"that" before a need or must clause
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. 21 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
The neuron fires on "that" after a noun when the description that follows says something is needed, required or must be handled, like "issues that are common" or "a risk that needs to be managed." Turned up, it produced a stream of "need to have to", which fits its job of anticipating that kind of wording.
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.
-
for sale.⏎- Land auctions.⏎- Local authorities who may have parcels of land they are prepared to sell for development purposes.⏎- Utility companies may have surplus land that is for sale
-
to ensure that the symptoms are not being caused by other types of stressors or conditions. This way, the patient can be sure that there are no underlying causes that could become worse.⏎Ignoring the symptoms associated with dizziness and blurry vision can be
-
lorinated biphenyl (PCB). While these compounds have helped to protect the personal health and safety of humans, they create risks and concerns to the environment that are constantly being researched.”⏎An investigative study by the Associated Press in 20
-
material heating to its ignition temperature. Since the balance of oxygen flow and heat loss needs to be carefully balanced, spontaneous ignition is not common but is still a risk that needs to be managed Some businesses have had unsafe work practices for many years before they
-
Every data-driven trend comes with vulnerabilities that, if left unchecked and unprotected, can cause great damage.⏎In such cases,
-
and presented in the Global Report on Communication for Empowerment (2010). The findings from each of the five pilot countries reveal a number of issues that are common to more than one pilot country, some highlighting emerging trends, and others
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 am going in the 205.50.50.55.5.5.5.5.5.5.5.5.5.5.5.5.5.5.5
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
tools for electrical engineers, technicians and hobbyists. They help to visualize how a circuit works and provide essential information about its components and interconnections. But understanding these diagrams isn't always easy. To get the most out of them, it pays to
-
that it uses so many items from recycle bin!⏎Guitar craft is a great alternative to purchasing low-quality plastic toy guitars. And cheaper for sure! All you need are few simple items you probably already have at home.⏎Make an amazing ABC DIY
-
added. … Baking powder is a combination of sodium bicarbonate and powdered acid (like cream of tartar). Because it already contains an acid, all it needs is moisture and heat added to activate it.⏎Can vinegar and baking soda be mixed
-
. I cover more reasons in this post on why seeds are better than buying seedlings.⏎Seed starting can sound intimidating! All the supplies and tools you think you need, when and where to start seeds and how to take care of them until they
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
- L17:2255 excites
- L19:2648 suppresses
- L18:522 suppresses
- L18:1162 suppresses
- L20:1015 suppresses
- L19:1298 suppresses
- L17:2087 suppresses
- L3:1914 suppresses
- L19:945 weakly suppresses
- L20:2035 weakly suppresses
- L12:1678 weakly excites
- L17:257 weakly suppresses
- L15:2624 weakly suppresses
- L20:1205 weakly suppresses
- L20:386 weakly suppresses
- L18:2301 weakly excites
- L17:2643 weakly suppresses
- L19:2676 weakly excites
- L19:2168 weakly suppresses
- L17:2818 weakly suppresses
- L19:1681 weakly excites
- L18:2474 weakly excites
- L14:2208 weakly suppresses
- L18:2719 weakly excites
Feeds later layers
- L24:764 strongly excites
- L23:1876 strongly excites
- L23:1500 excites
- L25:1678 excites
- L22:1422 excites
- L27:1123 excites
- L24:1041 excites
- L26:1276 excites
- L22:2551 excites
- L24:2481 excites
- L24:2268 excites
- L23:103 suppresses
- L23:1542 excites
- L22:2891 excites
- L24:1571 excites
- L27:1493 excites
- L22:2322 excites
- L26:1955 weakly excites
- L24:2503 weakly suppresses
- L25:1181 weakly excites
- L22:2148 weakly excites
- L25:1211 weakly excites
- L22:3032 weakly excites
- L24:2784 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
- need
- 需要
- 需
- 必须
- needs
- needs
- must
- must
- need
- 需要用
Suppresses
- omite
- <Type
- anesthesia
- 新闻网
- KeyName
- enna
- ocre
- rew
- درجة
- **)