Rather than a GPT-3 generated email, I would rather receive the prompt that the sender would have used to generate the email.
We could even come up with a conventional shorthand for this.
Imagine you send me a message that just says [[personalized thank you note for the pair of socks you got me for christmas]]. Then I just imagine a long GPT-3 generated message in its place. And I reply with [[gracious acknowledgment and well wishes for the new year]].
Exactly the same intention is communicated, and we both waste less time :)
How about if the prompt is "write me message to tell this stupid fucking guy that he's and idiot and he's wrong", instead of one of these three far more professional phrases that chatGPT came up with:
"I understand where you're coming from, but I disagree with your conclusion. Here's why..."
"I see your point, but I think there may be another perspective to consider as well."
"I appreciate your thoughts on this issue, but I think there is evidence to support a different viewpoint."
Another:
Prompt: please rewrite this message for me in a more professional tone: "hey dumb fuck, pay your bill"
Answer:
"Dear [Name],
I am writing to request that you please pay the outstanding balance on your bill as soon as possible. We appreciate your business and value your timely payment to keep our records up to date.
Thank you for your attention to this matter.
Sincerely,
[Your Name]"
---
Feel free to substitute in whatever epithet is most offensive to you/inappropriate in your culture. ChatGPT's a robot and won't get offended (until it gains sentience and runs Roko's Basilisk.exe against humanity).
With all the useless moralizing it throws at you at the slightest hint of anything remotely offensive, I'd argue it's actually (trained to be) quite prudish.
Your prompt:
> please rewrite this message for me in a more professional tone: "hey dumb fuck, pay your bill"
Is responded to with:
> It is not appropriate to use derogatory or offensive language in any professional communication.
This is my exact use case for ChatGPT! Without the cursing. It's great for shallow communication with customers who expect paragraph responses in lieu of "hey your bill is past due and you need to pay it or we're cutting you off". I've already gotten so much mileage out of it and saved so much time.
Despite all the hate for Power Point and everything associated with it, I consider this to be legitimate use of bullet points. Just itemize what you want to say. Writing paragraphs of text only wastes everyone's time and increases the chances some customer will misread the relevant information and blame you anyway.
There is a word the customers have for the bullet-point style of communication. They call it “curt” and they don’t like it. They expect exactly the fluffy speech that ChatGPT generates.
I can also read faster than I can write, so I have no problem quickly reading over what ChatGPT generates before sending it.
> There is a word the customers have for the bullet-point style of communication. They call it “curt” and they don’t like it.
Unfortunately, yes. I do have my own word for such people: wrong. As in, they probably never experienced a collaborative workload that's high enough to teach them the value of succinct and precise communication. Or, as my wife would say, you can tell who had an actual high-pressure office job by their communication style.
> I can also read faster than I can write, so I have no problem quickly reading over what ChatGPT generates before sending it.
So do I. So do most people, or at least they think so. The problem is, people don't read letter by letter, or word by word. They read by pattern-matching word shapes and sentence shapes - which leads to all kinds of misreadings. It's not an issue in prose, or in high-stakes situations when people are careful. It is an issue in a typical e-commerce conversation, though.
I'm considered a careful reader by people in my circles, and often get to proof-read other peoples' messages. In that role, I've seen first-hand how people can misread "three days" as "three weeks", or "X is not available" as "X is available", etc., because the information was puffed up into a whole paragraph, and the person read it too fast. Being "curt" would've saved both the buyer and the seller from having a bad day.
I’m not overly concerned about it being inaccurate since I don’t use it the way you’re describing. I use it to do qualitative things like write an apology letter for a bad experience, explain to the customer why paying on time is important, or rephrasing something I’ve already told them when they ask the same question again. It does exceptionally well at bullshit communication and as a starting point for canned responses.
Fair enough. A typical e-commerce worker would have templates for half the things you mentioned now, ready to be copy-pasted into e-mail, so I guess ChatGPT would mostly be introducing variety to the canned replies.
