Hacker News new | comments | show | ask | jobs | submit login
Write Emails Faster with Smart Compose in Gmail (blog.google)
215 points by devhxinc 44 days ago | hide | past | web | favorite | 190 comments

This is basically autocorrect on the level of phrases. If that's true, this will have an enormously homogenizing effect on the way we speak (spell check alone has caused an actual shrinkage of the English lexicon [1]).

I wrote extensively about this for Mondo 2000: (http://www.mondo2000.com/2018/01/17/pink-lexical-goop-dark-s...)

[1] https://www.nature.com/articles/srep00313#references

From the paper you cited:

> In our analysis we treat words with equivalent meanings but with different spellings (e.g. color versus colour) as distinct words...

I think it's a bit of a stretch to use that to say that English lexicon is shrinking. The number of different spellings for the same words is, yes, but that's not necessarily bad. Other articles suggest that English vocabulary is still steadily growing: https://www.theguardian.com/media/mind-your-language/2016/fe...

I think you made good points regarding QuickType though.

This is happening everywhere. We're being made more and more homogeneous. The same applies to other aspects of our life, e.g. music, films etc. The general rule is what is popular is safer, so it's being promoted more. In this case what is less popular is not even given a chance to surface - no matter if it's popular radio or YouTube recommendations.

I feel like we are becoming less homogeneous. When I was a kid everyone watched cable television after dinner and everyone watched all of the same shows. Thursday night was Friends. Friday night was Full House and Family Matters, etc. Now I find everyone I know watches all kinds of shows, many of which I've never even heard of. Not to mention YouTube which is whole new layers upon layers of content. In an odd way, I actually kind of miss the "bonding" cable television used to give the populace (but not really).

I just chose shows as an example. But I feel the same happens with reading, music, technology, politics, etc.

Personally I think you’re both right. We are becoming less homogeneous as a society, but the people within those subgroups interact much less than they used to, at least in person.

Yes, I think this is an accurate observation. There seems to be more variation on the fringe and more homogeneity closer to the center (or over large areas).

Thanks for this. Sometimes I feel like I'm taking crazy pills when I read something about how some thing or another is worse today than it was "back in the day". It basically never meshes with my experience of that time.

Local musical scenes used to have their own sounds, but thier influence has fallen away to more distant influences. If we all listened to POP top 20, then this would have homogenized music. Instead, there is much more variety and musical subcultures than ever before.

Globalised phenomena (Cola, Marvel, McDonald’s) can have localized meanings (trashy/hyped/cool/classy/masculine/feminine). It’s not an either/or thing. In sociology this is dubbed glocalization.


that is watering down the point i think. with companies becoming more and more like amazon, google, and apple, who care about statistics than individual customers, and many, many other aspects of modern society, we are just turning humans into numbers. everything we do is channeled and controlled, which isn’t negated by mcdonalds serving dim sum or whatever in china or other “glocalizations”.

> being made more and more homogeneous.

It might be the way around. This kind of intellectual homogenization might indicate an emergent entity coming to being - some kind of (new?) societal organism, sort of like national state, but I don't know what it actually is this time. We might never learn.

No, it's on a different scale. I'm talking about things like an anthill or a tuna shoal kind of emergence -- a collective animal.

An ant is a phenomenon that emerges from the interaction of genes as they react to the environment in the same way that an anthill is emergent from the collective behavior of the individual ants as they react to the environment.

Memes would be the basic building blocks of any organism that emerged from the manner in which replicated ideas collectively reacted to the environment. Arguably this describes human culture, though I think that's a long way from exhibiting independent intelligence (but maybe we'd be incapable of recognizing it).

Presumably something could emerge from the collective interaction of such meme-based organisms. Though like with ants, where the phenomena of an anthill can also be understood directly in terms of genes (all the ants are siblings), I suspect it could really just be reduced to the interplay of mutually reinforcing memes.

I wonder what the analog to delineating the internal and external environment is in the meme-based universe....

It's funny that you use those examples. In 'The Selfish Gene' where the term 'meme' originated, it's used to describe any 'things' that Darwinian forces act on. So things that persist and spread in the way that genes do, but aren't genes.

One of the examples used to illustrate the point was beavers' dams. So 'meme' always did exactly mean persistent phenomena like anthills and tuna shoals (although 'collective animals' would be a subcategory, it referred to more than just those). The word has recently been semantically narrowed into meaning 'a picture on the internet with some words on it'; sadly, in my opinion. The original meaning was much more interesting and useful as it ties in with Darwinism, or the theory of evolution applied more broadly and to things which aren't necessarily biological organisms.

Apparently Dawkins isn't bothered by it himself though and likes the current usage (or so I read somewhere).

I feel these constructs were always present (compare the behavior of a pack of wolves vs a lone wolf) but in our age they're large reinforced by the "pack" being orders of magnitude larger. See how lynch works in social media: someone gets accused of something, then an army of disgusted users carries on the attack, which is not only virtual today. Individual conversation in these circumstances is not possible: you're not talking to individuals, but to a Mob, an animal with it's own behavior and very limited reasoning abilities.

The communication with bots and computer systems that are designed to interact with humans and equipped with the so-called "AI" is of similar nature: it seems there is some exchange, but it's very shallow and limited by the abilities of the receiving end - you can never transmit anything new, anything that the system is unable to understand. Continued use of such systems might have hard to predict consequences.

I'm not sure how true your observation really is, but I sympathize with the sentiment. My personal fear is a future where the only freedoms left are to work and consume.

Maybe now that you are older and wiser you are seeing the patterns.

Next steps: AdWords in email suggestions. “M” expands to “Meet you at Starbucks (TM)”.

