Pinned post
Hey, I wrote a blog - I think it's fun, hope you do too!
A tiny implementation detail in low-precision arithmetic could be biasing your AI training 😲View quoted post on X
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 addressed with one simple fix 🎲
Check it out in the link below! 👇
