iOS spellcheck’s massive failure: a symptom of Apple’s bad AI?

scoble-robert

We talk to Robert Scoble about iOS spellcheck: is it really that bad (yes), why it’s that bad (privacy?), and how this will impact Apple’s competitiveness against Google and Android. This is episode #4 of Tech First Draft, interviews with sources that I do live on social, publish later on the TFD podcast, and also form the basis for a story on Forbes.

Who is Robert Scoble?

  • He was a futurist at Rackspace
  • He was an evangelist for Microsoft
  • Scoble was an early (and now recovered) blogger
  • He’s currently chief strategy officer for Infinite Retina, a spatial computing company

Topics and key quotes

In this episode we talk about:

  • Apple’s spellchecking, which seems to be getting worse
  • That’s the opposite of what we expect from tech
  • AI is supposed to use big data to get better fast … or at least steadily … but Apple’s approach to privacy might be inhibiting this

Some key quotes from this episode of Tech First Draft:

  • “80% said it got at least a little worse than before than two years ago”
  • “Swipe is a good new feature that comes with this system, but it’s an immature system. It hasn’t been trained very well.”
  • “I had dinner with the guy who runs Siri at Apple. He said … Google is beating us … we see Google’s learning at a faster rate.”

Listen, watch, or subscribe

Listen in here, or subscribe on any of the major podcasting distributors:

And if you like it, please consider giving Tech First Draft a rating and review.

Or watch on Youtube:
If you prefer, you can watch the full show here as it was broadcast live. Subscribe to my channel to get notified when I go live.

And … a full transcript of the call

John Koetsier: IOS spell check: is it getting worse over time?

This is Tech First Draft: tech news before it’s news. My name is John Koetsier. I write for Forbes. I consult with tech companies. If you’re on live social right now, you’re riding along on an interview for a story. If you’re hearing this on the podcast afterwards, welcome.

We’re talking about Apple’s spellchecking … it’s pretty bad. Maybe getting worse.

My wife got a new phone and she’s wondering, was it the phone? Was it iOS? Why is my spellchecking so bad? This is the opposite of what we expect from technology, especially AI. Supposed to use big data to get better fast, or at least steadily.

So our guest today ran a poll on his Twitter account with more than 400,000 followers. We’re going to talk about those results. He was a futurist at Rackspace. He was an evangelist for Microsoft, an early blogger, and currently chief strategy officer for Infinite Retina, a spatial computing company.

Robert Scoble, welcome.

Robert Scoble: What’s up? I got a new iPhone too!

John Koetsier: Excellent, excellent, wonderful. So thankful that you could make some time. I know you’re super busy starting up a new company doing amazing stuff, and you could join us today.

Robert Scoble: Thank you very much for having me on. Yeah, it’s so funny you know, you laid it out right, that we think that AI is going to make our lives better, but sometimes it makes it worse before it gets better, and this is one of those times, I believe.

John Koetsier: So you ran a Twitter poll, you ran it on iOS spellcheck, and you asked people, is it getting worse? Is it getting better? Is it about the same? What were the results? Or before we even ask that, why’d you run that poll?

Robert Scoble: Because I noticed it’s way worse on my iPhone and I’ve tried several things to fix it, including deleting all the settings and deleting all the history and trying to reboot everything.

And by the way, my new iPhone, I set up clean. I didn’t update it from the old backups. So I am starting out with a fresh OS install. And I thought, oh, I’m seeing a lot of bugs in the spellchecker where it’s putting capitalization where it doesn’t need to go, where it’s switching words a lot more often than it used to. Where it’s misspelling things and where it’s changing things very, very consistently.

One of them is if you use a swear word, it changes it to ducking, right? And it’s like, yeah, that’s a planned bug.

 

You know, one of them is if you use a swear word, it changes it to ducking, right? And it’s like, yeah, that’s a planned bug. Somebody clearly doesn’t want you to write a bad word on iOS. But there’s so many other bugs that aren’t that way and are frustrating. And so I asked my readers, are you seeing the same thing? And sure enough, I think 80% said it got at least a little worse than two years ago. And so something is up.

And there were a couple of really interesting technical answers, that Apple’s new privacy stance of keeping everything on your phone and not comparing it to stuff on the cloud, is what’s behind this. The spellchecking system is just not as mature as it used to be, and to try to protect more of our privacy or whatever, they added a new feature, which actually is a really cool feature, which is swiping your keyboard.

