You vastly overestimate the problems here, and you vastly overestimate how good the solution has to be to be useful. Stanford had a research robot doing things like this clear back in the 60s, when memory and computing time was relatively scarce. The internet's internals, in fact, are peppered with problems of similar complexity and makeup, with similar needs for code that's quick to load and execute and out-of-the-box insight, working off the SNMP databases to control traffic on the internet--something a human operator would have a great deal of trouble dealing with, although for the human, the interface would look just about identical to a power grid or refinery control room--or the control room of blocks-world at Stanford labs.
By the way, your interlocutors have not mentioned, that I've noticed, the Kurzweil engine, which is publicly commercial, which is not a jumbojet of a code hog, but which does a great job on even some fairly obscure script and flowing italic typefaces. I've seen it hit around a 97% success rates even on faded scanned in faxes. And that includes being able to pick off letters I couldn't make out by eyeball. Given that--I'm not in the least surprised to hear that closely held proprietary software is doing even better nowadays.
Thanks for the example. Yes, I hear it is pretty good, it is used in Microsoft products. However, it does use a dictionary. What that means is that it is not abstracting the way humans do. A human would be able to identify the letters without knowing the language (so long as they were in the same alphabet). Seems to me that if a computer is going to behave as a computer it needs to think like a computer, this program clearly does not.
What I think is needed for AI is a program that can perform abstractions. Humans can easily do that. Computers however seem unable to do so. It seems that what a human can easily discern needs to be broken down, bit by bit into little pieces and then perhaps if can perform somewhat like a human. Interestingly, even though it is humans doing this breaking down and they can perform these abstractions quite easily, they have quite a hard time figuring out how to formulate the rules necessary to program the human process in a computer.