unclear: mid-phrase word in mixed text

What this neuron does

This neuron fires on a word in the middle of a phrase, and the words it picks are very different from one another, so no single theme stands out. It goes quiet or negative on digits and codes inside addresses and product numbers. Turning it up made the model's text drift toward a repeated theme, which does not clearly tie back to the neuron.

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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

1. #1 peak

   G sites can differ from cell to cell as different nucleic acid sequences are used to create

2. #2 peak

   -12 structures for a end of the year 2010 total of around a half

3. #3 peak

   45° angle to it. When the sun is lower in the sky, the shadow will be

4. #4 peak

   2 years old to the almost uninhabited stretches of prairie in western Kansas. His

5. #5 peak

   ClippedText property is set to false⏎and also when passing an area of interest via GetText

6. #6 peak

   31-kilometre-long collider; like the Large Hadron Collider, it promises to recreate

Steering

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.

Most negative on

Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.

1. #1 peak

   provider:|| Postal: Katajanokanlaituri 6B, 00160 Helsinki|⏎Web page: http://www.wider.unu.edu/⏎More information

2. #2 peak

   7Availability: In StockCBD Stock No: WW265132<|endoftext|>Ophthalmologists work in hospitals and clinics on land and aboard ships.⏎

3. #3 peak

   Drive, Headington, Oxford, OX3 8QQ⏎Back to previous page<|endoftext|>A waterspout is a cyclonic wind event

4. #4 peak

   have an idea for an applied, data-centric note of no more than 3,000 words, please send a one-paragraph abstract to email@example.com for consideration.⏎This article discusses how the U.S. Department of Housing and Urban

Which neurons in earlier layers feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.

Activation histogram

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 lens

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 layers instead.