"that" as a pointing word
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. 6 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 "that day" or "that work." Turning it up did not make the model's text show this pattern, so the steering is unclear.
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.
-
also be rocky and have a composition much like Earth’s. Some might even harbor life. Over 99% of those alien worlds stay undiscovered — however that is about to vary.⏎With NASA’s new exoplanet-hunter house telescope
-
time throughout the summer to learn and complete work; even just 15-30 minutes a day will allow students to keep their skills sharp and knock out that summer work so they can enjoy their break worry-free.⏎How can you take this
-
agency.⏎Although they were in a hurry to complete the structure, the builders halted the work on Sunday for a time of worship and Bible study. It happened on that day that a Ho-Chunk man named Four Legs (a different individual than the
-
a very important gap for students experiencing race-related stress<|endoftext|>Western tradition gets upheld as a lone bastion of freedom, the shining city on a hill. To make that
-
under even greater strain. To help ease that pressure, Bruce and other scientists are working to understand and help improve control of crop
-
and is in part driven by the ‘diabetes epidemic’.⏎... the ‘healthier world of tomorrow’ requires that scientists today build the knowledge base to create that ’healthier world of tomorrow’.⏎Equal access to a healthy lifestyle⏎A
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 a question about the problem, and I have a problem with the problem. The problem is not only about the problems, and the problems are in the same with each other. The problem is not just about the problems. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is an element , where is some object of , which exhibits representability of via the Yoneda lemma. That is, any element induces, in natural bijective fashion, a natural transformation⏎and is universal if is an isomorphism.⏎
-
. If the ingredient list is not available on the product, check<|endoftext|>Help support New Advent and get the full contents of this website as an instant download or CD-ROM. Includes the Catholic Encyclopedia, Church Fathers, Summa, Bible and more all
-
of Arts in Kyoto, Japan. This provides evidence that exposure to certain patterns of speech can influence one's perceptions of musical rhythms. In
-
little to salvage for homeowners caught in harm's way.⏎Climate change will disproportionately affect California’s most disadvantaged and isolated communities, according to a recent report from the Pacific Institute.⏎By looking at a broad array of factors – from social indicators such as income
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
- L3:1003 excites
- L4:851 excites
- L4:468 excites "that" after "than" or "like"
- L4:2051 weakly suppresses
- L0:2841 weakly suppresses
- L5:2747 weakly suppresses
- L1:2213 weakly suppresses
- L1:875 weakly excites
- L4:1762 weakly suppresses "this" in time and context phrases
- L2:2259 weakly excites
- L5:1537 weakly excites
- L5:1561 weakly suppresses
- L4:1042 weakly suppresses
- L4:1748 weakly excites
- L5:1446 weakly suppresses
- L2:2794 weakly suppresses
- L2:2850 weakly excites
- L2:2615 weakly suppresses
- L1:2172 weakly suppresses
- L1:1036 weakly suppresses
- L1:146 weakly excites
- L3:2518 weakly suppresses
- L3:81 weakly excites
- L0:2204 weakly excites
Feeds later layers
- L10:965 excites the word 'that' introducing a clause
- L15:470 excites the word 'this' after a preposition
- L12:521 excites
- L7:1600 weakly excites
- L16:715 weakly excites
- L18:814 weakly suppresses
- L24:668 weakly excites
- L11:2338 weakly suppresses
- L17:655 weakly excites
- L18:1393 weakly suppresses
- L14:2489 weakly excites
- L17:221 weakly excites
- L8:1182 weakly suppresses
- L16:896 weakly excites
- L19:2547 weakly excites
- L13:1007 weakly excites
- L16:2861 weakly excites
- L21:323 weakly excites
- L25:274 weakly excites
- L13:2397 weakly suppresses
- L13:1376 weakly suppresses
- L25:2529 weakly excites
- L10:2276 weakly excites
- L11:810 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
- sett
- 坊
- omm
- opr
- hc
- raj
- hh
- der
- 结论
- hed
Suppresses
- -mile
- -opacity
- _each
- -fontawesome
- -average
- ctica
- cente
- -density
- �权
- each