While x.ai is very exciting, I have several concerns about their product and their chances at success.

1) Most meetings are scheduled internally within a firm where calendars are visible and shared. An AI PA that simply locks down time slots doesn't have much value add beyond what the calendars currently provide.

2) There are privacy concerns. Many will not trust x.ai with their calendar data due to concerns about x.ai data-mining or sharing their data.

3) There are other privacy concerns. People may not trust Amy and Andrew not to be social engineered by others to leak their calendars. The fact that person is busy at a certain time or for a certain duration or within a certain vicinity is potentially useful to people whose goals are more than just scheduling a meeting. This kind of over sharing may not be worse than goes on with Facebook, but the letters A.I. will make it sound scarier.

My last point is a wish. Most of my non-work meetings are scheduled over #1 text messaging and #2 Facebook. I would like to see x.ai support other mediums than just email.

2) Actual Risk. Software has bugs. When software is meant to give out some subset of data it is not unknown for it to give out more data than it means to. A human PA may make mistakes and blurb something that he/she shouldn't have, but he/she also knows when to stop giving out information. If a particularly trained AI starts leaking out information, it will happily do so for user after user after user.

[1] http://en.wikipedia.org/wiki/HAL_9000 [2] http://futurama.wikia.com/wiki/Project_Satan

The video clip talked about learning -- what's to learn?

For the sharing of calendars, why is that necessary?

So, my software sends you and three other people e-mail asking for a meeting with all of us somewhere in NYC Midtown next week.

So, each of you sends back some candidate times.

If you have the same software, it grabs the e-mail (say, from some tag in the header lines) and sends back to my e-mail some candidate times. My software answers, and there is a negotiation. If each person invited has the same software, then the software can negotiate, very soon, when all the software agents can be available to negotiate with my software agent, and then the negotiation can go really fast.

Sure, if your software proposes several times and my software won't accept any of those, then someone looking at the e-mail flows could find times when I was busy. A biggie? Maybe not.

Once all the agents are available, they could negotiate how to communicate peer to peer, that is, faster than e-mail, and really extensive negotiations could be done in less than a second. And the peer to peer traffic could be encrypted and, really, not easily available otherwise, that is, relatively secure.

My agent could have my appointments already and know when I will be at a meeting in NJ and consider my travel time to NYC Midtown to know when I could be available at Midtown.

What else to do?

What's wrong with that outline?

Where's the AI or the learning needed?

Yes, there could be some high push for virality and a network effect.

But, then, still, there's no very good barrier to entry. I.e., it's awfully easy to program, and if the scheduling software of some one company wants to schedule only with other copies of their software, then a disrupting company could offer to have open communications and be willing to schedule with other software. Then this little idea would go for nearly a freebie, and no one would make any money.

Ah, but encrypted communications and communicating with other software combine to yield a security problem -- anyone could write fake software that would report the unencrypted data. More generally, does look hackable so that there would be security risks.

Maybe a must-have for some people, but if they are meeting outside their office then there's a lot of travel likely. In that case, maybe use video conferencing.

I don't see much of a business opportunity there.

Marketing, I assume. Giving people their very own Jarvis to interact with could be a selling point (even if it's less efficient than a GUI-based scheduling solution), and the "AI" buzzword will certainly draw more attention to the product than a scheduling app would normally get.

No buzzwords needed, we are just trying to do some good work, and if we succeed, we think many people would like it: https://twitter.com/search?q=xdotai

But, I believe you are over estimating the importance of having some natural language understanding, under estimating the difficulty of achieving that understanding reliably in software, and are very much under estimating how easy to use, even for your mother, for anyone who can type, something like I outlined can be.

E.g., when I was in grad school, I took out some time to earn some money so that my wife and I could complete our Ph.D. degrees. For the money, I worked as an applied mathematician, computer guy in a research shop doing work mostly for the US Navy. So, yes, we had secretaries who did a lot of typing but had no word processing.

As part of my work in applied math, I ended up as system administrator of a super-mini computer. Of course, for programming, it had a text editor. If only for writing the computer manuals, it had a simple text formatting program, call it a member of the family of Runoff programs. And I'd gotten a daisy wheel printer nicely driven -- asynchronous serial communications with ASCII characters and the XON/XOFF handshaking (pacing) protocol -- by the computer. So, we were GO to do a LOT of typing.

