words after comparisons and contrasts
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. 17 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 small connecting words like 'than', 'in', 'from' or 'the' that set up a comparison or contrast, such as 'cleaner than' or 'differs from'. Turning it up gave loose, unrelated text, so the steering results do not clearly reflect this.
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.
-
.⏎Quantum computing combines computer science with quantum physics. These super computers store their data using sub-atomic particles known as qubits, instead of silicone chips used in traditional machines. Silicon chips store data in binary code – a series of ones and
-
large distance from the Sun, 40 times further than the Earth. It is orbiting in a plane that is far away from the nearly common plane in which the other planets are orbiting (know as the ecliptic). Moreover,
-
of Edinburgh and Copenhagen to investigate if they could outperform conventional machines for specific applications.⏎They examined whether quantum computing’s noise rates would prevent the new systems outperforming classical computers on specific tasks.⏎Experts developed a technique to compare how classical algorithms and
-
’s Mouth Cleaner Than a Human Mouth? The general answer to this question is no, a dog’s mouth is not technically clean, and is not actually any cleaner than a human mouth.⏎Is saliva of dog harmful?⏎Recent case reports demonstrate the harms
-
. The bugs look quite similar and have the same life cycle, nesting areas, and impact on humans.⏎The primary thing that differentiates the tropical bed bug from the common bed bug is their habitat. While the common<|endoftext|>448 pp.,
-
all the other stars. The nearest star to our solar system is Proxima Centauri. The light of that star takes 4.25 years to reach Earth. That’s how long it takes light to travel 40,2
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 "the" is the "the" is the "the..." is the "..." is the "..." is "..." is the "..." is "..." is "..." is "..." is "..." is "..." is "..." is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
two major types of roots, fibrous roots and taproots. Taproots have one major root that grows very deep into the ground, while fibrous roots have many roots that grow in all directions.⏎Image: DonNichols/Getty Images⏎The
-
, exposing them to hazardous and toxic material. They face harassment and discrimination because of a negative perception of their work. Women and children are especially vulnerable to violence both within and outside their homes. Substance abuse is high and accessing affordable healthcare is a huge
-
64-bit or 128-bit encryption key that must be manually entered on wireless access points and devices and does not change. TKIP employs a per-packet key, meaning that it dynamically generates a new 128-bit
-
round-the-clock care that's given by professional health care providers such as nurses, therapists, or aides under a doctor's supervision.⏎- Intermediate care: Also provided by professional health care providers but on a less
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
- L16:427 excites
- L16:2812 weakly excites
- L13:1598 weakly excites
- L12:1142 weakly suppresses
- L10:1933 weakly suppresses
- L10:1453 weakly suppresses
- L10:2837 weakly suppresses
- L7:2594 weakly excites
- L16:2763 weakly suppresses
- L1:2882 weakly suppresses
- L15:1338 weakly suppresses
- L4:1760 weakly suppresses
- L10:2939 weakly excites
- L11:1288 weakly suppresses
- L12:1726 weakly suppresses
- L15:1869 weakly excites
- L11:2824 weakly suppresses
- L13:756 weakly suppresses
- L15:369 weakly excites
- L2:1259 weakly excites
- L2:1754 weakly excites
- L7:1261 weakly suppresses
- L0:1583 weakly suppresses
- L15:2613 weakly suppresses
Feeds later layers
- L18:439 strongly excites
- L18:257 strongly excites
- L18:206 strongly excites
- L18:667 strongly suppresses
- L18:2662 strongly suppresses
- L18:1088 strongly suppresses 'and' or 'the' joining two parallel items
- L19:522 suppresses
- L19:666 suppresses
- L20:165 excites
- L20:524 suppresses
- L18:362 suppresses
- L18:438 suppresses
- L18:255 suppresses
- L18:449 suppresses
- L19:2484 suppresses
- L18:301 suppresses
- L19:2897 suppresses
- L19:2298 suppresses
- L18:213 suppresses
- L19:519 suppresses
- L18:684 suppresses
- L20:294 excites
- L18:2953 suppresses
- L19:631 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
- ched
- itional
- zenie
- Williamson
- строй
- lement
- zburg
- exterity
- chts
- pection
Suppresses
- timing
- 稠
- 腊
- aure
- lique
- ủng
- uja
- sausage
- sequ
- oric