Aside from what they say, they still can’t get their iOS AI based spelling correction up to par.
Why Apple believes it’s an AI leader–and why it says critics have it all wrong
131–140 of 217 posts
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#132(Googler here) I feel article presents a false dichotomy/view (imho borderline disingenuous impression) that Apple is doing on-device, and everyone else is doing on-cloud.
Android also does on-device. A lot of the time, things start out on-cloud, until whatever machine learning model can be shrunk to run on-device at the right performance. So you see for example, text-to-speech and speech-to-text start out as cloud calls, and now they're on-device. Google Translate ran in the cloud, but now for some languages, it happens on-device.
Things like Google's "Live Transcription/Caption" on Android wouldn't work if it wasn't on-device.
Apple similarly went to the cloud for Siri speech recognition and TTS until they could run it locally.
For other things which need large models, there is Federated Learning to preserve privacy. Google Keyboard has been using Federated Learning for some time now.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#133Earlier quoted context omitted.
I know exactly what you mean! This is even more annoying in German, where every noun is capitalised.
Why would it be more annoying? Seems like it would be harder to notice this error in German.
It's more annoying because when you type a non-noun there is a higher chance that it gets mistakenly autocorrected to something capitalised (as every noun in German is capitalised versus only proper nouns in most other languages). I hate this flaw of iOS as well.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#134Earlier quoted context omitted.
I'm bilingual and write approximately the same amount of messages in both languages using iOS. Autocorrect always suggests words from BOTH languages mid sentence. Like what the hell, they should be able to infer which language I'm using just from the first word.
That sort of makes sense. I am not sure of it's prevalence, but at least for Indians, it's pretty common to code switch while writing (even more so while speaking).
Yes, I only speak one language fluently but I'll still arbitrarily code switch sometimes.
Human language is actually just one vast system, it isn't best understood as a bunch of nicely compartmentalized independent languages. So in a way it would actually be more wrong for an AI model to insist that "Beacoup money" isn't a reasonable thing for the user to want to write, just because those two vocabulary words don't co-exist in a single defined language. The user may or may not know that, but they certainly don't care.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#135Setting aside all the other work Apple is doing in this area, can someone explain why iMessage autocorrect is unacceptably bad (to put it diplomatically) in the era of weapons-grade language models? Its behavior is totally bizarre. It’s like an underpowered chess engine that makes flagrant blunders: capitalizing random words in the middle of a sentence (rock -> Rock), the context sensitivity of an actual rock, forcin…
To be fair, the android keyboards (like Gboard, googles own) are also bad. It's a tough problem for sure, but they haven't gotten better for the last 6-7 years at least, in my experience, and that is the point of the parent post I guess.
The worst thing for me is that I frequently miss the space character and type a period instead, and Gboard can't understand that it's a mistake AND won't autocorrect and suggest words for what I'm typing after that period.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#136Setting aside all the other work Apple is doing in this area, can someone explain why iMessage autocorrect is unacceptably bad (to put it diplomatically) in the era of weapons-grade language models? Its behavior is totally bizarre. It’s like an underpowered chess engine that makes flagrant blunders: capitalizing random words in the middle of a sentence (rock -> Rock), the context sensitivity of an actual rock, forcin…
Compare Google and Apple's privacy policies. Autocorrection and autocompletion is far easier if you're willing to hoover up vast amounts of user data to train your models and provide personalisation. It's a similar story with Siri and Google Assistant.
Are you saying this is an unsolvable conundrum? You can't have good autocorrection with massive privacy violations?
I'm not sure I believe that.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#137Setting aside all the other work Apple is doing in this area, can someone explain why iMessage autocorrect is unacceptably bad (to put it diplomatically) in the era of weapons-grade language models? Its behavior is totally bizarre. It’s like an underpowered chess engine that makes flagrant blunders: capitalizing random words in the middle of a sentence (rock -> Rock), the context sensitivity of an actual rock, forcin…
Many teams at Apple are pure junk. When Steve Jobs basically fired iCloud folks in public, lot of people thought he went too far but little they knew how bad that team really was and still is. These folks are now center of the Apple universe but can't build a simple app that can efficiently sync the photos on Windows. Other teams are iTunes (one of the worst widely deployed Windows app ever), Apple Music (just look a…
As incredible as it is, the problems extend way beyond just syncing on Windows.
Sync works completely randomly on mac between the iPhone and and OSX, - you often have to wait "a long time", and since Apple also has the completely idiotic concept of "everything has to be magic" you never get an estimate, or an error message, or a progress bar, so like everything on a mac, you have to just fiddle, disconnect, reconnect, wait, and just hope and pray that things will work out.
And then we haven't even touched on the their productivity apps like Numbers and Pages which for a long time had features in the templates which weren't available in new documents, so you had to start from a template, completely insane.
Still use Mac and like it for the most part compared at least to Windows which can't even get search right without showing randoms ads and internet results to you - another pretty crazy thing to push to production for a billion people.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#138Did the interviewer not know enough about ML to even challenge the bogus statements.
Yes, apple’s strategy is more privacy protecting.
But holy hell, yes larger data sets are going to be more accurate and the resultant model doesn’t have to run in the cloud, it can run locally.
AI training in a data center with large data sets, ship the model to local devices to execute.
That is ALWAYS going to be better and more accurate than what Apple is doing.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#139Earlier quoted context omitted.
This also happens to me on Android. Working in ML, this is something to point to when people hype it up. We've come a long way, but we still got a long way to go. As for specific errors, does rock capitalize when "the" is before it? (The Rock) The models learn from how users write. But when you think about it, it is actually fairly difficult to predict what words are going to be, you can test it with your friends. Ha…
Could autocorrect not wait until it sees more of the sentence?
Having to monitor several words back to fix errors that the system adds is exhausting. I wish I could turn it all off, or at least revert to the system from the early days of iOS.
Re: Why Apple believes it’s an AI leader–and why it says critics have it all wrong
#140Earlier quoted context omitted.
Compare Google and Apple's privacy policies. Autocorrection and autocompletion is far easier if you're willing to hoover up vast amounts of user data to train your models and provide personalisation. It's a similar story with Siri and Google Assistant.
Are you saying this is an unsolvable conundrum? You can't have good autocorrection with massive privacy violations? I'm not sure I believe that.
> I'm not sure I believe that.
So, what would make you believe it?