the word 'the' after a lead-in phrase
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. 13 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' that starts a noun phrase after a short opening, such as 'what's the story' or 'in the school'. Pushing it up or down made the model's replies repetitive or off-topic.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
<|endoftext|>FROM ISSUE NUMBER 15 ~ SPRING 2013 GO TO TABLE OF CONTENTS⏎The Art of Limited Government⏎
-
shoulder dress.⏎So, what’s the story behind this elusive spray-on garment?⏎Who created the dress?⏎The dress technology
-
experience one or more of those problems.⏎Even five years ago, the increase in morbid obesity was noticed by pediatricians to a point that Dr.
-
said.⏎- They throw out food in damaged or outdated promotional packaging (think holiday cookies) that is still edible.⏎Farms aren’t excluded from the report’s waste list.⏎The NDRC found that growers often overplant and have
-
the current one occurred in both cases.⏎“We estimate that the probability to have such a heat or higher is generally more than two times higher today
-
women experience domestic violence⏎Two girls that Kopila supported are in the school⏎Introduction KOPILA- Nepal: ‘kopila’ means flower bud
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 not, but I have not. The problem is, I have not. Maybe that's a problem for me to share, but the thing is, the something for example, the something for example, the problem, but the something
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
str, 'p', strlen(str));⏎printf ("'p' found at position %d.\n", pch-str+1);⏎printf ("'p' not found.\n");⏎- Compare two blocks of memory (function⏎- Locate first
-
for design projects. Projects are developed from concepts learned in Wood and Metal Design, Interior Design and/or Intro to Architecture. New technologies such as a CNC Router, CNC Plasma CAM, laser
-
iasens_), etc.<|endoftext|>The American Meteorological Society⏎) promotes the development and dissemination of information and education on the atmospheric⏎Atmospheric sciences is an umbrella term for the study of the atmosphere, its processes, the effects other systems have
-
time.⏎You should give examples from dynamic typed languages instead, e.g. clojure.⏎Here's one example:<|endoftext|>The commerce of the ancient world was
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
- L11:43 strongly excites
- L8:7 strongly suppresses
- L8:10 strongly suppresses
- L8:4 strongly suppresses
- L8:16 excites
- L8:3 suppresses
- L8:14 suppresses
- L8:13 excites
- L11:2 excites unclear: blank before a year or number
- L5:110 excites
- L7:2 suppresses
- L7:122 suppresses
- L5:50 suppresses
- L5:84 excites
- L6:7 excites
- L6:12 suppresses
- L9:8 suppresses
- L0:1301 excites
- L6:3 suppresses
- L0:1224 excites
- L6:13 suppresses
- L8:5 excites
- L5:158 excites
- L7:110 excites
Feeds later layers
- L15:351 excites
- L14:1335 suppresses
- L14:1254 weakly suppresses
- L14:1896 weakly suppresses
- L14:2039 weakly suppresses
- L14:1572 weakly suppresses
- L15:1141 weakly suppresses
- L14:933 weakly suppresses
- L14:423 weakly excites
- L18:1366 weakly suppresses
- L18:688 weakly excites
- L14:2991 weakly suppresses
- L14:2372 weakly suppresses
- L18:2956 weakly suppresses
- L14:97 weakly suppresses
- L15:1537 weakly excites
- L23:638 weakly suppresses
- L14:644 weakly suppresses
- L15:2425 weakly suppresses
- L15:1430 weakly excites
- L15:611 weakly suppresses
- L19:1444 weakly suppresses
- L18:847 weakly suppresses
- L17:3043 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
- engeance
- MOZ
- meld
- setResult
- étr
- 更新
- ining
- simult
- empre
- .ma
Suppresses
- individual
- atório
- цию
- ocab
- читыва
- вор
- 级别的
- utilus
- acente
- ...]⏎⏎