'back' after a verb or noun
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. 24 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 'back' when it follows a verb or noun, as in 'backup', 'built back' or 'grown back'. Turning it up did not clearly make the writing mention 'back', so the steering evidence is mixed.
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.
-
employers may keep the SDSs in a binder or on computers as long as the employees have immediate access to the information without leaving their work area when needed and a back-up is available for rapid access to the SDS in the
-
959, but the name peaked in popularity in 1973, in Nebraska.⏎Martin has an old-fashioned ring to it, and the numbers back up that impression--the name peaked in 1964 and was especially
-
on a site's soil conditions. Root systems usually are shallower in heavy clay soils.⏎In some situations, it can be detrimental to amend or replace heavy clay backfill soil, because it can trap water around the tree's roots and also can
-
received information on their change in heart rates and blood pressure, I added this to the spreadsheet.⏎As this was happening the data was also being displayed in a graph back in our classroom about 100m away. On completion of the laboratory
-
each site, fish, macro-invertebrates, and benthic composition information is then quantified following standardized protocols. By relating the data collected in the field back to the habitat maps and bathymetric models, BB is able to model and
-
later, man would eventually take those first steps on the moon, meaning the United States had beaten Russia in the space race, but it also led to new changes back down on Earth.⏎“A lot of the things we developed to engage in the space
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 small family (family worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth worth
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Dry mouth⏎- Increased thirst⏎- Eyes that appear sunken⏎- Skin changes -- for example, if you touch or squeeze the skin, it doesn't bounce back the way it usually does⏎- Urinating less often or having dark yellow
-
igious affiliation within Cameroon boasts a whopping 70% Christian. 21% is of the Islamic faith, and 6% still maintain indigenous belief systems dating back to the cultural inception thousands of years ago. Islamic believers are concentrated in the
-
the actual practice (conducted in the arena).⏎These two continually learn from each other and evolve.⏎The beginnings of the science⏎Although the theories and logic used date back to Aristotle, the modern application began in the 1960′
-
to heat pumps and solar thermal, we must look at the period needed to cover their upfront cost through the savings on energy bills, which is known as the “payback time”.⏎With the existing incentives and prices, an average middle-income family
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
- L21:2530 strongly suppresses
- L21:2076 suppresses
- L21:1525 suppresses
- L14:2160 excites
- L23:2231 excites
- L17:1898 suppresses
- L21:2998 suppresses
- L19:3036 excites
- L18:3044 suppresses
- L7:196 excites
- L15:1563 excites
- L23:2861 excites
- L22:2937 weakly excites
- L22:2615 weakly suppresses
- L23:1195 weakly suppresses
- L19:1951 weakly suppresses
- L17:2799 weakly suppresses
- L19:1871 weakly suppresses
- L11:1688 weakly suppresses
- L5:2709 weakly excites
- L23:364 weakly excites
- L23:1182 weakly suppresses
- L7:2996 weakly suppresses
- L22:1872 weakly excites
Feeds later layers
- L27:1173 strongly suppresses
- L26:2189 suppresses
- L26:1281 suppresses
- L25:2242 weakly suppresses
- L25:1799 weakly suppresses
- L25:2159 weakly suppresses
- L26:2629 weakly suppresses
- L27:1179 weakly suppresses
- L27:2930 weakly suppresses
- L26:499 weakly suppresses
- L25:2679 weakly suppresses
- L27:928 weakly suppresses
- L25:2304 weakly suppresses
- L25:1660 weakly suppresses
- L26:2010 weakly suppresses
- L27:1686 weakly suppresses
- L26:1637 weakly excites
- L27:1820 weakly suppresses
- L25:2062 weakly excites
- L27:2032 weakly suppresses
- L27:798 weakly excites
- L26:3018 weakly excites
- L25:1012 weakly suppresses
- L25:275 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
- 门户
- idable
- umu
- умент
- porto
- 缢
- uisine
- umen
- senal
- ulpt
Suppresses
- back
- back
- Back
- Back
- -back
- (back
- →back
- _back
- BACK
- backs