the letter 'l' inside words
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. 26 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 letter l when it appears in the middle or start of a word, such as in lazaretto, Lorikeet or Llangynhafal. Turned up, the model produced repetitive, garbled text, so the steering output does not show a clear theme.
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.
-
of St Basil originally selected a name for their leper hospital. The name eventually gave rise to the terms "lazar", meaning a leper, and "lazaretto", a leprosarium, or Leper House. The
-
⏎Lorikeet feeding habits⏎Lorikeets are brightly coloured parrots of the family Psittacidae and sub-family Loriinae (l
-
had gradually swell In his mind. The chronological accounts testify his previous encounters from the very beginning of the narrative “My first mistress was a woman of… To “l had resided but a short time in Baltimore” to “Very soon after I
-
a restaurant for leftovers or as a takeout container from a fast food place⏎- egg cartons – either the styrofoam ones or cardboard ones (latter is preferable as you can plant these in the ground and they will rot away
-
, having outlived the other apostles — all of whom, according to legend/history<|endoftext|>The
-
ancestors. This work of art belongs to the old time of Koh Kok Thlork.⏎After the year 1979, the authority of the Take
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: bothogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogogog
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Equine Granulocytic Anaplasmosis (Ehrlichiosis)⏎|Deer (or Black-Legged) Tick⏎||Tear drop-shaped; the size of a pop
-
. With time we check more numbers and become more convinced that the conjecture is true. Currently, it has been checked up to the power 4.6 ⋅<|endoftext|>Picture Yourself in Orbit: Are You Really Weightless?⏎An object
-
0.97. The expected number of powers of 2 without zeroes starting from the power N is cxN/(1-x) ≈ 40 ⋅ 0.97N.⏎Let us look at A00
-
secretpassword';⏎mysql> GRANT ALL PRIVILEGES ON database_name.* TO user_name@localhost;⏎mysql> FLUSH PRIVILEGES;⏎Step 5: Installation and Configuration of Drupal⏎Open
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
- L19:322 strongly excites
- L17:134 strongly suppresses
- L21:325 excites
- L18:420 excites
- L20:1710 suppresses
- L16:117 suppresses
- L0:94 excites
- L2:1518 suppresses
- L19:1639 excites
- L23:1760 suppresses
- L22:3039 excites
- L6:2953 excites
- L24:2164 excites
- L5:1294 excites
- L16:55 suppresses
- L25:1316 suppresses
- L0:662 suppresses
- L2:1866 excites
- L1:23 excites
- L23:2074 suppresses
- L16:177 excites zero digit inside dates and codes
- L10:1349 weakly excites
- L7:904 weakly excites
- L1:31 weakly suppresses
Feeds later layers
- L27:1827 weakly suppresses
- L27:1602 weakly suppresses
- L27:2749 weakly suppresses
- L27:2772 weakly suppresses
- L27:2906 weakly suppresses the letter r in the middle of words
- L27:1897 weakly excites
- L27:914 weakly excites
- L27:2899 weakly suppresses
- L27:1528 weakly suppresses
- L27:2550 weakly suppresses
- L27:698 weakly suppresses
- L27:1084 weakly excites
- L27:2962 weakly suppresses
- L27:2421 weakly excites
- L27:556 weakly excites
- L27:1164 weakly excites
- L27:2777 weakly suppresses
- L27:1604 weakly suppresses
- L27:2337 weakly excites
- L27:418 weakly excites
- L27:1250 weakly suppresses
- L27:2618 weakly excites
- L27:1265 weakly excites
- L27:3011 weakly suppresses
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
- al
- alink
- 恰恰
- 至於
- enting
- 每人
- 用微信
- quello
- 煦
- indemn
Suppresses
- og
- ine
- oad
- ess
- odash
- ike
- iations
- ий
- ogh
- igkeit