If you write a word your spellchecker doesn't like, and you accede to the spellchecker, then it's not actually a word. Not to you. An utterance is only a word if you believe it in it and are confident enough to defend it against the opinion of a dumb machine.

So this alleged culling of the lexicon by the spellchecker is largely a shakeout of non-words.

Firefox's edit widget is underlining "spellchecker" in red, but I'm hitting [reply] anyway. I don't think it needs a hyphen that badly.

wait, so according to that, old words that are still written in old manuscripts/stones down which but we don't remember the meaning of, simply because they weren't used for too long, would not be words? I tend to disagree...

They are effectively like words in a foreign language.

I've heard arguments going both ways. That the internet and connectivity is going to dramatically slow language drift as well as accelerate it. I think the real answer is that we are rapidly changing how we talk but to a homogenous standard.

I really wonder how global language will change over the next century. I would really like to know if English continues to borrow from other languages or all these dictionaries are going to stop that.

That looks like a nice page! As feedback: it lacks any kind of meta-information in human-readable form about what it is. The best I found was

<meta property="og:description" content="Mondo 2000 The original magazine of cyberculture returns. Watch out for your overcoats!" />

Which was actually more helpful than the twitter or facebook bio texts :P

It's worse than that though; at least in Allo, it began suggesting smart replies before you even touched the keyboard.

This is an opportunity to stand out by using less common phrases.


agreed. auto suggest or auto correct for searches means we only search for things Google recognizes

I feel the same way when writing golang code.

Holy crap, everybody burn you're dictionaries.

> you're dictionaries.

that's why we need them :-P

Google Mail would suggest: Holy crap, everybody burn your dictionaries.


Unless they meant "Holy crap, everybody burn: you're dictionaries!"

Reads like a Fahrenheit 451 / 1984 mashup. The perfect dystopia.

The interesting thing is that autocorrect tries to fix many alternate spellings actually included by dictionaries (e.g. "judgement").

I find that even when I set my country to UK and my language settings say 'British English', I'm still getting various autocorrections to US English to the point where I'm maintaining a list of dictionary overrides for words that us Brits spell differently (or have a distinction for), and there are plenty of words that it claims don't exist when a cursory dictionary search elsewhere proves otherwise.

It's mostly manageable on a laptop because it's easy to go back and edit text and make your computer behave. On a phone it's ducking^Wducking^Wfucking infuriating because you can automatically get the same correction a good two or three times before you can get the word you want, and you are very actively fighting a tool that is trying too hard to help, to the point it will deliberately misinterpret your keypress if it thinks it has a better idea of what you wanted to type.

Recently my Google keyboard's auto-correct has developed an annoying tendency to insert apostrophes when they're not needed (e.g. "its" -> "it's", "were" -> "we're"). Even manually deleting the suggestion only works until the next time I do need to write the contraction, whereupon it resumes "correcting" the non-contracted word every single time.

Only fix I've found is to manually make the word a shortcut for itself (e.g. shortcut "its" for "its") so that when I type the word it inserts itself and is not auto-corrected.

I think the trend will lead to illiteracy. Why learn to read and write when your computer can just speak to you and transcribe everything you say to it?

I recently saw a signature board for honor students posted at a high school. The signatures from the early 1990s were all in cursive and every signature had its own style. The signatures from 2017 were mostly printed in block letters and looked like something a first or second grader would have written. It was pretty sad.

I don't think lack of handwriting skill has anything to do with writing ability, which is generally understood to be the ability to express ideas in text. Of course keyboard skills have supplanted handwriting. Perhaps dictation will replace keyboard skills soon. The method of written creation is irrelevant to the output.

Yes, except what we are talking about here is the computer doing the work of composing content, not merely transcribing it.

It is relieving the human of the work of actually thinking about what they want to say and how they want to say it-- that's the thing that the human should be doing, and needs to know how to do to be an effective thinker and an effective participant in society.

How a person speaks and writes tells you about how they think-- that is what language does; it opens a window into the mind of another person. With this tool, I am no longer looking into the mind of my conversation partner, I am looking into the mind of some sophisticated statistical aggregate of a huge population of Google users.

This kind of thing is a huge insult and injury to human communication.

It seems unlikely that the brain-to-text rendering method has no effect on what is written.

My phone-thumb-typing voice is different to my laptop voice is different to my handwritten voice. I can’t prove this but I believe it is true.

And of course this new development is actually proposing phrases. This seems really new.

I tend to agree. It wasn't that long ago that kids writing in cursive couldn't wield a quill and blotting paper effectively. Or a wax tablet before that.

At least, until some Doomsday scenario arrives.

I don't think it's irrelevant. Composing text deliberately by hand forces a more measured pace of thought and development of ideas. I've even heard educators say that it's not worth teaching arithmetic anymore because everyone has a computer, or a phone with a calculator. But if you don't grasp the concepts of addition and multiplication that are developed by actually doing it over and over you're never going to be able to handle the level of abstraction needed to progress to algebra, geometry, trig and calculus. Spelling is also not given much importance these days, with "creative" spelling being acceptable at least through the elementary grades. Even in high school it isn't given a lot of weight.

If you don't learn the meaning of words and how to spell them, and how to construct a grammatically correct sentence without the assistance of a computer, you become mentally weak just as you become physically weak if you never exercise your muscles.

Students are losing the ability to think. The ability to think is what is being fundamentally eroded by this sort of automation. Nobody tries to solve problems. They just reflexively google the question, or ask Siri or Alexa, and accept the first answer without giving a lot of thought to whether it even makes sense. I've caught myself doing it.

More and more we are having our worldview formed not by critical thinking and development of our own opinions, but by what Google and other big media companies tell us.

With all due respect, you sound like Homer bemoaning the advent of writing and it's effect on memory.

