unclear: word stems in technical text
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. 15 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 the middle pieces of longer technical or specialist words, such as 'roasting', 'math', 'orthodontist' and 'writing', in informational articles. It goes negative on words like 'only' and 'this' after 'to do' or 'can'. Turning it up did not produce a recognisable theme in the model's replies.
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.
-
, with thousands of compounds being created and broken down. Green coffee is loaded into the heated roasting chamber and undergoes drying as the first stage of the roasting cycle. During the first few minutes, the beans desiccate most intensively
-
crawl, stand and eventually walk.⏎Why is tummy time important?⏎Infants spend a lot of time lying on their back, especially a few weeks after being born. If they’re not lying on their backs, they are being carried or sw
-
to Graphs⏎A formal introduction to graphs may be found in most discrete math textbooks. The purpose of this section is not to provide such an introduction. Rather
-
/dev and /devices directories in response to dynamic reconfiguration events or file system accesses are handled by devfsadmd, the daemon version of the devfsadm command. This daemon is started by the service management facility when a system is bo
-
on a TV monitor) to perform the surgical procedure. These laparoscopic procedures offer benefits to the conventional open surgery including:⏎- Reduced Risk of Infection⏎Single Incision laparoscopic surgery
-
Care of Your Teeth with Braces⏎When you have braces, you'll have to be very good about following the brushing and flossing instructions your orthodontist gives you.⏎You can't eat hard, sticky or gooey foods such as
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: to have a set of information, I need to know the information, I need to know that, I need to know that, I need to know that, I need to know that, I need to know that... (the information,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ism. We must remind them of what they knew as toddlers: mistakes are a valuable part of the learning process. Innovation and discovery can only be fostered if we give our children the freedom to fail. Schools can build
-
in the right hands. It can determine the age, position and heart rate of the foetus, along with an indication of its overall health. But to do this consistently can take many years of practice.⏎Meanwhile, in developing countries, “a
-
kinds of deception are supported by visual information?⏎These are the types of questions we are attempting to answer as we build our typology of visual misinformation. To do so, we are using the data from the CrossCheck French election project and First Draft
-
<= R5+1; R6 <= R6-R3 would be held up in a non-scheduled pipeline, because the second operation based on R2 has to wait for the load from memory into R1, thus halting the CPU
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
- L7:2637 suppresses
- L12:1522 excites comma after a 'before...' clause
- L8:2965 suppresses
- L13:2948 suppresses
- L11:2506 excites
- L12:3018 excites
- L13:487 weakly excites
- L14:835 weakly suppresses
- L11:1234 weakly excites
- L11:2107 weakly suppresses
- L12:1496 weakly suppresses
- L13:1256 weakly suppresses
- L11:1986 weakly suppresses
- L14:3019 weakly excites
- L14:2633 weakly suppresses
- L12:2909 weakly suppresses
- L12:268 weakly excites
- L4:195 weakly excites
- L11:1825 weakly excites
- L8:1006 weakly suppresses
- L14:174 weakly excites
- L11:2004 weakly suppresses
- L10:2681 weakly excites
- L13:533 weakly excites
Feeds later layers
- L20:2888 suppresses
- L20:2529 suppresses
- L18:2494 suppresses
- L18:96 suppresses
- L19:2249 suppresses
- L17:1347 suppresses
- L17:2521 suppresses
- L17:863 suppresses
- L18:1563 suppresses
- L18:2441 suppresses
- L16:3061 suppresses
- L16:2788 suppresses
- L17:381 suppresses
- L18:1932 suppresses
- L16:2284 suppresses
- L18:2514 suppresses
- L20:2134 suppresses
- L24:1041 suppresses
- L17:1113 suppresses
- L22:1336 suppresses
- L16:514 weakly suppresses
- L18:1003 weakly suppresses
- L19:2139 weakly suppresses
- L21:2805 weakly suppresses 'the' right before 'subject' or 'source'
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
- aurant
- otify
- anguard
- inally
- ummings
- ngạc
- etCode
- ington
- .getResult
- .HttpServletResponse
Suppresses
- 的前提
- requires
- require
- requires
- require
- Requires
- 都需要
- Require
- 需要
- prerequisites