M, a personal digital assistant inside Facebook Messenger
wired.com
wired.com
>"Can you make me dinner reservations?"
or
>"Can you help me plan my next vacation?"
I'd really love to better understand who is actually asking those types of questions in such a vague fashion, and what their use case is. When I'm picking something as simple as a restaurant, I typically want options, I want to read reviews, I want to consider distance, parking, attire, etc. While their AI/human trainers might be able to handle this level of complexity eventually, the actual phrasing of the question would likely be much more complex than "can you make me a dinner reservation." Doubly so for something like a vacation which has a lot more moving parts.
But I respect that I'm reflecting on a sample size of one...me. So I'd love to hear from others with more insight into the data around this. Are people actually searching with such generalized queries when it comes to tasks like this? Do most people not sweat the details of things like which restaurant to eat at, or where to spend hundreds or potentially thousands of dollars on a vacation?
Not trolling, serious question.
I think the idea with a conversational interface is that it's succinct and on-demand. You receive the most relevant information directly in as simple of an interface as possible (arguably).
It's much faster for me to hit a few filters on things like prices and locations. Distance is just a simple ".2 miles away" text on the box, which shows an image and snippets of reviews. People are more and more visual.
I don't think a conversational interface _replaces_ a visual one.
It's that the initial query can be complicated, and it allows you to get into that 5-6 tier deep part of your search that you would have gotten to by using 5 filters and scrolling through 50 results.
Set timer for 5 minutes
Add eggs to shopping list
Play Clocks by Coldplay
Weather
They do so entirely based on past shopping history.
Currently the only annoyance is logging in once a week to check if there's any adjustments we'd like to make. But it's good enough that if I don't feel like it, I'll just take my chances (you can also add "always include X" and "never automatically add Y" rules, which is part of the reason why that works...) and most of the time I get what we need.
I never want to go back to putting together my own shopping list from scratch.
This is how I want these type of services to work. I don't want to have to talk to them. 99% of the time, I'd prefer them to be invisible to me, and make things that used to be an annoyance just disappear.
But the one way it could be better would be to make that one last interaction disappear more often: Not having to log in to make changes. Being able to just say out in the air that I want to add eggs, would be great, and in that case I'd want it added based on past preferences: If I've bought eggs before, and I'm not specifying, just add the quantity and brand I usually order. If an alternative I've also ordered is cheaper or on offer, ask for confirmation if I'd be happy with that instead. If I haven't bought eggs before, pick a brand based on my past overall purchase history, and ask for confirmation, or simply add some - if I'm asking to "add eggs" rather than "please recommend me some eggs for my grocery shopping", I probably don't care.
Over time they learn your preferences so they don't need to ask location for example next time.
Your right though people aren't likely asking such generic things in the first place, but rather something like "can you book me a great mexican place for dinner tonight, 2 people, has parking and casual attire somewhere with great yelp reviews"
Then they send you the best options they found (and the benefits of each one and price range) then you reply back option 1 and they book it.
For me, a lot of what you're doing here is the work that should be done by a machine. Considering "distance, parking, attire, etc." is basically what we have simplex method for.
But I agree the questions seem very vague in the context. To run such errands successfully, the program would have to know much more about your preferences than current iterations of personal assistant software do. And/or hold a dialog with you, asking for details and proposing options.
I guess it just seems incredibly inefficient compared to checking a couple boxes and reviewing a list along with a map or other visual aid.
To make this useful, perhaps you could set up some kind of saved preferences. For example, let's say I'm setting up a business trip. I like hotels that are within 1 mile of the conference center, and they have to be at least 3.5 stars and up. Provided they meet those criteria, the cheapest option is acceptable. I also need a plane flight that has no more than 1 layover, and that layover cannot last longer than 90 minutes or less than 45. I am willing to pay up to 25% more for a nonstop flight. The flight must arrive the day before the conference, but it can depart on the day the conference ends.
Setting up those criteria for each individual search would be irritating and a waste of effort, as they don't change from trip to trip. However, if I could say something like "Let me tell you about my criteria for choosing a location for a business trip.", and then go into detail once, that might work. Hell, I'd be perfectly happy setting up the details on a website. Then, the next time I said "I need to set up a business trip", all it would need to ask is the conference center and the dates of the conference.