> Composing text deliberately by hand forces a more measured pace of thought and development of ideas

Penmanship is orthogonal to the ability to write by hand. I for example have atrocious hand writing and can still easily write my thoughts down, and often do. Other's will have trouble with my written notes, but they're not meant for external consumption, I type for that purpose. Cursive versus print style writing are mostly interchangeable for writing down ideas, and learning to write well in cursive takes up to a few years. It may not be time well spent.

> I've even heard educators say that it's not worth teaching arithmetic anymore because everyone has a computer, or a phone with a calculator.

I know a number of teachers, and try to follow trends in US education, and I've never encountered this idea. Are you sure you aren't constructing a straw man here?

> Spelling is also not given much importance these days, with "creative" spelling being acceptable at least through the elementary grades. Even in high school it isn't given a lot of weight.

At least in the US, this is not the case. Any state that uses the Common Core follows these [1] standards, including phonics, building words from stems, and differentiation of homophones.

>If you don't learn the meaning of words

Who's saying anything about this? This tech looks like it just guesses at the banal pleasantries you type at friends and colleagues 1000s of times per year. I know I write the same phrases almost to exhaustion, I'd love to autocomplete the trivial stuff, and spare my hands an wrists. *note there's a great accessibility story here, where people with manual impairments will be able to answer emails more easily.

> Students are losing the ability to think.

Are they really? I'm not sure there's any evidence of this actually happening.

> More and more we are having our worldview formed not by critical thinking and development of our own opinions, but by what Google and other big media companies tell us.

How is this any different than mass media dissemination of information from the previous generation, when everyone watched the same 4 TV channels and read the same 1 or 2 newspapers.


I'm just old enough to remember the ridiculous "hunt and peck" style of typing nearly everyone around me employed in the 90's.

They may have had fabulous penmanship but I could tap out Morse code faster than they could type.

Why sad? Would you prefer we all learn how to use a feather and ink? or perhaps chiseling stuff into rocks?

I only write my bespoke contributions with my gold-plated Lamy chisel on Italian white granite.

I think children these days do not appreciate the art of text.

What's wrong with carving symbols into bamboo? Kids these days.

I don't think a native English speaker would benefit much from this feature. But as someone for whom English is a second language, I am sure that I would have appreciated this when I was less fluent in English.

I am now trying to learn French, and I find myself searching to the most trivial sentences in linguee.fr. Learning grammar and vocabulary is one thing; knowing the idiomatic way to express one's message is an entirely different beast. Things that might seem obvious to a native speaker, e.g. the difference between "I'm fine" and "I'm good," are completely unapproachable for a non-native speaker for the first few years of using the language. Now that I am fluent in English, I would find this feature annoying. But I wish I had access to something like this 10 years ago.

This has been one of the most consistent themes since the first prototypes we tried. It is particularly helpful for non-native speakers, but the large majority of native speakers find it helpful too.

This is why I've taken to watching a lot of TV in languages I'm interested in. Challenges with such:

1. A program I'm actually interested in and with which I can engage. And access to same.

2. Differences between: Spoken and written language. Dialect versus "standard usage".

3. Passive as opposed to active participation. (Also, I seem to have forgotten my original point 3.)

Even when I'm not entirely certain, what I've gained through this gives me a starting point in having an idea of what the/a proper idiom may be, which I can then take to the Web if I need further research/verification.

I wish I'd had access to such resources when I was learning my first languages. Instead, I was largely limited to the classroom and the professor for such exposure. (In most of the U.S., and with other than Spanish and perhaps family and friends, foreign languages are -- were, certainly, a generation ago -- "far away".)

And music. I happened across some that I enjoyed. If nothing else, this helped give me more sense of the sound of the language.

Yes. On a smaller scale, iOS predictive text is helping me learn a new language. Unexpected suggestions become new things to learn about.

Here is something funny. Google's Japanese Input on Android does English dictionary lookup when you're in romaji mode. And, here is the kicker: you have to use exact spelling, or no dice.

The exact spelling for Japanese is a lot more reasonable, though. In English there are lots of spellings for the same pronunciation. In Japanese there is only one spelling for each pronunciation. If you get it wrong, then you are writing a different word and there is very little you can do in the general case.

Of course the problem with romaji is our habit of spelling things wrong :-) To highlight the irony, "romaji" in Japanese is ローマ字 (literally roman (Roma) characters (Ji)). The ロ is "ro", the マ is "ma" and the 字 is "ji". But there is one character left! ー extends the "o" sound for the "ro" for one extra "beat" (Japanese is a rhythmical language). I don't even know the correct transliteration for this other than "rōmaji". For Wāpuro rōmaji I think I would enter "roumaji", but I think this is not actually correct romanisation.

Even to this day I mispronounce that word because of stupid romaji :-)

My comment is about English lookup when the keyboard is flipped into romaji mode, and the user isn't composing kana.

Interesting perspective, but a) I think it mainly draws its word choice from your own past messages anyway, and b) wouldn't an integrated translate button be more useful to this purpose?

So why exactly are people suddenly up in arms about this?

How's it different from opening up your android phone and using google's spellcheck, with algorithms that predict your next word when you're typing?

Never seen such atrociously negative comments to what's essentially a cool feature you can disable.

PS: How many of the outraged people here are actually voting with their wallet and paying for something like Fastmail? (https://www.fastmail.com)

Two big problems.

One, AI mediation of language sees issues like bias amplification (AI learns our biases then suggests them back to us). Google itself wrote about this recently [1].

Two, spell check has a homogenizing effect on spelling [2]. If you extend this to the level of phrases, that will homogenize not just spelling but the very way we talk.

[1] https://developers.googleblog.com/2018/04/text-embedding-mod...