So now there’s several ways to put text into the system. You can actually click a microphone and talk and it turns your voice into text. And it’s pretty good at that. And you can hunt and peck typing, and then you can swipe a word. So as you swipe a word, it turns it into a word as you’re swiping.

And actually swiping can go really fast, but I’m noticing it brought with it a bunch of new errors and it’s not real accurate. So even though it’s really fast to swipe through a sentence, you have to go back and edit it, and then editing it changed as well with this latest iOS.

So it’s a bit frustrating at the moment.

John Koetsier: Yeah, I could totally see that. It took me a while to get used to the new way of editing and moving the cursor. And by the way, I noticed that they switched that a couple of times where it went one way and then they added the other way back and so there was some back and forthing there.

Let’s talk about that privacy safe way.

So we know that Apple’s positioning itself as the privacy big tech company, in opposition to Google, which makes money off of ads. And in opposition to Facebook, which vacuums up all of your data to make money off of ads and other things like that. So Apple’s not there, doesn’t do ads, but when it’s doing all that AI on-device, on that special machine learning chip, is it learning fast enough? Is it only using your data to learn?

If you have to prime the machine for yourself, that’s challenging.

Robert Scoble: AI doesn’t mean it necessarily learns only from you. What it’s doing, AI machine learning is a very good pattern recognizer right? This is why they can do swiping because it can predict when you’re going to hit a certain pattern of letters and then it’ll put it in that word. It’s really good at that. And the training comes down from the cloud onto your local device.

This is why switching that swear word, that F-word to ducking happens consistently on everybody’s device that I’m on.

And then some of it is learning your own pattern and what your favorite words are and what your sentence structure is. And in fact, you can see this starting to happen with Google mail, when it’s starting to auto-answer your email based on your pattern of answers, right? Starting to learn a little bit from you and from the crowd as well. And so it’s a really complex set of technologies that are at play here. And right now we’re getting, I believe they just turned on a new system with this new keyboard.

Swipe is a good new feature that comes with this system, but it’s an immature system. It hasn’t been trained very well, and because they’re doing so much more locally it’s not able to learn from the crowd I predict, than it used to, and that’s frustrating right now.

But Apple is building, I call it a privacy temple.

John Koetsier: Yes.

Robert Scoble: I got to go to Jerusalem and visit where Paul the Baptist grew up and in the middle of the little town where he grew up was a temple, was a building, and the building wasn’t just the church. Everybody thinks temple is church. No. It’s where the government was. That’s where you paid your taxes. That’s where your shopping mall was. It’s where you went and bought your food for the day. That’s where your entertainment was that you took your family to in the evening, and stuff like that. Well that sounds like Apple services, right?

John Koetsier: Two thousand years ago!

Robert Scoble: Podcasting, music, TV, all in one building. All with one services. The floor is your identity. So who are you? It knows that, and then your behaviors are in that temple and are protected from government intrusion. My son’s going to be a cop, so if he wants to get a subpoena to get your data, Apple has got to be able to give him a certain set of data. But it’s trying to do more things on the local phone so you can’t get that that data. Which is, what are you typing?

Apple could put all of that data up on the cloud and make it work really well, but then they would have a huge privacy problem because advertisers and police and governments could get access to that data, and [they are] trying to protect that really well by keeping more of it on the chip phone… it’s an interesting thing to watch.

“Google is beating us.” And I said, so how do you know that? He goes “Well, we instrumented Google, we instrumented ourselves, and we see Google’s learning at a faster rate. The AI is ingesting more data at a faster rate than Apple’s is.”

 

And then you mentioned you wanted to talk a little bit about Siri’s role in this, and you know, three years ago I had dinner with the guy who runs Siri at Apple. And I said, what are you learning about being at Apple with the Siri team? And he said “Well, I’ve learned one thing. Google is beating us.” This was three years ago. And I said, so how do you know that? He goes “Well, we instrumented Google, we instrumented ourselves, and we see Google’s learning at a faster rate. The AI is ingesting more data at a faster rate than Apple’s is.”

And sure enough, three years later, it’s very clear that Google’s AI is ahead of Apple’s AI, that Siri understands more of my utterances and answers more of my questions, and it has more services behind it. It learned more about the world.

It’s still not perfect because you ask either one of those things a question like “How many people are checked in at the San Francisco Ritz Carlton?” There is an answer to that on Foursquare, right? Siri and Google understand what you say. They put the words together correctly. There is an answer to that on Foursquare. Foursquare does have an answer to that, they know how many people are checked in right now at any Ritz in the world or any location, and Foursquare has an API, but neither of them have been hooked up to the API.

