👋 Hello! I've, been working on something quite exciting over the past year at #GoogleDeepMind and I can now share part of it with all of you.
"#Orbax: Distributed Checkpointing with #JAX" describes how at Google we have designed a standardized checkpointing solution for high-performance distributed ML systems.
How do you enable fast recovery after a preemption? Especially for massive models? And how about resharding? We often think about how to optimise the transferring of data from host to accelerators. But how about storage?
Highly recommended read. And the best thing? The library is open source 🎉
https://t.co/NBJ1UbyVe4
@dhh Your whole thesisis is built on a comparison between an ethnic minority to a group of animals.
Do I need to remind you why this is bad, and how badly it went last time we did this in Europe
@dhh "Leading by example" only works if the example actually works. No developing nation is going to sacrifice its GDP and comfort just to get a pat on the back from Brussels.
@Pirat_Nation can you elaborate on how this contributes to thousands of tonnes of carbon emissions globally? How's this different than downloading any other thing as part of a software you are using
@ivanburazin GChat is the dive bar of corporate comms: no frills, slightly depressing lighting, and you’re only there because you’re already in the building. Slack is a $20 cocktail in a glass shaped like a bird; it’s fun until you realize you’re paying for the decor with your focus
@sama If it’s just fast and cheap, it’s a better tool for the tasks we already do. If it’s smarter, it starts doing the tasks we don't even know how to describe yet.