[2] https://www.nature.com/articles/srep00313#references

To add a third, I have a problem with my personal emails being used as a corpus of text to train these models.

Maybe they're using only my emails to train a model for me, and turning that feature off disables them doing so. Maybe they're using everyone's email to train a generic model.

Of course, they've been (machine) reading your emails to provide ads since forever, so in that sense maybe this isn't really that new. That said, many people habituated to email reading for ads back when Google's brand image was riding higher.

We're training Smart Compose from highly anonymized global corpus of consumer email.

We've stopped using emails for Ad targeting entirely, and the enforcement is quite strict, as reported in June: https://www.nytimes.com/2017/06/23/technology/gmail-ads.html

> highly anonymized global corpus of consumer email.

I read this as confirmation that, if I were a Gmail customer, that you would indeed be using my personal emails to train your models.

For me, I then have to wonder what "highly anonymized" means in this context. Does this means that users' emails are put through yet another processing pipeline to separate emails into anonymous / not-anonymous buckets or to redact personally identifying information?

[H]ighly anonymized global corpus of consumer email.

What do you mean by "highly anonymized"? What specific privacy notion are you using, and what is the (demonstrably unavoidable) tradeoff you are de facto threading between utility and privacy?

You don't have an email in your profile, so I'm just gonna hijack this comment. :)

I can't enable smart compose. My general tab doesn't have "Enable experiments" at the bottom. Is this getting rolled out in phases?

I logged out and logged back in just in case, but no dice.

My email is in my profile if you need it. Thanks!

Rolling out over the course of the day. We'll be at 100% in a couple hours. Thanks for your patience. :)

Sweet, thanks! I'll just keep hitting reload till I see it. ;)

Yup, I forgot where I read this, but I believe all Gmails are eligible for the model. Again, forgot where I saw this, but Google takes great precaution to protect your data from human inspection, but being used in the corpus is fair game.

I think if you are a (paid?) enterprise user, they do not do this.

On the other hand, sussing out what is and is not fair game is quite a lot easier with a service such as Fastmail.

Doesn't have to be "enterprise". Anyone can pay $5/mo for GSuite (https://gsuite.google.com), which includes gmail and a ton of other tools. Furthermore, it's on your own domain rather than @gmail.com, which is not only better (you truly own the email address), it's also a hell of a lot more professional and makes it easier to migrate to other services if you wish to.

To put it bluntly, if anyone here is complaining and has an @gmail.com address, email me and I'll be happy to help you buy your own domain and create MX records to either gsuite or fastmail. But I just can't hear the empty negativity anymore from consumers who say "no! bad google!", yet, by using the free product, keep making their emails available for the AI models and keep signaling to Google that an advertising business model works better than a paid-per-user one.

Why can’t I buy gmail and just use regular @gmail? Would love all the features of GSuite, but a custom domain name I would pay even more to avoid. Unless you’re running a business, it looks bizarre. For some in the development industry / those who write technical blogs I can imagine that being useful at times, but still......

Go to https://gandi.net and buy a domain name you like with your name in it. I recommend something with your last name. <lastname>.me is a good one, or something more clever if you like.

It doesn't look bizarre, trust me. It looks extremely professional. Even on CVs, this alone will get you past some filters. And the domain name is $5-10 USD / year, it's nothing.

Get a personal domain name, e.g. "taylor@bushinka.me" if your name is Taylor Bushinka.

With @gmail.com, you have lock in. With @yourdomain.com, you don't.

I've been meaning to do this for ages. Thanks to you, I just did.


Let me know if you have any tips for getting the most out of my new subscription!


Go to https://admin.google.com - there's a lot of cool stuff you can do there that isn't usually possible in a regular Gmail account, including for example recovering emails and drive files up to 3-4 weeks after they've been deleted.

That's a generous offer, and one that I'd echo. If anyone is unsure where to start, or wants some advice, I'll be glad to help you with domain registration and email hosting MX record setup.

I don't think this is true. When our large university switched to gmail, all I could find was that they said they would not use our emails for targeting advertising. Seems like that leaves a lot of other possibilities like profile creation (they just won't use it to target ad at you), machine learning of all types, cross correlation with my personal email for who knows what, etc.

Paid domains can choose if they want us to use their data for training. Many do since it improves the quality of features they receive.

This is correct.

Don't use Gmail if you don't want Google reading your mail?

I'm a paying Fastmail user. I switched over after they (EDIT: Gmail) started announcing compatibility-breaking features. Processing my emails to suggest content is one of a litany of reasons that I'm glad to have switched my email to Fastmail.

In this particular case, I think people are (probably rightly) concerned about how the machine learning models here are used to suggest content. I would guess that most of the people who are unconcerned with the Android keyboard's privacy model either haven't thought too much about it, or have come to accept it in small increments.

At this point the only reason I haven't ditched Gmail is because that's the email address I've used everywhere. Changing email addresses is even more problematic than changing physical addresses these days.

The reason people are so upset is that it has a negative effect on the world even if people have the option to turn it off (and don't), and it has virtually no real benefit. It's the latest in the trend of tech companies trying to solve problems that don't exist, and creating new ones in the process.

At best, you'll never know if the person on the other end of the line outsourced their response to Google. At worst, people's mental capacity for expression and nuance will start to atrophy because they couldn't be bothered to manually relate to another human being.

I've recommended it elsewhere in similar threads, but a nice middle ground between self-hosted email (lots of work), and a straight-up Gmail/Yahoo/Fastmail/Hotmail/Etc account is to buy your own domain and set up your MX records in accordance with Fastmail's (or whoever's) instructions.

I've ported my email address / domain across several hosting providers now. I don't have to retrain anyone.

