'the' before a body part or organ
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
This neuron fires on the word 'the' when the next word is likely to name a part of the body, such as 'the lung', 'the liver', 'the neck' or 'the arteries'. Turning it up or down produced rambling text with no obvious body theme, so the steering results do not confirm the pattern.
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.
-
unborn fetus.⏎SEE ALSO: 10 Ways Mercury is Toxic to the<|endoftext|>Featured Photo: Crystalline caffeine. (Credit: Annie Cavan
-
to offer them.⏎There’s only scant data on how many patients who recover from Covid-19 are left with long-term fibrosis, or scarring of the lung. But studies on SARS and MERS, relatives of the novel coronavirus
-
have the option of joining the expedition for two weeks.⏎HOW WILL YOU HELP⏎You<|endoftext|>Adding and formatting mathematical equations in your papers can be a real pain in the neck. This is especially the case with complicated expressions involving matrices, calculus, bin
-
global temperatures rise and extreme heat waves become more common, so does heat illness. But while doctors have known for decades about the adverse effects of some medications on how the body regulates temperature, little is being done in the U.S. to better warn
-
appetite in the elderly.⏎- Hepatitis or chronic liver disease. Loss of appetite is known as one of the first symptoms associated with both hepatitis, or inflammation of the liver, and chronic liver disease.⏎- Depression. Psychological issues like depression,
-
life-threatening invasive cancer.⏎The researchers investigated a common type of early-stage breast cancer called ductal carcinoma in situ (DCIS). DCIS tumors are confined to the milk duct, and, though small, they are detectable by mammography.
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 two people the two people the two the two the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
shall be done with a soft but strong measuring tape just behind the shoulders. The person measuring shall always stand on the left side of the sheep, slide the end of the tape down the right side, pull it up the left side and pull the
-
to make your ball bounce down a path or trail really quick, just add a Wind force field (Shift+A -> Force Field -> Wind). Then set the strength of that force field to a point where you can see your ball move in the direction
-
raised by a pull of 100 lb on the free end of the rope. To raise the load 10 ft, however, the free end of the rope must be pulled 30 ft.⏎Disregarding friction, work
-
ave. A single sheave mounted in a block and fixed in place simply changes the direction of force exerted on the rope passing over it. If the end of the rope that ordinarily would attach to the load is passed around a second, unf
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
- L20:47 excites
- L20:2520 excites
- L19:768 excites
- L20:837 suppresses
- L19:1361 excites
- L20:1494 excites
- L19:1817 excites
- L19:1181 excites
- L20:469 excites
- L18:1342 suppresses
- L19:656 suppresses
- L20:1972 suppresses 'of' in 'of course'
- L8:636 suppresses
- L19:2283 suppresses
- L19:1953 excites
- L20:219 excites "the" after time or place prepositions
- L19:2306 excites
- L19:685 excites
- L19:1056 excites
- L19:1785 weakly excites
- L20:762 weakly excites
- L17:429 weakly excites
- L18:1691 weakly suppresses
- L19:705 weakly suppresses
Feeds later layers
- L27:127 strongly excites
- L26:507 strongly suppresses
- L27:399 strongly suppresses
- L24:49 excites
- L23:94 excites
- L22:199 suppresses
- L25:2531 suppresses
- L22:171 weakly suppresses
- L27:994 weakly excites
- L27:121 weakly excites
- L26:251 weakly excites
- L25:1796 weakly suppresses
- L26:2278 weakly excites
- L25:2497 weakly suppresses
- L24:2538 weakly excites
- L27:2537 weakly excites
- L27:1464 weakly excites
- L27:1382 weakly suppresses
- L24:1050 weakly suppresses
- L23:1011 weakly excites
- L25:530 weakly suppresses
- L22:1589 weakly suppresses
- L24:837 weakly suppresses
- L24:1366 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
- ".$_
- zich
- 己
- {_
- thereof
- .jetbrains
- -neck
- HomeController
- 在家里
- 在那里
Suppresses
- the
- the
- The
- _the
- →the
- The
- .the
- THE
- -the
- ,the