mid-clause words in technical prose
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. 17 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a word in the middle of a sentence that is about to be completed by the rest of the clause, such as 'means', 'changed', 'has' or 'that'. Its strongest moments come in plain explanations of how something works. Turning it up or down gave rambling or number-heavy text, not a clear theme.
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.
-
or mod)⏎operator to find the remainder one gets when two numbers are divided,⏎say like a % b. If a remainder is ever zero, it means that b⏎divides evenly into a, or a is divisible by b.⏎
-
hash of that data; on the receiving end, we re-hash the received data and check that the computed hash equals that sent; if any of the data has changed then (with overwhelming probability), the computed hash value will no longer match the original
-
server to the browser.⏎Like most other servers, Apache multiplexes its operations so that it can⏎start processing a new request before it has finished working on the previous⏎one. On Unix systems, Apache uses a multiprocess model in which
-
data file.⏎In the sub-matrix format a portion (the block size) of a row is followed by further portions of subsequent rows before the remaining data values of the row are found.⏎Of the two, the sub-matrix format provides for
-
completing a circuit.⏎This process continues until you turn off the light bulb.⏎What is Contact Current?⏎Contact current is a phenomenon that occurs when you touch something that has current flowing on it. As soon as you touch the object, the current goes
-
hormone insulin then facilitates the uptake of glucose into cells, where it is metabolized into energy for immediate use.⏎To conserve fuel, you body stores excess glucose not needed for energy as a compound called glycogen. Through a process
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 will start a new project, and I will write a project report, then I will write a project report, then I will write a project report, and then I will write a project report, and then I will write a project report
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
When will the redistricting process begin?⏎The Census Bureau was delayed in its decennial count because of the COVID-19 pandemic. Typically, the bureau delivers data to the states by April 1.⏎This year, the bureau warned states
-
extensive research program as the industry addresses impacts and works toward 2025 industry goals⏎Every five years, Field to Market: The Alliance for Sustainable Agriculture publishes its National Indicators Report (NIR). The latest
-
image representing the strength and solidarity of Londoners emerged on Twitter. It’s not uncommon for workers on the London Underground to write messages of national unity on tube signs following tragic events, and the March terrorist attack seemed no different. The image proceeded to
-
Science article published this month, Americans should expect a tough flu season -- and an early one. While the article admits that scientists and doctors struggle to predict flu seasons each year, general consensus indicates that the 2017 - 2
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
- L12:1702 excites
- L16:2363 excites
- L15:1829 excites
- L15:1113 suppresses
- L15:560 excites
- L14:2104 suppresses
- L14:1347 suppresses
- L13:708 suppresses
- L16:2375 suppresses
- L16:2332 suppresses
- L15:1641 excites
- L15:2724 suppresses
- L15:720 suppresses
- L14:2571 excites
- L13:563 excites
- L14:797 suppresses
- L12:937 excites
- L14:1520 excites
- L15:1879 excites
- L13:1153 excites
- L11:311 suppresses
- L16:1748 suppresses
- L13:2527 suppresses
- L15:689 suppresses
Feeds later layers
- L19:553 strongly excites
- L19:2563 strongly excites
- L19:1862 strongly suppresses
- L19:894 suppresses
- L18:2594 excites
- L19:2475 excites
- L21:1093 excites
- L24:60 suppresses
- L20:2968 suppresses 'more' or 'not' before a likelihood claim
- L22:2399 excites
- L18:1319 suppresses
- L22:1616 excites
- L19:2254 excites
- L18:3047 excites
- L24:1770 excites
- L19:2536 excites
- L24:778 excites
- L18:1005 excites
- L24:252 excites
- L26:886 excites
- L22:745 suppresses
- L19:2133 suppresses
- L21:958 excites
- L18:535 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
- 现在的
- 現在
- 如今
- now
- 而现在
- -now
- 现在
- 今
- 算了
- (now
Suppresses
- usually
- often
- sometimes
- 必ず
- 注明出处
- seldom
- usually
- yapılır
- 时常
- lops