Compression is not information loss, but a test of prediction—and a path to AGI
yunta_tsai · x · 2026-07-23
People often treat compression as information loss, but this post argues the opposite: good compression separates signal from noise while preserving the structure of the domain.
It uses video compression as an analogy, saying the model spends compute on predicting motion and only budgets the remaining bits to encode it. The closing claim is that if a system’s predictions consistently beat the observed universe, that is AGI.
More from AGI Musings
- Sycophancy can stop an AI from disproving your conjecture — joshwhiton · 2026-07-23
- Matt Yglesias debates AI’s labor-market impact on a new podcast episode — binarybits · 2026-07-23
- Agentic AI is moving from assistant to systems that can do part of the work — RiciglianoMarceloE · 2026-07-23
- A long post argues AI alignment should maximize human autonomy through truth — GlenBradley · 2026-07-23
- Research explores antibodies that may help protect people from cancer — EricTopol · 2026-07-23
- Allison Duettmann imagines emulated time and many lives beyond one human lifespan — juanbenet · 2026-07-23