'could' after a noun subject
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. 3 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 "could" when it follows a noun that is the subject of a clause, such as "the camera could" or "citizens could". Turned up, the model moved toward news and quiz-style text.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
to power outlets meant that the camera could be plugged in directly in the wall outlet in the different measurement rooms.⏎The light
-
arnessed as an energy source. Researchers had believed that keeping the donor and acceptor layers as pure as possible was the best way to ensure that the excitons could travel unimpeded, so that solar cells could capture the maximum amount of energy
-
⏎or a theoretical ideal -it is a fact. Of it a statesman has said: 'They⏎had demonstrated to the world in actual practice that difficulties could⏎be resolved by discussion as they could not be resolved by force. Moreover,⏎
-
with endurance and resistance exercise -- to preserve lean body mass<|endoftext|>Scientists Engineer An Opioid That May Reduce Pain With Less Risk⏎Once people realized that opioid drugs could cause addiction and deadly
-
aufort , Hyde , Craven and Carteret .⏎The primary reason for establishing counties (or precincts) was to provide local seats of government where citizens could record documents, such as deeds or wills, and participate in court proceedings.
-
regeneration process in his book, The Body Electric. Dr. Royal Rife’s revolutionary work in the 1920’s documented and demonstrated how precise frequencies could destroy specific organisms such as cancer cells and viruses, while others could be used to
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: what is a new study that says that the new study has found that the first thing that has been thought about is something that says that the first thing that has been thought about something in the past that has been in the past. What is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Luke 9:58 German Bible<|endoftext|>Both her and Romeo were blinded to everything except each other and did not think before doing, for example, Juliet should have ensured that Romeo found out her plan, before she did it. She loved Romeo
-
By 1738 Albemarle and Bath Counties had been dissolved and the 14 precincts then in existence became counties, a designation that has remained since the seventeenth
-
improved medical care and better nutrition will have played their part leading up to the 1950s, but after this time there were some specific factors that have contributed to not only better chances of survival post birth, but also the introduction of
-
towns there was an organized community life and Jewish public institutions. Nitra, a major administrative and economic center in the Middle Ages, had a longstanding Jewish community that is considered one of the oldest in Slovakia. A document from the year 11
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
- L2:240 excites
- L1:2801 weakly excites
- L0:2025 weakly suppresses
- L2:1737 weakly excites
- L2:6 weakly excites
- L2:1321 weakly suppresses
- L1:2814 weakly excites
- L2:41 weakly suppresses
- L2:2115 weakly excites
- L2:209 weakly suppresses 'also' after 'not only ... but'
- L0:2544 weakly suppresses
- L1:146 weakly excites
- L2:1899 weakly excites
- L2:1302 weakly suppresses
- L0:840 weakly suppresses
- L1:391 weakly excites
- L1:83 weakly excites
- L2:2037 weakly suppresses the digit 5 in decimals like 1.25
- L2:497 weakly suppresses
- L1:1340 weakly suppresses
- L0:810 weakly excites
- L0:1217 weakly suppresses
- L2:1784 weakly suppresses
- L0:1087 weakly suppresses
Feeds later layers
- L7:153 excites
- L5:536 suppresses
- L4:110 excites
- L8:2719 suppresses
- L4:124 excites
- L17:2060 excites
- L6:669 excites
- L4:113 excites
- L27:2982 excites
- L15:2901 excites
- L4:1294 weakly suppresses
- L4:457 weakly excites
- L11:425 weakly excites
- L7:734 weakly suppresses
- L16:2202 weakly excites
- L8:2738 weakly suppresses
- L11:2896 weakly suppresses
- L11:731 weakly excites
- L6:535 weakly suppresses
- L25:1861 weakly excites
- L9:1550 weakly excites 'to' after 'matters' or 'means' phrases
- L5:2231 weakly excites
- L14:2949 weakly excites
- L13:2472 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
- instruments
- →TokenName
- typingsSlinky
- stdafx
- leggings
- sourceMappingURL
- swore
- icons
- 出路
- goggles
Suppresses
- uted
- rium
- etration
- uta
- eten
- 례
- Citation
- แล
- acted
- 质