This episode explores how a computational physics startup uses Herbie, an open-source tool from the University of Washington, to automatically detect and rewrite imprecise floating-point expressions. Hosts Lucas and Luna walk through a concrete example: a Naive Bayes classifier whose probability calculations were losing precision in the fifth decimal place due to catastrophic cancellation. Herbie rewrote the expression in under a minute, reducing error from 12% to 0.003% with no change in model logic. The discussion covers why floating-point errors matter in modern machine learning, how Herbie applies numerical analysis heuristics and Monte Carlo sampling, and the trade-offs between readability and precision. Relevant for engineers working on scientific computing, GPU kernels, or any system where tiny errors compound over millions of iterations.

#FloatingPoint #NumericalAccuracy #Herbie #NaiveBayes #CatastrophicCancellation #MachineLearning #ScientificComputing #OpenSource #UniversityOfWashington #Precision #GPU #ComputationalPhysics #Startup #TechnicalPodcast #BusinessAndTechnology #FexingoBusiness #BusinessPodcast #TheTechnicalCoFounder

Keep every episode free: buymeacoffee.com/fexingo