At the end of the day, I fully expect large-n Hanoi and all these things to end up as yet another benchmark. Like all the needles-in-haystack or spelling tests that people used to show shortcomings of LLMs and that were actually just technical implementation artefacts and got solved pretty fast by integrating that kind of problem into training. LLMs will always have to use a slightly different approach to reasoning than humans because of these technical aspects, but that doesn't mean that they are fundamentally inferior or something. It only means we can't rely on human training data forever and have to look more towards stuff like RL.
People also like to forget that from the dawn of modern computing and AI research like 60 years ago all the way to 7 years ago, the best models in the world could barely form a few coherent sentences. If LLMs are this century's transistor, we are barely beyond the point of building-sized computers that are trying to find normal life applications.
I have a blind friend that's been getting value from the Meta Ray-Ban spectacles. He used to have spectacles that let a human see where he was going and offer suggestions/directions/descriptions/etc. but replaced those with this somewhat more private solution and found various compelling (in his descriptions to me) use cases. I personally think AI is going to land on "coding" and "accessibility" successfully and fail in most other domains.
I know the NoSQL world had some progress in this space. RethinkDB was going in this direction around a decade ago if I recall. I would really like this from a modern relational database or data warehouse though. Polling sucks.
Schema-less is a bit wrong term, since you always end up with one but, indeed, schema-on-read vs schema-on-write discussion for individual use cases is far from a settled thing, even though the zeitgeist is that schema-on-write won sometime in the 70s.
You say "prototyping", but I would generalize that to "faster evolving" in the long term. Of course, not without tradeoffs.
I personally have zero issues writing a custom marshaller when needed for any the schema-less document collections I have. Constraints your application has on the data are a superset of db schema anyways.
Just to add to sibling comments, one killer feature for me is ChangeStreams[1]. It's miles ahead of what Postgres[2] offers, and it enables really interesting use cases. Some of my services built around ChangeStreams end up not doing a single query to the DB. Data is right there in the program memory, indexed how I need it to be, and program is immediately reactive to any changes in the DB.
OpenResty with Redis on the same system - both Lua-powered - is a god-like power.
Ability to do basically control and rewrite all your ingress AND egress and have some state via Redis over UNIX sockets.
What's not to love? After all, it's what powers the bulk of internet.
Granted, it sounds dreamier than it is, since a lot of the contraptions you could come up with might be better placed at application layer. But having operational ability to do these things ad-hoc, security advantages of manipulating at your own infra edge or performance boosts, sure does come in handy!
We did this on production (still runs at webscale) but we abandoned dev effort on it for ${reasons}. I _loved_ it though :) Hook Lua on the HTTP event processing stages and add icing on the cake with Lua modules to hand-off non-trivial stuffs..
