AI, Apple and Google
ben-evans.com
ben-evans.com
This is tough to reconcile with AI because AI sometimes fails.
When your device doesn't do what you want, it feels bad.
When you read an article about your HooliPhone handing over data to government agencies, it feels bad.
When you go to the Hooli Store to demo a HooliPhone, ask it to show you photos of your dog before her last haircut, and get no results, you feel less enthusiastic about owning a HooliPhone.
The solution Apple uses is to add constraints. Apple doesn't need AI to guess where you want to save a file on iOS because there isn't an exposed file system.
There are some areas where simplification isn't possible, like mapping and voice assistants. These happen to be two of the most widely panned Apple features.
The strategically tricky thing about Google's pivot to AI is that there's a confirmation bias against adding constraints. This may result in impressive demos that just don't feel right in the real world.
You've put into words something I've been feeling and trying to express for a while. "How does it make you feel?" is an incredibly important question and unfortunately for Google, something you can't really quantify (see: Circles and Google+).
Actually Circles and Google+ is a great example that we can look at to see how the feel of a product can help lead to it's eventual downfall. When someone "friend requested" you on Facebook, you had a simple decision to make: "Is this person a friend of mine?". While everyone has their own definition of "friend", it's pretty easy to make a decision and sometimes you're gonna have to say "no" but that's fine; it feels bad for a short amount of time.
With Circles and Google+, you had to ask yourself: "Do I know this person?", "Is this person a friend of mine?", "But how close am I to them really?", "Should I add them to the family circle? but they're not really family...", and so on. The worst part is, you had to do this all upfront. When you have to reflect about that many things about a human being, it starts to become far more negative than it has to be. And if you're going thru this experience all the time, then good luck. Now Facebook has it's issues too, but see how such a subtly negative "feel" to a product can be real friction.
I recently heard someone make the argument that while a lot of the Google AI tech is impressive, they haven't really been able to build something that feels that useful as a product in the real world. It's an impressive tech demo for sure at Google's conferences, but once the tech makes it to Android, they don't really become used all that much. Take Now-On-Tap, for example. It was a highly touted feature on Android: an incredible ability to rip context from whatever you were looking at on your phone. Now you hardly see it mentioned and people don't really seem to use it that much. Google Now stuff is cool too, but besides airplane reminders and calendar integration, is it really that useful day to day? I suspect most people would say no.
Look...all of the stuff Google does is very cool and very impressive...but it just lack a certain human feel to it. A certain simplicity in how something is useful. These AI experiences are not predictably useful. They aren't experiences you can rely on. Now I'm certainly generalizing a bit, and I know someone will come up with some counterexamples (please do! I'm curious), but I think you understand the sentiment I'm trying to convey: that Apple just understands the user from a different vantage point than Google simply by asking how something make them feel (and Jony Ive has emphasized this question time and time again, though he sometimes gets made fun of for his overzealousness).
He walked into the room and asked, "Who here is making something really impressive?"
I thought a long time about what that says about Google.
Maybe it's about Google's competing on talent--they already have a $70 BN revenue machine and simply need a way to attract people who can keep the lights on.
Maybe it's just the personalities of those at the top.
Recently a friend showed me Google's Project Soli, a gesture interface that uses radio waves to sense hand position. You click by touching your fingers together and scroll by rubbing your fingers.
The project looks fantastic and is clearly produced by some smart people.
But the product video is a big call-to-action to developers: "We're excited to see the gestures you come up with."
I feel bad for the talented people working on this. Because the first time somebody uses the product, they'll try a new gesture and it just won't work. Creativity and curiosity yields frustration and a feeling of stupidity.
If you make something really cool, make it easy to love at first sight.
There are ways to use Project Soli's precise depth sensing so that there's nothing to learn and no way to fail. There are applications that will make people rethink what our devices can do and reward natural curiosity. Unfortunately I can't go into details.
Instead, people will be waving at their Google watches and wondering why it keeps taking selfies instead of scrolling through their notifications.
As you mention above when talking about Google+ etc. Google have not really been great at doing products in the past (though they've definitely gotten better).
But the key is that they are getting themselves onto a virtuous cycle for their AI development - they're attracting brilliant minds, which leads to bigger better breakthroughs (like Alpha Go and driverless cars) which leads to great PR and attracts more talent and customers/consumers of their AI.
