LLMs specifically are fine with random bits flipped for the results to be 'creative'.
That's not exactly how LLM temperature works. :). Also that's on inference, not training. Presumably these would be used for training, the latency would be too high for inference.
alt Hacker News
That's not exactly how LLM temperature works. :). Also that's on inference, not training. Presumably these would be used for training, the latency would be too high for inference.