'the' before river names
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. 23 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 a river, or a body of water, is about to be named, such as 'join the Seine' or 'banks of the Mississippi'. It stays quiet on most other uses of 'the'. Turning it up gave odd text with no clear link to rivers.
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.
-
enters France northeast of Hirson, 9 miles (15 km) from its source, and flows generally southwestward, watering the Paris Basin, to join the Seine River at Conflans, after a course of 188
-
). It is also one of the “finest detours in France” network (Plus beaux détours de France). Stretching out along the banks of the River Ornain, it was, for many years, the capital of the
-
field, was the birthplace of modern Manchester.⏎It was established by 78 AD at a site overlooking the confluence of the rivers Irwell and Medlock.⏎This was a timber fort with earthen ramp
-
of Saint-Gobain, Compiègne, Hallatte, and Chantilly. Near Compiègne it receives its main tributary, the Aisne; and below the industrial town of Creil, it is joined
-
en gravure kaart] Schenckenschans, 1635.⏎Plan of the fortification of Schenkenschans, located between the rivers Rhine and Waal, including details on the landscape. During the Eighty
-
3, and 4 (teen team) will work between the remaining walls of a Roman villa on Poggio del Molino (Windmill Hill), overlooking the Mediterranean Sea. The villa was built at the start of the 1st century
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: how to test a real-time API that is only public in a private network only when you need to, and only when you know that you are only testing, and that you know that you are only testing. I thought that this might be
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Magnet or Ferromagnetic Material. The Stepper Motors are manufactured with steps per revolution of 12, 24, 72, 144, 180, and 200, resulting in stepping
-
that they were propelled primarily or entirely by sail. The length of the round ships ranged from 26 to 48 feet (8 to 14.7 meters), and they were between 8 and 16 feet (2
-
of them 10,000 km (6200 miles) away, and one 18,000 km (11,100 miles) from the blast. These stations monitor seismic waves, infr
-
said Lopez at the site, where the temperature reaches 45 C (113 F) at this time of year.⏎The first investigatory work was done
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
- L5:1546 excites
- L20:1308 suppresses the digit 3 inside years
- L20:361 excites
- L20:931 weakly suppresses the digit 6 inside years
- L1:1556 weakly excites
- L11:223 weakly suppresses
- L17:30 weakly suppresses
- L17:106 weakly excites
- L17:213 weakly excites
- L18:750 weakly suppresses
- L21:614 weakly excites
- L4:498 weakly excites
- L5:676 weakly excites
- L19:427 weakly suppresses
- L10:1349 weakly excites
- L16:177 weakly suppresses zero digit inside dates and codes
- L15:1452 weakly suppresses
- L2:2647 weakly suppresses
- L14:1541 weakly excites
- L18:242 weakly suppresses
- L16:75 weakly excites
- L19:1269 weakly suppresses the word 'about'
- L12:2888 weakly excites
- L20:3023 weakly excites
Feeds later layers
- L24:1968 strongly excites
- L25:2442 excites
- L24:2840 suppresses
- L26:1485 weakly excites
- L26:1221 weakly suppresses
- L24:1460 weakly suppresses
- L26:697 weakly excites
- L27:2417 weakly excites
- L27:1473 weakly suppresses
- L25:1974 weakly excites
- L24:2466 weakly suppresses
- L27:2151 weakly suppresses
- L26:1273 weakly suppresses
- L27:586 weakly excites
- L26:200 weakly excites
- L26:2915 weakly excites
- L25:1796 weakly excites
- L27:127 weakly suppresses
- L25:2345 weakly excites
- L27:2444 weakly suppresses
- L27:2852 weakly excites
- L26:1621 weakly suppresses
- L26:1253 weakly excites
- L25:1598 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
- odies
- 上的
- odem
- abel
- 呼和
- asses
- 上の
- asher
- 意义上
- thes
Suppresses
- leccion
- warts
- SPATH
- endar
- thought
- eguard
- prend
- ppelin
- umpy
- orias