digits inside years like 19xx
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. 5 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 is weakly active on the middle digits of years written out in full, such as the '8' in 1986 or the '7' in 1972. Its strongest examples are only a little above the typical range, and turning it up gave scrambled text with no clear link to years.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
.⏎In the Congo-Kinshasa Province, the Sisters believe that protection of the environment is an essential and
-
aint-Laurent, known as Yves Saint Laurent was a French fashion designer, one of the greatest names in French fashion
-
s would have said racism didn’t exist, but the progress was palpable and undeniable.⏎In 1986, “The Cosby Show” was the most popular show on television, but
-
- THE CREATIVE TEAM⏎- THE CREATIVE ARTISTS⏎- IMPLEMENTING THE DESIGNS⏎- PRODUCTIONG MANAGEMENT / STAGE MANAGEMENT / TECHNICAL DIRECTION⏎- PRODUCTIONG CREWS⏎
-
Jerusalem on the shroud. The weave of the linen has also been linked with first-century Syria.⏎In 1988, three laboratories in Oxford, Zurich and Tucson each used radiocarbon dating
-
commemorating the electrification scheme in 192<|endoftext|>David Bourget (Western Ontario)⏎David Chalmers (ANU, NYU)⏎Rafael De Clercq⏎Jack Alan Reynolds⏎Learn more about PhilPapers⏎O
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 have a problem with the same old one. I have to think about it again. It is the same old problem with a new problem. The same old problem is something I have been thinking about. I think I have to think
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
were collected, studied, and enjoyed widely in the eighteenth century. Interest in collecting reproductive prints stimulated new developments in printmaking techniques, including the mezzotint process and color printing. In a domestic setting, prints might be assembled in a
-
with a curved circular path was framed by rhododendrons and a see-through vault of birch. Beyond the entrance gate, the nave began with a broad central path, flanked by ‘columns’ of ascending flowering cherries.⏎
-
bee. Covering his more than 20-year career, Bright Common Spikes: The Sculpture of John Bisbee, features 20 large-scale abstract sculptures of nails that are welded and forged into various organic configurations. The exhibition will
-
crowds that hadn't eaten in three days. I can still see Schim [sic], a wisp of a man, standing on the jeep silhouetted against the darkling sky quietly photographing
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
- L3:202 excites
- L3:366 excites
- L3:993 suppresses
- L3:99 excites
- L3:855 excites
- L3:1303 weakly suppresses
- L3:204 weakly suppresses
- L4:83 weakly excites
- L3:950 weakly excites
- L2:695 weakly excites
- L3:1380 weakly suppresses
- L3:550 weakly excites
- L2:1043 weakly suppresses
- L4:2326 weakly excites
- L4:1201 weakly excites
- L4:914 weakly excites
- L4:322 weakly excites
- L3:2988 weakly excites
- L1:1703 weakly excites
- L4:84 weakly excites
- L3:836 weakly suppresses
- L1:928 weakly suppresses
- L4:1793 weakly suppresses
- L0:1866 weakly excites
Feeds later layers
- L13:42 suppresses
- L12:33 suppresses
- L8:1803 excites
- L22:2520 excites
- L16:480 weakly suppresses
- L14:8 weakly suppresses
- L6:634 weakly excites
- L16:203 weakly suppresses
- L16:125 weakly suppresses unclear: mid-word and mid-phrase fragments
- L6:964 weakly excites
- L21:846 weakly excites
- L12:11 weakly suppresses
- L16:1308 weakly excites
- L8:436 weakly excites
- L16:444 weakly suppresses
- L6:1928 weakly excites
- L15:17 weakly suppresses
- L21:2466 weakly excites
- L6:1030 weakly excites
- L10:11 weakly suppresses
- L7:1029 weakly excites
- L24:804 weakly excites
- L17:2378 weakly excites 'yes' answering a question
- L6:18 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
- setC
- _attached
- Opens
- CAF
- flexDirection
- 嵴
- legalization
- caf
- Platforms
- 렛
Suppresses
- issant
- agan
- iss
- essim
- 利
- 团
- 相当于
- 联想
- issan
- 分