seems like microsoft has multiple layers of ‘safety’ built in (Satya Nadella mentioned on a decoder interview last week). My read on what’s going on is that the output is being classified by another model in realtime which is then deleted if it’s found to violate some threshold.
> Second, then the safety around the model. Ad runtime. We have lots of classifiers around harmful content or bias, which we then catch. And then, of course, the takedown. Ultimately, in the application layer, you also have more of the safety net for it. So this is all going to come down to, I would call it, the everyday engineering practice.
https://www.theverge.com/23589994/microsoft-ceo-satya-nadell... is the full interview