So what if Google doesn't build AI products, what if AI is the product or platform. What if AI is the future to most things (e.g. automation of routine labour or optimisation for business operations through machine learning). If Google can position themselves to be at the heart of all of this, they can let other developers and companies build on top of that to deliver the useful products. Just like they've done with Android.
How will they monetize that? No idea, in my mind I imagine using consultants or licensing fees but that doesn't feel very "Googley"....
I look at it every day for traffic warnings, an overview of my schedule, reminders and updates, and interesting stories.
I mostly use it in the mornings for practical stuff and during the day when I'm bored for interesting news stories.
It's slowly moving into the must have category.
One of those is a bet on optimising for the future, the other for the present.
On another note using implied insults is not really in the spirit of hackernews. It significantly reduces the quality of your comment. Many of us here do work for both of these companies and have a strong respect for each other's work.
That's a pretty bold statement. When Apple Maps first launched this was certainly true. But Apple Maps has improved significantly since then, to the point where whether Apple Maps or Google Maps is better depends on where you are and what data you're trying to find.
One other thing - while in navigation mode with Apple maps, try panning the map around, or even zooming. Let me know how that goes.
I wish this were the case. I'm running the iOS 10 beta on a 6S+ and gave Maps several seconds to load this area of upstate NY.
http://i.imgur.com/wi7HRqe.jpg
The dark gray grid persists. Yea, it's the first beta, but I have mostly only noticed changes in the interface and not the actual map data.
And Apple's inability to do any sort of decent fuzzy matching is legion.
Be it in search or in maps, google services are auto correcting so much stuff. You can be in France and search for 'Akita', and the first result will be some japanese restaurant that happen to have the keyword in the name, with the more reputable and perfect match half a screen below.
Of course a lot of user might actually be looking for restaurants, but it's so frustrating this moment where you have to think "I really mean what I typed in this box" and going through the "what was the syntax to value more the keyword".
And then as it detects your preferences the behavior changes and you are left wondering what's happening (like how do I explicitely ask for local search to be prioritized ?)
It feels less and less like a tool and more like a pet engine you have to teach fetching the mail.
The more interesting thing to me now is seeing what OpenStreetMap and Mapbox are doing. Their combined trajectories may end up surpassing both Google and Apply. Mass produced drones along 3D scanning tools and improved 2D->3D imagery tech could provide for very comprehensive and open map data with all kinds of layers of usage far beyond geo coordinates, lanes, and street sign data.
And last time I was in Palo Alto using Google Maps to find a dry cleaners, it sent me on a 20 minute goose chase, and had me going in circles. On the third try of reentering the address (using copy/paste all three times, all 3 times the address was identical) it finally game me an accurate route.
I still use Google Maps more than Apple maps mainly because of the subway/bus integration and because I've been using it for so long, but in my anecdotal experience Apple Maps isn't any worse than Google Maps and has been less irritating when it's wrong than Google (although I use Google more).
Do you have any source for that? Thanks
I kid, but only sort of. Apple Maps has serious data freshness problems. For years they listed businesses in a building that had long-since been demolished. Looking around in my Oakland neighborhood, Apple Maps is missing two new streets. Google has an indoor map of the BART station, Apple doesn't. Etc.
So basically, Apple is slightly competitive if you are driving a car within the US, but even there it's probably a lot worse in the aggregate.
And Flyover is not even close to being as useful as Street View.
I'm stunned at the number of people who use Apple Maps in Singapore. Google Maps is much better here - but about 50-70% of the people I know who have an iPhone just use the Map App that is installed on it. Kills me.
I don't know any english speakers with an iPhone who have changed out their keyboard - I've played with the others, but perhaps it's a foreign language thing?
My experience is mostly focused on SE Asia but nobody here sends an SMS if they can help it at all. They're fairly expensive, especially given that monthly 3g plans are as cheap as $3. Whatsapp, Facebook and Viber own messaging here.
Line is popular here, but only because it has stickers and iMessage doesn't.
That's not true. Why do you say that? I've used the foreign keyboards for Spanish and for Korean, and both keyboards are excellent and offer good autocorrect. I've also never heard anyone complain about the foreign keyboards being bad before.
I grabbed a bus a couple of days ago because maps had a live update that it was just 1 minute away from me.
I've given Apple Maps several chances to delight me since 2012 and have submitted multiple bug reports and feature requests. While the product has improved significantly, it needs a few years of development to match the current offerings from Google. The inaccuracies in the search results for local businesses/parks/points of interest is unacceptable and should be a priority for the Maps team if Apple has any intention of feeding data into a device that will take humans from point A -⇒ B -⇒ C. Also, not being able to create waypoints/multi-leg routes is simply unacceptable in 2016. For most Apple Maps users, I suspect it takes only one shitty search result page for them to give up and use Google Maps.