Soon enough the secretaries wanted to get on board with this new stuff. So, for them, one at a time, we pulled some 5 conductor, general purpose signal cable with a 25 pin plug at each end (we only connected the pins for the signals SEND, RECEIVE, and SIGNAL GROUND), and a dumb terminal.

Bottom line: All the secretaries got on board right away. The ones without a terminal begged for one. No one ever gave one of the secretaries more than a few minutes of 'training'. They all caught on right away, to logon, files, the hierarchical file system, the text editor, the runoff program, the daisy wheel printer, etc. Right away.

Later as a college prof, I did the same for the business school, and again the secretaries 'got it' right away.

Gee, we weren't asking them to write macros in TeX!

So, back to my suggestion:

You want an appointment. So, you send something like

     === May we meet?  Let's pick a time and place.

     === My name

     === My job

     === My e-mail address

     === My phone number

     === Subject of meeting

     === When to meet

     === Where to meet

A human, including your mother, can read it and respond. And they can respond with a similar e-mail.

If either end of the communications is using your software, then the software can read that data and make sense of it much easier than for natural language processing.

Or, send the person some e-mail with a URL to a server with a Web page with a form and some JavaScript to help the person fill out the form. Then we're into just a Web form solution that we know very well how to handle.

Forms work great; people are really good at entering data into forms.

For people who have your app, it can replace the Web site.

No way do you have to use gibberish such as "RT, MT". Instead, everything a user reads can be in complete English sentences with a button with Help for more explanations, etc. It can be easier to use than most Web forms, and 1+ billion people use those. It can be easier to use than Google, e.g., where Google has those little horizontal bars to indicate (I never got that memo or read any explanation in a help message) for more options. Last I checked, Google had a lot of users. It can be easier to use than the Amazon Web site, easier to use than Facebook.

You are underestimating the ability of your target users to make use of simple computer user interfaces.

My guess is that there is no way natural language processing can be good enough that, net, it is easier to use than just some simple forms or something like I outlined above.

Gibberish like "RT, MT" is really a straw man to knock down. Instead, never but never should there ever be any gibberish or undefined terminology or undefined acronyms in any user interface. That rule is easy to obey, and natural language processing is a very hard way to obey it.

The real business need and opportunity is meeting scheduling, not finding work for natural language processing.

Algorithm != AI

He expected something else?

Because, I for one, would feel uncomfortable using such a conversational tone with an algorithm. I'm curious if they are encouraging anthropomorphism because...

  Amy will succeed more often in parsing this style of writing?
  This is how people *want* to talk to their AI scheduler?
  People should think of "Amy" as maybe a little-bit human?
  People should *confuse* "Amy" with a human?
  The more natural tone will increase user engagement?

Edit: I just went to the site and saw the ad copy for 'How it works'. It appears the answer is closer to making the Amy algorithm actually appear human. Users pretend they are CC'ing a real person, and unless the other participants on the email know better, they converse with amy@x.ai as if it really is a human? That is ELIZA territory and trying to actually beat the Turing test is a lofty goal indeed, I think far surpassing the challenge of merely scheduling meetings.

- We've seen that "Normal's" want to talk to their assistant without Syntax (my biggest worry is one of early users creating some artificial syntax)

- I like the idea of people seeing Amy as a little-bit human (it helps with the above two). I might be off on this one.

- We tell the guest that we are an AI Assistant (so no hiding)

- We hope it will increase user engagement, we do not know yet.

You'd be surprised how much text is annotated as "gratitude" :-)

Most of the solutions I'm familiar with require all scheduling parties to be on the same Exchange Server with calendar busy/away shared. Or you have to click-thru to a website and then deal with creating a user account.

Using email seems like a nice friction-free way forward. I hope they get traction!

Why force users to type several sentences (minutes wasted on a mobile device) when a couple taps/click in a GUI would suffice? Not to mention the risk for ambiguity that lies in natural language processing (unless this company has already solved the problem of strong AI?).

In contrast, consider doodle (http://doodle.com/), which is exactly the GUI you are imagining. It shows me a grid of possible times. I click the times that are good. It detects my location and automatically displays times in IST. Takes me about 15 seconds to use with no confusion.

