'that' after a noun, introducing a clause
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. 4 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 follows a noun and starts a describing clause, like 'the data that was being sourced'. When turned up, the model's writing drifted into rambling, repetitive text rather than anything clearly tied 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.
-
of this activity meant that we had to leave the standard classroom and venture into the school gymnasium.⏎With this is mind I wanted to ensure that the data that was being sourced within the session would be instantly displayed for the students to analyse.
-
is good for your garden soil, it’s not the best thing to add to your compost pile. This is because it takes a very long time for the bacteria that eats dog waste (and its nutrients) to be broken down. To make things
-
up to date, call Dr. Shaheedy's office.⏎As you age, your body's immune system will also change. This may mean that the antibodies that were produced by previous vaccinations may have weakened or the viruses they were designed to prevent
-
, there was not enough-if any-on daily Roman life, which, to me, was and is much more interesting. Some people might argue that all the information that you gain from museums is in books, but where do the pictures and some of
-
fiscal impact of the budget’s policies over the next ten years. The CBO does not generally look at the long-term impacts of these policies, so the numbers that ultimately shape the public debate are ten-year numbers. This introduces a serious bias into
-
The separate video card plugs into a slot on your board and provides your computer with a second sub-processor called a GPU (graphics processing unit). All the calculations that are needed to turn the computer code into the picture on the screen are given over
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 number of times of a system is 100. The system has a lot of factors, which can affect the performance. The number of times of a system is 100, but the system has a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ier has made waves with its fresh look at adolescent sexuality⏎Almost half of American adolescents report having had sex by the time they graduate from highschool — a fact that has raised concern among adults, and parents of teens in particular, about the emotional
-
the system. Incorrect data sometimes produce an erroneous result that reduces the quality of data. To handle this problem web Miner's has introduced several
-
economic stimulus packages in 2014 to bolster growth, including reforms to environmental regulations in order to spur investment in Peru’s lucrative mining sector, a move that was opposed by some environmental groups. However, in 2015,
-
) and Maila (later called Mary), with his first wife, who died in 1822. Hirsch later married Rebecca Haas, with whom he had two children, Vogela (later called Fanny) and Loeb
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
- L3:1529 excites
- L2:1863 weakly excites
- L3:2916 weakly excites
- L1:1562 weakly suppresses 'that' after reporting verbs
- L3:1968 weakly suppresses
- L0:1389 weakly excites
- L2:1564 weakly suppresses
- L2:1781 weakly suppresses
- L2:1962 weakly suppresses
- L3:676 weakly suppresses
- L1:2932 weakly suppresses
- L3:1485 weakly excites
- L3:89 weakly excites
- L1:2006 weakly suppresses
- L3:1316 weakly suppresses
- L1:186 weakly suppresses
- L1:642 weakly suppresses
- L2:2023 weakly suppresses
- L1:175 weakly excites
- L2:1588 weakly excites
- L2:258 weakly excites
- L1:2660 weakly excites
- L2:2306 weakly excites
- L2:552 weakly excites comma inside a stacked adjective list
Feeds later layers
- L5:1989 weakly excites
- L15:1344 weakly excites
- L11:429 weakly excites
- L15:3032 weakly excites
- L10:1477 weakly excites
- L9:75 weakly excites 'what' or 'how' after see or understand
- L13:354 weakly excites
- L24:1782 weakly excites
- L18:2995 weakly excites
- L8:742 weakly excites
- L7:1968 weakly excites
- L21:215 weakly suppresses
- L23:440 weakly excites
- L9:865 weakly excites
- L6:988 weakly suppresses
- L15:925 weakly suppresses
- L7:1607 weakly excites
- L8:1825 weakly excites
- L20:2651 weakly excites
- L16:986 weakly suppresses
- L10:1058 weakly excites
- L8:276 weakly suppresses
- L10:1116 weakly excites
- L8:465 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
- 斫
- emens
- purchasers
- xic
- Configurer
- authorize
- 税务局
- .fasterxml
- pear
- �
Suppresses
- ``⏎
- lahoma
- effectiveness
- bole
- HERO
- ỏi
- ANTE
- Ć
-
- ICC