22 comments

[ 2.9 ms ] story [ 58.2 ms ] thread
"Eerie" "creepy" "scary"... You keep using that word. I do not think it means what you think it means…

Did someone pay gizmodo to write a faux-praise article, or is this shlock par for the course?

It's Gawker Media, of course shlock is their bread and butter.
How private is it? Does it send keystrokes?
As far as I've read, all keyboards have full access, so they send all info and keystrokes back to their servers
It will be interesting to see what happens to language as these technologies get better. Would you rather type out your ideal word, or tap the suggestion that is close enough?

If this type of good prediction grows widespread it could be a force toward standardization/avoidance of language drift. it could also discourage more in depth conversation, which may be less likely to get favorable auto-suggest.

I do it a lot for Japanese. You need to select a suggestion to get the kanji right, and from there on you're already looking at the suggestion box, so might as well choose from there if it's close enough.
This is a really interesting idea that I hadn't thought of. It could have implications for the ability of people to express their ideas, and thus on their imaginations. It reminds me a little of Newspeak. Not that I'm suggesting that SwiftKey's neural net is some sort of Orwellian thought-controlling tool by design or mistake, but it's an interesting idea that as we get the benefit of accurate autocorrect (I don't actually use autocorrect on any of my devices), we start to lose both the individuality of expression, and then the imagination that fuels that expression.
Keyception.

It was a dark and stormy night, and I will be in touch with you to discuss the position with you and your family are doing well and that you are not the only one I have to say that I am a beautiful person and I love you and I love you and I love you (repeats)

What is the best way to get the money to you and I will be in touch with you to discuss the position

Who will be able to get the job done and I will be in touch with you

Why I was not able to get to the point where I want to be a part of the team and the team will be able to get the job done

It's interesting to see how simplistic these are at the moment
Probably just a very limited set of bootstrap training data. I'm curious what keyception would look like after a few months of using it.
Only english, or also for other languages?
English, from what I can see.
I wonder if it can take advantage of Qualcomm's Zeroth neural processor in the Snapdragon 820.
Swiftkey is great except it does not work at my iPhone. Before neural network enhancements, they should work on the basics.

https://www.google.com/search?q=swiftkey+iphone+issues

https://support.swiftkey.com/hc/en-us/articles/201310402-Why...

As per your first link, that's a problem with iOS and the brain dead way they implemented third party keyboards, not Swiftkey. All of them are a little bit glitchy.
Same here. Tried if for a long time until I couldn't bare it anymore and rolled back to the iOS default keyboard.
Does anyone know of a predictive keyboard for the desktop, something geared for natural language and not just IDE's code autocomplete.
Not for English, but for Japanese and Chinese it is fairly common to have predictive elements.

Even the standard IME in at least OSX has predictive elements.

Thanks, I didn't know that. Would love to have something similar for other languages, though.
Swiftkey's flow still sucks compared to Swype.
I thought that people already stopped using Swiftkey.
It's worth noting that neural networks has been used for Chinese and Japanese keyboards for some years now.

Here's a nice (and easy to read, even if you don't know how written Japanese works) paper on the structure of a Japanese IME called Simeji by Baidu:

Wu, Xianchao, Rixin Xiao, and Xiaoxin Chen. “Using the Web to Train a Mobile Device Oriented Japanese Input Method Editor,” 2013. http://www.aclweb.org/anthology/I/I13/I13-1172.pdf