I doubt it'll ever work reliably. Language is subtle and irregular lexically and semantically speaking. You can't conclusively define a language well enough for there to be any determinism in interpretation of it. If you look at how humans react to language which is carefully crafted to have meaning other than what is obviously written or uses idiomatic or metaphorical speech rather than simple conceptual speech then it gets even dirtier.
What pops out is a statistical approximation of the information that went in. If you include a metaphor which is obvious but has been contextually adapted then it'll throw the weights off and garbage comes out every single time.
Statistical approximations also have a significant error. The current error margin is 30-40% on the best models on even basic things. That is large enough with low enough confidence that validation is required for every output, which is likely more costly than doing the work yourself in the first place, at least from experience.
Add to that, the cost of generation and execution of the models grows exponentially whereas the error decline is almost linear, suggests it'll never converge on a usable product. We're already hitting a hard wall now and investors are losing interest in propping it up.
Really though, if you assume that 1 in 3 things the summary says is bullcrap and you have to Google everything and find a source and check it yourself in case it's talking bullcrap, is that a viable technology?
Nope!
And trust it you must not from day one. Unless what you are creating or using it for is of zero value.
www.reddit.com
