Slashdot Mirror
Siri Co-founder is Surprised By How Much Siri Still Can't Do (qz.com)
In an interview with Quartz, Norman Winarsky, a founder of Siri, suggests that Apple may have given Siri an overly ambitious collection of responsibilities and hasn't made the feature reliable enough. From a report: And while vastly improved from its earliest days, Siri still isn't a sparkling conversationalist. "Surprise and delight is kind of missing right now," said Winarsky, now a consultant and venture capitalist. Winarsky acknowledges that some of this disappointment stems from the sheer difficulty of predicting the pace of major technological advancement, which Bill Gates once summed up as the human tendency to "overestimate the change that will occur in the next two years and underestimate the change that will occur in the next 10."
But part of it is also likely because Apple chose to take Siri in a very different direction than the one its founders envisioned. Pre-Apple, Winarsky said, Siri was intended to launch specifically as a travel and entertainment concierge. Were you to arrive at an airport to discover a cancelled flight, for example, Siri would already be searching for an alternate route home by the time you pulled your phone from your pocket -- and if none was available, would have a hotel room ready to book. It would have a smaller remit, but it would learn it flawlessly, and then gradually extend to related areas. "These are hard problems and when you're a company dealing with up to a billion people, the problems get harder yet," Winarsky said. "They're probably looking for a level of perfection they can't get."
But part of it is also likely because Apple chose to take Siri in a very different direction than the one its founders envisioned. Pre-Apple, Winarsky said, Siri was intended to launch specifically as a travel and entertainment concierge. Were you to arrive at an airport to discover a cancelled flight, for example, Siri would already be searching for an alternate route home by the time you pulled your phone from your pocket -- and if none was available, would have a hotel room ready to book. It would have a smaller remit, but it would learn it flawlessly, and then gradually extend to related areas. "These are hard problems and when you're a company dealing with up to a billion people, the problems get harder yet," Winarsky said. "They're probably looking for a level of perfection they can't get."
Siri has a "founder"? Not a programmer? Do other apps have "founders", or just Siri?
Really, Siri is just Hal by another name. They should have just called it "Hal-9000" and left it at that.
My guess is that the deep learning neural net that Siri is running isn't deep enough or neural enough. They just need to add more neurons and make the network deeper and the problem will be solved.
Seriously
Siri might have been able to point this link to the right article.
Modern appy apps like Appy don't have LUDDITE founders or LUDDITE programmers. They have appers who app apps while apping other apps!
Apps!
but she doesn't want any part of it.
As far as I'm concerned it can think of a really big number, double it, and fuck off that much.
Do not want.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
The problem is that the public has been led to believe that the capabilities of Siri (and its counterparts by other companies) wouldn't fall too short of those displayed by HAL 9000 and the Star Trek computer. When you start interacting with them under such expectations, you are bound to be sorely disappointed. Siri et al. remain gimmicks good for grins and giggles, and not really much else - just about anything they can do, people can do themselves, probably more efficiently. Plus, the things we would really want for them to do remain well beyond their capabilities. And the AI community still has the chutzpah (or recklessness) to carry on coming up with exuberant forecasts.
Considering the quality of the code coming out of apple these days.
"Surprise and Delight"
The crap that Apple (even ex employees) touts makes me want to puke. Please, just stop bullshitting.
In general this technology, is just the command line interface all over again, with some rudimentary natural language parsing, with a default fail over of googling the question.
The problem with All the Digital Assistants is that it doesn't really get context. So it comes up with silly answers to questions, because the context of the question isn't place in concern.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
But donut worry! The NRA met with Trump and their hand is well and firmly up his puppet-hole. Siri will be able to buy guns via Amazon stating 2020. FREEDUMBS!!
a negative Apple story posted by msmash? WHAT DAY IS IT?!!
Get over it.
Data science is about shaping databases to better match phenomenon - often VERY badly, but good enough to work for business or solving some immediate problem at hand with the resources at hand. I've worked at it, and it's powerful and amazing in its own ways - but it's not neuroscience, at all.
Siri has some lovely canned responses, shaped to match common human inputs, and improved based on what new common inputs come in, largely by adding more human inputs rather than really dynamically generated content.
Data science can help you shape an estimate to match previous responses better, can shape a curve to match an exponent better - but it isn't neuroscience.
