'ur' and 'ud' after a capital H
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. 2 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
This neuron fires on the second piece of words that start with a capital H, such as the 'UR' in HURRICANE or the 'UT' in HUTCHINSON. Turning it up gave wandering text that did not mention weather or storms.
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.
-
them. (We’ll start saving). For now, his contribution to the national weather conversation is this Hurricane FAQ.⏎LET’S START SMALL. WHAT IS A HURRIC
-
If the cyclone keeps growing and sustains 74 mph winds, it graduates to hurricane.⏎ONCE WE CALL IT A HURRICANE, HOW DO WE CATEGORIZE IT?⏎We look to the
-
a typhoon if it happened in the northwestern Pacific.⏎WHAT’S THE DIFFERENCE BETWEEN A HURRICANE AND A TROPICAL STORM?⏎Wind speed. When tropical cycl
-
gasoline, if the bottles are stored in your garage) or from foods that have gone bad in your<|endoftext|>SHORT HISTORY OF STINNETT, TEXAS HUTCHINSON COUNTY⏎April 6th, 2010⏎
-
6 types of lepton, 3 of which are negative, the other 3 neutral.<|endoftext|>From: Nicholas Holton, River House, Inc., Huron
-
more than 400 varieties of biscuits and employed more than 5,000 people.⏎Huntley & Palmers employed the latest technology but their success was due more to inspired
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 problem of the problem is the same as the problem of the problem of the problem... It is the same as the problem. So, the problem of the problem of the problem. So. So. So. So. So.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ons.(Figure⏎of Consensus Splice Sites) A majority of these searches are based on algorithms that look for complete genes.⏎The Gene Finder site from the⏎Baylor College of Medicine has a variety of options to search for introns and
-
Cooper.⏎Buy the hardcover edition of⏎I Heard the Willow Weep for $6.00. (Retail price
-
, it might be useful to assume that your search path for⏎cd starts from one of those folders.⏎For example, I spend a lot of time working in⏎/var/www. Instead of typing⏎cd /var/www/example.com all the time
-
appropriate office for investigation.<|endoftext|>Skip to main content⏎Create interactive lessons using any digital content including wikis with our free sister product⏎. Get it on the⏎Pages and Files⏎Democratic Teaching Methods⏎Hidden Histories, Dramatic Events⏎
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
- L0:1301 weakly suppresses
- L1:2522 weakly excites
- L0:2788 weakly suppresses
- L0:1825 weakly suppresses
- L1:53 weakly excites
- L1:900 weakly excites
- L0:1224 weakly suppresses
- L1:2110 weakly suppresses
- L0:178 weakly suppresses
- L1:2683 weakly excites
- L1:1935 weakly excites
- L0:2879 weakly excites
- L0:1464 weakly excites
- L1:413 weakly suppresses
- L0:559 weakly suppresses
- L1:929 weakly excites
- L1:1335 weakly excites
- L1:2505 weakly suppresses
- L0:2095 weakly excites
- L0:134 weakly suppresses
- L0:1259 weakly excites
- L1:1304 weakly suppresses
- L0:2105 weakly excites
- L0:424 weakly suppresses open parenthesis after a name or term
Feeds later layers
- L5:1417 weakly suppresses
- L14:1969 weakly excites
- L6:2213 weakly suppresses
- L20:663 weakly excites
- L4:849 weakly excites
- L19:1968 weakly excites
- L10:2946 weakly excites
- L5:3039 weakly suppresses
- L25:1273 weakly suppresses
- L4:2931 weakly suppresses
- L19:2311 weakly excites
- L22:761 weakly suppresses
- L23:1495 weakly suppresses
- L10:1361 weakly excites
- L9:1379 weakly suppresses
- L13:1924 weakly excites
- L14:2807 weakly excites
- L24:1747 weakly suppresses
- L17:1019 weakly excites
- L8:3063 weakly excites
- L3:1318 weakly suppresses
- L16:457 weakly suppresses
- L24:23 weakly suppresses
- L20:1909 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
- })(
- pipes
- десят
- amaha
- riott
- сел
- DateTimeKind
- plers
- _resolver
- consect
Suppresses
- ように
- for
- far
- ratio
- far
- bağlantı
- LAY
- 鲤
- 脲
- edly