unclear: mid-phrase words in mixed prose
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. 10 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a word in the middle of a phrase where the next word is fairly predictable, such as the 'of' in 'memorandum of points' or 'without' in 'prohibited without permission'. It shows no single topic, and turning it up or down did not produce a clear change in the writing.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
who described a lawyer as 'a person who writes a 10,000 word document and calls it a brief.'⏎A brief usually contains a memorandum of points and authorities. Points and authorities explain why the law authorizes the judge to
-
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⏎Young Canada Works<|endoftext|>From the Publisher:⏎Over roughly
-
etermine value using Look-up table method⏎Create Volcano (or Atoll) martix. It is an extension of Bullseye matrix problem.⏎Who has power to do everything in this world?⏎Close MATLAB with keyboard⏎Find the treasures in
-
’s assume that we have a more general problem. There are n questions, and a(n) is the smallest number of times we need to take the test to guarantee that we can figure out the answers. First we can try all combinations of answers
-
pointing in toward the nose.⏎crossroad: a road that crosses another road.<|endoftext|>If you are pondering how to make your computer virus, you’ve come towards the right place. There are many tutorials and steps on making a malware. However
-
which has more advantages. It is cheaper and more practical. This study aimed to examine the effect of cognitive therapy restructuritation in decreasing depression score of the mental disorder patient of Banyumas public hospital. This study
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: "a bit of a philosophical problem of the sort of thing as it is actually taught in schools... the matter of fact being that a man is not all the more fully formed from the world at large in fact to be the same as he
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
be compared to a wise man who built his house on the rock……⏎Revelation 3:20 'Behold, I stand at the door and knock; if anyone hears My voice and opens the door, I will come
-
a sapphire stone, in appearance in the likeness of a throne, appeared above them. And He spoke to the man clothed in linen and said, “Enter between the whirling wheels under the cherubim…⏎…Then I
-
giveth of his bread to the poor.” – Proverbs 22:9⏎2 Corinthians 9:6-8 – “But this I say, He which soweth sparingly shall reap also sparingly; and he which
-
and statutes written down for Abraham, Isaac, and Jacob to keep, and yet they did the will of God because of
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
- L9:1625 suppresses
- L3:159 weakly suppresses
- L8:277 weakly excites
- L3:615 weakly suppresses
- L9:115 weakly excites
- L9:17 weakly excites
- L6:1110 weakly suppresses
- L3:2075 weakly excites
- L8:254 weakly excites
- L3:1869 weakly excites
- L7:705 weakly excites
- L7:359 weakly suppresses
- L3:2761 weakly excites
- L0:1412 weakly excites
- L8:786 weakly suppresses
- L2:299 weakly excites
- L4:2228 weakly suppresses
- L8:2049 weakly excites
- L9:1293 weakly suppresses
- L7:2120 weakly suppresses
- L0:2625 weakly suppresses
- L1:49 weakly suppresses
- L0:137 weakly excites
- L1:1956 weakly suppresses
Feeds later layers
- L15:646 suppresses
- L11:393 suppresses
- L12:887 suppresses
- L13:1513 suppresses
- L11:1384 suppresses
- L18:819 suppresses
- L12:576 suppresses 'of' in fixed phrases like table of contents
- L12:434 suppresses
- L16:200 weakly suppresses
- L16:592 weakly suppresses
- L18:867 weakly suppresses
- L13:450 weakly excites
- L15:227 weakly suppresses
- L24:1787 weakly excites
- L13:1929 weakly suppresses
- L11:796 weakly suppresses
- L20:272 weakly suppresses
- L25:643 weakly excites
- L19:1982 weakly suppresses
- L18:775 weakly excites
- L22:157 weakly excites
- L15:364 weakly suppresses
- L14:2288 weakly suppresses
- L22:490 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
- 声
- 精神
- PT
- NK
- gt
- 软
- (uint
- PDO
- (',',$
- 最后
Suppresses
- STRU
- ,))⏎
- contemplated
- esda
- ,");⏎
- testified
- ----------- ⏎
- ...,
- MainThread
- ..."⏎