ELIZA and her informational descendants like Siri aren't immitating humans - they're selecting from a data set of mixed repeating inputs and canned responses, with a few lexical alterations for effect.
They're not systems fooling humans - they're humans fooling humans using sliced up prerecorded clips.
Almost all of artificial intelligence and even business intelligence is like that - focused on satisfying expectations to some percentage, not on actually modelling absolute truth. As long as customers are indicating improvement, managers give the thumbs up, it continues.
It's very much more stage magic than anything else - under the hood, it's ugly framework and empty air, but dressed up to show the illusion just where it can be seen.
Which makes sense - if you're spending millions, and millions, and millions on it - you expect some stagecraft, I mean "modern professionalism" painted on top to pretend every dollar was spend with perfect wisdom.
Siri-ously though - it's a cool extension of previous technology, and a neat way to present it to cell-phone users and the like. But it's also cheesy use of such tools, and shouldn't be taken as more than window dressing to other tools as it is. The promise of 'virtual assistants' from the mid 90's is nowhere closer with this. Wolfram Alpha might be closer - but they're mostly in the same bag of oddball interfaces.
Ryan Fenton
There's certainly no AI there. These kinds of apps are currently just expert systems with manual tweaks and a search fallback.
Actual conversation I had with Siri.
... "I can't answer that" ... "I can't answer that" ... proceeds to phone my wife. ... "Your wife is right beside you."
"Siri, what is my relationship status?"
"Siri, what is my marital status?"
"Siri, phone my wife"
"Siri, where is my girlfriend?"
That last one has me confused. Did Siri know that I am married and thus girlfriend means spouse? Or did she try to warn me that my wife was right next to me.
---
Programming is like sex... Make one mistake and support it the rest of your life.
All my life, we've been expecting voice recognition "real soon now". And it always flopped. You had to shout really slowly and carefully to get the system to recognize maybe half the words you said.
Then along came Siri, and finally there was a commercially available system that was good enough with normal speaking tone and pace (mostly) and... it dropped the ball miserably at doing simple stuff with the recognized speech.
I haven't tried Siri in ages, so maybe they've improved it recently, but I already gave up on using it because of how dumb it was. For example:
I could ask Siri for directions, say to my hotel, and she would understand fine. But if I asked for gas stations along my route, or restaurants near my destination she wouldn't do it. My old Tom-Tom could do that fine, you had to push the touchscreen as it had no voice capabilities, but it did it great. Siri could understand my voice, but could not do what my Tom-Tom could. To add insult to injury, Siri's canned response indicated that she understood what I was asking for (to use my route or destination as a search location instead of my current position), she just wouldn't do it.
Another time, I wanted to call my wife from a rental car (my regular car has its own voice recognition that works better for this). So I asked Siri to call [wife's name]. She didn't understand, fine, my wife has a weird foreign name. So asked Siri to call [our last name]. She found 2 people with that last name in my contacts (myself and my wife ) and asked me which one I wanted to call. Great! I responded "[wife's name]" Siri then asked "what do you want to do with [wife's name]?" Siri you just asked me which of 2 people I wanted to call! Oh well, I responded "Call her". Siri didn't understand what "Call her" meant and looked up websites related to "Call her". Now even back in the '80s when playing Infocom text adventures like Zork, you could type commands like "hit troll" have the game respond "what do you want to hit the troll with?" and answer "axe". The game remembered just fine that it asked you to fill in some info, and was ready to plug in the new info into what you were doing a few commands ago. But Siri couldn'do that, 30 years later.
She found 2 people with that last name in my contacts (myself and my wife ) and asked me which one I wanted to call. Great! I responded "[wife's name]" Siri then asked "what do you want to do with [wife's name]?" Siri you just asked me which of 2 people I wanted to call! Oh well, I responded "Call her". Siri didn't understand what "Call her" meant and looked up websites related to "Call her". Now even back in the '80s when playing Infocom text adventures like Zork, you could type commands like "hit troll" have the game respond "what do you want to hit the troll with?" and answer "axe". The game remembered just fine that it asked you to fill in some info, and was ready to plug in the new info into what you were doing a few commands ago. But Siri couldn'do that, 30 years later.
Oh, if I call the wife and forget to specify "mobile" Siri will tell me there are five phone numbers, which do I want to call, reads the entire list without letting me interrupt, and promptly forgets what she was doing after reading the list.
