I wish someone would clone a service like this and make if self hosted. I am completely uncomfortable with sharing my email data with a third party company.
This sounds incredibly hard to achieve with email contents - nothing identifiable at all? Does it do things like strip names out of email bodies? It says they learn contacts under priority stuff (which seems sensible) - so they do that locally?
Looking at their other features - what are they sending messages to the cloud for?
This has nothing to do with AI... Not long ago everything needed to have "built with JavaScript" stamp, now it must have AI to sell, but ...does it have electrolytes?
There's tons of areas where AIs, or rather, basic ML, could be used but for some reason isn't. I'm thinking about e.g. adaptive interfaces, the kind that can really be learning your habits on the fly. Think e.g. public transit app / widget that takes into account your precise location, as well as day of the week, time of the day, etc. to learn your habits and prioritize which timetables to suggest you by default (e.g. it's Wednesday evening, so you're probably coming back from work, so the default timetable will be the bus home at the stop next to work).
I tried to do exactly that for an university project, it's not very hard. Even simple methods can give pretty good results.
Google Now is about the first case I ever saw someone finally working on it, but it doesn't seem like it's a priority for them. My charitable guess is that they're afraid of confusing non-tech people.
I disagree with your assertion "There's tons of areas where that AIs, or rather, basic ML, could be used". The issue with AIs or ML or any probabilistic prediction of what user wants is that it takes away control. E.g. sorting your apps by usage. It's a solution in search of a problem. If you rearrange my apps by usage, you are asking me to search my apps everytime. It breaks my patterns and habits. And we are animals for patterns.
There are some definite places where AI/ML/Adaptive algos help. Examples would be Facebook search, Apps recommendation, music recommendation, smart playlists, autocomplete and search etc. where any good tech team worth its salt is already using these techniques.
It's also the reason why Knowmail or Inbox by google are not great. I go through my mail in a set pattern. I code those patterns by labels or folders and clean folder by folder. I don't need to context switch between each mail of similar definitive categories. In these mail solutions, I make decision of the context on each mail. Is it a support email or a collegue email or a mail from my priority mailing list or from a developer for tech problem?
And that's why I'm thankful that AI is not being used in the tons of places that you refer. You will put me in a world where I'm living by someone else's patterns that mine own.
At the risk of generalizing (but reinforcing the patterns thinking), it's a common pattern with Engineers to apply a solution on all kinds of problems without challenging the efficacy of said solutions for those problems.
I guess some people think about it differently than others. I don't see a properly applied ML as taking away control - I see it as reinforcing your control. ML as applied to UIs should be about your conversation with the app (and nothing else, no cloud bullshit please).
I haven't seen any actual adaptivity in Inbox by Google. The interface is static, it's just different than ordinary GMail. Is there something I've missed?
To address your example of sorting apps. This is a dumb idea mostly because - as you said - it interferes with habit-forming. Not that anyone cares these days, I can't find many examples where people would remember to not rearrange stuff like context menus pointlessly. But I think it could work as an addition. Instead of resorting your icons, just have a (small) area with, say, 5 apps most relevant contextually. This is something I'd actually pay to use. Hell, long long time ago I backed a project for an Android tablet homescreen that was supposed to rearrange visible widgets depending on your location and time of day. Great idea, but they fucked up execution (honestly, for such a project MVP is not enough, it gains utility per feature added in a superlinear fashion). I'd pay for something like this again.
> At the risk of generalizing (but reinforcing the patterns thinking), it's a common pattern with Engineers to apply a solution on all kinds of problems without challenging the efficacy of said solutions for those problems.
Hey, I'm not asking for a product like this to be built for general population. GenPop has plenty of flashy shiny apps already. I wish that someone would build a tool for a subset of engineers who think the way I do. The market is big enough to facilitate that. And apparently unwilling to cater for niches.
> I haven't seen any actual adaptivity in Inbox by Google. The interface is static, it's just different than ordinary GMail. Is there something I've missed?
If you turn on smart filters, it'll try to club promotions / payments / forums etc. smartly. It'd be right 90% time. But my threshold would be less than 1% failure.
If you open an email in Inbox, it'd recommend quick action replies like "Thanks, got it." or "Lets meet" etc.. It was on spot a lot of times, but I always wanted to add something more to the message and that's why I didn't use them.
> This is something I'd actually pay to use. Hell, long long time ago I backed a project for an Android tablet homescreen that was supposed to rearrange visible widgets depending on your location and time of day
I used One Plus2 briefly which had an area for smart apps. Since I could never predict what would be there, I never used it. I went there a few times to see if it would shorten time to action.
iOS has a recommended app at bottom left of lock screen. I use it purely as discovery. I've never found acceptable hit ratio of recommended app in my situational context and now I've blindness to that particular place.