Obviously this doesn't help you now, but maybe you can start training people on to your new brundolf@superfancydomain.org email address to avoid this problem in the hypothetical future when Fastmail becomes dystopian and gross.

The best time to start using your own domain for email was 10 years ago. The second best time is today.

Get your own domain, set up an email forwarder on it that points to your Gmail, and start shifting your contacts and accounts over to it. Then if down the road, you decide you want to switch, most of the work will already be done. I spent about 18 months on Gmail using my own domain before I switched to FastMail, and it made the transition much less painful.

I think HNers are overestimating the amount of thought people put into composing emails absent this feature.

How many times does it take for people to understand that incremental "cool" features are exactly what leads to a dystopian future? Facebook added one cool feature after another until suddenly people realized that Facebook knew everything about you; only then were people alarmed. This is so obviously, blatantly terrible it's beyond belief that anyone could be in support of this. It's literally bending the way you think to fit more closely into a form field, rounding out the edges on human interaction. Do you really think that most people will take the additional time to write out their real thoughts, or just go with the autogenerated approximation? Just wait until you start typing "can you go to the supermarket and pick up" and the thing suggests Doritos, because that's exactly where this is going.

Are you seriously not seeing the difference between [human] giving information to [company], versus [company] making a suggestion to [human]?

I never liked Facebook; I never thought any of Facebook's cool features made up for their creepy business model. I do like Google and Google Mail because I pay for the service and the relationship is extremely different. But that's outside the point anyway. You're giving information to the company by simply using gmail (receiving and sending emails through gmail). Autocomplete is a feature that doesn't gather information (at best a yes/no)...

Seriously, wtf is dystopian about autocomplete? Do you write your code in nano or something?

And how much thought do you put into an email anyway? Outside of development ones, all my emails are extremely standard (the few times I even bother sending any). I won't use this for development emails. What is the problem?

I never understood the need for autocorrect or even spellchecking if I'm not writing a report or anything official. Am I a luddite for thinking that there are a lot of technologies that provide very superficial improvements to our lives while opening the possibility for serious problems (see other comment on parent about incremental changes that may be not harmful by themselves)?

A report, anything official, anything professional, or really anywhere you don't want to appear to make typos left and right.

Furthermore, if you're dyslexic, autocorrect is essentially assistive technology.

Now, do I need autocorrect when I'm texting my mom? No, I don't, but that doesn't mean I want to be turning it on/off every time I switch context. What harm does it do by staying on? Hide my true nature as a poor speller?

I'll let you in on a secret: My handwriting looks like shit. Digital text has allowed me, for the past two decades, to hide my true nature as a shitty handwriter.

I pay for FastMail. Voted with my wallet two years ago.

If I get any impression someone is writing me messages with this feature, I'll filter them into my trash. I want to speak to people, not Google.

I got news for you, a large chunk of the HN userbase is on mobile and has been using predictive keyboard input for years.

A large chunk of people on the internet, in fact.

You're speaking to robots right now.

>If I get any impression someone is writing me messages with this feature, I'll filter them into my trash. I want to speak to people, not Google.

This seems like an extreme overreaction and quite antisocial.

Sending me robot-generated messages is quite antisocial!

See scrollaway's response above. A lot of the text you already read is robot-generated. More than half the words in this comment were generated via autocorrect or auto complete in some fashion. It's still useful and thoughtful though.

> A lot of the text you already read is robot-generated. More than half the words in this comment were generated via autocorrect or auto complete in some fashion. It's still useful and thoughtful though.

Just to be clear, I don't find these things reassuring, it just makes me think you don't care about communicating very much. Please accept that some people find machine-generated communication to be gross and don't want to be subjected to it, whether or not you understand.

On the other hand, I care about communicating so much that I want to focus solely on the thoughts, not the form. So I defer the form (such as spelling and completing phrases) to a tool, so that I can focus on the content. This is essentially the same thing that writing tools like Hemingway do. They force you to focus on content, whereas autocomplete lets you focus on content.

Saying you don't want to be subjected to machine generated communication to me says that you don't want to read anything written in the past 10 or so years. Practically everything written in that time frame is in some way machine generated or edited.

Like I said, just because something is in part machine generated doesn't make it not thoughtful.

I think trying to suggest spellcheck and your email client writing sentences for you as the same thing is incredibly disingenuous. One fixes the content, one generates it.

I'm suggesting that autocomplete of words and autocomplete of short phrases aren't substantively different, yes. Especially given that a lot of mobile keyboards already essentially do this.

I can type "I" and then my phone's keyboard will suggest "be there at" as the phrase completion it does this a single word at a time, but in practice it does let me complete a phrase. If I don't want to say that, I'm capable of not using the autocomplete suggestions. There's no substantive difference with Smart Compose except that I don't have to confirm between every word, so its slightly more efficient.

I'm not sure what you think smart compose is doing, but "writing sentences for you" isn't it.[1] So no, I think you're either misunderstanding what this is, or perhaps you're being disingenuous by describing it as something that writes sentences for you.

[1]: Granted given the Duplex demo and other work in the NLP space, I expect that in certain contexts a tool that given an instruction like "set up a meeting with John this week" that schedules it over emails (like x.ai[2], which has been around for quite some time) is totally possible, but Smart Compose isn't it.

[2]: https://x.ai/

> I want to focus solely on the thoughts, not the form.

I think this is a dangerously stupid way to think about linguistics.

I mean in certain contexts absolutely. I certainly wouldn't use smart compose to write a novel, but then I don't think it would be very helpful for creative writing where or writing where my focus is on tone.

