"and" joining two actions
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. 6 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 "and" when it links two actions or verbs, such as "the markets rallied, the press swooned, and" or "succumb and die." Turning it up did not produce any clear theme in the replies.
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.
-
Union directives. It is for technical evaluation only. This device has not been authorized as required by the rules of the Federal Communications Commission. This device is not, and may not be offered for sale or lease, or sold or leased until authorization is
-
point higher than it is now. Yet, after Nixon’s announcement, the markets rallied, the press swooned, and, even though his speech pre-empted the popular Western Bonanza, the
-
it to Willamette’s 1929 freshman orientation. A month later the stock market collapsed — his bank failed, his savings account was wiped out and his scholarship was withdrawn for lack of funds. “I thought the world was coming
-
leads to another classic hallmark of Alzheimer's, neurofibrillary tangle in nerve cells. After several decades of exposure to these insults, neurons ultimately succumb and die, leading to progressively damaged learning and memory centers in the brain.⏎The mice
-
wartime, when a nation's survival is at stake, the reasons for secrecy are most apparent, the secrecy restrictions imposed by the government are most widespread,[*] and acceptance of those restrictions by the citizens is broadest.[†] In peac
-
labor conflict that existed in the past two decades. The AFL and the CIO merged and formed one organization which represented “35% of all nonagricultural workers
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 need to a more comprehensive understanding of the role of the *first* and *second* parts of the problem. In particular, the first part of the problem, I suppose, the need to a more comprehensive understanding of the role of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Tejel (Syria’s Kurds: History, politics and society).⏎Geographically, Kurds in Syria live mostly in non-contiguous regions of the country—as is apparent on the map. Around 30-3
-
probably because of water column mixing induced by hydraulic dredging. Overall, 185 phytoplankton taxa were found, of which ca 50% were present in both lakes. No significant differences in species
-
with a trained eye.⏎For example, in the image shown below, the bottle’s cast shadow is clearly incongruous with the shape of the bottle (as is the shadow on this cover of Time Magazine). Such obvious errors in a shadow are
-
ure, supplemented with need-based fertiliser applications. The most significant aspect of SRI is that the fields are not kept submerged under water all the time, as is usual in rice farming, but are allowed to remain just wet without flooding.⏎The
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
- L1:1696 weakly suppresses
- L4:762 weakly suppresses
- L4:261 weakly excites
- L4:822 weakly suppresses
- L4:670 weakly suppresses
- L0:3035 weakly excites
- L0:1873 weakly excites
- L0:1729 weakly excites
- L3:569 weakly excites unclear: mid-phrase function words
- L2:370 weakly suppresses
- L2:2739 weakly excites
- L4:438 weakly excites
- L3:814 weakly excites
- L4:631 weakly suppresses
- L1:1121 weakly excites
- L0:356 weakly suppresses
- L0:618 weakly suppresses
- L2:2999 weakly suppresses
- L5:460 weakly suppresses
- L3:415 weakly excites
- L4:486 weakly excites
- L3:2674 weakly suppresses
- L3:1924 weakly suppresses
- L4:2355 weakly excites
Feeds later layers
- L8:132 excites
- L9:1903 suppresses
- L27:717 excites
- L7:93 excites
- L18:857 excites
- L19:1982 excites
- L23:298 excites
- L12:434 excites
- L21:1622 weakly excites
- L10:1449 weakly excites
- L13:625 weakly excites
- L18:509 weakly excites
- L20:837 weakly excites
- L18:867 weakly excites
- L21:655 weakly excites
- L18:1519 weakly excites
- L8:615 weakly excites
- L7:428 weakly suppresses
- L17:960 weakly excites
- L17:28 weakly excites
- L19:1501 weakly suppresses
- L18:920 weakly excites
- L17:2213 weakly excites
- L9:272 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
- they
- others
- oster
- they
- AssignableFrom
- 公开发
- convo
- other
- demás
- There
Suppresses
- affen
- 료
- ect
- ",-
- istingu
- foresee
- /watch
- isseur
- roducing
- reib