'other' in 'and other ...' lists
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. 12 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 reacts to the word 'other' when it follows 'and' in a list, as in 'roads, buildings, and other infrastructure'. It is steadily strong on that word. Turning it up did not make the writing obviously about lists, though.
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.
-
silver plate, including crosses, chalices, and other spoils of the Church might be as much more. These movables were before
-
.⏎(ii) Similarly, the boycott of British institutions posed a problem. No alternative institutions were set-up in their . place. So, the students, teachers and other professionals finally had to resume their studies and jobs in government institutions like schools, colleges
-
ulates that rice husk ash could replace up to 20% of the cement typically mixed into concrete for the construction of skyscrapers, bridges, and other structures, dramatically reducing carbon dioxide. Every ton of cement manufactured for use in concrete
-
/French (N.Y. &N.F.) is the third supporting idea. * Amerindians forced to do unpaid labor to build roads, buildings, and other infrastructure in towns through the Encomienda System (S.W.) * The N
-
lowercase," they told us. There would be dire consequences, such as bad handwriting, nervousness, stuttering, and, in general, the self-defeating attitudes that often accompany the too-early introduction of formal
-
to schedule any kind of task you can imagine. It's useful for sending emails about system or program statistics, performing regular system maintenance, running backups, or doing any other task you can think of. There are similar programs on other operating
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: the time of the hour (or the hour) and the time of the hour (or the hour) and the time of the hour (or the hour) and the time of the hour (or the hour) and the time of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Writing⏎- HR Interview Questions⏎- Computer Glossary⏎- Who is Who⏎Demand & Elasticities⏎The 'Law Of Demand' states that, all other factors being equal, as the price of a good or service increases, consumer demand
-
of more/less entrants into the market, and the changing of consumer preferences. The shift can either be parallel or nonparallel.⏎Extension of Demand⏎Other things remaining constant, when more quantity is demanded at a lower price, it is called
-
vessel, although Inmarsat does provide broadband service for maritime communications using all 14 satellites. Regular terminals also cannot be used on aircraft.⏎Join thousands of others with our weekly newsletter⏎The 4th Era of IT Infrastructure: Supercon
-
could possibly test the hypothesis; and must document in the hypothesis just HOW the hypothesis will be tested. The tests or experiments must be rigorous enough, that there is a big chance that his or her
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
- L10:1219 suppresses
- L6:2654 suppresses
- L3:2029 weakly excites
- L10:2724 weakly excites
- L7:1714 weakly excites
- L11:706 weakly suppresses
- L6:1376 weakly excites
- L7:2691 weakly excites
- L3:2209 weakly excites
- L6:1977 weakly suppresses
- L3:1147 weakly suppresses
- L8:641 weakly excites
- L5:2356 weakly excites
- L9:557 weakly excites
- L5:2560 weakly excites
- L9:285 weakly suppresses
- L10:985 weakly suppresses
- L6:527 weakly suppresses
- L2:2088 weakly excites
- L8:1137 weakly suppresses
- L7:601 weakly excites
- L1:1302 weakly suppresses
- L5:1327 weakly excites
- L8:838 weakly excites
Feeds later layers
- L14:980 excites
- L14:1485 excites
- L18:1850 excites
- L19:1754 weakly excites
- L13:2409 weakly excites
- L15:2824 weakly excites
- L15:2393 weakly excites
- L27:922 weakly excites
- L13:2234 weakly excites
- L20:1708 weakly suppresses
- L13:1546 weakly excites
- L19:2121 weakly excites
- L13:2808 weakly excites
- L16:1212 weakly excites
- L15:2137 weakly excites
- L15:1372 weakly excites 'other' after 'any', 'all' or 'or'
- L17:2187 weakly excites 'etc' ending a list
- L15:609 weakly suppresses
- L14:2521 weakly suppresses
- L15:126 weakly excites
- L14:1821 weakly excites
- L20:2104 weakly suppresses
- L15:3069 weakly excites
- L15:2322 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
- other
- Other
- Other
- other
- 其他
- aside
- yên
- 诸
- 其它
- blend
Suppresses
- rese
- #w
- aurant
- isu
- wers
- ikt
- printed
- tere
- cus
- atis