But for otherwise mechanical things ("hey can we set up a meeting tomorrow", "do you want to grab lunch at ____ today"), the form is mostly irrelevant. It's the kind of thing where if we weren't in a formal context, I'd write as "hey u wanna grab fud at ____ later". I'm not losing anything by having a tool formalize that to "Hey do you want to eat lunch at ____ today." Again, we already do that a lot. As far as I can tell, this is just doing that a bit faster.

Good lord that's scary. Rapidly ushering in the future where none of our words are our own.

Anyone ever read Roald Dahl's short story The Great Automatic Grammatizator ? 10/10 strongly recommend, about an engineer who develops a computer-like machine able to compose fiction.

Googling "the great automatic grammatizator pdf" gives a .doc version as the top result, for me anyway.

Oh no, more autocomplete!

I don't see it as much scarier than tab-completion while programming or on a phone. As long as you're watching over it and fixing mistakes, it's just a convenience.

Tab compiletion while programming is hugely different then this completion and phone completion is somewhere in the middle.

While I am programming I usually know what method I want, I might not know the exact name but I know what it does and when I see it I select it. There may be a rare case where I want to see what methods are available but that is just a convient way to read "documentation". However this is guiding your sentence as you write it. I suspect that most people would have different phrasing if using this feature then they otherwise would.

If it wrote the entire email for you, and you watched it to see there were no mistakes, I suppose that would just be a convenience too. But at some point you cross a line and your voice is lost. That line draws closer, inch by inch.

At that point it shades into ghostwriting, which can be done well or badly. If done well, it should be imitating how you'd write it yourself.

"David Ryan is the designer of ELOPe, an email language optimization program, that if successful, will make his career. [SPOILER: ELOPe is essentially Smart Compose in Gmail] But when the project is suddenly in danger of being canceled, David embeds a hidden directive in the software accidentally creating a runaway artificial intelligence.

David and his team are initially thrilled when the project is allocated extra servers and programmers. But excitement turns to fear as the team realizes that they are being manipulated by an A.I. who is redirecting corporate funds, reassigning personnel and arming itself in pursuit of its own agenda."

"Avogadro Corp: The Singularity Is Closer Than It Appears" (3 part book series)

https://smile.amazon.com/Avogadro-Corp-Singularity-Closer-Ap... (Book one)

I despise the homogenization of thought that these tools lead to. My words are my own, even my typos.

I don't think there's any basis for the argument that autocomplete tools lead to "homogenization of thought", and therefore no reason to "despise" them.

I don't think typos communicate any information of value.

The "My address" autocompletion gave me that feeling too.

People are saying this is 'basically the same as autocorrect or predictive text'. It's not. Autocorrect doesn't make any creative decisions for you, whereas this does.

That is to say: we think at the level of words, not letter-by-letter. When I make a typo, autocorrect corrects what my hands do to match what my brain is thinking. My brain still has primacy. This thing sits at the level of words and even sentences: if it's autocorrect, it's working to correct what my brain is thinking. Which is creepy and sad.

It's a little bit more like predictive text I admit. But because predictive text only suggests one word at a time, there's little semantic meaning to a suggestion and it's rare that I have my thoughts distracted or changed because of it. It's still largely a convenience tool. Suggesting a full sentence is shaping the direction of your thought, which is very different.

I'm still horrified that Google has put this out.

This is awful. We need _less_ mediation and commodification of our personal interactions, not more. What is the use of this? At best this is a solution searching for a problem, at worst it's an attempt to standardize our communication in a way which makes semantic meaning easier to analyse.

Agreed. Rather than something customers asked for, this feels like something driven by the culture at Google: "AI all the things" and "build new things" to get promoted.

You should receive the amount of corporate email I do. I don't want my email responses at work to be rich personal interactions. I want them to take me the least amount of time while still being useful.

Agree. While predictive keyboard is really helpful on mobile (the Windows Phone one was amazing) because it's inconvenient to type on a small on-screen keyboard, and some daily-life interactions are repetitive, on the email side it is the reverse that's apply: most (personal) mails are a priori different in content and a physical keyboard does not justify external helping system. Especially that being a Goole product, you can be sure they will reused everything you type to know more about you.

When the majority of written communications are composed from a few selected branches, the potential storage and transmission savings are huge!

These types of tools worry me, because I put a lot of value in organic spontaneous conversation.

Does anyone remember the old days of Google search maybe 10-15 years ago? You would put in a search query, and often you would find something interesting in a serendipitous manner.

These days the machine tries to figure out what you like, and what makes them the most money / is most popular and keeps feeding you that and nothing else.

You get trapped in a filter bubble, and nothing ever seems to change in your searches. The same sites, the same quality of content (often low), the same authors.

What ever happened to running into the blog of someone who is totally unknown, but put a lot of effort into researching and creating an amazing and informational blog post? Quality well above average, maybe it was their only post. One topic they spent years researching and distilling for the world to see.

This is what you will lose in your conversations, just like it has been lost in search.

Side note:

How is this blog design by Google considered "good"? Between the dropdown from the top when you scroll and the stickied footer, about a 1/3 of the page is easily readable (shoved into about 500px).

I'm not a designer. But this stuff is bad. How is this greenlit at Google?

You're right the related articles stickied at the bottom is super-distracting, it draws the eye, and has to be dealt with before you can continue reading. It's an interruption plain an simple - not as bad as a layered modal - but it still introduces an element of work to the reading experience.

The really frustrating thing is it's for related articles and it's shown first thing on the page. How about giving me a god damn minute to read the article I clicked on before prompting me to read something else. Related articles are fine. But how about putting them at the bottom of the page? I guess this is an interface a team of geniuses creates when they optimize for engagement instead of content.

This is definitely an anti-pattern.

Screenshot for reference: https://imgur.com/a/NMEOBgb

Lol. This is exactly like the plot of the book Avogadro Corp. where a Google-like company creates an email-composing system which then leads directly to the singularity as the program pursues it's own agenda.

