> The trainers are a group of a few dozen human contractors working in Facebook’s headquarters. These humans, as Marcus explained it, review each of the AI’s responses and decide whether they’re good enough to pass along, or whether they need revising. When the trainers revise, the AI watches and learns: “With every single one of these interactions, that data is fed back into that AI instance that will then use the data to learn how to automate more and more answers or how to get better and better at those things,” Marcus said.
The unfortunate side effect of this is that the training elements won't work on a larger scale. Yes, the underlying AI can, but the training system can't handle thousands of requests per day.
People previewing M post things like this (https://twitter.com/neilkod/status/654442151837831168?ref_sr...) where M does "magical" things like lowering Comcast bills, but this is all stuff that trainers would either have to intervene for or train some system for. The result is that people beg for it to go public, expecting this level of assistance -- even though M won't have nearly the same level of "intelligence" on the backend.
Google Now, Siri, and Cortana could all have the same level of "usefulness" if there was an army of "trainers" on the backend, constantly adding new features for very odd requests, but they wouldn't respond quickly and probably wouldn't be available.
They do claim the quality of responses is improving. Will it get to the point where they can send responses without trainer approval? Who knows, but if they can get to that point for a majority of responses then it might work.
The only thing that has me confused is why Apple, Amazon, Google and MSFT aren't doing this .. or at least one of them. It seems like the right approach, but I'm surprised by the lack of competition and wondering if there is a reason for it.
I guess they're in wait and see mode. I guess theoretically, if it works out you could always throw 100s of millions at contractors to "train". My only thought though is that there are a lot of automated toolsets being built that aren't typical deep learning. Things like APIs / screen scraping / automated voice / etc etc. You can't build that sort of stuff over night.
They've been doing significant, important stories for at least the last 2 years, after having built up what is now maybe one of the largest investigative teams with several Pulitzer winners:
Buried underneath the ghastly clickbait, they've actually got some pretty decent journalism and long-form articles. The good stuff's hard to sort from the crap, unfortunately.
(A while back there was an exemplary piece of in-depth analysis on some space phenomenon. I've just spent a good 20 minutes trying to find it, so I could post it as an example, and have failed.)
Not to mention an article that isn't total crap!? It is pleasantly surprising to read an article from buzzfeed and have it not be dredged up sludge of the internet.
> Wit.ai works bottom-up. Let your end-users express their intents freely, and have your app learn from that and adapt. Not the other way around.
> Each item in your Inbox is something you can act on in order to improve the accuracy of your Wit.ai app. A good practice is to start having beta users test your app as soon as possible, even if the initial configuration is the bare minimum with just a few intents, and then use the Inbox to improve your app.
> The trainers are a group of a few dozen human contractors working in Facebook’s headquarters. These humans, as Marcus explained it, review each of the AI’s responses and decide whether they’re good enough to pass along, or whether they need revising.
So the AI of the future is being trained by a group of stressed-out and underpaid humans screening machine output as quickly as possible. How could this possibly go wrong?
The unfortunate side effect of this is that the training elements won't work on a larger scale. Yes, the underlying AI can, but the training system can't handle thousands of requests per day.
People previewing M post things like this (https://twitter.com/neilkod/status/654442151837831168?ref_sr...) where M does "magical" things like lowering Comcast bills, but this is all stuff that trainers would either have to intervene for or train some system for. The result is that people beg for it to go public, expecting this level of assistance -- even though M won't have nearly the same level of "intelligence" on the backend.
Google Now, Siri, and Cortana could all have the same level of "usefulness" if there was an army of "trainers" on the backend, constantly adding new features for very odd requests, but they wouldn't respond quickly and probably wouldn't be available.