So they fail and they go to their search engines. Siri goes to Bing, and Google goes to Google, and give you a stupid answer. And so there’s a separate problem there that we know Siri is being rebuilt. And I think this new keyboard is part of that because the keyboard is really important input into Siri. And the ability to talk to anything on Apple’s iPhone and get text out of it is really an important feature. Works really great with these new AirPods by the way.

John Koetsier: Yeah. Yeah. I’ve got them too.

Robert Scoble: My precious.

John Koetsier: Yes. The challenging part. I mean…

Robert Scoble: We could do a whole show about headphones too you know.

John Koetsier: Yes we could.

Robert Scoble: Which you’re not using anymore because of AirPods.

John Koetsier: The challenging part with Siri in learning slower than Google, it’s just, you’re falling farther and farther behind. You’re competing with one hand tied behind your back because you’re not aggregating all the data. So you’re trying to do this complex thing of some aggregated data that you can take, some personal data that you can take, and put it together, and there’s huge failings in there. And we’re all seeing that I think. So in 2018 we thought Apple was gonna really get their AI in gear because they hired Google’s head of machine learning and AI, and we expected some major improvements. We haven’t seen it yet. I mean, do you expect it’s going to get some results?

Robert Scoble: Google has gotten better in the last year. You can now have conversations with a Google home device.

John Koetsier: Yes.

Robert Scoble: And that’s pretty cool. Siri has gotten marginally better, partly because the microphones in this new iPhone are better. When I do videos now on this new iPhone, it’s definitely getting better audio, and the audio and Siri understands me better. I noticed my wife is less frustrated this year with Siri than she was before because she has an Iranian accent and it’s gotten better there.

But it’s still, it’s not satisfying for those two reasons.

One, it just doesn’t know much about everything and it hits these edges where it doesn’t work at all. It gives you a stupid answer. And we’re going to see … I just saw a talk about an NLP from an NLP expert. NLP is the natural language processing that converts your voice into text basically in the first step of Siri. And he says there’s a real revolution underway that in the next two to three years, you’re going to see a real shift in how these things understand you and the context that they’re able to understand you with.

You know, we’re having a podcast context right now, right? And it could pick up on some of those nuances and give us different answers than if we were talking in a grocery store or talking in a church or talking in a school or talking at work.

John Koetsier: The real question there is, will Apple and will Siri, and autocorrect, which is where where we started, will that participate in those benefits? Will their model enable them to grow faster?

Robert Scoble: I know there’s smart teams and teams of smart people working on this so I’m assuming it’s going to get better from here. I sure hope it doesn’t get worse because it’s frustrating as it is right now, right?

John Koetsier: Yeah, it is. It is. The other question is that, I mean, Apple has a different model of releasing technology than many other tech companies, right? Many other tech companies, it’s just slowly getting better over time. You may not even hear about it. They make a big announcement, but it’s been getting better over time. Apple likes to, I’m not saying they have waterfall development processes, but they like to improve, improve, improve them, then release a big chunk and then leave it for half a year. I hope that’s not what they’re doing here.

Robert Scoble: That’s not really true anymore. I mean, iOS 13 has had nine updates so far. And…

John Koetsier: That could be seen as a bad thing 🙂

Robert Scoble: Well, you know, I’d rather they fix the bugs than they leave them there for another year, right? But they do this kind of bringing out one big package of functionality every year and then fix it kind of a thing, and they’ve announced that they’re going to try to change that in the next year. We’ll see if that’s true or not.

You know, when I worked at Microsoft, we all knew they tracked how many bugs were in each release and they knew they were never going to get to zero when they shipped. And in fact some releases had a lot more bugs than other releases. We all know those famous releases like Windows Vista or Windows i7 right? Windows ME…

John Koetsier: Yeah, ME, the famous ME.

Robert Scoble: Right? And so you see technology is immature at some level and gets shipped with a lot more bugs than they probably would like it to ship with. Particularly these new technologies with AI that that relies on huge amounts of training off of data that may or may not have been trained or set up properly.

You know, one of the companies that Apple bought was … an AI company. I saw the founder speak at the data science conference a couple of years ago, and he said, we trained it to see the difference between huskies and wolves. And it was mostly accurate, but once in a while it would throw an error. And AI is not supposed to throw an error if you trained it properly on husky and wolf and it’s supposed to say this is a husky, this is a wolf right?

It actually trained on the snow in the picture and not on the face of the dog, and it turns out that’s fairly accurate because huskies are usually in snow, wolves are on grass.

