"this" pointing back to a thing
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. 14 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 "this" (or "these") when it points back to something just mentioned, like "this converter" or "these proteins". When it was turned up, the output did not show that pattern clearly and became repetitive and unrelated.
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.
-
sub-processor called a GPU (graphics processing unit). All the calculations that are needed to turn the computer code into the picture on the screen are given over to this second sub processor and it is dedicated to this task. This hand-off of this
-
to cook something slowly and gently in water that is just below the boiling point. Usually eggs are what's being cooked when "coddle" is used in this sense today, but when "coddle" was first being
-
by...⏎Yersiniae cause severe intestinal infections. Studies using Yersinia pseudotuberculosis as a model organism aim to elucidate the infection mechanisms of these...⏎Researchers from the
-
3D printers. They often use 8mm lead screws as demonstrated in the video below. The concepts explained above are all shown in the first few seconds of this video:<|endoftext|>Tourette's Syndrome (cont.)⏎IN THIS ARTICLE⏎Treatment
-
(Anderson) saw potential in his work, and began mass-marketing them. He created several versions over the years, the last and most popular of which was this one created in 1941.⏎Christ at Heart's Door. The
-
have to fill in the value of Arabic numeral and it will show the same value in Roman numeral. You can convert any desired Arabic numeral to Roman format by using this converter. You can start from number 1 and go up to any possible number
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: I have to be able to to get the the to get to the to get to to get to. I have to to get to to. I have to to. I have to. I have to. I have to. I.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
1991-1995⏎The materials in this archive are copyright © The Manitoba Library Association. Reproduction for personal use is permitted only if this copyright notice is maintained. Any other reproduction is prohibited without permission Copyright information for reviewers
-
three arcs above and to the right of the cluster core, and the fourth is down and to the left, partially obscured by a foreground star. Each arc is that
-
and animals lost to the eruptions. The government will buy endangered cows on the volcano, the president said.⏎Many of those who live on its slopes risked their lives by staying or returning to feed their cows during lulls in the volcanic activity
-
. How quickly you respond to your computer's symptoms can make all the difference in how critically it is affected. Initially, don't worry about identifying the culprit. That will come later. For now, focus on isolating the problem before it does
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
- L0:1035 excites
- L5:125 excites "of" after all three, both, ten
- L10:167 suppresses
- L6:1801 weakly suppresses
- L10:965 weakly suppresses the word 'that' introducing a clause
- L10:1404 weakly excites
- L12:521 weakly excites
- L2:602 weakly excites
- L12:1065 weakly excites
- L9:493 weakly suppresses
- L6:1218 weakly suppresses
- L4:52 weakly excites
- L9:249 weakly excites
- L6:61 weakly suppresses
- L8:217 weakly suppresses
- L8:636 weakly suppresses
- L4:968 weakly excites
- L2:26 weakly suppresses
- L4:731 weakly excites
- L1:1318 weakly excites
- L4:468 weakly excites "that" after "than" or "like"
- L3:229 weakly excites
- L6:1135 weakly suppresses
- L7:224 weakly suppresses
Feeds later layers
- L16:284 excites
- L17:985 excites
- L15:795 suppresses
- L19:832 excites
- L20:271 suppresses
- L18:1228 excites
- L21:511 excites
- L21:592 excites
- L18:602 excites
- L18:1278 excites 'a' or 'most' before a quantity
- L16:2520 excites
- L15:736 excites
- L15:1746 excites
- L23:203 excites
- L18:77 excites
- L17:1027 excites
- L22:194 suppresses
- L17:1592 suppresses
- L16:1711 excites
- L20:337 suppresses
- L18:141 excites
- L22:901 excites
- L19:1707 excites
- L21:124 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
- shortened
- ened
- ening
- two
- 复杂的
- huge
- strange
- extensive
- particular
- four
Suppresses
- рам
- iota
- 看电影
- arium
- -go
- /movie
- 纳
- èmes
- lessness
- cci