Okay, I guess I've got the contrarian take this time around. As a CEO/founder, I spend an astonishing amount of time writing emails that are mostly the same. Things like the following:

Dear [name],


I'm writing to check in on [$$non-automatable follow-up action$$]. [Have you been able to take care of this yet? Let me know if I can be of assistance.]

[best regards / thanks]



For me, things like Gmail's one/two-sentence responses on mobile are _honestly_ a godsend. Things like Smart Compose are similarly incredibly valuable. I'm not trying to be the world's best orator, I'm just trying to bang off the dozens of emails I need to get taken care of each day as quickly as possible.

Seems like you want a mail merge feature, which does exist for emails.

Does this kill EasyMail (YC W18) ??


Interesting that the comments on that link for this technology are much more positive than the comments here.


I don’t understand the sentiment of the comments here (I’m no google fan!). Isn’t that just the natural evolution of word prediction like on a smartphone?

Yes. But let's extend that evolution -- if an AI can pretty much write a solid email, don't we basically become like a safety driver in a self-driving car? Not alluring.

Let the AI handle the form, let the human handle the content. If this leads to better or more efficient communication, all the better.

In the car analogy, as long as people get where they want to be, what is the issue?

But here we're letting AI handle the content.

The issue is agency -- like the God of the margins, our agency being relegated to margins at the very least raises some questions for me!

A Natural Intelligence in the form of a human assistant can write a solid email for an executive that doesn't have time, based on what the executive tells them to write. This is an assistant for everyone, not just the super rich.

It's an interesting analogy, since this tool could potentially keep you safe from making costly errors. And, as an email-composing safety driver, you can always take the wheel when you need (or want) to.

I think one problem is convenience. It's convenient to use this feature, so we find ourselves speaking spontaneously less and less -- and generating less new language.

Aha, by training a machine learning algorithm on everybody else's emails, you can now write what everybody else wrote.

It's a new generation of Clippy! I'VE GOT THIS, THANKS!

"I apologize for such an algorithmically-generated letter letter - I didn't have time to write a non-algorithmically-generated one."

― Mark Twain's Gmail

I'm the PM for this and the author of the article. Happy to answer any questions, with the exception that I can't comment on future roadmap.

Hi, thanks for offering to answer questions. Did you consider the language homogenisation question and if so, how did you answer it?

It's a great question and one of my early concerns as well.

For most people email is a utility. The diversity of language in the middle chunk of the distribution isn't very high today and the biggest complaint most people have about email is how much time it takes. Most of our users aren't writing poetry, they are doing every day business transactions and we can help them be much more productive. Think how many times you've written "hope to hear from you soon", or some equivalent.

For the tails of the distribution Smart Compose is not helpful. We address this primarily with a triggering model, we aim to only show suggestions when we're quite confident that you're in the 'just getting things done' mode.

And if you're the type of person who always has a lot of personality in your emails, the feature probably isn't a great fit for you. Today it's opt-in, and there will always be a setting even if we turn it on by default.

Okay interesting. Did you get stats on how diverse the majority of users' language is?

I think there's also a second aspect. If I understand correctly, you're referring to the homogenisation of a single user's language (where 'Hope to hear from you soon' loses variation). I think the second aspect is homogenisation of language across many people. Did you look at tailoring the suggestions? In England there must be 30 ways to say goodbye, and this slang arguably is part of regional identity (although I don't know how much made it across to email).

I go out of my way to avoid banal phrases in my emails.

Banal pleasantries cost nothing (less than nothing with this new feature), and at worst are ignored and at best make the receiver feel better about you and the important text that follows.

They can have a cost if they dilute the core intent or message of the email: the reader has a higher cognitive overhead before understanding how to reply or what action to take. I spend a lot effort stripping out fluff and filler fron my mails, as a courtesy to the recipient!

No kidding.. their demo gifs look like an April Fools joke to me.

I was thinking this might actually be a useful feature for me - if Google suggests an autocompleted phrase I was about to use, I could think of a better, more interesting alternative.

Smart reply/compose might be the most dystopian thing I've seen come out of Google, which is saying something. Beyond the obvious reduction in conversational authenticity, it threatens to homogenize our voices themselves.

I would imagine that after you've been saying "close enough" for a while, the smart replies start to warp the way you phrase things mentally, instead of the other way around.

This comment already seems quaint after having seen the Duplex demo.

Wasn't this being done by a YC combinator funded company, doing it as a gmail plugin, within the last few years? did they get bought out? or are they now competing against google?

I wrote a tool that helps you do this anywhere by using Google's autocomplete API: https://github.com/nathancahill/Anycomplete

It also supports DuckDuckGo, so for the privacy conscious it should work fine.

Unlike the complaint that Google's product shrinks the lexicon, Anycomplete actually expands your lexicon by allowing you to type words/phrases you have an idea of but don't know how to spell.

I could write emails faster if, when composing an email, the "New Message" box did not cover up the email that I am reading and wanting to respond to. There used to be a function to tear off the "New Message" box, but I guess that was too confusing for people? Now I have to open another browser window if I want to look at an email and compose one at the same time.

I'd love to know if there is a good work around or trick to have the compose window open up in a separate window.

This is just extended auto-complete with the unintended reminder of how predictable/cliche your writing is.

Only if your recipient is a bot too. Having said that could work very well for ‘business emails, where essentially we’re writing in the same tone than personal email’

PS: I’m loving the idea and let’s hope it can get tailored to ones individual needs

This reminds me of waitbutwhy's article on neuralink[0]. Language is an inefficient means of communication. It require niceties and good grammar just to get a simple point across, especially with unfamiliar people. The question is, do we really want to optimize communication?

