Pinned
A tiny implementation detail in low-precision arithmetic could be biasing your AI training 😲
This interactive deep dive from Graphcore Research's @Awfidius uncovers a subtle failure mode in stochastic rounding that only appears when randomness is limited, and how it can be

