Discussing the 2017 Scaling Laws Paper

teortaxesTex · x · 2026-07-19

This post discusses the 2017 paper "Deep Learning Scaling is Predictable, Empirically" and its author composition. The conversation notes that: - Back in 2017, it wasn't surprising that Baidu wasn't doing much "foundational AI" work. - It was also pointed out that there are no Chinese authors on this "Scaling Laws" paper, though it's unclear if this should be interpreted as a stronger signal. The attached image shows the paper's arXiv page. Its core conclusion is that generalization error in deep learning scales as a power law with training volume, meaning increasing model size, data, and compute will systematically improve results.

Original post →

More from Research

Research channel →