the word 'exception' or 'exceptions'
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. 3 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 'exception' (or 'exceptions') when it comes right after words like 'the only', 'is no', or 'an'. When it was turned up, the written text drifted into unrelated topics and got repetitive, so the steering shows no clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
,” you do it. I’m usually not a fan of historical fiction, particularly novels set during wars, but Joseph Bruchac’s Code Talker is an exception, a book I’m so happy to have been recommended.⏎During WWII, the
-
article Oils, Fats and Health).Varieties⏎General: Oils and fats are pretty uniform at 120 calories per tablespoon. The only exceptions are semi-solid products sold whipped with a lot of air - fewer calories by the
-
. When an athlete breaks the law, the league tries to avoid making a statement with their response. Michael Vick, Donald Sterling, and post-video Ray Rice are exceptions. Leniency—what we saw in the initial response to Rice’s behavior
-
s. Words ending in -quy don’t follow this rule (as in soliloquies).⏎Add -s to most nouns ending in -f.⏎Exceptions: In some cases, change the -f to -v and add -es
-
953, before there was a reasonably complete edition of his writings. Sure, there were plenty of Lincoln biographies before those dates, but with only one exception (an authorized multi-volume door-stopper by his former White House aides), there
-
indication of the part of the wall upon which they labored, such places being on that side of the city nearest their place of abode; the only apparent exception being, perhaps, where they repaired more than 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: I have been working on this for a while, and I have been thinking about it in my mind, so I think that this is a good experience. I have been thinking about this for a while, and I think that this is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
don’t like to eat food contaminated by pesticides,” said EWG president Ken Cook. “Our shopper’s guide to pesticides in produce gives consumers easy
-
As RECOGNIZED GENERALLY BY BIBLICAL SCHOLARS, TREATED AS A SCIENCE, DERIVED INDUCTIVELY FROM AN EXEGESIS OF MANY PASSAGES OF SCRIPTURE. By CL
-
eighth edition of its Shopper’s Guide to Pesticides in Produce with updated information on 45 popular fruits and vegetables and their total pesticide loads. EWG highlights the worst offenders with its new Dirty Dozen Plus list and the cleanest
-
latest USDA tests show we have much more work to do.”⏎EWG researchers analyzed annual pesticide residue tests conducted by the USDA and federal Food and Drug Administration
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
- L2:111 suppresses
- L2:75 suppresses
- L2:49 suppresses
- L2:2328 weakly suppresses
- L2:1321 weakly suppresses
- L2:549 weakly suppresses
- L2:2161 weakly excites
- L1:2143 weakly suppresses
- L2:122 weakly excites
- L2:1494 weakly suppresses
- L2:2893 weakly excites
- L0:1301 weakly suppresses
- L0:1224 weakly suppresses
- L1:575 weakly excites
- L0:1509 weakly excites
- L2:1717 weakly excites
- L0:178 weakly suppresses
- L0:1577 weakly suppresses
- L1:491 weakly excites
- L2:489 weakly suppresses
- L2:493 weakly suppresses
- L1:2532 weakly suppresses
- L1:125 weakly excites
- L1:3004 weakly suppresses
Feeds later layers
- L4:873 excites
- L19:1495 weakly excites
- L8:1795 weakly excites
- L16:1837 weakly excites
- L5:2285 weakly suppresses
- L23:1648 weakly excites
- L6:621 weakly excites
- L4:2684 weakly suppresses
- L21:2371 weakly excites
- L4:2771 weakly excites
- L5:1801 weakly excites
- L8:2874 weakly suppresses
- L20:1497 weakly excites
- L4:1539 weakly excites
- L6:2895 weakly excites
- L10:2279 weakly excites
- L16:2652 weakly suppresses
- L11:1560 weakly suppresses
- L5:805 weakly excites
- L4:2018 weakly suppresses
- L8:1680 weakly excites
- L11:3055 weakly excites
- L8:1438 weakly suppresses
- L4:2566 weakly suppresses
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
- 例外
- imation
- ika
- imon
- htub
- imals
- imate
- im
- (exception
- let
Suppresses
- 起来了
- itest
- arta
- 针
- stakes
- usalem
- -anchor
- altet
- ROSS
- UGH