'the' in 'what are the' questions
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 'the' in questions that start with 'what are the' or 'what's the', just before the thing being asked about. Turning it up did not produce a clearly related change in the model's 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.
-
. Do you grab a disposable one or do you get one that you can recharge? For that matter, why can’t you recharge all of them? What’s the difference? That, turns out, is a matter of science and cost.⏎How
-
V. That’s 10% of the intended voltage!⏎So what are the traditional options? Well, the best one is to have a separate set of wires
-
?⏎- Kidney stones⏎- Polycystic kidney disease⏎- Diabetic Nephropathy⏎- Renal tubular acidosis⏎What are the symptoms of
-
system?⏎8. Why are program counter and stack pointer 16-bit registers?⏎9. What does it mean by embedded system?⏎10. What are the different addressing modes in 8085?⏎11
-
Best guess: the water rushing over the microphone of the hearing aid makes a loud rushing sound.⏎If I can’t swim with my waterproof hearing aids, what’s the point?⏎Your waterproof hearing aid is more durable in the case of accidental water sub
-
most frustrating behaviors a dog owner can experience. It not only disturbs your sleep but also gives you the sense that the dog is not comfortable. So what are the reasons why dogs growl?⏎Here’s The Sort Answer To Why Your Dog Grow
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 need to be a student of a "simpler" student, who can write a small, formal letter. This letter, as a student of a "simpler" student, who can write a small, formal letter,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the first generation of transgenics came from observing the way that agrobacterium infects plants
-
for anything;⏎There is no such a thing as a free lunch, someone has paid for it. Government doesn’t have money of its
-
is often diagnosed during a visit to your doctor. Blood pressure is measured using a cuff around your arm and a device called a sphygmomanometer. Your doctor may ask you to sit quietly for five
-
and silica), is connected to the wheel and/or the axle. To stop the wheel, friction material in the form of brake pads (mounted on a device called a brake caliper) is forced mechanically, hydraulically, pneumatically or electrom
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
- L5:125 excites "of" after all three, both, ten
- L5:1372 excites
- L4:128 suppresses
- L2:2172 suppresses
- L4:52 weakly excites
- L3:615 weakly excites
- L3:159 weakly excites
- L5:1747 weakly suppresses
- L4:812 weakly suppresses
- L2:1126 weakly suppresses
- L4:2786 weakly excites
- L4:1743 weakly suppresses
- L4:2589 weakly suppresses
- L5:2649 weakly excites
- L3:1456 weakly suppresses
- L2:2250 weakly excites
- L1:1147 weakly excites
- L4:731 weakly excites
- L1:1967 weakly excites
- L5:2619 weakly excites
- L3:585 weakly suppresses
- L2:516 weakly suppresses 'had' after has or have
- L3:473 weakly suppresses
- L4:2365 weakly suppresses
Feeds later layers
- L14:727 excites
- L9:459 weakly suppresses
- L7:15 weakly excites
- L10:272 weakly excites
- L24:1887 weakly excites
- L25:655 weakly excites
- L16:485 weakly suppresses
- L10:459 weakly suppresses
- L8:1003 weakly suppresses "the" or "of" before a noun phrase
- L7:135 weakly suppresses 'of' after 'issue' in citations
- L24:108 weakly excites
- L9:266 weakly suppresses
- L18:264 weakly suppresses
- L18:370 weakly suppresses
- L20:321 weakly excites
- L27:1696 weakly suppresses
- L26:981 weakly suppresses
- L20:1314 weakly suppresses
- L14:2955 weakly excites
- L10:1132 weakly excites
- L20:229 weakly excites
- L22:237 weakly excites
- L14:672 weakly suppresses
- L16:2207 weakly suppresses
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
- iology
- gings
- 把手
- ник
- 要说
- ToShow
- otypes
- forc
- iggers
- śl
Suppresses
- ourn
- lưới
- isted
- ently
- -mask
- così
- ATE
- adan
- apsed
- ivered