unclear: small words in maths and how-to text
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 fairly consistent: several of them 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
The positive side is weak and fires on small words like 'a', 'the' and 'to' in maths and instructions, such as 'percentage of a whole number'. The much stronger side is negative and fires on words like 'starts' in graph definitions. Turning it up gave unrelated, repetitive text.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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.
-
Jersey Turnpike. He was 86.⏎There are several different ways to solve problems involving percentages, decimals, and fractions. Watch as find the percentage of a whole number.<|endoftext|>1. How does the U.S. manage the environmental legacy
-
, it can also be modified manually in the top menu.⏎Now, the formula to get the percentages is basically: the letter and the cell number of the general amount, divided by the letter and the
-
Strategies and Tips⏎Use this lesson to show how exponential functions and geometric sequences are related.⏎- In a geometric sequence, terms are found by multiplying the same constant to the previous term.⏎- For exponential functions, when the input variable is increased by
-
function can be used to determine, based on the employee's name, the position in the array, or table. Check the array in that position and see if the record is there.⏎Hash functions have a big science behind it because as you may
-
make use of a mark percentage calculator to calculate your marks. If you’re unsure of how to write a number, try a calculator.⏎To find the percentage of a number, you should first know how many decimal places it falls under. You can
-
to each other (parallel to the axis of symmetry as well).⏎: A parabola is the set of all points in a plane that are equidistant from the focus and the directrix of the parabola. [Adopted
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 a participant in a whole (i.e., a total) change in the sense that I have gone from the beginning, so I have gone to the end, so the change in total is zero. So, for example,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a large role in our graph algorithms. A path is some sequence of edges that allows you to travel from one node to another. A cycle is a path that starts and ends at the same node. Sometimes cycles are useful, but they often represent
-
flow are what keep a business going. What good are record revenues if you’re spending more money than you’re taking in?⏎Sidebar – My buddy Steve Chou owns a successful online wedding
-
. How many new shapes can you make?⏎Can you shunt the trucks so that the Cattle truck and the Sheep⏎truck change places and the Engine is back on the main line?⏎In how many ways can you fit two of these yellow
-
negative, which sum to zero.⏎The positive cycle may move charge carriers, such as electrons. But then the negative cycle pulls the charges back to where they started. To reliably control the quantum information, an asymmetric light wave is needed.⏎“The optimum
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
- L12:2854 excites
- L9:1623 suppresses
- L6:2090 suppresses
- L13:2770 suppresses
- L13:2309 excites
- L6:1145 excites
- L13:990 suppresses
- L12:2541 excites
- L11:2410 excites
- L14:2445 excites
- L8:1526 excites
- L13:2816 weakly suppresses
- L9:2391 weakly excites
- L7:2309 weakly excites
- L13:2942 weakly excites
- L10:141 weakly excites
- L7:536 weakly suppresses
- L7:245 weakly excites 'even' in 'and even' lists
- L11:951 weakly suppresses
- L12:2192 weakly suppresses
- L11:557 weakly excites
- L7:513 weakly suppresses
- L11:543 weakly excites
- L6:2517 weakly excites
Feeds later layers
- L16:2990 excites
- L16:2123 suppresses
- L16:2756 suppresses
- L19:746 suppresses
- L18:2955 weakly suppresses
- L16:1376 weakly excites
- L21:2233 weakly suppresses
- L18:969 weakly suppresses
- L18:1710 weakly suppresses
- L17:1726 weakly suppresses
- L17:2351 weakly suppresses
- L16:2894 weakly suppresses
- L20:1768 weakly suppresses
- L22:2541 weakly suppresses
- L17:2887 weakly excites
- L20:1362 weakly suppresses
- L22:2920 weakly excites
- L23:2539 weakly suppresses
- L18:2910 weakly suppresses
- L18:2643 weakly suppresses
- L20:2576 weakly suppresses
- L23:1897 weakly suppresses
- L18:2879 weakly excites
- L16:1279 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
- 分辨率
- PLEX
- ⌠
- urgently
- 部分内容
- ropdown
- 紧迫
- 慑
- 霸道
- urgent
Suppresses
- cancelled
- vanish
- canceled
- ancellation
- 消失
- neutral
- canc
- vanished
- cancel
- Cancellation