unclear: quantities and measured changes
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. 13 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 reacts to words in the middle of number-based phrases, like a fraction, a rate of change, or a count of things. It is a loose pattern, and turning it up gave rambling text that had nothing to do with numbers.
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.
-
not until June 15, 1956 that one was adopted by the tribal elders. It consists of seven stripes, the central being 1/6 the size of the other stripes. The hoist consists of a triangle edged
-
4% of strokes worldwide, the authors write. They note that the risk of heart attack and stroke drops by about half for every 20 mm/Hg reduction in systolic blood pressure (the top number of your blood pressure
-
lic acids. The rate at which this occurs depends on a number of factors. Temperature is perhaps the most critical one, since the rate of oxidation doubles for every rise of
-
than merely memorising them, the researchers introduced words the birds had never seen before.⏎The pigeons correctly identified the new words as words at a rate significantly above chance.⏎According to the study’s first author, Dr Damian Scarf of the University
-
| Permalink⏎"crochet school - Lesson 9", "crochet school", "learn to crochet", "treble crochet", crochet⏎| Digg This⏎| Save to del.icio.us<|endoftext|>Back in the early 2
-
, or they will invest elsewhere.⏎The following is the WACC calculation formula:⏎WACC = E/V × Re + D/V × Rd × (1 - Tc)⏎Re = cost of equity⏎Rd = cost of debt⏎E
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: 10000000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2. A company is offering to pay you one penny on your first day at work and each day after that your salary would triple. Assume you work five days a week.⏎a. Fill out the tables representing your daily salary for your first two
-
and girls. From six years onwards, you detect small differences<|endoftext|>An average size tree produces enough oxygen in one year to keep a family of four breathing.⏎Three trees planted in the right place around buildings can cut air-conditioning costs up to
-
to activate it.⏎Can vinegar and baking soda be mixed?⏎Here are some recipes to try. Freshen your sink by mixing one part of baking soda with two parts of vinegar. This mixture unlocks an eff
-
Yuan replaced the Fǎbì in 1948 at a rate of 1 Gold Yuan to 3 million Yuan Fǎbì. That same year, the Yuan Renminbi
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
- L11:95 excites
- L12:1686 excites
- L11:726 suppresses
- L12:2169 suppresses
- L11:1036 suppresses
- L11:805 suppresses
- L12:1141 excites
- L12:983 suppresses unclear: mid-clause words in expository prose
- L11:2493 suppresses
- L11:2502 excites
- L12:2937 weakly suppresses
- L8:1827 weakly excites
- L10:1252 weakly suppresses
- L11:1775 weakly excites
- L11:2812 weakly excites
- L11:1889 weakly suppresses
- L12:1350 weakly excites
- L11:374 weakly excites
- L12:2085 weakly suppresses
- L12:1446 weakly suppresses
- L10:2487 weakly excites
- L10:114 weakly excites
- L12:390 weakly suppresses
- L11:2534 weakly suppresses
Feeds later layers
- L14:2850 suppresses
- L14:258 excites
- L16:1154 suppresses
- L17:1804 suppresses
- L25:646 suppresses measurement words before a value
- L21:1243 excites
- L18:2465 excites
- L22:2844 weakly suppresses
- L14:2303 weakly excites
- L16:2643 weakly excites
- L14:3003 weakly excites
- L15:2789 weakly excites
- L15:1423 weakly suppresses
- L14:365 weakly excites
- L14:1706 weakly suppresses
- L18:2584 weakly excites
- L14:1507 weakly excites
- L16:1491 weakly suppresses
- L17:481 weakly excites
- L16:2402 weakly excites
- L16:2998 weakly suppresses
- L14:914 weakly excites
- L16:2106 weakly excites
- L14:2832 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
- esome
- IDA
- newList
- \a
- ows
- adow
- fixed
- TINGS
- azing
- ANGED
Suppresses
- veces
- оличество
- RefCount
- salt
- Promotion
- sem
- inness
- reviews
- يون
- 盐