Cheng Lou Compares LLMs to Sudoku Iteration
Engineer links repeated model steps to filling puzzle cells and wider reasoning.
TLDR
Cheng Lou, who worked on React.js, Facebook Messenger, ReasonML, and ReScript and now works at Midjourney, replied that LLMs and diffusion models iterate through trials and errors. He noted that a sudoku with one extra cell filled remains valid, so iteration can yield more data or signal. He added that the approach may extend to thinking in general, not only to producing nice results.
Combined views
28.1K
4 Sources, first seen 28d ago