"in some" opening a qualifier
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. 4 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 "some" when it comes right after "In" or "At" to open a sentence like "In some cases" or "In some countries". Turning it up gave wandering text that did not clearly reflect this.
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.
-
are published in a state’s code are grouped by subject matter into Titles, as in Title 11 of the United States Code (bankruptcy laws). In some states, including California, Texas and New York, the term "code" may
-
have often been without and are looking for a provider to assure their future financially. We have also learned to patiently build trust instead of expecting it. At times even enduring undeserved blame for not being the expected hero, meeting their needs as
-
protection[change | change source]⏎Copyright laws usually protect owners of copyright beyond their lifetime, such as until 50 years after the author's death. In some countries this period has been extended to 70 years. When the period of
-
involve an index of software patches and upgrades for key parts of the software system. Experts also recommend vulnerability scans that look for risk factors across complex software systems. In some cases, this involves a multi-phase process that consists of multiple scans and remediation
-
2 to 18 inches of soil, although the depth depends on a site's soil conditions. Root systems usually are shallower in heavy clay soils.⏎In some situations, it can be detrimental to amend or replace heavy clay backfill soil,
-
7799 for IT security and the ISO 9000 series for management, establish guidelines and general principles that codify industry best practices. In some cases independent certification
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 need to be able to find the input to the first line of the given here. So, 0 1 2 3 4 5 6 Okay, I need to be able to find the input for the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Bagh, Sheesh Mahal, Saif-ul-Maluk Lake, Mohenjo-daro, Minar-e-Pakistan, Pakistan Monument.⏎The main city
-
Indus Valley Civilisation of Harappa and Mohenjo-daro from the Bronze Age. Nowadays, the Thar Desert is the
-
Atmospheric Release Advisory Center in Livermore also assisted with the 1986 Chernobyl nuclear power plant disaster and the 1979 Three Mile Island accident
-
, sulfur, and microorganisms. (Crude oil with lots of sulfur in it is called “sour,” as opposed to “sweet.”) If enough of these contaminants collect in a steel pipe, they’ll work together to eat
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
- L3:122 excites
- L1:1340 weakly suppresses
- L1:1029 weakly excites
- L3:464 weakly excites
- L3:186 weakly suppresses
- L3:1945 weakly excites 'to' in 'person to person' phrases
- L2:1176 weakly suppresses
- L3:1394 weakly excites
- L3:512 weakly suppresses
- L3:914 weakly excites "any" in negated sentences
- L2:549 weakly suppresses
- L1:2801 weakly excites
- L3:50 weakly excites
- L2:1127 weakly suppresses
- L0:1870 weakly excites
- L3:342 weakly excites
- L1:279 weakly excites
- L3:3060 weakly suppresses
- L3:691 weakly excites
- L2:1315 weakly excites 'of' after nouns like 'variation' and 'section'
- L2:2328 weakly suppresses
- L2:2834 weakly excites
- L3:2181 weakly suppresses comma inside numbers
- L2:1124 weakly excites
Feeds later layers
- L11:522 excites
- L7:2733 excites
- L8:2231 excites
- L15:2651 suppresses
- L18:2896 excites
- L16:2830 excites
- L7:561 suppresses 'all' or 'every' before a group
- L21:331 excites
- L13:2339 excites
- L15:869 excites
- L21:1472 weakly excites
- L19:2909 weakly suppresses
- L14:1763 weakly excites
- L5:85 weakly suppresses
- L16:311 weakly suppresses
- L8:981 weakly excites
- L9:352 weakly excites
- L14:1175 weakly excites
- L11:158 weakly suppresses
- L12:942 weakly suppresses
- L17:847 weakly excites
- L6:220 weakly excites
- L19:2973 weakly suppresses
- L10:2867 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
- leet
- mund
- 需求
- RequestId
- greSQL
- 挑战
- ye
- 来电
- entimes
- 钽
Suppresses
- IMO
- %+
- Główn
- 深交
- /=
- Comparable
- endforeach
- xious
- closest
- สะ