If Knowmail guys are reading this, I'm not trying to discourage the product. I think Smartness should be in "Search mode" and not "Action mode". Action mode is when I want quick decision making and less evaluation. Search mode is when I'm open for back to back snap evaluations and unsure of next steps. i.e. don't interrupt my natural patterns. Ofcourse, I'm just one data point.
> If you turn on smart filters, it'll try to club promotions / payments / forums etc. smartly. It'd be right 90% time. But my threshold would be less than 1% failure.
I see. It's about as accurate for me, but I don't care much about the 10% of misclassified e-mails so I didn't even notice.
> If you open an email in Inbox, it'd recommend quick action replies like "Thanks, got it." or "Lets meet" etc.. It was on spot a lot of times, but I always wanted to add something more to the message and that's why I didn't use them.
I totally forgot about it. I.e. I've read an announcement of that feature, but I'm yet to see it. I don't know if it isn't available in my region yet or if I just dismissed it and forgot about it. Thanks for reminding me about it.
RE apps, I haven't seen it done right yet. I still think it can be done well, but developers need to stop doing it via trivial heuristic (frequency of use) or by hooking it up to a global recommender. Hell, maybe I'll write a proof of concept over the weekend (and who knows, maybe I'll even prove myself wrong, and discover why the idea is pointless).
> I think Smartness should be in "Search mode" and not "Action mode".
I agree with that. I think ML would be useful in "Search mode", or "Learning mode". Actions definitely should be static, so that one can have a chance to develop habits that speed up things.
> iOS has a recommended app at bottom left of lock screen. I use it purely as discovery. I've never found acceptable hit ratio of recommended app in my situational context and now I've blindness to that particular place.
Are you sure this is what you think it is? For me that app is always just an iCloud continuity app from my Macbook - like Safari or something.
I do have Siri app suggestions on the spotlight screen, those are typically useless.
You're right - looks like the same space is used for continuity/handoff and for the suggested apps - I've never actually noticed a suggested app like that.
I guess the original point was that machine learning is nothing like actual intelligence, it is closer to a process of artificial model creation than anything else. "Artificial intelligence" is a term that's completely abused these days.
Well, machine learning might not enlighten us to understand actual intelligence. There is no reason why actual intelligence can't simply be modeled by (and function like) a giant neural network. Many of us hope there is some form of symbolic intelligence deeper than just stochastic models or neural nets, but we really don't know.
> Machine Learning was certainly considered a part of AI then
That's part of the problem. The definition of AI is extremely loose. If by AI we actually mean "replicating the process of human intelligence through algorithms" then Machine Learning is certainly not it.
That's a question of terminlogy. Nowadays AI seems to include machine learning as a subfield. It doesn't matter much. But AI is a sexy term again (after the years of the AI winter), and companies market machine learning under AI now. Also, classic AI books now include ML, not just symbolic stuff, logic, planning and graph search.
While Machine Learning and Unsupervised vs Supervised algorithms are getting a lot of buzzwords -- language needs to exist to communicate to users what is happening.
No, this isn't 'Her' level AI prioritizing your email, but unsupervised Machine Learning is often swapped with AI to help people understand there is no staff looking at your email but there is still adapting rules customized to you, automatically.
I understand your frustration that is often felt about 'big data', 'AI', 'Machine learning', etc.
it's cloud based as they note SaaS: Conducts the “heavy lifting” algorithmic analysis of the anonymized information and generates an updated mathematical model for the client.
Somebody please explain to me what could possibly be there that would require actual "heavy lifting"? From skimming the features, nothing there seems to require much in terms of computation and data.
It's also a general question - a lot of the new ML-ish startups seem to do trivial calculations that should be totally doable locally, but for some reason they do them on my butt. Is it because (cynical view) butt-first is the easiest way to monetize users further (e.g. through data) and lock them in, or (less cynical view) it's just like with the big data - people drink the Kool-Aid and think they need a Hadoop cluster to do what could be done equally fast with a shell script?
It is also automated filtering and prioritization, as well as its aimed for corporate with individuals receiving hundreds of daily emails, thus cannot filter them semi-manually.
Furthermore, there are next-best-action features, as to save time, keep organized and reduce noise...to move to appropriate folder automatically, clean an email until a later (more appropriate time), mute yourself from a conversation until named, etc.
