unclear: clause starters in old texts
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 responds weakly to connecting words that start a clause, such as 'But' and 'And', mostly in older or archaic English. It is not a clear pattern, and turning it up did not produce an obvious change.
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.
-
with flower, but he also sometimes spelled them the same.⏎It fears not policy that Hereticke,⏎Which works on leases of short numbred howers,⏎But all alone stands hugely pollitick,⏎That it nor growes with heat,
-
spokes, and their hubs were all cast. —1 Kings 7:27-33 NASB excerpts⏎A wise king scatters the wicked,⏎And drives a threshing wheel over them. —Proverbs 20
-
re de Kerguelen, Appellee par M. Cook, Isle de la Desolation [on sheet with] Plan du Port Palliser, dans la
-
he died of the plague. One of his best-known works was the Libellus de abbreviaturis Hebraeorum, tam Talmudicorum quam Masoritarum et aliorum rabbinorum (Paris,
-
above tools to ask questions. Asking questions is a very good way to learn a language or start a conversation.⏎|What is this called?: ieu naon nyarios na|⏎|Why is it expensive?: naha ieu awis⏎
-
in accordance with Section 107 of the Copyright Act 1976.⏎Fair son, the first thing I would teach thee is to set thine heart to love God; for unless he love God none can be saved.”⏎From
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: It is the 2013, 2014, 2015, 2017, 2016, 2017, 2018, 20
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
institutions; focus on race.⏎Discrimination and Disparity within Correctional Institutions⏎Despite America’s long standing battle for equality, disparity and discrimination are alive and well. The following is an attempt to address the existence and expose the role disparity and
-
adequate for bamboozling curious amateurs.⏎The source code (with examples) is here: hideinfo.cpp.⏎This software comes with no warrantee of any sort, use at your own risk!<|endoftext|>Mind Ransomware is one of those
-
, you can infuse the milk first with spices.⏎For example, you can add:⏎- bay leaf or kaffir lime leaves – for an extra herbal kick⏎- oregano, basil, marjoram – for more Mediterranean flavours
-
:23 "But this is what I commanded them, saying, 'Obey My voice, and I will be your God, and you will be My people; and you will walk in all the way which I command you, that it
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
- L8:432 suppresses
- L7:1005 excites
- L7:462 suppresses
- L6:1186 excites
- L8:176 weakly excites
- L7:642 weakly suppresses
- L8:756 weakly suppresses
- L2:369 weakly suppresses
- L7:1685 weakly suppresses
- L1:2448 weakly excites
- L8:639 weakly excites
- L3:2563 weakly suppresses
- L2:35 weakly excites
- L0:2616 weakly excites
- L0:1236 weakly suppresses
- L1:2803 weakly excites
- L6:2902 weakly excites
- L8:2268 weakly suppresses
- L0:2558 weakly suppresses
- L5:1786 weakly suppresses
- L7:533 weakly excites
- L4:2968 weakly suppresses
- L9:169 weakly suppresses
- L2:2763 weakly excites
Feeds later layers
- L11:1527 suppresses
- L13:2100 suppresses
- L14:1138 suppresses
- L13:339 excites
- L12:526 excites
- L16:2843 suppresses
- L12:1779 suppresses weak end of document signal
- L14:3017 suppresses
- L11:1785 suppresses
- L12:885 suppresses
- L19:198 excites
- L17:1026 weakly suppresses
- L12:2931 weakly suppresses
- L16:1372 weakly excites
- L13:2017 weakly suppresses
- L14:1808 weakly suppresses
- L17:693 weakly suppresses
- L12:1648 weakly suppresses
- L20:669 weakly suppresses
- L13:2640 weakly suppresses
- L15:1752 weakly suppresses
- L12:1372 weakly suppresses
- L14:1121 weakly suppresses
- L17:564 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
- levator
- cket
- kee
- vious
- ına
- ""},⏎
- olland
- 舐
- peated
- ,),⏎
Suppresses
- RequestId
- acher
- '])[
- tex
- arem
- 阴影
- NotFoundException
- bars
- amu
- .createTextNode