'or' in either-or choices
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. 11 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 'or' in phrases that offer two options, like 'prove or disprove' or 'publish or perish'. It is switched off by words like 'changed' and 'change'. Turning it up did not produce a clear theme in the model's text.
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.
-
th Report talked about how wrongful incarceration, conviction, and prosecution are defeating principles of justice. In order to curb this, the notion of ‘innocent until proven guilty
-
their] former glory.”⏎Supersymmetry and Q-balls are thus far only theoretical, although Cox asserts that he and the physicists at CERN will prove or disprove their existence by 2017. However, as Anthony Kauf
-
crisis are wide ranging, but typical center on the preference for “novel” findings in science and the pressure on investigators (especially new ones) to “publish or perish”. My favorite reason places the blame for the entire crisis on p-values.⏎
-
years of age, who had already failed to respond to at least two antipsychotic treatments, to participate in a 12-week, double-blind, randomized study – the most rigorous of clinical trial designs. After initial assessments, the patients
-
nation.⏎According to the Title I Director’s Conference (February 2003), “highly qualified” means that all teachers must have full State certification or have passed teacher licensing exam, and hold a license to teach, and that certification
-
sheep.⏎As such merino production inherently has the potential of being fully sustainable providing it is operated within the carrying capacity of the ecosystem.⏎They must operate, or be in the process of implementing, a grazing management strategy based on long term sustainability.⏎
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 thinking about this: I have been thinking about this: I have been thinking about this, and I will now share the following. The following is the content I have been thinking about: I have been thinking about this. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
“triggered” using a web bug. Each time it is requested, it checks to see if the current wiki page (containing the web bug) has changed and needs re-indexing.⏎On the first release containing the indexer<|endoftext|>Publications
-
things that background applications do, namely,⏎read its configuration files, go into the background, shut down when told to,⏎and restart in the case of a configuration change. Because it is an HTTP server,⏎it must be able to listen for incoming
-
loaded ? Or what to do if we would like to selectively update only parts of a page. For example only the stock item in an HTML TABLE whose rate has changed ?⏎Let's assume that we have a Java web/application server from which we want
-
devfsadm command. This daemon is started by the service management facility when a system is booted.⏎Because the devfsadmd daemon automatically detects device configuration changes generated by
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
- L9:2920 suppresses
- L10:2051 suppresses
- L10:135 weakly excites 'or' in alternative requirements
- L8:1086 weakly suppresses
- L7:2974 weakly suppresses
- L6:1588 weakly suppresses
- L3:85 weakly excites
- L8:1667 weakly suppresses
- L10:14 weakly suppresses
- L9:1348 weakly excites
- L10:1208 weakly suppresses
- L5:2602 weakly excites
- L2:1766 weakly suppresses
- L3:1832 weakly excites
- L10:2399 weakly excites
- L0:2820 weakly excites
- L9:2467 weakly excites
- L7:2435 weakly suppresses
- L2:2992 weakly excites
- L10:64 weakly suppresses
- L6:2368 weakly suppresses
- L4:1312 weakly suppresses
- L10:2076 weakly excites
- L4:3025 weakly suppresses
Feeds later layers
- L13:1981 excites
- L18:2284 suppresses
- L13:2961 suppresses
- L17:539 excites
- L16:1678 suppresses
- L18:2643 suppresses
- L24:954 weakly suppresses
- L12:1744 weakly excites
- L13:3051 weakly excites
- L26:1958 weakly suppresses
- L16:2997 weakly excites
- L16:2314 weakly suppresses
- L17:1916 weakly excites
- L12:839 weakly excites
- L19:1219 weakly excites
- L18:2027 weakly suppresses
- L14:605 weakly excites
- L22:1555 weakly suppresses
- L23:1908 weakly excites
- L22:2213 weakly excites
- L21:1642 weakly excites
- L16:2871 weakly suppresses
- L17:2673 weakly suppresses
- L14:2623 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
- 蔚
- dez
- iska
- Förder
- ensem
- icone
- mieux
- heit
- skies
- BEST
Suppresses
- changes
- 变化
- 那次
- 的变化
- ModifiedDate
- 修改
- 갱
- 直接影响
- 改变
- changed