double bar table cell separators
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. 6 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 responds to the doubled vertical bar "||" that separates cells in flattened web tables, such as a label followed by its value. Turning it up produced unrelated rambling text, so the steering does not confirm the pattern.
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.
-
,4) and (2,3). Then the table would look like this:⏎1 2 3 4⏎1 T T F T⏎2 T T T F⏎3 F T T F⏎4 T F F T<|endoftext|>
-
100%|⏎|20 years||100-170 beats per minute||200 beats per minute|⏎|30 years||95-162 beats per minute||190 beats per minute
-
60 direct from Amazon (22% off) Amazon page|⏎|Call number||QC7.H62 2005|⏎|External links||This entry has no external links. Add one.|⏎|Through your library||
-
|⏎|SPRING 2016||ARAB 200 AND ARAB 300|⏎|FALL 2016||ARAB 100 AND ARAB 210⏎This projection
-
11)|| 0.805⏎very high · 43rd⏎|Currency||Euro (€) (EUR)|⏎|Time zone||EET (UTC+2)|⏎|-||Summer (DST)||E
-
is a full breakdown.⏎15 Hours of Mathematics Instruction⏎|Day 1 – Algebra 1||3 hours||–|⏎|Day 2 – Geometry||3 hours||Algebra 1 Quiz|⏎|Day 3 – Arithmetic||
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 math in the curriculum. The set S = {0, 1, a, b} The set S = {0, 1, a, b} So, the set S is: S = {0, 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
VE Articles!⏎Assessment of Social Cognition in Non-human Primates Using a Network of Computerized Automated Learning Device (ALDM) Test Systems⏎Institutions: Aix-Marseille University.⏎Fag
-
amplification and RNA-Seq.⏎19 Related JoVE Articles!⏎Detection of Rare Genomic Variants from Pooled Sequencing Using SPLINTER⏎Institutions: Washington University School of Medicine, Washington University School of Medicine, Washington University School of
-
Radio-telemetric System to Monitor Cardiovascular Function in Rats with Spinal Cord Transection and Embryonic Neural Stem Cell Grafts⏎Institutions: Drexel University College of Medicine, Heidelberg University Hospital, Veterans Administration Medical
-
and high-dose irradiation of a precisely defined target volume (tumor). It is usually performed in a session lasting about 30 minutes. Radiosurgery involves treating tumors in the head or body. The tumor cells
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
- L4:1647 weakly suppresses
- L0:2932 weakly excites
- L5:3023 weakly excites
- L0:2576 weakly excites
- L5:2603 weakly suppresses
- L5:392 weakly suppresses
- L1:2329 weakly excites
- L2:977 weakly excites
- L2:1657 weakly excites
- L0:1347 weakly excites
- L2:237 weakly excites
- L5:2154 weakly excites
- L2:2635 weakly excites
- L1:2312 weakly excites
- L1:2683 weakly excites
- L1:1831 weakly suppresses
- L4:14 weakly excites
- L2:3003 weakly suppresses
- L4:637 weakly excites
- L0:1286 weakly excites
- L2:908 weakly suppresses
- L5:2368 weakly suppresses smell and taste words in health text
- L2:484 weakly excites
- L5:533 weakly excites
Feeds later layers
- L20:1521 weakly suppresses
- L14:97 weakly excites
- L17:541 weakly suppresses
- L15:959 weakly suppresses unclear: sentence ends in how-to text
- L18:2607 weakly suppresses
- L15:2208 weakly suppresses
- L13:812 weakly suppresses
- L8:2303 weakly suppresses
- L22:794 weakly suppresses
- L23:2523 weakly suppresses
- L16:1794 weakly excites
- L7:1717 weakly suppresses
- L17:1801 weakly suppresses
- L11:37 weakly suppresses
- L7:2702 weakly excites
- L12:338 weakly excites
- L21:2703 weakly excites
- L15:2734 weakly excites
- L21:1458 weakly suppresses
- L22:2357 weakly excites
- L9:1238 weakly excites
- L8:1187 weakly excites
- L7:1220 weakly excites
- L11:1473 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
- 書き
- ophys
- aciones
- ény
- ujet
- templ
- apphire
- álido
- 惚
- ẫu
Suppresses
- ока
- wick
- ()</
- role
- ch
- ki
- 肩
- ::
- ordinates
- ://%