unclear: mid-name word pieces
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. 8 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 piece of a longer word or name, such as the middle of 'retainer', 'Renminbi' or 'Beijing-Tianjin-Hebei'. The examples don't share a clear topic, so it is hard to say what it is really for. Turning it up made the model's writing wander without a clear theme.
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.
-
loose.⏎Because the teeth only can be moved slowly, you'll need to wear braces for about two years. After that you'll need to wear a plastic retainer over your teeth.
-
9 cities across China due to severe winter smog. The affected cities, which include the capital Beijing, are located in the regions of Beijing-Tianjin-Hebei and of the northern Yangtze River delta - including the Jiangsu province
-
onomy and Epidemiology⏎What do the fields of astronomy and epidemiology have in common? You might think nothing. Those two departments are often not even on the same campus at most universities! However, they have<|endoftext|>By Cliff Ellis⏎The building
-
a targeted school. This practice raises the cost of achieving program objectives, such as increased attendance rates, because it provides transfers to many children who would have attended school anyway. Take-home rations programs are
-
took place and a new Yuan Renminbi was introduced at a rate of 1 new Yuan to 10,000 old Yuan.⏎The Renminbi in Foreign Exchange⏎During the command economy, the Chinese Yuan Renminbi
-
0% of behavioral intention, social norms are rarely even significant and always account for a very small percentage towards the variance of exercise intentions. ...⏎Not exactly what you need?<|endoftext|>Member of the Turkish National Allergy and Clinical Immunology Association. Lect
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 careful that I do not make a mistake, but not necessarily the most important one, and even if it is a small part of the work of a person who is not a professional person in a professional field. What is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Perron J, et al. Association between lifestyle habits and adiposity values among children exposed and unexposed to gestational diabetes mellitus in utero. Diabetes Metab Syndr. 2019;13(5):
-
the beams. This could lead to problems such as heart or lung damage, or even an increased risk of another cancer later on. Doctors do everything they<|endoftext|>Variety of electromagnetic radiation with frequencies between a few and a few hundred quadrillion oscillations
-
is through meticulous and time-consuming detective work that we slowly begin to understand how these ships were built, modified, overhauled and used," Pulak told Live Science."By such means, we try understand the minds of the shipbuilders and their
-
bengalensis), a critically endangered species which has lost 99.9 percent of its population due to diclofenac poisoning. Source: Scientific American⏎One of the most important stories in Jane Goodall's book concerns the three
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
- L4:1454 excites
- L5:133 excites
- L5:198 weakly excites unclear: weak "and" in lists
- L7:49 weakly suppresses
- L7:359 weakly suppresses
- L6:1047 weakly suppresses
- L7:1549 weakly suppresses
- L4:499 weakly excites
- L2:781 weakly suppresses 'time' in 'for the first time'
- L1:2316 weakly suppresses
- L6:1394 weakly suppresses
- L4:267 weakly excites
- L1:2301 weakly excites
- L6:1284 weakly suppresses
- L0:1191 weakly excites
- L5:158 weakly suppresses
- L2:3008 weakly excites
- L5:94 weakly suppresses
- L1:675 weakly excites
- L6:878 weakly suppresses
- L7:726 weakly excites
- L7:52 weakly excites
- L5:46 weakly excites
- L1:2430 weakly excites
Feeds later layers
- L10:2328 suppresses
- L10:2924 suppresses
- L19:2878 suppresses
- L10:1142 weakly excites
- L11:2796 weakly suppresses
- L13:1421 weakly excites
- L10:2721 weakly suppresses
- L9:2186 weakly suppresses
- L18:820 weakly excites
- L10:1316 weakly excites
- L13:852 weakly excites
- L10:2036 weakly suppresses
- L10:2463 weakly suppresses
- L11:1645 weakly excites
- L10:1086 weakly excites
- L9:1509 weakly excites
- L17:2627 weakly excites
- L9:1520 weakly excites
- L13:1386 weakly excites
- L19:894 weakly suppresses
- L10:603 weakly suppresses
- L11:159 weakly excites
- L9:2882 weakly suppresses
- L11:1608 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
- 两国
- 而来
- 两地
- osto
- ソー
- gren
- .requests
- tr
- 握手
- 請求
Suppresses
- iveness
- illum
- aned
- /column
- inar
- явление
- 法宝
- liv
- onna
- cef