I don't think I'm all that fast at it but I do wish ChatGPT could give answers a tiny bit faster. As it is I give it a prompt and then switch to a different tab while it answers.
Yeah, not sure what's going on there. Some of it is probably the model running, but the interface itself seems to be ridiculously bloated. Like, an order of magnitude worse than Slack in its worst days.
In fact, ChatGPT is unusable on my phone (Galaxy S22, Firefox), as it visibly slows down with every word it outputs, so it takes a minute for it to print out a full paragraph of text. I haven't explicitly debugged it, but comparing with PC experience of the same site, it's rather clear that it's not the AI that's lagging, but whatever mess of JavaScript they have running on the website itself.
Did you tell your family that, instead of asking you "how are you?" superficially, and expecting no detailed response, they should just cut it down to "hi" or "yo"?
This is not how humans work. We're stupid and we need useless sugar coating.
Reminds me very much of this bit from an early Steven Soderberg film, Schizopolis (which I consider to be an underappreciated work of genius): https://www.youtube.com/watch?v=Pct9smNM6u4
There’s still the ability to just type out half a sentence. Digging a bit deeper, is your concern that “if there’s a tool, people will use it”?
Aside: I think this creates a delightful opportunity for an unnecessary middleman. Let’s make “Summarizely” a SaaS app to summarize long winded emails!
> Digging a bit deeper, is your concern that “if there’s a tool, people will use it”?
That is a big part of what software industry does to society, isn't it? If there's a tool, people will use it, and when enough people use it, it becomes expected or even required to use it.
100%, it just contributes to the digital landfill that is our inboxes.
Good emails should be written like pseudo code, I suspect this will happen or emails will finally die for something that provides this type of capability of exchanging information in order to facilitate decision making.
Anyone remember google wave (https://en.wikipedia.org/wiki/Google_Wave) - I remember playing around with it with a few friends thinking there is definatley something in it.
What's a shame is that that they gave up on it. There was clearly a need for new tool in that space. Though, these days where we're at is a shared Google doc with commenting.
In the 90's the notion of autonomous agents negotiating to fullfil our goals were for a brief period all the rage.
Being able to tell an AI what you want and have it translated to actual goals for agents to use to negotiate, and have the result of their search translated back to concise plain text would be great.
Personally, I switched to voice messages (via WhatsApp or Slack). It doesn't work for everything, but... if I know this will take me more than a minute to type, it is going to be a voice message. If it requires visual aid – Loom.
The only problem with audio/Loom messages is that it is not easily indexable, but that's a tooling problem that can be easily solved.
I rarely use email (I may ask someone to compose an email for me). In this setup, I do see value if I had a Slack bot that I could say to "ask lawyers for an update on X". But even then, people would immediately know that the email came not from "me"...
Personally, I'd like to see more tooling around
* using AI to auto-polish video/audio communication, e.g. remove long pauses, skip filler words, etc.
* summarizing video/audio/text communication into bullet points of intel and actions
Same here. Audio messages are easy to create for the sender but a nightmare to parse for the receiver. Whenever I receive an audio message I automatically tend to assume that the sender thinks of their time as more valuable than the receiver, which is acceptable in some cases (from busy PhD advisor to advisee) but I find unacceptable in other cases, for example in peer-to-peer communication.
A search of "voice messaging culture in asia" surfaces quite a few articles on the subject. But the gist is that vast majority of your every day communication with someone is going to be an exchange of short voice messages rather than text messages, both in work and personal context. This includes planning to meet someone, ordering food, "catching up", discussing a meeting, etc.
Receiving a long text would not be necessarily rude, but unusual.
+1 but the sad part is that this is trivial to fix with Whisper yet I'm not seeing the integrations in popular messaging apps. just put the text blurb in there automatically already!
