unclear: verbs and prepositions in news prose
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. 16 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
The neuron fires on assorted words in news and science writing, like 'developing', 'by', or 'when' just before a clause continues. No clear pattern stands out. Turning it up gave unrelated text.
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.
-
graph progress of conservation measures. If monthly water use figures are provided, use can be graphed monthly.⏎- Writing, art and technology skills can be practiced in developing water conservation messages for posters, the school<|endoftext|>Agile is
-
just to evaluate where women are in terms of representation and equal pay for equal work, but also to consider the potential for jumpstarting climate negotiations and the green economy by strengthening women’s leadership in these areas.⏎Our elected
-
recalled. The study could lead to new treatments for conditions like depression, post-traumatic stress disorder (PTSD) and even addictions— all of which can involve memories that are intrusive and potentially destructive.⏎The intriguing experiment involved 42 severely
-
wandering close to a supermassive black hole and finding itself mercilessly ripped apart by this beast weighing millions to billions times more than the hapless star. It doesn’t happen very often. But when it happens, it’s pretty spectacular. And
-
Alaska’s permafrost is now essentially melting beneath our feet. We know the problem; we all drove on wavy Alaskan roads with bumps and dips caused by thawing permafrost.⏎On a larger scale, oil companies and commercial fishing
-
the speed at which your brain can function, while hours spent sitting at a desk causes back, shoulder and neck pain. Then there is the decreased sleep quality experienced when using tech too much, which causes many more health concerns. Basically, too much
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 a new project that is a new one. I have a lot of questions and I need a lot of questions. It is a new one. I have a lot of questions, and it is a new one. It is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.⏎Unlike medicine, there is no instant gratification with public health. It’s like Michelangelo and the Sistine Chapel. It took him 4 years to paint its ceiling. 4 years! It might not take that long for public health
-
can add in more and more challenges.<|endoftext|>A post-mortem examination, also known as an autopsy, is a step-by-step examination of the outside of the body and of the internal organs by a doctor known as a
-
, they are more prepared to grow up and obtain a decent job. When water is readily accessible, a family does not have to waste time and resources hauling buckets from the nearest stream, but
-
on obstetrics. The very first ultrasound scan during pregnancy was carried out in Glasgow in 1956, using a machine that was originally designed to detect industrial flaws in ships
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
- L15:908 excites
- L15:45 suppresses
- L15:149 suppresses
- L15:1537 suppresses
- L15:121 suppresses
- L12:1889 excites
- L15:2564 weakly suppresses
- L5:598 weakly suppresses
- L13:1407 weakly suppresses
- L15:252 weakly excites
- L15:90 weakly excites
- L15:1369 weakly suppresses
- L8:1677 weakly suppresses unclear: mixed, mostly math and instructions
- L12:2259 weakly suppresses
- L14:2101 weakly excites
- L8:193 weakly excites
- L14:2133 weakly excites
- L15:1327 weakly suppresses
- L15:2198 weakly excites
- L15:680 weakly excites
- L7:191 weakly suppresses
- L12:684 weakly suppresses
- L4:909 weakly suppresses
- L2:1610 weakly excites
Feeds later layers
- L17:1151 excites
- L17:2907 excites
- L17:2057 weakly excites
- L18:989 weakly excites
- L19:783 weakly excites
- L21:12 weakly excites
- L21:296 weakly excites
- L20:2030 weakly suppresses
- L21:475 weakly suppresses
- L20:638 weakly excites
- L19:1156 weakly excites
- L18:1959 weakly suppresses
- L17:138 weakly suppresses
- L17:807 weakly suppresses
- L19:2087 weakly excites
- L19:2074 weakly excites
- L17:1315 weakly excites
- L26:2548 weakly suppresses
- L24:690 weakly excites
- L19:1567 weakly excites
- L21:715 weakly excites
- L19:2059 weakly excites
- L19:1055 weakly excites
- L27:2706 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
- 用
- ien
- indirectly
- palm
- ospace
- thermal
- cover
- Phys
- cribe
- Sup
Suppresses
- vik
- FRING
- sticks
- rette
- ]-->⏎
- cedures
- Rede
- 与众
- viders
- .cljs