Our next IPI Talk will be on Saturday, 6th of June at 17.00 London time zone.
Live stream link: https://youtube.com/live/XY_YJiukBe8?feature=share
Title: A Fourth Kind of Compression — Decomposing a Fixed Corpus Through Frozen Instruments, and Its Bridge to the Second Law of Infodynamics
Abstract: Information is physical, countable, and I will argue compressible in a way that has not yet been named. I will present three linked ideas and the systems that run them. Derived Data Abundance (DDA): a fixed corpus of real data, projected through a panel of N frozen, approximately-independent measurement instruments (embedders), yields up to N + C(N,2) (+1) structured supervisory signals per input - no synthetic generation, and structurally outside the model-collapse regime that degrades recursive AI training. Meaning compression is the ratio view: structured signal extracted per unit of raw data, which I propose as a fourth category of compression alongside bit-, weight-, and activation-compression. Teleological Constellation Training (TCT): the same frozen panel becomes an alignment target - a "constellation" of identity centroids, each in its own measurement space, never flattened into one vector - and a generator is trained to land inside it, with every output verified geometrically at inference. Where this meets information physics: a well-designed panel is built to maximise differentiated information per instrument - each must add new bits, and a redundant instrument is noise and is retired (Shannon information made an engineering contract, measured in bits against a real outcome). And aligning a generator to a frozen constellation drives its output distribution toward a high-symmetry, low-entropy configuration around the identity centroid - the preferred low-entropy state your second law of infodynamics predicts, here observable in a learning machine. I will show this running across three production systems (panels of 7, 13, and 21 frozen instruments, spanning video, code, and civic matching) and present measured results: a voice reproduced at 0.961 mean speaker-embedding similarity (encoder-matched - I will state the cross-encoder caveat plainly) and a style model that holds character under adversarial prompting (public weights). I will close with a concrete joint experiment: measuring the entropy of generated outputs collapsing toward the constellation across training, open to scrutiny by IPI.
Speaker: Chris Royse
Bio: Chris Royse is the founder of Leapable.ai and socialmedia2.com and the author of "Teleological Constellation Training: Multi-Embedding Decomposition as Meaning Compression" (2026). His work treats a panel of frozen, independent embedders as measurement instruments — decomposing a fixed corpus into structured supervisory signals (Derived Data Abundance) and using the same frozen panel as a geometric alignment target for identity-locked generation. He has applied the method across text, video, and civic matching, reproducing a voice at 0.961 speaker-embedding similarity and training a style model with public weights. He also wrote "The Symmetry of Knowing," accepted to the 2026 AEJMC conference.
6th of June @ 17.00 London time zone. Live YouTube talk and debates via online Teams after – a link will be emailed to the IPI members.
