unclear: end markers and clause commas
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. 15 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 is weak and mixed. It fires on document-end markers and on some words after an introductory phrase, and it is most negative on commas that close such a phrase. The steering outputs were repetitive and do not reveal a clear meaning.
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.
-
, all of which are dependent on biodiversity. They are predominantly public goods with limited or no markets and do not command any price in the conventional economic system, so their loss is
-
i>⏎;; Now insert one character at the beginning of the buffer: (goto-char (point-min)) ⇒ 1 (insert "Q") ⇒ nil⏎<|endoftext|>A bead is a small bauble or jewel with a hole through its middle for
-
After Effects using precomps!⏎Check back next week to learn the difference between frame rate and shutter speed, or view more lessons at the Vimeo Video School .<|endoftext|>Some time back I presented the untold story of Kebara 2, the
-
an error has occurred. The ldexpf() and ldexpl() functions do not set errno.⏎See attributes(5) for descriptions of the following attributes:<|endoftext|>The
-
tool. Follow this same procedure to use them. They take measurements fixed on their grid direction, even if the tool endpoints are not perfectly aligned to the grid.<|endoftext|>Khalistan Issue: A brief Analysis⏎Author: Isha Pandey⏎
-
and a highly centralized economic system caused Albania to lag far behind its neighbors in terms of economic development.⏎After Hoxha died in 1985, his hand-picked successor, Ramiz Alia, who became party leader while retaining
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 been thinking about the ways in which the choice of a single person can be interpreted as a way of understanding the world in which we live. This is, in other words, a way of understanding the world as a single person,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
date now remembered as Kristal Nacht, or the Night of Broken Glass.⏎That night, as Nazis broke the windows of Jewish stores and burned their synagogues, Goldner’s uncle was arrested along with more than 20,00
-
ons of England organized numerous uprisings. In the face of such strong and well-organized opposition, King John was forced to renounce certain rights and to grant a charter of liberties
-
led armies at Budapest, despite nearly succeeding, they failed and thus the door was open to cross the Oder⏎River and assault Berlin. As this was happening, British and Canadian forces were slogging their way towards the Rhine (taking the
-
the governors around him. The kings of Babylon were exceedingly prideful and rash. They had little regard for life. In the midst of such evil kingdoms and governments, the gatekeepers of the Lord influenced kings, governors, and peoples. There is
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
- L9:652 excites
- L14:1845 suppresses
- L14:1492 suppresses
- L12:2016 suppresses
- L13:1732 suppresses
- L12:1529 weakly excites
- L13:1226 weakly suppresses
- L14:1658 weakly excites
- L13:2051 weakly suppresses
- L13:1511 weakly excites
- L12:2994 weakly excites
- L12:2924 weakly suppresses
- L13:775 weakly excites
- L14:655 weakly excites
- L14:531 weakly suppresses
- L12:1536 weakly suppresses
- L11:2876 weakly suppresses
- L13:109 weakly suppresses
- L11:2879 weakly suppresses
- L14:2316 weakly excites
- L11:182 weakly excites
- L11:2517 weakly excites
- L8:1041 weakly suppresses
- L13:2465 weakly excites
Feeds later layers
- L16:2501 excites
- L16:2368 excites
- L16:939 excites
- L18:1724 excites
- L16:2988 weakly excites
- L22:2321 weakly excites
- L16:3043 weakly excites
- L18:1897 weakly excites
- L20:1205 weakly excites
- L16:2949 weakly excites
- L16:1674 weakly excites
- L18:1629 weakly excites
- L17:2322 weakly suppresses
- L17:3019 weakly excites
- L19:2858 weakly excites
- L18:2349 weakly excites
- L16:1527 weakly excites
- L16:890 weakly excites
- L16:2984 weakly excites
- L22:1477 weakly excites
- L23:1906 weakly excites
- L21:1645 weakly excites
- L20:1745 weakly excites
- L18:2736 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
- nhật
- rels
- 成
- иш
- [first
- elist
- aravel
- unft
- onna
- UICollectionViewCell
Suppresses
- eph
- eus
- strup
- 본
- 傧
- theValue
- 盲
- 了自己的
- .CreateDirectory
- alone