"that" pointing back to a noun
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. 3 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 "that" when it points back to something just mentioned, as in "for that product" or "at that time". Turning it up made the writing wander without a clear link to this pattern.
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.
-
On June 26, the Chinese observed "an extraordinary star" that was described as greyish-yellow, and was "as large as a pellet." On that same date, David Rittenhouse (Norriton, Pennsylvania, USA)
-
“tracing request” to restore links with family members across the border via that platform.⏎[F5] The following day State A’s forces commence arrests of people
-
, and the effect may be determined only through the analysis of process establishment data for that product.⏎Initial temperature is normally determined using a dial, electronic or other type of
-
, while on the other, individual teams and firms can track the development of open source software through tools created for that purpose. There is a multitude that is available on the market, many of which
-
on September 10, then received the name of Hurricane Humberto at the first report of the National Hurricane Center (NHC) the next day. At that time the hurricane was located about 310 mi (500 km
-
bars denote the size of the inflation change in each item relative to its historical norm, measured by the standard deviation of the seven-month change in the inflation series for that item since 2000. (We go back only to 2
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 need to know 10,000,000,000,000,000,000,000,000,000,000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ia in human culture⏎In the last quarter of the 18th century, Professor Jacob Seydelmann of Dresden developed a process to extract and produce a more concentrated form for use
-
the university. The Female Law Practitioners Act had been passed in 1896, and on 10 May 1897 Ethel Benjamin gave effect to the new legislation when she was admitted as a barrister and
-
that it competed with the figural component for the viewer’s attention, and eventually landscape developed into an independent genre. The paintings of German artist Albrecht Altdorfer mark a significant step in this evolution.⏎The majority of Altdor
-
92 she passed the university junior scholarship examination.⏎In 1893 Ethel Benjamin enrolled for an LLB degree at the University of Otago. She
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
- L2:2922 weakly suppresses
- L1:1318 weakly excites
- L2:2170 weakly suppresses
- L2:126 weakly excites
- L0:1966 weakly suppresses
- L2:1215 weakly excites
- L2:977 weakly excites
- L2:2692 weakly excites
- L0:3033 weakly excites
- L2:527 weakly excites
- L1:685 weakly excites
- L1:2801 weakly excites
- L0:2691 weakly suppresses
- L0:444 weakly suppresses
- L0:288 weakly excites
- L0:1841 weakly suppresses
- L1:1029 weakly excites
- L0:2283 weakly suppresses
- L2:1481 weakly excites
- L2:853 weakly suppresses
- L0:149 weakly suppresses
- L0:1312 weakly suppresses
- L2:872 weakly excites
- L1:894 weakly excites
Feeds later layers
- L4:468 excites "that" after "than" or "like"
- L11:199 excites
- L4:731 excites
- L4:270 excites
- L17:1816 excites
- L4:968 weakly excites
- L7:562 weakly excites
- L12:521 weakly excites
- L8:147 weakly excites
- L17:2011 weakly excites
- L4:1105 weakly excites
- L22:2014 weakly excites
- L9:1231 weakly excites
- L9:82 weakly excites
- L17:221 weakly excites
- L15:470 weakly excites the word 'this' after a preposition
- L5:9 weakly excites
- L4:962 weakly suppresses
- L20:2038 weakly excites
- L15:537 weakly excites middle digits of numbers
- L18:731 weakly excites
- L13:459 weakly excites
- L15:1957 weakly excites
- L4:1762 weakly excites "this" in time and context phrases
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
- 他认为
- 他就
- \Eloquent
- ”的
- aisal
- 之后
- 他会
- 他也
- 也好
- 但是他
Suppresses
- лиз
- orb
- ibility
- viron
- erable
- stile
- 完整性
- =&
- ancell
- shore