “default” directory writable by installer.⏎#
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
- L7:312 strongly excites
- L6:142 excites
- L3:528 excites
- L6:270 excites
- L4:1088 suppresses
- L7:401 suppresses
- L6:324 excites
- L7:193 suppresses space before day after month
- L6:59 suppresses
- L7:23 suppresses
- L6:61 suppresses
- L6:29 weakly suppresses
- L0:1926 weakly excites
- L2:1739 weakly excites
- L6:45 weakly suppresses
- L7:396 weakly excites
- L4:1527 weakly excites
- L7:59 weakly suppresses
- L3:472 weakly suppresses
- L4:260 weakly excites
- L4:319 weakly suppresses
- L7:13 weakly suppresses 'the' after 'member of' or 'one of'
- L5:8 weakly excites
- L6:152 weakly excites
Feeds later layers
- L14:254 suppresses
- L9:1557 suppresses 'the' after 'disease of' or 'infection of'
- L17:609 suppresses
- L13:536 suppresses
- L10:67 suppresses
- L9:145 suppresses
- L11:55 excites
- L16:342 suppresses
- L10:342 suppresses
- L12:72 suppresses
- L11:577 suppresses
- L12:71 suppresses
- L9:682 suppresses
- L18:598 suppresses
- L15:376 excites
- L9:172 excites
- L14:382 suppresses
- L11:329 weakly suppresses number or citation fragment split
- L10:121 weakly suppresses
- L11:2805 weakly suppresses "to" or "for" linking into a following phrase
- L15:2427 weakly suppresses
- L16:844 weakly suppresses
- L9:270 weakly excites
- L13:317 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
- eral
- ดวง
- хват
- 网讯
- pan
- ẳng
- notation
- compet
- 离
- cribing
Suppresses
- aukee
- Georg
- Augusta
- urent
- ftware
- Benedict
- iled
- .getOwnPropertyDescriptor
- aurant
- Dank