Hacker News: SerdarGlhttps://news.ycombinator.com/user?id=SerdarGlHacker News RSShttps://hnrss.org/hnrss v2.1.1Thu, 08 Oct 2026 05:13:37 +0000<![CDATA[New comment by SerdarGl in "Dust: Pretraining Transformers Without Backpropagation"]]>At massive scales 0th order methods will parallelize better than backprop especially along depth, u can train very deep models pipeline parallel without bubbles

]]>Tue, 06 Oct 2026 02:23:54 +0000https://news.ycombinator.com/item?id=49973430SerdarGlhttps://news.ycombinator.com/item?id=49973430https://news.ycombinator.com/item?id=49973430