A self-published paper on Zenodo claims memory technology that stores 447 TB per square centimeter with zero energy needed to retain data. Author Ilia Toli says single-layer fluorographane can encode information in the orientation of individual fluorine atoms on a carbon scaffold, with thermal bit-flip rates around 10^-65 per second. He projects a full-scale system could hit 25 PB/s throughput using near-field mid-infrared arrays. If real, it would address the memory wall strangling zettaflop supercomputer performance. That's a lot of ifs.
The credibility problems are hard to ignore. The paper has 53 revisions and lists a Gmail address for contact. Toli claims affiliations with CSU Global, University of Central Florida, and San Jose State University, but no faculty directories or institutional profiles confirm any of these connections. More troubling: "fluorographane" returns zero results in scientific literature. "Fluorographene" is a real compound. The foundational material in the paper might be a typo.
The working prototype uses a scanning probe to read and write bits. The paper doesn't specify the probe type. That technique gives you incredible density but glacially slow I/O. The 25 PB/s throughput projection assumes array configurations that haven't been built, using read-write methods that haven't been demonstrated at scale. The gap between what exists and what's projected is enormous.
The memory wall is a real bottleneck for AI systems. We need genuine breakthroughs in LLM inference. A claim this consequential, from an author with no verifiable academic presence, self-published without peer review, revised 53 times on a preprint server, with a potentially fictional material name, doesn't meet the bar. Extraordinary hardware claims need extraordinary proof, and the proof here is thin.