Awhile back, I tried to phone my mom. She has ONE number listed, and an easy name, but Siri still couldn't do it right. If I'd actually been driving instead of parked, I probably would have crashed into something while yelling obscenities at the phone. I ended up taking the phone out and dialing the number manually.
Slow down, cowboy! It has been 4 hours since you last posted. You must wait another few hours.
I suspect that the Siri founders painted a vision where all these capabilities were just 6 months away, with "hockey stick" revenue curves and smattering of the usual lies that startups use to inflate their valuation. Now they act surprised that Apple can't make it do all these things...
A.
What I love is how Siri has built in logic to detect swearing and tell you not to use such language, but it can't do basic things like keep the context of the question that it had literally just asked you a moment before. Brilliant way to spend your programming resources Apple... Make sure Siri has a witty comeback if I use profanity because of how crappy it is at not understanding context that it literally created by asking me a question.
---
Programming is like sex... Make one mistake and support it the rest of your life.
Ahhh yes. You mention ZORK, and instantly "open the mailbox andread letter" to my wonderful childhood.
Because true AI is generations away from current technology architecture?
Task programming (ie, the "task" of converting speech to text) is in no way, artificial intelligence. The ability to recognize patterns (when an algorithm is designed to recognize patterns) is in no way, AI either. Don't let parlor tricks fool you. We have nothing that approaches AI today. Nothing.
It has flopped because there is NO AI. There are no systems that have any comprehension of your context, your experiences or your insight. That is AI. Everything else is a "parlor trick" to try and make you believe it's intelligence.
I don't think SIRI needs AI to be more useful. Once it does the 'hard part' of converting speech to text, it just needs to use existing (non-AI) technology to make it more useful.
The two examples I gave were already on the market technology (the second one for over 30 years) that Apple could have trivially copied, but didn't.
They're probably looking for a level of perfection they can't get.
Give it up; the jobmeister is gone.
He happens to be six feet underground. Too soon? You saw me standing alone...
That Star Trek TNG *DIDN'T* have the computer act all knowing (except in a few episodes about computers gaining consciousness, or some of the Voyager eps with the neural packs..) You had to tell it 'Computer, .' Sometimes it wouldn't respond, sometimes it would ask you to rephrase the question. While it certainly had processing functionality greater than the current offerings, it for the most part perfectly mimicked where we are today, sans the 'holodeck programming' and a few other things which we have the technology for today, but not integated into voice applications yet.
I listen to one album on my daily motorbike ride. Morning and evening, it's the same. I ask Siri to play it.
"Siri, play Beach House"
I get one of these responses:
>>> "Hmmm. I'm having trouble finding that. If you're not on wi-fi or 4G please connect." (You'd think Siri would know my network status, and it *is always* connected when she says this).
>>> Plays Led Zeppelin
>>> Nothing
>>> Starts playing from where the album left off.
>>> Plays the album from the beginning. YAY!
When they came for the communists, I said "He's next door. Take him away. Goddam commies."
It's not "how much Siri can't do". It is "AI, or, I would rather say, what's being called AI, can't do". AI, in its current shape, is nothing but a glorified expert system which works fine in well defined domain. That has nothing to do with human intelligence. So, the usual sequence of events is (i) you create something you call AI, market it and sell it. (ii)Profit! (iii) Then some gullible customers start wonder why that system isn't quite intelligent.
It does come up with quips and jokes at time. Just can't pick where it does. Seems to happen mostly after hours.
I'm not an Apple guy, but I was pondering making my long (8 hour) drive more pleasant.
Would love to be able to speak to a device and get it to read back web pages to me, basically browse via voice. Could be quite useful.
No it has flopped because the implementation of it consistently doesn't look at human factors.
Have you used Hound? It's WAY better than Siri at core. It just needs support of a phone maker. Has anyone bought it? NOPE. Cause they don't care about the design.
It's not the lack of AI. It's the lack of Real I.
Agreed. A swear should be great data that the previous interaction needs work. instead they throw the data away and at the same time piss off the user, who is already angry.
It should be able to handle that. But it can't. So here's a trick. Put a contact in for your mother like 'mater natura yourlastname.' Fill in the first two words int he first name and the last name right in the last name.
When you ask to call a nice long name like that, she'll get it right.
It's 'Ann' and 'Bob' she can't handle.