the word a before long
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. 25 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 "a" when it is followed by "long" or "while", as in "for a long time". The logit lens points the same way, toward the word "long". Turning it up made the model repeat "long" over and over.
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.
-
is easy to find, looks beautiful in astro landscapes, and it contains a fantastic assortment of objects to photograph.⏎If you’ve been reading MilkyWayPhotographers for a while, you might think that the previous paragraph sounds a little familiar – well,
-
sharp drop in the number of Purple Martins in Ontario and the reason behind this is not clear. Figuring out where they go to escape our cold winters will go a long way in helping us understand why their numbers are dropping. To start answering
-
Aluminium is more affordable regarding prize. The advantage of stainless blades is that they are more durable than aluminum therefore once you buy them they are likely to serve you for a very long duration.⏎Various things should be put into consideration before purchasing a propeller
-
, but the animals are sometimes abused by careless owners or harsh conditions.⏎- While the use of animals in entertainment, largely against their will, has been around for a long time, there are more and more anti-cruelty laws protecting animals
-
_of_algorithms)⏎> 1 get a positive integer from input⏎> 2 if n> 10⏎> 3 print "This might take a while..."⏎> 4 for i = 1 to n⏎> 5
-
Gear, talks about sophisticated gear manufacturing methods and how they can help solve common gear-related problems.⏎Noisy gear trains have been a common problem for gear designers for a long time. With the demands for smaller gear boxes transmitting more power at higher rp
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: how can I not make me more...? I am not sure where to start. Let me go to the map... meosh, meosh, meosh. The result is that I have not made me more... meosh. The
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
still in development, the technology has far-reaching implications in both aeronautics and robotics. Cory was also named Boeing Engineering Student of the Year yesterday at Hamphsire, England's Farnborough Airshow. Click here to read an
-
. Children who have fibromyalgia may also have pain or cramping in the lower abdomen with constipation
-
of antihistamines include drowsiness which is compounded in people with fibromyalgia. Decongestants are stimulants which the doctor says can
-
children and adolescents. Estimates suggest that juvenile-onset fibromyalgia affects 2% to 6% of schoolchildren, mostly adolescent
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
- L21:2777 excites
- L22:1046 excites
- L23:143 suppresses
- L23:3034 excites
- L23:1982 weakly excites
- L3:1545 weakly suppresses
- L3:2356 weakly excites
- L18:2374 weakly suppresses
- L12:1880 weakly excites
- L10:13 weakly excites
- L23:380 weakly suppresses
- L13:504 weakly suppresses
- L22:1027 weakly suppresses
- L16:838 weakly excites
- L10:1385 weakly suppresses
- L24:712 weakly suppresses 'long' in 'long-term' and 'long-lasting'
- L5:1072 weakly suppresses
- L19:1712 weakly excites
- L3:494 weakly suppresses
- L16:1802 weakly excites
- L13:2687 weakly excites
- L15:2715 weakly suppresses
- L3:249 weakly excites
- L1:1467 weakly excites
Feeds later layers
- L27:1178 excites
- L26:1469 weakly excites
- L26:887 weakly suppresses
- L26:892 weakly suppresses
- L27:1830 weakly suppresses
- L26:2030 weakly suppresses
- L26:2721 weakly suppresses
- L26:2451 weakly suppresses
- L26:3002 weakly suppresses
- L26:1930 weakly suppresses
- L27:1583 weakly excites
- L27:797 weakly excites
- L26:2546 weakly suppresses
- L26:3007 weakly excites
- L26:2606 weakly suppresses
- L27:2653 weakly suppresses
- L27:1733 weakly excites
- L26:2337 weakly suppresses
- L26:568 weakly suppresses
- L26:2060 weakly excites
- L26:1674 weakly excites
- L27:1590 weakly suppresses
- L26:1655 weakly suppresses
- L26:2836 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
- 长
- long
- Long
- Long
- 長
- .long
- long
- _long
- (long
- →long
Suppresses
- 飨
- ender
- 匀
- xED
- бур
- Jaguar
- ус
- ser
- urus
- +)\