 

And he said, we ripped apart how the AI built its neural networks or its algorithms. And we found out, oh, it actually trained on the snow in the picture and not on the face of the dog, and it turns out that’s fairly accurate because huskies are usually in snow, wolves are on grass. You usually see a wolf on the grass. You usually see a husky in Alaska on snow. And so it was fairly accurate. Because even though it wasn’t seeing the right thing, but once in a while it would see a husky on grass and go, hey, that’s a wolf.

John Koetsier: Yup.

Robert Scoble: So it shows that this stuff can go bad and go bad in a way that’s really hard to figure out and unpack. This is why autonomous cars are taking so long to finish off. Because you can make it see paint on the street really well, but then all the other stuff that happens on the street … are rabbits crossing the road? What do you do? You got a lot of complexity that needs to be trained into the system and trained properly without bias.

John Koetsier: It shows you how dumb smart systems are still essentially.

Robert Scoble: They’re very dumb. Remember, a computer is still a light switch that’s on and off, on and off… it’s just now we have billions. I mean, my Tesla has 21 billion transistors in it’s car. And so it has a lot of light switches to turn on and off, but the technology doesn’t have the history and hasn’t gone through the maturity to know what to do on all situations. We’re better at handling certain ambiguous things. You know, I went to journalism school. I learned how to type, I learned how to spell, and so it frustrates me when my technology doesn’t know the same thing.

John Koetsier: Exactly. And I mean, I guess the one thing that we can say in Apple’s defense here is that they are trying to do something different. They are trying to apply AI and machine learning in a privacy safe way. That’s a good thing. That’s a wonderful thing. That’s going to be a good thing longterm if they can get there.

But it is a massive competitive disadvantage on user experience, on a common user experience thing to be typing and you’re trying to type ‘Sharon’ and it gives you something that is totally unlike that. It’s the husky in the snow, it was training on something that was not relevant.

Robert Scoble: By the way, if you do hit problems and you’re trying to improve it, go to settings. Unfortunately my camera’s too bright here, but go to ‘Settings’ and go to ‘General’ and then go to ‘Reset’ and you can reset the keyboard dictionary. That did improve things a little bit for me. It’s still not perfect. It still is giving me some other errors and it doesn’t fix the ducking problem, you know?

But it might help you a little bit, and it resets the dictionaries that learns off of your pattern, says you’re swiping or typing on a keyboard. It’s a little frustrating right now. I would expect it to get better over time. Hopefully a lot better soon.

John Koetsier: I hope so too. So I have to ask since we’re on… I know you’ve got the Model 3, and I know you’ve driven tens of thousands of miles to lots of the state parks and national parks, and the Tesla has driven you for 90-95% of those miles.

I still have people telling me there’s no such thing as self-driving, and I understand it’s not there, but it’s growing.

Robert Scoble: It depends what you mean by self-driving. You know, there’s five levels to self-driving and nobody does level five yet. It drives everywhere by itself, without a human in the car. Even Waymo that’s probably the leading example, only drives in Phoenix, Arizona and Mountain View, California and one or two other places.

Waymo has to drive a mapping car down every street and every lane to be able to map out the street in enough detail to get the kind of accuracy that they need.

 

I talked to the Waymo team just a couple of weeks ago about that, they have to drive a mapping car down every street and every lane to be able to map out the street in enough detail to get the kind of accuracy that they need. Now Tesla takes a completely different approach, which is we’re going to put the cameras on normal people’s cars and an Nvidia card or an AI chip. And we’re going to let them map their own way around the world, and we’re going to aggregate all that data and do training off of that data, and then send back down new features. And you know, in the last couple of months, my car starts switching lanes automatically, for instance.

John Koetsier: Nice, nice.

Robert Scoble: A new feature, I didn’t have that when I first bought my car. So getting better, you know, when are we gonna be able to just go to sleep? Hmm. I already sleep, but I’m taking a risk, right?

John Koetsier: Yes, you are.

Robert Scoble: It doesn’t see potholes very well. It doesn’t see like if a ladder gets dropped on the freeway right in front of you. And it doesn’t see if somebody is walking on the freeway very well. Going 80 miles an hour, it’s really hard to detect that and figure out what the heck that person is doing on a freeway when they’re not supposed to be there.

In fact, this is what Uber… Uber killed somebody, and their AI made the same mistake. They assumed, oh, there’s not supposed to be any pedestrians in this stretch of roadway, so we won’t even look for pedestrians. And there was a pedestrian there and obviously tragedy. Tragic.