Until it supports these kinds of detailed requests, it doesn't make sense to use these kinds of services in the way they market them - you'll end up using it in the same limited way you could use Siri. For example, if you've already decided what restaurant you want, you might say "Make me a reservation at Dorsia for 7:30 this evening" instead of the examples you provided.
Just my 2 cents.
I used to fly in to the Bay Area very often on business. At first the office manager arranging things would ask me details about which airline and which flights I'd prefer after listing the options, and which hotels, describing address and location and how near they were the office. Possibly e-mailing me a bunch of links for me to look at. But after just a few trips it was down to "is flying out on the 2.30 on Wednesday and returning on the 3.15 the following Thursday, ok? [she know when I preferred to fly, and she'd implicitly have ensured they were the right code to maximize my chance of an upgrade] Your usual hotel is full, is the Sheraton ok?" [no addresses necessary - we'd boiled it down to 2-3 preferred hotels within walking distance of the office].
I'm thinking "get me a dinner reservation next sunday with patio seating for 5 in the east village at an upscale tapas place".
As I mentioned elsewhere on this page, my thesis around conversational interfaces isn't that they start off broad and use more Q/A to refine your query. That's slow, and people are visual.
Rather, their power lies in the user being able to express a complex query in one go - which is equivalent to tapping 10-15 filters and scrolling through results - ideally combining data from sources that aren't limited to one service.
You can now execute related actions to your result set through the same interface, without needing to shift to a single purpose app that would allow you to take the action, but for most purposes, won't keep your context.
It really is all about the interface and the efficiency. I have to wonder though at what point is adding all those filters more involved than checking a couple boxes and glancing at a map or some photos. I'm sure a lot of that depends on context (I can't do those things if I'm driving, but I can use voice recognition).
The other thing I'm unclear about is how such a recommendation engine can best present information about tradeoffs. In theory, each of my filters has a weighting, and that weight might be dynamic based on several other factors. Maybe I really want chinese, but the best match is further away or I know there will be lots of traffic, so I might be willing to compromise on thai, but only if they have that one dish I like. And a lot of it is seeing the options in the moment and making a snap decision. Really curious about the approaches to solve that type of problem.
_When the filters are across datasets and services that are hosted on different platforms, and there's no way one UI that allows you to access them._
Table timings are on OpenTable/Yelp, reviews are on Google/Yelp, traffic is on Google, rides are on Uber and Lyft, menus are on the web, there are recommendations you trust amongst your friends and blogs, and pictures on instagram - and you're on a messaging platform trying to coordinate with 4 other people.
At that point, whatever service helps you to narrow to 4 choices based on all of this data is a Godsend. It's about making decision taking easier.
> how such a recommendation engine can best present information about tradeoffs
That tradeoffs are still yours to dictate - you simply look at the results of your complex query and then use conversation/UI to refine. Faster than using 8 services to do this. Repeatedly.
Nope. I don't think anyone would ever leave it up to M (or whatever) to select the restaurant.
Anecdotally, if we'd ask a real person "where is a good place to eat" the chance we'd go there without more information is slim. And if we don't even trust people, trusting Siri will be a while.
What we're really doing with these questions is making our hunger known, and starting a conversation. We actually don't care that much about other people's thoughts, and we may not even have anything in mind yet as far as where to eat. We do care about how people feel if they are someone we care about, but the thinking part we love to do ourselves.
So to offer a service that "thinks" is rather misguided, and may even constitute a disservice. We already rejected the talking paperclip in 1996 [0]. It's failure wasn't it's intelligence, but in the value proposition itself. To have a paperclip presume to know better and to tell you what to do was not tempting. It's failure was it's existence.
Is it a glitch in the Matrix or is their pitch for Cortana identical?
> What is Cortana? Cortana is your clever new personal assistant.[1]
--
[0] https://en.wikipedia.org/wiki/Office_Assistant [1] http://windows.microsoft.com/en-us/windows-10/getstarted-wha...
No, its not a glitch that Siri, Google Now, Cortana, and now M have essentially the same pitch -- they are direct competitors intended to attach people to their respective platforms.
