Stumpokapow
listen to the mad man
Gary Whitta said:I would expect a ton of videos and stories starting Friday based on all the things Siri CAN'T do and calling it overhyped. Because as smart as it is, I imagine it's pretty easy to confuse it if you ask it really complex questions. We'll see.
As far as I can tell, Siri does the following:
STEP 1: Convert speech to text.
Its success at being able to do this will be measured in... its success at being able to do this. If I say "What's the oats of Chinese pigment octopus purses?" and it correctly parses the words, it's done its job there even though the phrase is nonsense. Given Apple's cautious rollout in terms of languages, I suspect this is half-decent. I live in an area notorious for difficult to understand accents (people here are frequently subtitled on national news, even when speaking English) so it should be a good place to test things
STEP 2: Contextually determine what I'm asking it to do.
There are only two answers here:
1) Manipulate an iOS system application. Its success at doing this will be measured by whether or not it correctly guesses that what I'm asking can be done by an iOS application, and whether or not it correctly translates the words into action. I am relatively certain that Apple has done a good job of getting Siri to understand synonyms, so I can say "Set up a reminder every Monday at noon for me to eat" versus "Set up a recurring reminder Mondays at noon for me to eat" versus "Remind me every Monday at lunch time to eat" and I'll bet it'll get all three correctly.
2) Asks the internet for information. Here's going to be the controversial bit. As best as I can tell they're using Wolfram for this. If you ask "How many Japanese Yen in a US Dollar?", it asks Wolfram or wherever, gives you the result. Siri isn't actually knowing or looking up the information, it's just querying an API and trusting the result. So if I ask something that Wolfram can't understand, it's not going to work. If I were to ask a complex question like "In general, how popular were re-elected US presidents 30 months into their term?", Siri could successfully parse the question, successfully realize that it needs to ask the internet, and still fail because Wolfram will fail.
So unfortunately whether or not Siri "succeeds" for a lot of people will depend on external services. Hopefully Apple has invested in as many key partnerships as possible in order to get the best results possible, if they're not willing to invest themselves into knowledge systems.
Marty Chinn said:But that's not any different than a language parser where it uses a combination of phrases to lead to a result.
Right, but that doesn't change the fact that it's a big deal if the combination of their software and their training data results in a language parser that understands the great bulk of all phrases people use to trigger a given result (or intelligently adapts to user feedback and learns as time goes on), while it's not a big deal if it simply understands one or two pre-programmed phrases.
One of the most common definitions of an AI is a Turing Test, even if it's possible to imagine a system that passes the Turing Test but isn't very intelligent beyond that.