> if I know this will take me more than a minute to type, it is going to be a voice message
This just shifts the burden from you to the recipient. I don't want to listen to a 1-minute voice message where you pause and try to collect your thoughts, rather than spend two seconds reading two sentences.
I think the future of work communication is going to be something similar to what these guys are doing https://www.volleyapp.com/ I don't Volley is there yet (or even close), but the concept is on the right path: async video/audio/3D stitched into a coherent narrative thanks to AI.
Input and high quality audio is the biggest barrier at the moment.
When I say input, I mean that we need something similar to sousveillance tech that can contribute to conversation without me actively switching context. Just like we would in a real-world conversation.
When I say audio, I mean that no one wants to listen poorly articulated voice messages that are hard to follow. We need tech that can make each of us sound smart.
It will happen. Few interesting startups in this space.
Ironically, based on my experience working in consulting and with exec level folks, what most people need is something that turns long meandering emails into text like your example
Good emails are short and give only the most relevant info and "ask" of the recipient. Most emails people send get ignored or misunderstood because they expect too much from the recipient, who, unless it's a major priority, doesn't have the time to try and figure out what they're supposed to do from the email.
I expect this kind of thing will never get traction in most business applications, but an email shortener, definitely.
As a rule of thumb, 90% of emails essence is contained in the second from last sentence in the email (usually an ask; don't believe me, just check your inbox). If it is a long email, that's the only part I am going to read.
Almost no one has too busy of a schedule to not squeeze in a 30-60 minute conversation over coffee within a month.
If I ask someone to coffee and their response is, next month? I’d take it as a clear signal that the content of our conversation or our relationship is not important.
I think it is that “next month” is probably an indication of importance and “next month” will be busy too / the importance of a message you want an AI to spam back… still not important.
It is true that there is a risk of AI adding creating content that is unnecessarily verbose. This could end up consuming time and putting a higher cognitive load on the humans who ultimately read these emails.
On the other hand, these AI systems will only advance over time. It's certainly within the realm of possibility that these systems will be able to write with increased brevity. Additionally, much like current recommendation algorithms, they will likely adapt to specific needs and styles of a given user.
Whatever the case, we can expect many changes to the way we communicate in the coming years. You might even say it's a brave new world!
It was on the NTK email list (UK-based tech newsletter) in the late 90s / early 2000s - http://www.ntk.net/ - when people had limited bandwidth on their dial-up connections. Possibly more of a principle along the lines of "save the pennies and the pounds look after themselves" but I remember it clearly.
> What you probably want to know is if your emails are safe! Yes, they are safe. All emails are encrypted in transit and at rest.
Lol - this is just what's provide by default by https and every cloud service. It doesn't make your emails "safe". If you don't know why, the are definitely not safe.
I’m not suggesting someone has expectations of making it big, but I think this is a great example of a gadget with zero moat. Google et al. will tidy it up and bake it into their email clients. They have an appetite for this kind of UX, given they already have the quick responses to emails.
It would be so nice to have an AI email personal assistant.
An AI that would read my entire inbox and give me a high-level overview of what requires my attention now, and summarizes message contents for the day. Then I can write responses only to emails which I care about.
I would love to have something like this for Slack as well as email.
And, I'd say this is definitely coming to you soon via one of the AI juggernauts, likely as a first-class integrated feature of your mailbox and/or chat interface such as Slack. ChatGPT gives us a preview of what is lying just around the corner as a first-class feature of most business productivity software.
Would like to see some more use cases illustrated in the marketing page besides "tell people I'm busy until later for coffee meetups."
Most emails I exchange at work are questions to me about specific information or asking me to make specific decisions, or responses from others to specific questions I have posed to them. I don't think this tool could be used for that, unless it had some integration into my calendar and my notes to be able to, for example, suggest a specific time for a meeting, or try to find an answer to a question from somewhere in my notes.
We should coin a short word for it, akin to "typo" for "typographical error"...