The issue here is one of trust, which is built on an individualized relationship over time. When I ask someone I know for a recommendation, I'm doing so because I already have a sense of their judgment. That's more the key here- build a history of reliable judgment. That's the goal.
Right. This is certainly one path and the path most seem to be on, and exactly the one that needs to be challenged. The key intuition here being that a judgement, which is a decision, is not an answer to a logical problem. A decision entails a will, and when our personal will is overridden by an animated paperclip, we close said program. Decision != Answer.
People don't necessarily want decisions made for them, but rather, they want assistance in making their own, or better yet, reasons to justify the decisions they've already tentatively made. "Reliable judgement" is the complete opposite of "a resource of intelligence". Certainly all of these assistants feature a little bit of both, but I keep sensing the urge towards the former. Worse yet, a decision is often treated as an abstraction that somehow justifies hiding everything that went into that decision, even though there is immense value in actually being told why. People have entire conversations over why to eat at some place as part of the process of sharing the decision to go there.
Even when used only as a resource, if only these robots wouldn't keep trying to read our minds or insist on telling us what to do. Maybe a handful of people will accept a robot's choice, but everyone loves more information.
Maybe we shouldn't be looking for some secret sauce that enables robots to make better generalized and rational decisions than humans. Maybe we should be building robots capable of assisting humans at better making their own personal and irrational decisions instead?
(edited for grammar)
What is "best" and to whom? Ideally the software would figure this out but I'd always be wondering if it was just going to TripAdvisor and grabbing the first result.
Another problem is that we don't always know what kind of food we want. There's an urban legend that someone actually called a restaurant "I don't care" so that boyfriends would have a place to go when asking their girlfriend for dinner.
That's the difference between talking to a real person (or good NLP) and a search query.
And if I know which pizza shop I want to order from, what's the benefit of adding an intermediary?
But until it elevates from 'digital assistant' level to just 'assistant' (ie. do all the work and just confirm with me before booking) it may not take off as they expect it to
That's why I use an intermediary. If that intermediary was being able to just say "I'd like my usual pizza/Chinese/burrito, but instead of X I'd like Y" and just have it confirm what it was about to do, I'd love that.
If want something new, or I'm somewhere I haven't been before, that's different - then I'll be spending time looking at the menu etc.
I actually don't want a machine to make decisions for me. I want a machine to do what I tell it to do, or present me with I formation required to make a decision.
Examples: if I need a dentist appointment or to schedule maintenance for my air conditioning, I'd like to tell a machine to set it up. Heck, I'll even tell it who to call and which days and times work for me.
If I'm looking for a restaurant, show me the options, give me their distance, top reviews, and some of their dishes. If I want reservations, I'll tell it when and for how many.
Ideally, I want a "Jarvis" from "Iron Man". I ask questions, it gives data in a digestible quantity, and then I can make a decision and tell it what to do. Obviously, such a system is not available (yet), and these inferior systems are needed in order to make progress, and get there...eventually..but sometimes I wonder if the focus is on the right outcome, or just the broad strokes cookie cutter solution that comes to mind first (restaurant reservations). Similar to how all JavaScript MVC frameworks demo a to-do app, and rails tutorials demo'd a blog (initially)...
I mean, seriously... How often do you not go out to eat because you are too lazy or busy to make a reservation? Now, how many times do you skip oil changes, or making calls to cancel your cable service, because you don't want to make time in your day to stop what you're doing, pick up the phone, and call?
If it told me it recommended the restaurants along with commentary like "you really liked X at another place, and this place has been voted to have comparable X, plus it is close by and you've had a long day and need to get up early tomorrow" that would be super useful and help me reach my own conclusion faster.
Bassed on my understanding of semantics and Knowledge engineering, this is doable.
I can't help but picture a large, fluorescent-lit room of jolly old British "trainers" in safari khakis running around admonishing misbehaving AI for telling bad jokes, all the while trying to juggle placing calls to the DMV and restaurants to make reservations for 700 million messenger users.
Same thing as Yelp, just a lot scarier.