John Koetsier: Yeah, yeah, yeah, yeah. And when you learn on lots of data, it’s hard to deal with anomalies. It’s really, really challenging. So the question I was going to ask you though about Tesla was, did you order a Cybertruck?

Robert Scoble: Yeah, I did, but …

John Koetsier: I figured you had!

Robert Scoble: I did, but rationally thinking about it, where the hell am I going to park that thing?

John Koetsier: I think it’s a very large vehicle. It’s bigger than a Ford F-150.

Robert Scoble: Yeah and I can barely fit my Model 3 in the garage. So it’s not going in the garage for sure. And then driving around Silicon Valley, it’s great on freeways, and it would be great with all the potholes we have, but…

John Koetsier: Don’t go to San Francisco!

Robert Scoble: Can I park it anywhere? I don’t know, I mean I see people with pickup trucks trying to squeeze in these small little spaces and they can’t open their doors…

John Koetsier: There is one good thing about Cybertruck. You can take a sledgehammer to the door, so you can park next to somebody and if they ding you with a door nothing happens.

Robert Scoble: There we go. And it’s really hard to break into the glass, which is what’s really happening in San Francisco where everybody’s getting their shit stolen because somebody comes by, breaks the window, grabs their iPhone and runs.

John Koetsier: True, true.

Robert Scoble: Don’t leave your stuff in your car! I don’t care if you have a Tesla or an old Chevy. You leave stuff in your car, you’re going to get it ripped off and your $400 windows go.

John Koetsier: I’m going to suspect that’s not a $400 window in the Cybertruck. I’m gonna suspect it’s more like a thousand dollar window, but…

Robert Scoble: The composites and plastics that are going on, I mean these new iPhones, most of the time when you drop them, they don’t break right? Once in a while you drop it, it will break. But the composites that are going on in any glass design are actually pretty cheap now. So it would probably not be that much more. But yeah, I’d expect it’s a little bit more expensive to replace, but you know, if you only have to do it once a month instead of three times. A friend of mine had three windows broken in one month.

John Koetsier: Horrible.

Robert Scoble: That’s 300 bucks a window. Even if you don’t leave anything in the car. They break it, by the way, just to get in the car and check if there’s anything in there.

John Koetsier: Yeah, yeah that’s horrible.

Robert Scoble: It’s really a problem in California, it’s not just in San Francisco either. Do not leave anything in your car ever!

John Koetsier: Yeah exactly.

Robert Scoble: We had a car ripped apart in Vancouver, Canada so it’s a problem all over the place.

John Koetsier: What, in Canada? Never!

Robert Scoble: Even in Canada you know. Who knew?

John Koetsier: Well Robert, I want to thank you for joining us. It was such a pleasure to chat with you. I really do appreciate it.

Robert Scoble: Hopefully our keyboards get better. You know, maybe Apple’s attitude is soon we’re just going to talk to Siri and we won’t need a keyboard, but I think we’re going to need a keyboard for at least five years, at least…

John Koetsier: I would tend to agree with you.

Robert Scoble: It’s all about typing on a keyboard. I’m a writer and I’ve actually written a lot on my book by talking to Siri or talking to the AI. And it’s pretty efficient at doing that, but it never looks right.

John Koetsier: No, and you have to do a lot of editing.

Robert Scoble: You have to do a lot of editing, particularly for punctuation and look. And a lot of writing is, do you have a bunch of sentences in a paragraph then one short one to make a point, you know, talking to the thing doesn’t do any of that. You have to come back manually and clean it up later. So even though the ability to speak to a computer is getting a lot better, it’s still not to the place where I can replace it yet.

John Koetsier: Yeah. The other problem, which I’ve learned from doing a lot of transcripts or podcasts, which I’m doing a lot of right now, is that a lot of us sound like Donald Trump when we’re talking. A lot of run-on sentences. It’s not as clean, as sharp, or as intelligent sounding as we’d like to sound. So yeah, there’s some work to be done.

Robert Scoble: There’s augmented reality coming for that.

John Koetsier: Excellent. Well, thank you so much for joining us on Tech First Draft. Whatever platform you’re on, please like it, subscribe it, share, comment. If you’re on the podcast later on, you like this, please rate it and review it.

Thanks again to Robert Scoble, and until next time, this is John Koetsier with Tech First Draft.

Subscribe to the podcast, or connect with me on YouTube or Facebook or LinkedIn or Twitter to watch them live: