The process I described would not in any way harm backwards compatibility though. They can still run the same process, but in parallel, so that iteration can happen faster amongst people who care the most.
I suppose you can tell people "hey this service shouldn't be relied on! We might get rid of it one day!". But that erodes trust - people just don't fully internalise what that means, they expect the same level of long term support from all AWS services. Soon the common sentiment will turn into "AWS changes its services all the time" when the truth is "AWS changes the services marked as experimental all the time".
The key is to not give wide access. Google's problem was that anyone could turn on "labs" and use them, and then get really upset when they went away. A few customers getting really early access to a feature under NDA would not be nearly the same.