the zero inside a year
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. 22 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 second digit of a year, mostly the "0" in numbers like 2009, 2001 or 1901. When turned down, the model got stuck repeating the word "should", which matches what this neuron tends to suppress.
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.
-
: “There is no reason for a landscape ash tree to die from emerald ash borer anymore.”⏎The destructive insect, first detected here in 2009, has been found in Hennepin
-
nature of what – surface, slush or seas – lay beneath, Titan continues to serve up unending surprises and mysteries. Since the summer of 2009, the tilt of its orbit around Saturn has caused its northern and southern
-
The following article refers to President Bush's deliberations about allowing funding for stem cell research. On August 9, 2001, Bush gave his qualified approval for such funding, permitting it for research done
-
other regulations, websites or to contact your legislators.⏎ADA Anniversary Tool Kit includes: Overview of the Americans with Disabilities Act and the ADA Amendments Act of 2008; ADA Findings, Purpose, and History; The ADA from a
-
1%) as influenza B. Of the sub-typed influenza A viruses reported, 51.8% were influenza A(H1N1)2009 and 48% were influenza A(H3N2).⏎(Flu
-
shots from a .32-caliber Iver-Johnson revolver to assassinate William McKinley. He was shot on September 6, 1901 in the late afternoon by Leon Czolgosz during a public hand
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: Should I should I should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should should
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
at the bottom left of the screen and click on it. It should open up after clicking.⏎· Type HydroShare.org into the browser and click on it – it should open up.⏎·
-
the Outlook window, choose the transferred particulars folder.⏎- Then open the email that you want to recall. You must double-click to open the communication. The Recall option isn’t available for a communication displayed in the Reading Pane.⏎- Now from the
-
2: Watch the child try to stand up with help from the furniture from just before the end of their first year, and finally walk with some help. They will bring objects to you, having built their confidence enough to carry things and interact
-
0.⏎- Inside the loop, we are printing the element for index i and decrementing the value of i by 1.⏎If you run this program, it will print the below output:⏎5 4 3 2 1⏎
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
- L20:2748 strongly suppresses
- L19:2624 strongly suppresses
- L21:2363 excites
- L21:2025 suppresses
- L19:2748 suppresses
- L18:2120 suppresses
- L21:2121 excites
- L21:1826 weakly excites
- L15:2327 weakly suppresses
- L7:1963 weakly excites
- L11:1372 weakly excites
- L21:1913 weakly suppresses
- L19:1678 weakly excites
- L18:2959 weakly excites
- L2:611 weakly suppresses
- L19:1700 weakly suppresses
- L16:2197 weakly excites
- L18:2036 weakly excites
- L5:2380 weakly suppresses
- L15:1468 weakly suppresses
- L17:1493 weakly suppresses
- L15:2076 weakly suppresses
- L4:2231 weakly suppresses
- L20:3048 weakly excites
Feeds later layers
- L27:1236 strongly suppresses
- L24:764 strongly suppresses
- L23:1876 suppresses
- L24:2481 suppresses
- L24:1897 suppresses
- L27:2484 suppresses
- L25:959 weakly suppresses
- L25:1181 weakly suppresses
- L23:2700 weakly suppresses
- L23:1173 weakly suppresses
- L25:2905 weakly suppresses
- L27:1867 weakly excites
- L26:1933 weakly suppresses
- L24:1571 weakly excites
- L23:1993 weakly suppresses
- L26:1276 weakly suppresses
- L27:2213 weakly excites
- L26:1494 weakly excites
- L24:471 weakly excites
- L25:2802 weakly excites
- L23:2104 weakly excites
- L25:695 weakly excites
- L25:2145 weakly suppresses
- L26:2014 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
- sätze
- meet
- apixel
- atur
- met
- uzzi
- él
- meet
- omet
- 样子
Suppresses
- should
- should
- 应该
- Should
- 应
- shouldn
- 应当
- Should
- .should
- 應該