opening parenthesis before an acronym
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. 26 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 reacts to an opening bracket that comes right after a full name and introduces its short form, like 'Finite Element Method (FEM)'. Turned up, it made the model repeat words like 'aur', which does not clearly match that pattern.
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.
-
, and living or life skills training.⏎- Eligible California residents may connect with the San Andreas Regional Center (SARC) to create an Individualized Program Plan (IPP). This is a contract between SARC and the individual that sets up an
-
from injuring themselves and others by requiring gun owners to keep guns locked up and unloaded, and ammunition stored in a locked location separate from the firearm.⏎Child Access Prevention (CAP) laws go one step further by imposing criminal liability to ensure that owners are
-
companies that includes Advanced Plasma Power (APP), is building a full-scale gas plant that will prove this technology is commercially
-
have their influence on one another visually mapped. This methodology was named Interpretive Structural Modelling (ISM). Over the next 20 years, the use of ISM to
-
Technologies.⏎Dr Copeland and co-investigator Dr Carl Sangan will assess the technical feasibility of producing high temperature radial turbines using the Selective Laser Melting (SLM) Additive Manufacturing (AM) technique with nickel super-alloys.⏎
-
, we enhance these assumptions by considering mechanically diverse layers with varying dips. We employed the Finite Element Method (FEM) to create numerical models simulating two chamber shapes: a circular form
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: there's one data sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera sera
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Heating, Refrigeration and Air-Conditioning Engineers (ASHRAE)⏎Structural Engineers Association of Kansas & Missouri (SEAKM)⏎Structural Engineering Institute (SEI)⏎Illuminating Engineering Society (IES)⏎American Society of Plumbing Engineers
-
Propulsion Laboratory manages the 2001 Mars Odyssey mission for NASA's Office of Space Science, Washington, D.C. The Thermal Emission Imaging System (THEMIS) was developed by Arizona State University, Tempe, in collaboration
-
Council (ESRC), the research was presented recently at a day-long higher education stakeholder event in Johannesburg, hosted by the team and the Council for Higher Education (CHE). The MOOC was launched at the same event, which was the first
-
mental illness for good.<|endoftext|>Researchers synthesized the literature on physical security planning from school-specific sources and related disciplines to discuss the policy and other challenges that local education agencies (LEAs) face in developing and implementing protection and mitigation measures for kindergarten-through–
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
- L25:2055 suppresses
- L24:600 excites
- L25:925 excites
- L23:625 suppresses
- L21:468 excites
- L20:1423 suppresses
- L19:1422 excites
- L22:309 excites
- L23:176 suppresses
- L24:120 suppresses
- L20:2497 suppresses
- L21:2936 suppresses
- L22:1393 suppresses
- L23:666 excites
- L22:982 excites
- L20:2479 suppresses
- L24:2046 excites opening bracket before an acronym
- L24:241 suppresses
- L21:1400 suppresses
- L19:2585 suppresses
- L12:642 excites
- L25:586 suppresses
- L19:1317 excites
- L25:2082 suppresses first-person 'I' starting a clause
Feeds later layers
- L27:563 strongly suppresses
- L27:1781 suppresses
- L27:2008 suppresses
- L27:273 weakly suppresses
- L27:2740 weakly excites
- L27:897 weakly suppresses
- L27:2870 weakly excites
- L27:913 weakly excites
- L27:536 weakly suppresses
- L27:760 weakly excites
- L27:630 weakly excites
- L27:2929 weakly excites the letter c splitting a word
- L27:915 weakly suppresses unclear: mixed word fragments and phrases
- L27:957 weakly excites
- L27:1009 weakly excites opening parenthesis before an acronym
- L27:2324 weakly excites
- L27:1942 weakly excites
- L27:1200 weakly excites
- L27:3033 weakly suppresses
- L27:1283 weakly excites
- L27:3038 weakly excites
- L27:516 weakly excites
- L27:2821 weakly excites
- L27:709 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
- arrêt
- aur
- 格尔
- clid
- Athen
- jem
- eton
- vrir
- utes
- öt
Suppresses
- sera
- DA
- da
- ва
- HA
- aura
- ha
- VA
- inda
- leta