If i say "find me a good ticket for Chicago tonight". How
will it know that i'm talking about the rock band that's
playing tonight in Paris, France, since there's absolutely
not a chance that a human being asks for a plane ticket for
such a big trip just a few hours in advance ?
I don't even need an assistant system to figure out based on context which I'm talking about: All I want is one that can simply ask clarifying questions. "Did you mean Chicago the city, or Chicago the band?"I had an amusing Siri interaction a few months ago where I asked "what language is spoken in Shanghai?" (I can never remember if it's Mandarin or Cantonese). Siri came back with a very well-reasoned, completely out-of-place answer of "Hindi": https://twitter.com/claymill/status/690386752683577344
> out-of-place answer of "Hindi"
it's using wolframalpha - which interprets 'Shanghai' as a movie made in India. this is probably because it doesn't have a language field for cities but does for movies.Funny, your "absolutely not a chance" interpretation was mine. Maybe there was a death in the family. Maybe you need to visit a friend short notice. Maybe I'm being a bit pedantic, but I was with you until that point. :)
I think you discount how often the plane ticket to Chicago might happen. On total number of requests spread across the planet, I think the plane ticket might be the more common one. The point is, we don't care about the rest of the planet, at least not nearly as much as we do about our local surroundings. The way people answer that query is by using their continually updated knowledge about the state of the local system. If the band Chicago isn't playing around your current location soon, then the correct way to interpret that query would be that you needed a plane ticket (as anyone around you would probably think). This is why I think Google is really in a better position to answer these queries, because the knowledge graph allows the context to provide the nuance you are looking for.
I agree completely. That's why the viv presentation left me a bit unsatisfied. I also had the exact same thought in the middle of their talk : "with google database of concepts, you could do so much more".
But even Google still needs to break the last barrier, and it's related to language itself and culture. Every translator and linguistic knows that a language conveys much more than clear information. It is a mental construct that evolved during thousands of years, and keeps evolving, is full of ambiguities than only culture, experience (and a bit of common sense) can solve.
The best example of this is to notice how bad automatic translators are. Guys at google tried to brute-force the problem, with a gigantic database of text, yet the result is absolutely miserable.
I suppose. I am fluent in two languages (also Spanish) so I'm in a position to be able to judge the quality of a translation, and I would say that pasting in blocks of text to translate them these days yields a result that's quite a bit better than miserable, thought indeed not flawless yet.
Maybe we should mostly think about AI products/tools as just another (weird) friend who is good at answering certain questions for us. At the end, we don't even expect any of our friends to be able to answer all of our questions. Indeed, we pick the right friend for right question.
It probably is. Some representation of the query goes into that module, and a strategy for answering it comes out. That's probably going to look something like what comes out of an SQL query planner, but with more references to external data sources, and possibly queries followed by more queries.
You should be able to say "Viv, explain what you just did" and get back a natural language representation of that program. Then tell it what you meant.
This would be a useful feature on a Bloomberg terminal.
Why? What type of queries can you perform on those terminals that would be useful to analyze in this way?
The Bloomberg interface is horrible and insane, and even relatively simple things like "ok well if I change the vol on this European equity option and you give me a new price...i still need to know which method you used to solve..."
There are plenty of other models in there that use much more complex pricing algos, and it would be useful to be like "Trinomial blah blah assume vol calculated as daily variance over last 30 days"
1. do we really want a centralised service for everything that has access to a lot of data within a walled garden?
2. Siri only works well with North American accents afaik (maybe all western accents?). Alexa's voice recognition is the best and I hope Viv does a similar job.
3. This is mostly a personal preference, but for online shopping I prefer using a lot of filters. This UI doesn't really help with that. Also, HN probably knows better but is the "computed plan" shown in the demo, different from an execution plan created by an RDBMS?
Apparently not[1], but I imagine it's gotten quite a bit better since then.
1: http://www.dailymail.co.uk/news/article-2053684/Scottish-iPh...
I think the key here is 1) evolution - as he said it will get better with time. Just like wikipedia was discarded by many intelligent people in its infancy, similarly now it may feel not so much revolutionary - but it will become something big. And 2) modularity/good design so it's easy to add "improvement" layers.
Some nodes/areas can be backed with deep AI, some don't need to. At the end you just wrap the whole thing in AI at the top to improve itself and we can slowly start calling Viv 21st century's god.
...or evil as some people will probably do. As granpa you'll tell your grandchildren how cool it was when you had to go to the market to pick flowers for your gf, now it's all automated ordering from robots growing flowers.
Viv has to disambiguate the intent with user profile and context, normalize the schema across partner databases, rank partners which support the same verbs, build a partner system which lets them add verbs and domain-specific NL training… there's a lot of grunt to make it work.
Two points :
1/ viv keeps mentioning a "breakthrough" in computer science because they managed to create a program that will scan a network of services and fill the parameters from the query. Now, those guys aren't jokers, so i presume there is definitely something fantastic behind, yet i can't help thinking that once you've "understood" the intent, it sounds quite close to a sql / graphql query planner.
2/ Which leads to the big question : guessing the intent. Every time i hear people mention "intelligence" or "understanding", i show them this : "hey siri, please do NOT set a timer for 4 minutes, i beg you". The fact is, it's just a trick. It recognize words, but it doesn't understand anything, it doesn't have concepts, knowledge or any experience. It's a dumb program that never had any life or sense to understand anything.
If i say "find me a good ticket for Chicago tonight". How will it know that i'm talking about the rock band that's playing tonight in Paris, France, since there's absolutely not a chance that a human being asks for a plane ticket for such a big trip just a few hours in advance ?
This is, to me, the big and interesting question that online assistants makers needs to solve. Now of course, viv is aiming for a product release this year, so they're building an intermediate solution, where developers will manually associate keywords to services, in an "easy" way. And it will sort of work.
Yet i'm still waiting for the real "big" advance.