'at' after a job title, before an institution
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. 21 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 'at' when it comes right after someone's job title and before the place they work, such as 'a researcher at Stanford' or 'a professor at the University'. Turning it up made the model write 'University of Arizona' and similar university names over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
morphine and binds to the same receptor, but the way it turns the receptor on is slightly different," says Dr. Aashish Manglik, a researcher at Stanford University School of Medicine who studies opioid receptors.⏎After Bohn's discovery,
-
said Guillaume Gronoff, a French astrophysicist at NASA Langley.⏎"And you think, 'OK, what I am looking at
-
twice as many goods to China each year.⏎"Both sides are yet to tap their trade potential," said Ravi Shekhar Vishal, an assistant professor at Sikkim University who co-authored a research paper on Nathula trade. "
-
kidney diseases, as well as provides information on a protein region whose interaction with filtered light chains is implicated in cast nephropathy”, says Professor Luca Jovine at the Department of Biosciences and Nutrition and corresponding author of the paper.⏎U
-
’s largest anti-masking campaigns, spearheaded by the Anti-Mask League of San Francisco, according to Howard Markel, a professor of the history of medicine at the University of Michigan and co-editor-in chief of The American Influenza
-
ensure that resources are allocated to patients with the greatest need,” says Natalia Trayanova, the Murray B. Sachs Professor in the Department of Biomedical Engineering at The Johns Hopkins University Schools of Engineering and Medicine and the project’s principal investigator…
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: when you really, really, really, you just need to get here, you need to get here, you need to get here, you need to get here, and you need to get here. That is, when you really, really
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Salby told the H<|endoftext|>Posted by Marie on .⏎Pilot Reem Ahmad's flight plan has her leaving Dulles International Airport and flying a Boeing 727 due east at 450 mph. There is a 65
-
the Advanced Research Core (ARC) of the Concussion Assessment, Research, and Education (CARE) Consortium. The consortium is a joint effort of the National Collegiate Athletics Association (NCAA) and the U.S. Department of Defense.⏎
-
th ed.). Boston, MA: Cengage Learning.⏎[TradingCoachUK]. (2018, July 6
-
October 2016, we will conduct a series of low-speed, low-altitude flights across the Great Pacific Garbage Patch in a C-130 Hercules aircraft, using expert spotters and state-of-the-art sensors.
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
- L19:2937 suppresses unclear: mixed possessives and fragments
- L18:1931 suppresses
- L19:2246 suppresses
- L18:3001 excites
- L17:3041 excites
- L17:978 excites
- L20:2623 suppresses
- L19:2135 suppresses
- L18:2529 excites
- L10:3047 excites second word of publication names
- L11:1994 excites
- L18:1451 excites
- L20:1661 excites
- L19:2255 suppresses
- L12:1853 excites
- L16:2427 suppresses
- L6:2759 excites
- L13:3005 suppresses
- L12:2889 suppresses
- L16:2852 suppresses 'in' after a noun or participle
- L18:2292 suppresses
- L13:894 excites 'of' after a person's name
- L20:2169 excites
- L17:1522 suppresses
Feeds later layers
- L22:2134 excites
- L22:838 excites
- L22:1083 excites
- L24:1210 excites
- L25:3053 excites
- L24:1424 excites
- L24:340 excites
- L25:875 excites
- L23:330 suppresses
- L24:2460 excites
- L22:2339 excites
- L24:1124 excites
- L23:348 excites
- L23:1501 excites 'at' in idioms like 'at work'
- L22:1981 weakly excites
- L27:2 weakly excites
- L26:404 weakly excites
- L24:2988 weakly excites
- L23:2719 weakly excites
- L22:1340 weakly suppresses
- L27:1886 weakly excites
- L26:23 weakly excites
- L25:2475 weakly excites
- L24:1953 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
- 中信
- Univers
- University
- University
- ATIC
- SIN
- institutions
- Univers
- universities
- university
Suppresses
- notated
- ozo
- Zombies
- 老人家
- OPSIS
- -mode
- izador
- subclass
- bootstrap
- 节目中