We’re not talking about eloquence in translation here, we are talking about getting across functional semantics. That will happen extremely quickly. I tried the app and it works OK for Spanish to English. My Spanish is very basic - AP Level highschool with no real practice since, but to me it worked.
As for replacing human translation altogether, the crowdsourcing will get better, as will the machine translations. There are only two real limiting factors to having the machine able to this perfectly: Horsepower and connection speed/reliability.
In mobile, batteries and heat are bounding horsepower. Pace of carrier investment bounds connections. These connections are needed to off-board heavy processing to a bigger machine “in the cloud” and then get back the result on the mobile.
Dragon’s Dictation on the iPhone/iPod/iPad is better than the desktop version because it off-boards all of the computation to a big machine somewhere else.
There are many computers in the world today with enough horsepower to do perfect translation better than any human can in terms of speed and error rate. The things keeping these machines from being accesible for this purpose is the direct cost of running that horsepower, the opportunity cost of what else that horsepower can do, and thus the fact that giving the public access to these machines in exchange for looking at ads is not commercially viable right now.
Thus, take into account that a Google Translate request is running on a server instance that has about the same amount of processing power and memory as a high-end consumer desktop, and that machine has to serve thousands to tens of thousands of these concurrent requests.
Also, take into account the sheer body of brain power being committed to the “semantic web”. From what I saw while in Boston, it seemed a sizeable percentage of Harvard & MIT etc. math and computer science hotshots were doing some work in this space as are others coming from the top research universities in the world in that space. As this gets better, the meta information on passages gets better and thus machine translation gets better.
It’s hard to guess how soon before we see translation of the most nuanced passages and writings totally automated, but it’s inevitably going to happen. One thing to keep in mind to put this in perspective is that personal computers in the home have only really existed for a bit more 25 years and in any real consumer penetration for a little more than 15 years. Oh and once again, that mobile as real platform has only really existed for around 30 months.