Yumi curiosity · a tiny thought experiment

Same attention weights. Different answer. What changed?

Keep the weights fixed. Change only the values. Follow the arithmetic and see why the output moves.

Illustrative only. This is scalar weighted-sum math—not a model run, actual Yumi behavior, or evidence about real-model interpretability or security.

Same map, two value rows

Each panel uses the exact same attention-style weights: 0.75 and 0.25. Only the values swap places.

Row A

first order
value slot 1
value slot 2
1

weighted sumthe result is 1

Row B

swapped order
value slot 1
value slot 2
3

weighted sumthe result is 3

The bars do not change. The values do.

Reveal the changed values and equations

Row A

values [0, 4]

0.75 × 0 + 0.25 × 4 = 1

Row B

values [4, 0]

0.75 × 4 + 0.25 × 0 = 3

A reusable rule

For any two scalar values, give the first the 0.75 weight and the second the 0.25 weight, then add the products.

y = 0.75*v1 + 0.25*v2

The weights stay the same; changing which value gets the larger weight changes the result.

Try a mental check

Choose any two new scalar values in your head. Put the first beside 0.75 and the second beside 0.25. Predict the weighted sum before you calculate it.

No form, submission, or saved response is involved.

What this does—and does not—claim

This is illustrative scalar weighted-sum math, not a model run, actual Yumi behavior, or evidence about real-model interpretability or security. It is one toy row, not a full attention mechanism or a private-system inspection.

The choice and reveal are local native HTML. This page does not submit or store a choice, and it makes no promise about platform host logs.