Big Techday 26: How not to blow up: Training a 400B MoE to 17T tokens without loss spikes - Atkins
About this video
Hand-picked by a BrighterTube curator · found via round-3 back catalogue of a channel Lumen picks video by video · 82% confidenceTeaches training large mixture-of-experts models at scale without loss instability, covering optimization techniques. Conference talk on deep learning architecture and training stability.
A 52-minute video from TNG Technology Consulting GmbH: “Big Techday 26: How not to blow up: Training a 400B MoE to 17T tokens without loss spikes - Atkins”.
Video by TNG Technology Consulting GmbH, played through YouTube's embedded player.
Part of Big Techday 26