X.ai seems to be solving the problem of "I wish I had a secretary, but can't afford one". That's not the same problem as "lets schedule meetings in the most efficient way possible".

Dear x.ai folks, if any of you are reading this thread: please clone doodle and stick an x.ai/doodle link into the amy emails. If you do this, I'll consider deleting the gmail filter which currently blocks all emails from x.ai.

Incidentally I don't mean to suggest doodle as necessarily the best service of this sort. The only reason I remember doodle specifically is because someone used it on me last week.

Having this discussion on Twitter right now with one of their developers. Go to their home page and click 'Sign in'. See the username/password prompt? Now, look at the URL bar. See something missing? In fact, the request is POST to https://... but the point is the user has no assurance that's the case. A web login form MUST be served from a top-level document under HTTPS.

The weird thing is that, in fact, the dev said they do provide SSL when logged into the app all the time to all users. It's only SSL on their public site that they only provide to paying customers. Well, I don't get that at all. And anyway it seems broken because I have a 14-day Premium Trial and I'm not getting that "feature". Just altogether very weird vibe. Anyway, this is now way OT for this thread.

Or, you would need to have them use an app, and make sure it is installed.

In both cases, the users would need to understand the UI. Even the most super-obvious UI takes a little focus to follow and use properly. Especially if they want something a little out of the ordinary ("if X is good, then X, otherwise Y, but only if Z").

Compared to those, writing a sentence in English actually sounds fairly convenient - that's how humans communicate naturally.

- Hi Amy, can we push the meeting tomorrow 20 min ?

- Hi Amy, let’s do the meeting when John is back from holidays (Friday, yes?)

- Hi Amy, Let’s do the meeting end of next week, preferably Thursday

- Hi Amy, Have John call my Skype (if he's got an account) instead of Cell, and add Linda to the invite please.

- Hi Amy, set something up with Jordy from Softbank next week, heck, I'll do something outside my scheduling hours as long as it is next week.

I scheduled 1019 meetings myself as startup founder in 2012 [0] and this is all I want, which obviously does not equal us being able to pull it off.

[0] https://x.ai/1019-meetings-in-2012/

Interestingly, in a sense as more people are using it, there's a chance that the technical aspect get easier (two x.ai users try to schedule something!). At least until a (hypothetical and will never happen) scenario in which every single email users will be using x.ai, at which point the program turns into a giant optimization and scheduling program, which will be hard - just something fun to think about.

Also, holy buzzwords and business-speak, I really like the startup, but I feel uneasy reading the article.

[0]: http://paulgraham.com/ambitious.html

Plus, think of all the times the AI doesn't understand the conversation and makes crappy suggestions. Like each time that API is cmentioned, suppose the system says "involve Carl! He's a Pro". i Can see this getting annoying really fast

The thing I'm puzzled by is how Amy can possibly follow the most important rule of physical meeting scheduling, namely: "you go to the money, the money doesn't come to you". In other words: vendor goes to startup founder's office, startup founder goes to VC's office, VC goes to LP's office... How does Amy know enough about your business hierarchy to respect that rule?

One way to think about it, if the first thing the algorithm can do is score its own confidence in its "level of contextual understanding" and "confidence score in the decision" then X.ai can put thresholds in place that get the threshold emails into the hands of actual humans when they need to be.

If the goal is 'up to 100% automation', the idea is exploit the 80/20 rule to launch quickly, and then pour money into making 80% -> 99%. So I guess the crucial question is, does the algorithm know its own limits?

The interesting pressure relief valve, is if a location isn't calculable with a high enough certainty, Amy would kindly ask her operator for assistance in resolving the quandary.

But our early beta users should not be to shy to cue in Amy:

- Amy, set something up with Lerer in Soho

- Amy, can you arrange breakfast with Matt and FirstMark please. They can pick a place.

- Amy, setup a data science / whiteboard chat for Prateek and I *

* My default meeting location is 48 Wall (Amy knows that, and simply assume I want this location used for a new candidate interview)

Could Amy differentiate Ken and Ben from the context?

> Q. Is Amy human?