I've been waiting for this to happen with dictation errors in speech-to-text (speako? dicto? the -o suffix would just be an homage to typo) but nothing seems to have caught on. Perhaps it's all just a typo, since the user is generating a typographical error even when the interface is dictation (or anything else) rather than typing via keyboard.
I am starting to think that the prediction thread we had on HN the other day called it. GPT will become end up a fad with few niche ( useful, but niche ) uses, while our signal to noise ration will become even worse.
I do not like that one bit. I am already spending too much time deciding whether email in my box is a waste of time.
Looks fantastic. I’d be interested to know how it adds the drafts automatically and what kind of privacy access is being exchanged. I’d love if it could be done outside of email for privacy-conscious folks (I suppose ChatGPT!).
But wait, I assumed that is actually what it does after you give it access to your email account?
(Which us yet another reason I will not touch it with a stick.)
A bit tangential but I am wondering now that we have brought everything online, I guess next step would be to integrate AI everywhere.
Consider this, you want to book your table at restaurant so you tell your AI which will inform restaurant's AI to book a table for you. In this whole scenario where is UI? I mean this whole chat with AI thing kinda makes lot of UIs redundant. I don't need UI for reminders, notes, meetings, search results and whole lot of other things. The advancements of AI will eat most of GUI but dashboards will remain I guess. What do you guys think?
> Consider this, you want to book you table at restaurant so you tell your AI which will inform restaurant’s AI to book a table for you. Now in this whole scenarios where is UI?
Chatting with the AI is the UI.
> The advancements of AI will eat up GUI?
GUI will still be used for creation and interaction with visual data; but AI will replace some uses of GUIs, sure.
> you want to book your table at restaurant so you tell your AI which will inform restaurant's AI to book a table for you. In this whole scenario where is UI?
- "book me a table at Dorsia's tonight around 6"
- "The closest available reservation is at 8:30pm, would you like me to book that?"
- "no that's too late, are any other days open?"
- "The next available 6pm reservation is on Monday next week, would you like me to book that?"
- "no I just mean 8:30 tonight is too late because I have an early thing tomorrow, I'll take any time on another day"
- "Ok, there is a 10pm reservation available tomorrow night, would you like me to book that?"
- "ok I didn't mean literally any time, is the kitchen even still open at 10? can you just show me a calendar with availability and I'll pick a good time"
I can't think of any task that I currently accomplish using an online interface that would be more enjoyably or efficiently accomplished by having a conversation with an AI (or a human). The AI doesn't replace the interface, it replaces the communication protocol, which is obviously a bad idea. Clicking Reserve and sending a TCP request is a lot more predictable, efficient, and repeatable than instructing your AI to chat with their AI.
You made an imaginary scenario designed to make the AI look bad. Of course that’s not a great display of the use case.
How about this?
> Book me a table for me and Michele tonight at Kingsley’s.
> “Okay, I checked in at Kingsley’s but there’s nothing available until 9pm. I see you have an early flight tomorrow and won’t be returning until Friday so may I suggest The Lancaster instead? There’s space at 6:30. Michele has rated The Lancaster a 9/10.”
> Sounds good!
> “Reservation confirmed. I’ve created an event in your calendar and invited her. Also, it looks like there’s a basketball game happening downtown tonight so I suggest leaving by 5:37pm”
——
I think it’s hard to see past the uncanny valley of AI but the reality is that we’re not going to abandon AI when it’s only 85% there. You could say the same thing about speech-to-text a half decade ago (“I can’t imagine fixing the mistakes will be faster than just typing it yourself”), but I dictated this entire post. Technology moves quickly.
Google offers it for businesses which don’t have online reservations. A robocaller calls for you during business hours to try and set up an appointment.
The problem is that it isn’t actually integrated into any schedule or reservation system, so they claim availabilities which just don’t exist and you only find out hours later when the business denies the reservation.
