Re:Understanding Positional Embeddings in Modern LLMs
by skew-aberration on Sunday November 15, 2026 @08:07PM
(#50512788)
Regardless, that line of reasoning would still seem counterproductive to me, even if it did predict electrons.
We have found particles and empirically identified their symmetries. To model the propagations of the particles' fields, we replace those discrete symmetries by corresponding continuous ones. Then, we are shocked to find that we can work backwards from the continuous symmetries to the discrete ones again!
The continuous symmetries are widely taken to be the fundamental thing, because they make the math of Lagrangian mechanics easier, and in QFT we can 'quantize the field'. I don't this stands up to philosophical scrutiny though. We don't actually have 'electrons rotated by 10 degrees', we only have 'electron field excitation with momentum p, rotated by 10 degrees phase'.
We have found particles and empirically identified their symmetries. To model the propagations of the particles' fields, we replace those discrete symmetries by corresponding continuous ones. Then, we are shocked to find that we can work backwards from the continuous symmetries to the discrete ones again!
The continuous symmetries are widely taken to be the fundamental thing, because they make the math of Lagrangian mechanics easier, and in QFT we can 'quantize the field'. I don't this stands up to philosophical scrutiny though. We don't actually have 'electrons rotated by 10 degrees', we only have 'electron field excitation with momentum p, rotated by 10 degrees phase'.