'for' after a span of time
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. 19 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
The neuron lights up on the word 'for' when it comes right after a length of time, as in 'for several weeks' or 'for years'. When it was turned up, the model kept drifting into talk about how long something lasts, such as 'for a long time' and 'a few days'.
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.
-
the city was almost destroyed. The conflict spread to outlying provinces.⏎By 1530, the wars of the Sengoku Period had been fought for over two generations. At that point, there were nine major
-
only hinted at by the book.⏎I have admired Dick’s work for years, but this is his best. He delves more thoroughly into many themes
-
has helped me to uncover student misconceptions and look into their thought processes regularly. A supervisor gave me the book in August, and it sat on my nightstand for several weeks. In my mind, it was going to be another book about visual
-
mental illness, affecting more than 350 million people worldwide, according to a Reuters video. Bob Holmes is one of them. "I struggled with that for many years, didn’t know really what to do, tried to pull myself through
-
England, in 1813 and made a career for himself as a renowned architect, engineer and surveyor. After living in Dublin for several years in the early 1830s, Lanyon moved
-
and challenges for these individuals truly is a spectrum that can range from a completely nonverbal kindergartner, to a professor of linguistics.⏎We’ve known for more than 20 years that the lion’s share of the risk for developing
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: A person who is not a regular, but a regular in a regular, and in a regular of a regular of a regular... a regular of a regular, and a regular of a regular of a regular... a regular of a regular,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Assessment's Biogeography Branch (BB) has been conducting research in Puerto Rico and the US Virgin Islands since 2000 and 2
-
process improvements are elastic and, unless rigorously monitored, drift back to their original state. Determin<|endoftext|>Archeological sites are places where people have lived or worked in the past and
-
<|endoftext|>The term “education technology” (or “edtech,” as it’s commonly known) refers to more than just learning via computer. Edtech has existed in various forms for as long as education itself has been around. At the beginning of
-
.⏎The Hasinai group of the Caddo tribe built a village in the area about 800 A.D. and continued to live in the area until the 1830s, when they migrated to the
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
- L15:1825 strongly excites
- L17:1842 strongly suppresses
- L14:2236 strongly suppresses
- L12:2377 excites
- L16:1277 excites
- L18:2735 excites
- L17:2544 suppresses
- L18:2437 suppresses
- L18:2130 suppresses
- L13:2868 suppresses
- L17:3010 excites
- L15:1520 excites
- L14:1088 excites
- L13:2890 excites "still" or "as long as" meaning continuing
- L18:2977 weakly suppresses
- L17:1041 weakly excites
- L10:2361 weakly suppresses
- L12:39 weakly suppresses
- L15:2665 weakly suppresses
- L7:1114 weakly suppresses
- L18:1505 weakly suppresses
- L16:2173 weakly excites
- L6:1343 weakly suppresses
- L12:2782 weakly suppresses
Feeds later layers
- L21:1284 strongly excites
- L23:2386 strongly excites
- L24:1778 strongly excites 'a' before a time span
- L20:1355 strongly excites
- L24:1394 excites
- L22:1699 excites
- L22:2704 excites
- L24:587 excites
- L20:3050 excites
- L23:2777 excites
- L27:1469 excites
- L23:1919 excites
- L22:1074 excites
- L20:2745 excites
- L22:2601 excites
- L20:1219 excites
- L21:785 excites
- L21:3058 excites
- L20:2376 excites "the" after "over" in time spans
- L21:848 excites
- L27:1131 excites
- L21:1797 excites
- L24:826 excites
- L21:676 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
- 一段时间
- 很长时间
- years
- 几年
- months
- 两年
- 多久
- 很久
- 太久
- 三个月
Suppresses
- uctose
- _language
- твержда
- ovich
- _scaling
- ��
- юр
- .language
- ítica
- _positive