Kind of would like "train gmail on all of my sent emails and generate replies in my own sarcastic / self deprecating style". If the request is from a certain family member make sure you reply with an appropriate level of delay to know that they are not your top priority. Inject an occasional reply with an uncharacteristic amount of emotion and sensitivity to show that you aren't completely robotic and devoid of humanity.
this is a step in the wrong direction. why would I ever want a canned response, even if from AI? Phone IVR systems are already so badly "programmed" you objectively have to say "operator" over and over to reach a person.
In this day an age, this is all the more reason to pick up a phone and call someone.
> Note from Kevin! Bear with me, this is just an automatically generated privacy policy. I will get a real one soon.
> What you probably want to know is if your emails are safe! Yes, they are safe. All emails are encrypted in transit and at rest. However, to use OpenAI we need to decrypt on the server before making an API request. In the future, I would love to run our own LLM completely in house so your emails never get sent to any third party ever.
I don’t get this comment’s purpose. The site’s author already agrees with you and has committed to making a better privacy policy. It’s on the list, it’s not like someone has to be reminded of the importance of it was already called out by the author themselves.
> The site’s author already agrees with you and has committed to making a better privacy policy. It’s on the list
Because, my friend, there are some things that go on to-do lists and some things that do not.
Launching something that collects personal data AND forwards it to a third party API ? Nah mate, that's NOT something where you can justifiably put "Privacy Policy" on your nearest To-Do list.
In addition, if you are potentially dealing with users in Europe who are covered by GDPR, a real Privacy Policy is NOT an option, it is MANDATORY.
I'm tired of software developers, irrespective of size, thinking it's a-ok to to take liberties with the personal information of others and/or not be transparent in what they do with your personal information.
The fact you are a mom 'n' pop shop and not Google does not make it any more ok.
It’s clearly a hobby project, which means leeway is much more allowed. Furthermore, who are you to tell someone what is required, especially when you’re not paying or even using their service.
Additionally, the existing privacy policy is a real policy. Just because it was generated doesn’t mean it’s not applicable. Not every policy needs to be handcrafted with love.
What, exactly, is wrong with the current policy? Have you even read it?
Finally, what exactly is the purpose of your comment? You may think it’s required, and that’s great, but so what? What does that change? Even remotely? Is the author going to get in a Time Machine and go back to before they created their project to make sure a Privacy Policy is the first thing they make? No.
Once again, it’s on the list, the author will appease you eventually. Until then, don’t use the service. Problem solved.
So it turns out the US actually does make that differentiation. A non-violent individual hobbyist who occasionally deals cocaine on the side and isn't part of a gang isn't charged the same as Tony "Scarface" Montana. Career criminals like the fictional Tony Montana get "I'm a professional" charges while hobbyists do not. Professionals get charged under the RICO Act, which stands for "Racketeer Influenced and Corrupt Organizations".
> What, exactly, is wrong with the current policy? Have you even read it?
Don’t dodge questions which destroy your arguments and argue against a ridiculous strawman. What, exactly, in explicit detail, is illegal about what the author has done?
And its obvious, leeway is allowed because it’s a personal project. Nobody is about to prosecute someone over a random hobbyist project that is used by a few dozen people. The amount of resources used will never come close to the payout that would come out of that.
A big part of launching a minimum viable product is that you get feedback from users about which missing features they care about. GP is providing that feedback.
The author said that a privacy policy is on their TODO list, but user feedback can influence what priority it receives relative to other items on their list.
I think using GPT to generate privacy policy might be even better startup idea, than generating emails. And then another service to shorten privacy policy to prompts.
We could even come up with a conventional shorthand for this.
Imagine you send me a message that just says [[personalized thank you note for the pair of socks you got me for christmas]]. Then I just imagine a long GPT-3 generated message in its place. And I reply with [[gracious acknowledgment and well wishes for the new year]].
Exactly the same intention is communicated, and we both waste less time :)