'least' in 'at least'
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 lights up on the word 'least' when it follows 'at' in the phrase 'at least'. Turning it up or down made the model wander into unrelated text.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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 your dreams.⏎Different industry sectors demand different employability skills – personal as well as technical. Most employer surveys do, however, note that personal skills have at least equal rating with technical skills or qualifications for example;⏎In the Confederation of British
-
ancient DNA from old bones, are revealing some of that lost history.⏎We now suspect, for example, that the bacterium has been infecting people since at least the Neolithic period.⏎Last year, scientists revealed that a Stone Age hunter-g
-
Others feel it too, and send you back more of the same.⏎SO…if you send out a sincere wish to help each other get through this or at least agree to disagree, people feel that too, and it can shift their state.
-
no way a<|endoftext|>If you’ve been around these few months, chances are good that you at least have a passing knowledge of what cryptocurrencies are. You may even have heard of names
-
via hands and surfaces*. But as there is no evidence that I know of that supports the fomite route for COVID-19, and we now have at least some (relatively weak) evidence against it, then it is surprising how much
-
. The tribade was the most common (vulgar) lesbian in European texts for many centuries. ?Tribade? occurs in English texts from at least as early as 1601 to at least as late as the mid
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 other people are busy with their hands, but with the limitations of time and creativity. Oh, I don't have time for this. It's a very common problem. Maybe the people are just getting confused, so they need to get
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the direction of which is reversed at regular intervals or cycles. Electricity transmission networks use AC because voltage can be controlled with relative ease.⏎Amorphous Silicon — A thin-film, silicon photovoltaic cell having no crystalline structure. Manufactured
-
In practical terms, the exact meaning of "surface" will vary according to the measurement method used. A satellite infrared radiometer indirectly measures the temperature of a very thin layer (about 10 micrometres thick) or skin of the
-
uses in everyday life is an intriguing work of art combined with science. Fiber optics has been fabricated from materials that transmit light and are made from a bundle of very thin glass or plastic fibers enclosed in a tube. One end is at a source of
-
ub that grows up to 900 mm, but some forms may reach a height of up to 2 m. It has very distinctive, long, thin, needle-like leaves that curve upwards, giving the plants the appearance of small pine
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
- L4:2624 suppresses
- L3:2794 excites
- L4:203 excites
- L4:1795 weakly excites "at least" phrase
- L4:1461 weakly excites
- L4:1115 weakly suppresses
- L4:2941 weakly excites
- L2:1899 weakly excites
- L1:2166 weakly suppresses
- L0:1979 weakly suppresses
- L0:1922 weakly suppresses
- L2:2565 weakly excites
- L2:1424 weakly excites
- L4:1779 weakly suppresses
- L2:1106 weakly excites
- L1:592 weakly excites
- L3:2529 weakly suppresses
- L3:2284 weakly suppresses
- L3:649 weakly suppresses
- L2:1219 weakly suppresses
- L4:2860 weakly suppresses
- L3:2746 weakly excites
- L4:1177 weakly excites
- L4:694 weakly suppresses
Feeds later layers
- L8:1736 excites
- L7:2434 excites
- L15:3063 weakly excites
- L18:2965 weakly excites
- L22:1612 weakly excites
- L8:1050 weakly suppresses
- L11:1004 weakly excites
- L7:1155 weakly suppresses
- L9:563 weakly excites
- L6:678 weakly suppresses
- L24:442 weakly excites
- L16:2836 weakly excites
- L16:885 weakly excites
- L16:1853 weakly excites
- L8:1914 weakly suppresses
- L10:1636 weakly excites
- L25:2468 weakly suppresses
- L22:2571 weakly suppresses
- L10:1006 weakly suppresses
- L19:1862 weakly excites
- L11:472 weakly excites
- L15:1916 weakly excites
- L10:476 weakly excites
- L14:2149 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
- ators
- gow
- Incre
- 蕉
- ious
- ackle
- 产出
- binations
- BR
- ments
Suppresses
- .unique
- pragma
- mpz
- AVOR
- 风口
- ENARIO
- seize
- phoon
- liğinde
- jeta