"g" in "e.g." examples
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. 1 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 "g" in "e.g." when it introduces examples in brackets, such as "(e.g. eagle shrieks)". Turning it up or down gave replies on unrelated subjects, so the steering results do not show what it does.
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.
-
, salinity) affected the physiology, reproductive capacity, and survivorship of marine organisms. Likewise, Comparative Physiology seeks to unify understanding of common physiological processes (e.g., osmoregulation, respiration) across many taxa; in this un
-
and by eagles and leopards in the Tai Forest.⏎Using transcriptions of these monkey calls gathered in field experiments involving playbacks of predator calls (e.g. eagle shrieks and leopard growls), the researchers found greater complexity in expression
-
and perhaps most importantly,⏎conduct research on those behaviors with the greatest environmental impact (e.g., transportation choices, particularly in the U.S.) rather than those that are more
-
not advisable. Additionally, homebound instruction may be used to supplement the classroom program for health-impaired children whose conditions may interfere with consistent school attendance, (e.g., students receiving dialysis, radiation or chemotherapy). The inability to attend school must
-
the solid Earth is composed, the processes that generate Earth's landforms, natural hazards associated with geologic processes, geologic time, and surface processes (e.g., glaciers
-
the performance of methods designed to eliminate edge-effect bias. We show that the reflection method solves the edge-problem for designs that use symmetric radial transects (e.g., straight-line and X-shaped transects centered at the sample point). This method
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 am not sure about this, but I do like the rest of the week. But I have been thinking about lately: "I am not sure about this, but I do like the rest of the week." But I have been
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
(i) Khadi cloth was often more expensive than mass produced mill cloth and poor people
-
(A) tails. This prevents oligo d(T) mediated mRNA isolation. Here,
-
(2) failing to provide food, care or shelter;⏎(3) abandoning an
-
Currents⏎An alternating voltage may be generated⏎(1) by rotating a coil at constant angular velocity in a uniform magnetic field as shown in Fig. (a).⏎(2) by rotating a magnetic field at a constant angular velocity within a
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
- L0:1224 suppresses
- L0:178 suppresses
- L0:1301 suppresses
- L0:1 weakly suppresses
- L0:13 weakly excites
- L0:595 weakly excites
- L0:2284 weakly excites
- L0:1404 weakly excites
- L0:16 weakly excites
- L0:1672 weakly suppresses
- L0:1062 weakly excites
- L0:1619 weakly suppresses
- L0:2362 weakly suppresses
- L0:1991 weakly excites
- L0:489 weakly excites
- L0:1705 weakly suppresses
- L0:210 weakly suppresses
- L0:2004 weakly excites
- L0:564 weakly suppresses
- L0:2184 weakly excites
- L0:1060 weakly suppresses
- L0:1249 weakly suppresses
- L0:1935 weakly excites
- L0:581 weakly excites
Feeds later layers
- L4:45 excites
- L3:1111 weakly excites
- L19:1797 weakly suppresses
- L19:1966 weakly excites
- L12:1146 weakly excites "but" after a concession
- L15:628 weakly excites
- L25:2822 weakly excites
- L5:1999 weakly excites
- L20:2253 weakly suppresses
- L11:2058 weakly suppresses
- L18:1090 weakly excites unclear: mid-sentence words before adverbs
- L22:2677 weakly excites
- L18:2938 weakly suppresses
- L8:1998 weakly suppresses
- L27:1565 weakly excites
- L20:104 weakly suppresses
- L27:548 weakly suppresses
- L15:3009 weakly suppresses
- L18:620 weakly suppresses
- L20:2344 weakly suppresses
- L3:1329 weakly excites
- L20:1743 weakly excites
- L5:1556 weakly suppresses
- L24:1527 weakly excites space before a number, often an age
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
- ellan
- fühl
- рен
- ул
- reon
- iron
- icana
- at
- 饫
- .NoError
Suppresses
- đích
- tracts
- getStore
- ög
- 目的
- .hxx
- _CTX
- colours
- prizes
- (co