I am actually okay with the privacy I give up when using Google Now, for example, because being passively informed about things I'm getting shipped to my house, about traffic conditions to/from my house, etc. is nice. Giving up privacy seems justifiable in those instances.
But FB suggesting restaurants, I just know that there will be money exchanging hands. FB being FB, will extort small business owners into paying them hand over fist to get considered for suggestions to the user.
It would be pretty cool to order in the chat, 'deliver me 10 burgers from this restaurant using doordash at 6pm'. I would rather not click through crappy websites/enter my cc every time.
However every time we do do that, our habits and conversations get written in stone (in multiple data sets being passed around and bought/sold everywhere).
I'm not sure what the server side component is - but I don't believe they would have itemized data. So Apple/Amex know that you go to Whole Foods, but don't know what you're buying. Obviously, credit card companies have always had that data anyways.
I count 4: Facebook, Messenger, Whatsapp, Instagram.
That said, they prefer and guide you at every point to use an FB account to sign up vs making it easier for you to keep the two de-coupled.
(1) Not for facebook and (2) I don't think that they would inevitably promote every new app to the top — I think they'd rather see how it grows organically first to determine how good it turned out to be.
The article is written by someone who doesn't know what he's talking about. The "cat videos" story from a while back ostensibly used Unsupervised training, that means, the Google team didn't have to tell the deep neural net what a cat looks like, it discovered the concept of "catness" by itself (there was a "cat" neuron in the top layer).
I'm wondering who writes all the AI articles I read every day. Such a detail was crucial for the cat story. It's easy to make a cat/non-cat classifier with a few thousand labeled images for each category. The hard thing to do is to take raw photos with no labels and still discover cats.
Building high-level features using large scale unsupervised learning http://arxiv.org/abs/1112.6209
From the abstract: Contrary to what appears to be a widely-held intuition, our experimental results reveal that it is possible to train a face detector without having to label images as containing a face or not.
The Cat detection thing was just a side product of learning to identify features of things in an unsupervised manner, but the news outlets locked on to that with titles such as "How Many Computers to Identify a Cat? 16,000" in NY Times.
Wasn't it amazing that they could distill the concept of cat from images with no help from external labels (human intervention)? They missed the core of the discovery by not understanding that.
The deep learning method is an unsupervised way to process raw input and transform it into useable features. This used to be done by a combination of domain knowledge and supervised training, but they could build an automated way to extract relevant features from images.
This opened the window for hope that one day neural networks will be easily applied to any new domain if there is sufficient raw data to build a deep network for it. In the past there was a need for a large investment in human based data labeling and how to extract the best features from raw data (also described as voodoo magic by the same researchers - it was hard, it was domain locked and expensive).
In fact that very network produced also millions of other "concepts", that is, classes of images, that have no direct interpretability in human terms. The "cat neuron" was a fun gimmick, but you're reading way too much into it.
It reminds me of a story Feynman told: https://haveabit.com/feynman/knowing-the-name-of-something/
Now imagine that reversed, that what we considered one thing could be considered 2 or more by the model, we had just never thought of them separate because we had no words to describe them.
There are many of these examples, where one language has one concept that's split among others in another language, and the speakers of the first language might never know the difference unless those words exist.
A good example is colors: https://eagereyes.org/blog/2011/you-only-see-colors-you-can-...
To use an analogy - if messaging apps are the new "browsers", then content accessed through them are the new "websites". What FB is doing is the equivalent of AOL in the 90s.
What then, is the equivalent of a search engine like Google/Yahoo, in that world?
Which is why I'll now plug the company I work for, MindMeld, since (i) we do that better than Wit and (ii) we are not feeding our data to an advertising team.
What if you could get a recommendation from your friend's friend without asking them, and without violating trust or privacy? This is what I am building today.
Congrats to ar7hur! Here's the original Show HN introducing Wit.ai: https://news.ycombinator.com/item?id=6373645
M: "How may I help you?"
User: "What are my options for deploying a Python/Django project and making sure it is setup for scalability from the start? Compare five hosting providers for me. No, I don't know what metrics I should look for. Please research these and let me know what they are when you deliver the report. I also need an objective evaluation of our project in order to determine the risks that might be involved in going with Python 3.x rather than 2.x in the context of the libraries we might need to use in the future. Analyze the nature of our application in order to determine what the applicable libraries might be. Also, go through PEP's and make me aware of anything that might be relevant to the above. You have one week."