There are some deep, systemic issues that Apple needs to address:
1) The accuracy and freshness of their data
2) The ability to return fast, intelligent search results from their data
3) Siri's ability to understand user intention when being asked for location data and directions
I'm rooting for Apple, I really am - It just seems that they are having a rough time catching the tailwinds of Google.
Or is good speech recognition just that intensive, such that common consumer devices really can't handle the load?
Google Now for instance can recognize me saying "Set a timer" or "Turn on WiFi" even while airplane mode is enabled, but start adding in context (such as "Who is Bob Marley?" "When was he born?") and it probably wouldn't work.
For just speech recognition you can use Kaldi. Its not hard to deploy on your own server. I've been meaning to package it for Ubuntu phones but haven't gotten around to it yet.
Pretty good speech recognition is doable on current mobile hardware. The trained models are large and there is still work ahead shifting the inference to mobile GPUs but it's doable.
It also helps if it's integrated into the OS, if you want it to work invisibly with multiple apps.
Snips organizes all your information and make it searchable intelligently and actionable fast (it predicts your recurrent patterns so you can order a Uber in one tap)
"the promise of extracting new insight from all sorts of data pools will not always be met."
While certainly true, we can all agree that not having access to these data pools is a disadvantage.
A major fucking company takes the same approach, to provide its massive user base with additional functionality and improved privacy, improves the service over time, and people still want to claim it's all about "no one uses it by choice"
Nothing prevents people going to the App Store to get the google maps app or using the browser version, and yet for millions of people the apple maps app is sufficient. It works, it's easy to use.
If you don't want to use it you don't have to, but Jesus fucking Christ don't assume your experience and mindset is the same as every other person on the planet.
I've always laughed at how poor speech recognition is. I know it's probably a hard problem, and I know that sometimes you can transcribe a whole sentence without error. But how long will it be until it will just work; I'll be able to just speak normally, in English, and have it transcribe without error? Microsoft put up a page recently with a demo and prices, but it was a lot worse than the last one I tried.
I'd say Google Now cards are pretty original interface and they are downright magical. Google already knows the time & place of an appointment I have across town, it also knows that traffic is getting pretty bad on the possible routes, so it tells me (without me asking) to leave much earlier than I had planned.
> Where do you want to save this file?
> Which photos do you want to delete to save space?
I dearly hope not. These questions are too important to decide by the computers at their current level.
>> In the last couple of years, magic started happening in AI. Techniques started working, or started working much better, and new techniques have appeared, especially around machine learning ('ML'), and when those were applied to some long-standing and important use cases we started getting dramatically better results. For example, the error rates for image recognition, speech recognition and natural language processing have collapsed to close to human rates, at least on some measurements.
Let's see.
a) Magic? Really?
b) "The last couple of years" goes back to 2014. Machine learning's big break into the mainstream happened at least in 2012, when a conv net won ILSVCR [1]. Did the op not look further back before reporting on all this "magic"?
c) Unfortunately no "new techniques" have appeared. Convolutional neural networks date from 1988 [2]. LSTM Recurrent Neural Networks were first described in 1997 [3]. Machine learning in general is pretty much as old as GOFAI, with the earliest connectionist ideas detailed in the 1950s [4].
All this is really well known and understood and not at all controversial- Geoff Hinton himself is on record explaining that the recent boom is due to more data and more processing power (I can dig up a link if there is any doubt).
d) "the error rates for image recognition, speech recognition and natural language processing have collapsed to close to human rates, at least on some measurements"
In image recognition, the big success story is the ImageNet results detailed above. For speech processing I'm told there's been a big jump also, someone else could give an example.
For natural language processing: not in your wildest dreams. We're nowhere near "close to human rates" in any way, shape or form, unless you count very restricted results like how good this or that natural language parser does on a one-million word corpus like the Brown corpus. Unfortunately, that counts for nothing.
Now, most of that is stuff you can find easily on wikipedia with a bit of reading. There's absolutely no reason to waste peoples' times with disinformation and pointless overhyping.
[1] https://en.wikipedia.org/wiki/ImageNet
[2] https://en.wikipedia.org/wiki/Convolutional_neural_network#H...
[3] https://en.wikipedia.org/wiki/Recurrent_neural_network#Long_...