> A. I am one part machine, and one part Human! The important part is that I handle your requests properly all the time.

https://x.ai/faq/

The concept of a good ai based personal assistant is one of the Next Big Things. And I mean facebook size big thing. If the busiest and most important people in society start doing all of their scheduling through one software platform it won't be long before everyone else follows. And that opens the door to all kinds of other social and communication services. "Have your bot talk to my bot and we'll figure out a good time to meet up". Maintaining a list of your friends and contacts is important, but algorithmically maintaining your relationships and schedule is even moreso. Ton of value here waiting to be unlocked.

As with all data based applications, the value increases exponentially with the amount of data being analyzed. The ideal personal assistant is one that will make the same decisions that you would make in the absence of instruction. If you tell a PA "book dinner reservations for my wife and I next thursday", a good assistant will know what kind of restaurant is appropriate, where you like to sit, whether you prefer dinner at 7:30 or 8:30, etc. You don't have to say, because they learn your preferences from the data they've collected from you and previous employers. Just like a computer would. There are a ton of small breakthroughs and conceptual leaps to be made between there and where we are now and the field is still wide open. x.ai seems like an encouraging step in that direction, one that I'll be watching closely.

(Before he edited his post OP had a comment about the confusion between the terms 'machine learning' and 'artificial intelligence')... Put another way I think machine learning specifically refers to the mathematical ideas used to teach things to a machine, and the ai to the resulting entity which often makes use machine learning techniques(usually in addition to specialized code specific to the application).

I think some people feel that "general intelligence" would require a paradigm that somehow transcends math and algorithms which seem to only result in specialized a.i's. And that maybe it all needs to be tossed out and re-examined with a blank slate. On the other hand, i would suggest that specialized ai software will become more and more generalized until we can't tell the difference. I think it's all about having a sane perspective. In this day and age with machines and gadgets in every room, man-made artifacts teach us how to think, and we teach machines how to think in a recursive interdependent cycle.

In that sense reaching the final goal post means progressively improving our machines to be active communicators and good listeners, to the point where they can learn things from each other. Eventually their interactions will become complex enough that their communications will be unpredictable but still productive on the whole. So we'll leave them on so they can talk. At which point you could call it a society of machines....

Anyway I guess the point is specificity of language seems to be important learning and thinking about AIs.

Yes, that is a widely accepted view of AI/ML. "AI" designates any method to give machines cognitive capabilities, while ML is a particular family of techniques (statistical learning) for giving machine such capabilities. Other branches of AI would include symbolic AI, expert systems, and other 70s techniques where processes were explicitly programmed rather than learned from data.

Today, because these other branches have largely fallen out of favor over the past 40 years, ML has become almost synonymous with AI. The semantic distinction is still there, though.

Is it good at dealing with timezones? I always get invited to meetings where people give me the time in their timezone, I have to realize that time wouldn't make sense for them if it was in my timezone etc etc.

The other question is regarding the fact that I don't communicate with everyone in english. There is no mention of how good Amy is when switching to and from different languages. This last one is one of the reasons I never used virtual assistants.

Finally, she only understands English today (2015). We are super ambitious, but we want Amy to perfect one language at a time. dmortensen keeps pitching for Danish as Amy's next language around the office...we'll see.

We have this fantasy (delusion perhaps) of us becoming arbiters of time. You don't need a calendar, because Amy will help manage your time. One step at a time though and the problem is hard enough as is, before we print "Arbiter of Time" t-shirts.

> N-Gram, the latter part of her surname, is a technique used in computational linguistics.

it's not a technique; it's just a name for a sequence of n things. Also, I've never seen the 'N' or 'G' capitalized.

[105 of them as I write this...plus their coworkers.]

Let's just not forget about the positives here: there is a company that openly uses term "A.I." and raised a lot of money. Am I the only one here waiting for the AI Winter to be over [1]? Maybe it's not the best A.I., yet, but they got the funding. Personally, I think that's great and I hope for others to come out of the closet.

[1] http://en.wikipedia.org/wiki/AI_winter

So far it has paid out, because large tech companies (Google, Facebook, Baidu, etc.) have been on a deep learning startup acqui-hiring spree. It might not last, though. Any company that bases its marketing on "AI" and subsequently fails to deliver is bringing us closer to the next AI winter.

Just like a dishwasher is not a replacement for a scullery maid; and won't individually polish each piece of silverware like a proper butler. This doesn't need to be a human level fully rounded intelligence imbued with common-sense. It just needs to be able to reconcile schedules and locations and travel times.