"that" starting a relative 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 lights up on the word "that" when it introduces a describing clause after a noun, like "the parts of the brain that control breathing." When it was turned up, the model fell into repeating "is" over and over.
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.
-
the decimal and cents to the right of the decimal. To add decimals, simply line up the decimal points of the amounts of money that are being added. When you have added the amounts of money, be sure to
-
Center expect these cancer vaccines first will be used as supplements to initial treatment plans that use surgery, chemotherapy or radiation. The vaccine then will sweep out any cancer cells that remain in the body.⏎“Hypothetically, once the immune system has
-
not need an exact number, you can⏎round a number to make things simple.⏎Here are rules for rounding numbers⏎to the nearest ten.⏎Make the numbers that end in 1⏎through 4 into the next lower number that ends in
-
the herniation occurrs. Without treatment, death is likely.⏎A brain herniation often causes a massive stroke. There can be damage to parts of the brain that control breathing and blood flow. This can rapidly lead to death or brain death.⏎
-
before you start adding any additional chemicals to ensure active chlorine production is optimised.⏎- Put all your pool equipment e.g. hoses, poles and any other accessories that may be infected, in the shallowest part of your pool so that when you
-
a scarce resource,” he says. “So, by using the wastewater you’re really repurposing something that you’d otherwise pay to be processed and converting<|endoftext|>Career and technical education courses utilize
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: What if the new government which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which which
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
problem?⏎Our tutors have indicated that to solve this problem you will need to apply the Conjugate Acids and Bases concept. If you
-
gone up since then. Then they adjust the gross spending numbers by this (“deflator”). This is good for the government because it is lower than what we believe to be the actual price inflation rate and makes GDP look better than it really is.
-
ervention. It therefore avoids wastage of resources. The objective was to compare the number of CT-scans done for minor head injury as compared to the number that would have been done if the CCHR was applied. Methods:
-
ruins of both these temple complexes cover a considerable area and are still very impressive. Nothing remains, however, of the houses, markets, palaces and gardens that must have surrounded the temples in ancient times. The principal feature in Egyptian social centers,
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
- L17:1361 suppresses 'that' after verbs like recommend
- L18:960 excites
- L19:1703 excites
- L20:1000 suppresses
- L18:491 excites
- L18:2031 suppresses
- L20:349 excites
- L18:1389 excites
- L16:2975 suppresses
- L19:622 suppresses
- L17:227 excites
- L18:2787 excites
- L19:547 weakly suppresses
- L16:948 weakly suppresses
- L18:1513 weakly excites
- L20:636 weakly suppresses
- L17:137 weakly excites
- L19:1115 weakly suppresses
- L9:865 weakly suppresses
- L15:1834 weakly excites
- L8:199 weakly excites
- L2:205 weakly suppresses
- L13:344 weakly suppresses
- L19:1175 weakly suppresses
Feeds later layers
- L23:30 strongly suppresses
- L22:228 strongly excites
- L23:127 strongly suppresses
- L22:479 strongly excites
- L23:320 excites
- L25:9 suppresses
- L26:669 excites
- L22:518 excites
- L26:53 suppresses
- L23:1013 excites
- L26:318 excites
- L26:467 excites
- L27:604 excites
- L25:698 excites
- L27:804 excites
- L25:586 excites
- L27:321 excites
- L23:220 excites
- L24:241 suppresses
- L27:382 excites
- L24:252 excites
- L27:565 excites
- L24:1439 suppresses
- L27:461 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
- is
- are
- could
- was
- may
- has
- 是一
- can
- includes
- were
Suppresses
- которое
- which
- whom
- Which
- Which
- which
- velopment
- who
- 哪一个
- 哪