M: "My responses are limited. Would you like me to find you a restaurant?"
User: "No. I've lived in this town all my life. I know where most restaurants are and I know the handful I frequent. I need help with real questions. I can get the latest weather report, I can find a restaurant, I can order pizzas, I can go to the drive-through if needed and I sure as hell am not going to plan a vacation for my family this way. What I could really use is having you run through seriously time-consuming research, summarize results and present them to me in an easy to consume form. What I could really use is having you save me from doing 40 hours of research across 100 websites. Food, the weather and vacations are not a problem."
M: "Ah, but there's a great new BBQ joint not too far from you"
User: "I'm vegetarian"
M: "My responses are limited. How would you like a thrilling and exciting hunting safari in Africa?"
User: ":-("
Article pitches aren't all inherently bad ideas for articles either. A good example from my industry that I'm pretty sure was from a pitch is this one from the WSJ [1]. The basic concept of regaining focus at work is a strong one that resonates with people right now, but all the blog post ends up being is an ad for the product.
[1] http://blogs.wsj.com/atwork/2015/04/19/the-office-chair-desi...
It's not in FB's interest to make honest recommendations. If Bob's Burgers is paying $1000/mo in FB ads, but Karen's Burgers keeps being recommended as the "good burger joint", how long before Bob stops buying ads? And why would Karen start buying FB ads since she's getting exposure for free?
How is this any different than the approach laid out by Google Search? AFAIK, Google isn't suffering in the "search ads" department.
But then the wise man, who had much to learn, discovered that ads pay per view and click, and just like nobody reeps in a barren land, just like nobody pees against the wind, nobody advertises where the users aren't.
The users, to much a surprise, were at the website which valued quality.
The old village wise man learnt that Internet is powerful only in consensus. That what worked at his tubewell, will never work with running water. What worked at his ice factory, will not work for refrigerator.
But it's all AI - no human assists :)
But I'm far away from launching it and it's only a side project. But it's cool to see so many in the space doing something I also want / wanted to do.
I know NLP is difficult, and frankly i hate doing it, but i want an expressive language to "speak" to an internal process i use (a bot), and NLP seems like the only solution. I imagine a rule based approach is best (for my simple needs), but i have yet to see any examples that come close to wit.ai.
Appreciate any replies :)
It's pretty good for a basic set and you can train more. Ultimately, you need something that is learning online and that will require an understanding of ML techniques such as CRFs.
[0]: https://en.wikipedia.org/wiki/Jeremy_Howard_(entrepreneur) [1]: http://www.ted.com/talks/jeremy_howard_the_wonderful_and_ter...
1. Motivated team in a larger company builds new, cool product (in this case Messenger) 2. It's good and becomes successful 3. The rest of the company wants to get in on that, think of ways to add value 4. A bunch of stuff gets bundled, some good, most bad 5. Some of the original team stay around, most get disillusioned and go work on something else 6. Eventually, the app becomes another iTunes
A few weeks ago I was trying to find toys for my son. I was most interested in "things for a 6 month old". They did have that filter, but it was 0 - 24 months. At this age a few months make a HUGE difference. I wish the box was a bit more fine grained.
But, yes, it would have _worked_ if our requirements weren't so strict :)
So it can spend my money in behalf of me?
Ohh, wait, can it do something else?
Apparently this is 4 images and it shows up like this: https://www.dropbox.com/s/1n8wkixfimkmvlr/Screenshot%202015-...
Can anyone do this? If so, how exactly?
I assume OP was going for the James Bond feel, but it is M.
Well, M is also a fictional character in James Bond - head of MI6.
Q seems fitting (in terms of James Bond characters) to be your personal AI.
Afaik, they haven't been shut down. It's actually even free now.
[0] not that it's a bad thing, but wondering it's more than just "facebook bought it". Funny as it is, I trust that companies will go on when facebook buys them as opposed to google or amazon.
We have been building Q ! :)