Andy Pernsteiner, VAST Data | Supercloud 4
summary
Argues that widespread synthetic outputs risk model collapse, so organizations must capture and curate and cleanse vast amounts of real-world data for training; notes cloud GPU hours (H100-class) are scarce and costly so teams push data prep off-GPU, using GPUs for ETL and training on providers like CoreWeave, G42 and Lambda; contrasts table formats (Delta/Iceberg/Hudi on Parquet) with unstructured needs and promotes GPU direct storage and a NAS-style platform that proves fast enough for large-scale deep learning.
AI-written summary · about
