For a long time we understood nothing of the direction of the MLS. Cool work on finding silent bugs when editing computational graphs using XLA. These silent bugs are some of the worst, technically everything works and nothing crashes, but the results are wrong. It should be like startups, fail fast, otherwise you will waste a lot of energy (and here there are millions of dollars in training). The authors built a different right to the execution automaton: it already chases different versions of the kernel collection in the fastest search, and now simultaneously compares their answers with the default fused kernel (and when it searches for a bug inside the graph - and with an unrealized standard). Almost free to look for problems. Here it is for TPU, but the approach is generally universal. Caught 142 bugs. A big result. XLA from Canals: Fixing TPU Workloads at Scale Ignacio Cano, Amir Saadat, Rishipal Singh Bhatia, Samuel Ginzburg, Patrick C. Toulme, Ioanniszkadas, Praveen Narayananan, Pratik Bha Edits Baha Edgee, Alexander Gumuller, Wang, Wangdben, Wang, Editor-in-chief of Peter Behann Behann Bendern Bendern Bendern Bendern Bendern Berkn Bendern Bendern Bendern Bendern Berkbn Bendern Behann Bendern Bendern Bendern Bendern Bendern Bendern Bendern Narkn Bendern Narkn Narkn Kah; What they did: The authors presented a verification and diagnostics system for Google XLA:TPU machine learning graph organizer for TPU machine code.