'too' meaning 'also' at clause end
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 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 'too' when it means 'also' at the end of a phrase, like 'dangers too' or 'do it too'. It is one of the most consistent word detectors we looked at. Turning it up did not bring out that word in the model's writing.
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.
-
to your eyes⏎- Remove or reduce eye hazards whenever possible⏎- Know the eye safety dangers at your work and make sure your coworkers are aware of these dangers too⏎- Replace your safety eye wear if it becomes damaged⏎Types of Eye Protection
-
to gain control — so a tool isn’t always calling the shots. (We get enough of that in politics, yes?)⏎It’s super easy to get started too.⏎Here’s a quick how-to:⏎- Get in a comfy position. You
-
, it becomes our problem too. The Aqua Dental clinic at Pearland assure you to be partners in your child
-
. To achieve this, our researchers continually develop state-of-the-art water purification technologies, and now, many of these innovations can be used here on the ground, too. During this webinar, three innovative technologies that were developed at NASA Ames Research Center
-
That was before scientists knew about the oceans on moons around other planets, before they knew about how tough microorganisms get here on Earth (and so maybe in space too?), before they started planning experiments to look for life on Mars, and before tech
-
the crosshairs of numerous governments for alleged data privacy and national security risks. It is already banned in India, and the US administration is seriously considering banning it too.⏎How can a platform that hosts and displays user-created silly videos be a data
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 power of a position, the value, and the remaining time. It feels like the perfect time to take action. And value is something you can't have, and the remaining time is something you can do, so there's no more
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
that uses a network or the internet to communicate.⏎User datagram protocol udp is a protocol that sends a datagram in its entirety from the host to the desired destination. Tcp
-
audio amplifier have, without a doubt, proved to be more electrically-resilient than their transistor counterparts. When faced by a short or open circuit, valves tend to be protected by their output transformer, whereas transistor-based amplifiers
-
and Impedance -- Capacitive"⏎- How out-of-phase AC voltages do not add algebra
-
30’s, UHF is an acronym for Ultra High Frequency because at the time 300 MHz was considered high frequency. They can be used when impedance mating is not required.⏎|
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
- L7:1908 excites
- L4:1458 suppresses
- L4:3031 excites
- L0:114 weakly excites
- L6:2358 weakly suppresses
- L6:107 weakly suppresses 'organisms' and 'Code' in reference contexts
- L7:2357 weakly suppresses
- L6:1825 weakly suppresses
- L3:2174 weakly excites
- L0:2060 weakly excites
- L5:2953 weakly excites
- L6:2301 weakly suppresses
- L6:2374 weakly excites
- L5:2291 weakly excites
- L7:2953 weakly excites
- L3:1825 weakly suppresses
- L1:972 weakly suppresses
- L5:1784 weakly suppresses
- L2:955 weakly excites
- L6:1652 weakly suppresses
- L0:2922 weakly excites
- L3:50 weakly excites
- L6:2288 weakly excites
- L6:2523 weakly suppresses
Feeds later layers
- L12:2970 excites
- L10:696 excites
- L11:339 excites the word 'also' before a verb
- L18:957 weakly suppresses
- L16:2179 weakly excites
- L12:2360 weakly excites
- L14:1963 weakly excites
- L10:2163 weakly excites
- L9:1497 weakly excites
- L22:1324 weakly excites
- L15:2573 weakly excites
- L15:220 weakly suppresses
- L10:343 weakly suppresses
- L10:1058 weakly suppresses
- L15:2755 weakly suppresses
- L12:1854 weakly suppresses
- L10:959 weakly excites
- L13:2797 weakly excites
- L18:2877 weakly excites
- L9:1593 weakly suppresses
- L11:1472 weakly excites
- L11:583 weakly excites 'both' after a paired subject
- L10:1752 weakly excites
- L10:2809 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
- RIA
- 了自己的
- icism
- rin
- 了他的
- rias
- 自己的
- themselves
- mee
- .experimental
Suppresses
- Protected
- 不动
- ablo
- sequelize
- ắn
- aira
- orque
- aler
- 开盘
- unks