'I' starting an acronym or word fragment
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. 22 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 a capital letter 'I' that begins a short abbreviation or a piece of a longer word, such as the 'I' in 'IUCN', 'ILO' or 'IBS'. Its effect on the written output was hard to see, because the steered text was unrelated to this pattern.
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.
-
edition of the International Basic Safety Standards. The edition is co-sponsored by seven other international organizations — European Commission (EC/Euratom), FAO, ILO, OECD/NEA, PAHO, UNEP and WHO.
-
security, and many other scientific domains.⏎"NVIDIA and IIT Delhi share the common vision of developing technologies that boost computing performance to exas
-
), the Federal Institute for Drugs and Medical Devices (BfArM) and IFA. 03/31/2020.<|endoftext|>Def
-
. The study, conducted on Drosophila melanogaster by scientists at the Institute for Research in Biomedicine (IRB Barcelona) and led by ICREA researcher Marco Milán, has reproduced each of the steps known
-
the species.⏎The Great Slaty Woodpecker is a vulnerable species on the IUCN Red List, due to habitat loss. Without proper conservation measures, its
-
In studies, up to 75% of IBS sufferers have been helped with a low FODMAP diet.⏎FODMAP
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 to be able to create a right powerlomate. I were not able to create a powerlomate until the right powerlomate of the right side, and I were not able to create a powerlom
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
political contention until it was taken under the control of the autocratic Roman emperors.⏎- List of Latin phrases⏎- Roman circus⏎- Idiot/idiocy<|endoftext|>Four European capital cities lie on Danube, Vienna, Bratisl
-
far stood outside the Standard Bank offices in Rosebank, Johannesburg, the Soweto Theatre in Soweto, the Iziko Museum in Cape Town, Maropeng in the Cradle of Humankind
-
) It is derived from the Hebrew Ebion, "poor," and was not given, as Origen supposes, for their low view of Christ. Id. at 6
-
could be written on this subject. I'm guessing the "multi-million line production code base" in question is the Dropbox code base, and maybe Dropbox has an idiomatic way of writing Python with lots of "isinstance()"s.⏎> In
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
- L18:205 excites
- L21:1492 excites
- L2:1463 suppresses
- L17:367 suppresses
- L20:2017 suppresses
- L1:45 excites
- L17:106 excites
- L10:1641 suppresses
- L19:551 excites
- L19:1241 excites
- L17:366 suppresses
- L20:342 suppresses
- L0:2262 excites
- L6:483 suppresses
- L16:503 suppresses
- L20:1228 suppresses
- L16:203 weakly excites
- L20:121 weakly excites
- L18:1442 weakly excites unclear: word fragments and digits inside numbers
- L0:47 weakly suppresses
- L1:985 weakly excites
- L1:1274 weakly excites
- L16:493 weakly excites
- L2:7 weakly excites
Feeds later layers
- L26:2485 excites capital I starting a word fragment
- L23:678 excites
- L24:2607 weakly excites
- L24:1116 weakly excites
- L27:2113 weakly excites
- L26:2469 weakly excites
- L25:785 weakly excites
- L26:1627 weakly suppresses
- L26:1959 weakly suppresses
- L26:1839 weakly excites
- L24:1468 weakly excites
- L23:2151 weakly excites
- L23:935 weakly excites
- L24:2219 weakly excites
- L25:2082 weakly excites first-person 'I' starting a clause
- L24:1972 weakly suppresses
- L23:2358 weakly excites
- L23:958 weakly suppresses
- L25:2583 weakly excites
- L27:732 weakly suppresses
- L24:420 weakly suppresses
- L24:2425 weakly excites
- L24:709 weakly excites
- L23:1494 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
- inis
- setting
- ampp
- Setting
- Setting
- believe
- Corinthians
- inion
- bero
- Dtype
Suppresses
- mes
- richt
- olate
- breakup
- olves
- net
- diff
- 侘
- nen
- 噔