"as" in "it seems as if" phrases
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. 8 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 "as" when it starts a comparison like "it seems as if" or "it felt as though", the word that sets up an imagined or perceived comparison. When it was turned up, the model drifted into rambling, hesitant sentences about things that seem or feel a certain way.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
his book:⏎The two entities that were supposed to write the regulations, the Pay Board and the Price Commission, wrangled and dithered. When it looked as though they were going to miss a crucial deadline
-
(the form is at the bottom of the page).⏎Esther – Day 3⏎Xerxes wanted to show off his beautiful wife! But it seems as though he did not give her too much thought until he had drunk too much!
-
Frankenstein’s monster, consisting of two identical M1 Max chips bolted together using a silicon interface that serves as a bridge. This clever design makes it seem as if the conjoined chips are in fact just
-
because there is seldom just one. In the beginning, actinic keratoses are frequently so small that they are recognized by touch rather than sight. It feels as if you were running a finger over sandpaper. There are many times the number
-
y⏎May 27, 2012⏎When the last American astronauts blasted off from the Moon in 1972, it seemed as if they were leaving behind monuments that would stand for all time. On a life
-
for she was lovely to look at.⏎As the story continues, pride creeps in.⏎Xerxes wanted to show off his beautiful wife! But it seems as though he didn’t give her too much thought until he had drunk too much!⏎
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 value of 22.3335. I need this value. Answer: 22.3335 34.55 Answer: 22.3335 34
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
typical limited partnership agreement may provide that a majority vote of the limited partners is necessary for the sale of assets or to remove a general partner. The partnership agreement determines whether the limited partners can vote on these matters.⏎If a limited partner assumes an active
-
populate our lives and inform our perceptions of the world. Explore these questions according to the expressions of artists from all over the world. Go beyond Western art to consider where art comes from and how it impacts all of us.⏎Contributors to this open
-
ethical caesura in the history of Europe, but there is no consensus when it comes to interpreting it. Some historians trace its origin to Hitler’s radical antis
-
. For example, a lake may have a good density of crappie, but few fish over 10 inches. The Preferred Rating tells an angler where to go to for a chance to catch
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
- L4:1254 excites
- L5:1807 suppresses
- L7:1368 excites
- L5:2256 suppresses
- L6:2313 excites
- L7:2058 excites
- L7:1795 suppresses
- L4:227 weakly suppresses
- L5:2026 weakly excites "to" after made or need-type verbs
- L4:730 weakly excites
- L6:502 weakly suppresses
- L5:2209 weakly excites
- L4:2051 weakly suppresses
- L7:2656 weakly excites
- L2:1601 weakly suppresses
- L5:1982 weakly excites
- L4:1474 weakly excites
- L3:2549 weakly suppresses
- L2:1963 weakly excites
- L5:2889 weakly excites
- L3:1390 weakly excites 'that' in 'so that' purpose clauses
- L7:2186 weakly suppresses
- L1:1857 weakly suppresses
- L7:1752 weakly suppresses
Feeds later layers
- L9:143 excites
- L17:2426 excites 'that' after nouns of belief or fact
- L14:2082 excites
- L9:288 excites
- L18:1389 excites
- L13:3 excites
- L14:313 excites
- L9:330 suppresses gap where a number was removed
- L9:75 weakly excites 'what' or 'how' after see or understand
- L21:2876 weakly excites
- L23:914 weakly excites
- L13:2243 weakly excites
- L10:707 weakly excites
- L18:579 weakly excites
- L25:1074 weakly suppresses
- L17:2718 weakly excites
- L19:1647 weakly excites
- L14:830 weakly suppresses
- L16:2210 weakly excites
- L27:925 weakly suppresses
- L16:948 weakly excites
- L19:2737 weakly excites
- L17:28 weakly excites
- L23:197 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
- that
- that
- что
- tô
- →that
- 责
- _that
- ไร
- しか
- 办
Suppresses
- pliers
- plier
- chool
- ounding
- inally
- endance
- opleft
- yscale
- oops
- erval