'for' after past participles like selected
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. 7 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 responds to the word 'for' when it comes right after a verb like 'selected', 'tested' or 'chosen', where it introduces a purpose or reason. Turning it up did not clearly change the writing in a visible way.
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.
-
in theory, target any protein that you have a ligand for and send it off for destruction. Building these molecules can be a challenge, and the choice of linking groups
-
and a Dictionary of controlled vocabulary.⏎The Writing Rules cover aspects of grammar and style.⏎The Dictionary specifies the general words that can be used.⏎These words were chosen for their simplicity and ease of recognition. In general, there is only one word for
-
arvesting has threatened its long-term sustainability, and there is a knowledge gap relating to the effect of harvesting on the ecosystem. After conifer species were selected for harvesting, the mixed Chinese fir, pine, and
-
summer as a thesis project for my M.A. degree in Anthropology at the University of Alaska. This site was selected for several reasons. First among the reasons for selecting this site was the fact that it
-
Everything will be tested for the first time at a Ugandan hospital, the Comprehensive Rehabilitation Services for Uganda (
-
and wear resistance. But, we need to always remember that tiles are not indestructible. If mistreated, or subjected to harsh conditions, or wrongly selected for purpose, the glaze on a ceramic tile will scratch or it can wear.⏎
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: 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
as a sports arena..⏎Do we have 2 moons?⏎After more than half a century of speculation and controversy, Hungarian astronomers and physicists say they have finally confirmed the existence of
-
mathematics, the liberal arts, and the value of a well-rounded education may find themselves echoing Daniel Dennett: “There’s nothing I like less than bad arguments for
-
from a quarry and shaped using bronze tools, then smoothed with sand in order to fit tightly together. No mortar was used in the construction. Even after many centuries of wear and tear, the precision is amazing; It would be impossible to slide a
-
Korean Peninsula, while the northern part re-emerged as Balhae, a successor-state of Goguryeo. After nearly 1000 years of rule, Silla fragmented into the brief Later Three Kingdoms, Silla,
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
- L5:3049 suppresses
- L3:381 excites 'for' in research institute names
- L6:2379 suppresses
- L4:1035 suppresses 'for' after verbs like 'fall' or 'click'
- L6:685 excites
- L6:368 excites
- L6:2718 weakly suppresses
- L4:677 weakly suppresses
- L6:439 weakly excites
- L4:1154 weakly excites
- L4:1023 weakly suppresses
- L3:159 weakly suppresses
- L6:1684 weakly suppresses
- L4:1525 weakly suppresses
- L5:1090 weakly excites
- L4:1013 weakly suppresses
- L1:1516 weakly suppresses
- L5:2251 weakly excites
- L5:2071 weakly suppresses
- L4:2748 weakly suppresses
- L0:2534 weakly excites
- L6:2907 weakly suppresses
- L4:1897 weakly suppresses
- L6:1880 weakly suppresses "by" introducing a deadline or means
Feeds later layers
- L17:3000 excites
- L24:480 excites
- L21:2338 excites
- L16:817 suppresses
- L15:2925 excites
- L17:1196 excites
- L26:298 excites
- L20:233 suppresses
- L18:622 excites
- L8:655 suppresses
- L27:303 excites
- L27:208 weakly excites
- L17:264 weakly suppresses
- L20:2948 weakly excites
- L9:1269 weakly suppresses
- L10:272 weakly suppresses
- L9:143 weakly suppresses
- L19:1840 weakly excites
- L10:459 weakly excites
- L18:277 weakly suppresses
- L22:1224 weakly suppresses
- L8:45 weakly excites
- L17:790 weakly excites
- L21:1813 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
- ilmington
- litigation
- áb
- ól
- 地理位置
- cleanliness
- 综合治理
- kính
- ót
- inking
Suppresses
- verse
- vers
- deselect
- hitch
- NSSet
- 饰
- Harness
- utter
- .setChecked
- หมด