The Mi Paradox

A signal in trained models that appeared to offer a route to better compression turned out to be largely an artefact of measurement.

Kind
Paper
Programme
FCN Compression
Published
27 June 2026
Status
Published

What we found

A signal that first looked substantial proved, once corrected, to be small and almost entirely unusable. Most of what first appeared came from how it was measured, not from anything compression could use.

Why it matters

We kept both the attractive signal and the correction that bounded it. A measurement that collapses under correction becomes a guardrail, protecting later work from building on a route that is not real.

The paper

The Mutual-Information Paradox in BF16 Weight Streams

Oliver Cooper, Forlais Group. Published 27 June 2026.

Read the paperPDF, 564 KBDOI 10.5281/zenodo.20961807CC BY 4.0

Cite this record

Forlais Research (2026). The Mi Paradox. Forlais Research Library, record FR-F-03.

@misc{fr-f-03,
  author = {{Forlais Research}},
  title = {{The Mi Paradox}},
  howpublished = {Forlais Research Library, record FR-F-03},
  year = {2026},
  note = {27 June 2026}
}