It feels disingenuous to optimize only your end of the conversation, but it could also be really helpful. Hopefully we find a more openly efficient means of communication soon.

[0]: https://waitbutwhy.com/2017/04/neuralink.html

Maybe this will cause us to think less about we write, or maybe it will allow us to quickly write the things that don't require much thought so we can think more about the things that are important.

So far I have found it helpful for the latter. It has not got in the way when I'm trying to communicate something that requires nuance or careful expression, but it has kicked in with suggestions when it sees I am heading towards some banal pleasantry that is expected at certain points in many emails - the sort of thing I don't want to have to think too much about.

It also seems to pick things that I would say - not sure if this is learned from my own mail, but if so then it is actually reinforcing my individuality in some way.

I wonder where the training data comes from? Is it from my account only, my organization account, or...?

If I type “Sure, the password is” will it auto suggest some passwords from other people’s previously sent emails?

If you use Gmail (or message people on Gmail), your messages have been part of Google's bulk training data since Inbox rolled out with Smart Replies in 2014.

In the new equilibrium of people sending lots of automated emails, how would the receiver of an email distinguish between a canned response and a real response? I suspect the incentive to properly signal your email as real can reduce adoption of "Smart Compose" for non-trivial communication. So that leaves us with use cases like scheduling meetings, quick updates or alerts - most of which are already automated in some extent through notification and logging.

Since many commenters seem to be concerned about the possibility of this homogenizing written communication. What if this system implemented something similar to an epsilon-greedy strategy where say 5% of the time it actually recommended a phrase outside of ordinary writing patterns, but was still grammatically correct and semantically the same?

I think if something like that was implemented this could be a cool way to introduce people to new and different writing patterns.

Why do I want this?

At this point we might as well do away with the pre-structured responses and just send single a single emoji as a response. It'd get the same message across.

I was happy to open the comments and see everyone writing approximately what I was thinking.

Almost made me feel there would be no point writing something similar myself.

I quite like this incremental approach towards AI, of providing a suggesting assistant rather than trying to solve the full problem at once.

Is this some kind of belated April Fool's joke?

The title reminds me how Google effectively crippled fast writing of emails a couple of years back, when they fouled up the Compose textbox with fancy JS/HTML hackery. Can't write & paste an email to the Compose box, can't invoke an "edit-text-in-external-editor" browser function on it. There's smart in simplicity.

Those of you complaining about this have obviously never had to deal with the unintelligible gibberish that can sometimes be spewed out by desperate customer support staff on tight deadlines. If this can even begin to help them write sentences that are readable and understandable then I'm all for it.

This is very neat technologically. I can't imagine it is easy to do predictive text and have it be meaningful. I suspect that that will be the case here. It will be useful for a narrow subset, mostly rote business things. Otherwise, it will just be an annoyance.

I love tech, but it's kind of scary, I think we are heading towards a dystopian future.

> Over the next few weeks, Smart Compose will appear in the new Gmail for consumers, and will be made available for G Suite customers in the workplace in the coming months.

Weird to roll out to free consumers first. What's the reasoning behind this?

Almost everything rolls out to free consumers first. The consumer version is like a regular stable release channel, G Suite is like an enhanced product built on top of the latest LTS.

Thanks for the clarification!

I think Google creating reincarnation of MSFT's clippy - solution looking for problem.

These "intelligent" hints of what someone think I likely want to express - quickly becoming a distracting nuisance.

Hopefully it'll be optional feature.

In their defense, I do like the quick answer buttons sometimes. There's lots of cases where "Okay, thanks!" or "Got it, will do" is all I need and if it offers that, great, one click and done.

The "smart" autofilling of details like address and subject matter are interesting and possibly useful, but I am disgusted by the idea of letting google talk for me.

Could anyone direct me to research material of how to implement a system like this for a single language, or a collection of documents?

Is this not going to be a part of Inbox?

What's the difference between this and auto complete? I think there is a bit of overreacting here.

Yet another reason to avoid Gmail.

An so it begins... To anyone who has read Daemon by Daniel Suarez this sounds very creepy.

"It looks like you're writing an email! Would you like help?"

Now where did they get all that training data...

Feels like auto-complete from any IDE - love it.

I want this for code, not email ...

I'm actually not opposed to this feature in gmail. For code, it already exists, and it's terrible. "Hey, don't you want to catch this error?" Then people catch an error without actually doing anything about the error condition, causing another error to occur later, harder to diagnose.

In code, this sort of thing makes it easy to write programs that are skimmed, not read, with lots of tiny cracks for bugs to hide in.

What's the point of humans

From "Echopraxia"

>"You think this matters"... [Valerie the transhuman]

>"You think so, too," Moore [baseline human] began. "Or-"

>"-you wouldn't have reacted," he and Valerie finished in sync.

>He tried again: "Were they under formal con...," they chorused. He trailed off, an acknowledgement of futility. The [transhuman] even matched his ellipsis without missing a beat.

Clippy 2.0.

I feel like I could type these robotic responses as fast if not faster than it takes Google to "help" me.

it's not for you. it's for grandma

I don't think so, I think this is for professional users. Grandma is not gonna realize that the slightly greyed out text is not already there, and also that tab actually writes out the text for her.

In my experience anything keyboard related is not a good solution for “grandma”. What I’ve seen work best is voice to text. My dad composes all of his messages with voice, as the alternative is first finding his reading glasses, then pecking out keys with only his index fingers at a rate of about half a key a second.

Sure he could abbreviate and autocomplete his way through a sentence... but then I’m not talking to my dad anymore; instead I’m talking to some k’th mean of human interaction.

Watch out. That's a stereotype, and it might be offensive.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact