I don't know if it's just me, but I very much prefer easy to skim emails to more dense ones.<p>At one point I got weekly project updates from a team I was working with, and one guy wrote dense, short, emails where I would have to read every sentence carefully to get a hang on what was going on. Another guy would write longer, fluffier emails but with bullet points and paragraphs in the same order:<p>> Hi everyone!<p>> {fluff}<p>> {general status}<p>><p>> - Bullet points of things done that week<p>> {comment about things done}<p>><p>> - Bullet points of things to do next week<p>> {comment about potential problems etc}<p>><p>> {fluff}<p>> {annoyingly long footer}<p>Just by parsing the number of bulletpoints, and the length of each bullet point (and the first word) I would get a surprisingly good grasp on how things were going, and what was hard/complicated (longer bullet point -> more complex), and very easy to read about exactly I wanted to know.
I like the idea of this product but<p>(1) sometimes a BIT of a personal touch goes a long way<p>(2) even in the examples you give it misses potentially important context. Eg In the board message example, it removes a reminder of the topics. How could you know those weren't necessary? Why is the number of members what it included?<p>Just doesn't seem ready for prime-time for this use.<p>My preference would be having it suggest shorter versions of each paragraph inline or something.
Have you tested this in marketing emails? that seems to be the profitable marketplace. Even if you can prove 5% increase in roi that's worth a lot to a lot of people!<p>There are some interesting studies showing short emails perform better but we haven't found that to be true it seems the message is the biggest factor. Some way to improve the message is even more valuable but also likely a much harder ML task
I think there is still a way to go here.<p>For example, the "really" in "really appreciate" isn't useful.<p>Also the last sentence in the "summarized" version reads: "If for any reason you haven't let, let me know..." I would also that that the "for any reason" is not useful, it is again implied.<p>You don't need "by now", it is implied.
I'm reminded of the recent post on HN showing that long, personalized emails reduced email conversions, and that short, pinpointed emails increased conversion. Long emails give the impression that one is either a program or desperate, and no one takes the time to read the entirety of every email they receive.
Based on some digging around, it looks like the app is made with Facebook AI's BART model [1]. The only summarization implementation I spotted was made by huggingface (of course). [2]<p>[1] <a href="https://github.com/pytorch/fairseq/tree/master/examples/bart" rel="nofollow">https://github.com/pytorch/fairseq/tree/master/examples/bart</a><p>[2] <a href="https://huggingface.co/facebook/bart-large-cnn" rel="nofollow">https://huggingface.co/facebook/bart-large-cnn</a>
I see the biggest opportunity here in summarizing books and blog posts. At least for me.<p>Too often, after reading a blog post, I feel like it was a waste of time. I'd prefer to read a super dense summary first. Then optionally select longer variants before reading the full thing.<p>Maybe There could be a compress function [1, 2, 4, 8, 16, 32, 64, ...]. Choose the compression level before reading.<p>Especially with books. I'm not sure how well this could work, but I'd prefer to read 100 books that way first, before deciding on the few I actually want to read.<p>I'd totally pay for this.
The issue with efficient summarization is the necessity of context. For a message to be effectively summarized you require a bit of "theory of mind"[0], you need to have at least a decent idea of what the receiver already knows. This is something that, especially when done for a single message at a time (with no other info), likely doesn't have a global solution.<p>Edit: I should be clearer, I believe that the solution shown here does not have this capability, and the effort of using ML to do message summarization is flawed in the general case.<p>[0] <a href="https://en.wikipedia.org/wiki/Theory_of_mind" rel="nofollow">https://en.wikipedia.org/wiki/Theory_of_mind</a>
I'm excited to see where this goes -- both this product and the general idea of using ML to reduce noise in interpersonal communications.<p>A human can read the shorter summary and add personal touches as they wish, but the first pass of what the model considers to be extraneous is interesting by itself.<p>I just put in a 300 word personal email and it spit out a 60 word summary that actually did extract the one concrete piece of news I was sharing, and just cut everything else. That's not precisely what I wanted in this case, but it wasn't a bad guess at all.<p>Side note: just finished a rewatch of Silicon Valley, I like that the sample text is from someone at Hooli, I wonder if it will change my .... to ... :)
Neat idea, I could definitely be more terse in my comms.<p>As an FYI, some bad English on the Hooli Onboarding Email example:<p><pre><code> If for any reason you haven't let, let me know and I'll make sure you get them.</code></pre>
Short emails tend to be either notifications, highly transactional (web based platforms are better for these as they can do tracking and trigger other outside actions), vague, or assume outside context.<p>These are the same problems as always-on corporate instant messaging platforms, just slightly improved by the cross-organizational functionality and store-and-forward.<p>IMHO organizations should strive to make email less like instant messaging and more like letters. Fully formed thoughts distributed less frequently and with more purpose.<p>In short, tax hassling others.
Seriously. I already write short mails and put bullet points which makeout much smaller. This year one of things I want to do is write big emails like paragraphs atleast two.
I've never written an email that was too long and then later realized it should have been way more concise. I generally know my intentions prior to writing the email, so if it needs to be short, I'll write it short. The reverse situation probably happens more often though (i.e I write a short email, realize it doesn't make sense due to lack of details, then I elaborate).
Super cool. Would add this as a company gmail feature.<p>However, I also write emails to reduce time spent in conversation, so while I have learned to be more terse, when I'm not, it's to provide references so that I can shorten or preclude a conversation.<p>An ML email shortener is a fantastic idea for business comms, with the caveat that short email culture can also reward vagueness.
Coincidentally I've been feeling like my emails are becoming way too verbose recently. I'm not sure what it is, but I'm just taking longer to reply in a way that conveys my message properly, or I'm finding myself editing a draft to death before I ever send it. I used to be able to just reply and move it to the next bucket.
Can I use this in conversations with my wife? So much talking. ;)<p>Joking aside - this is a great tool in corporate settings. In technology, one works with two types: Very verbose (tend to be introvert or detail oriented) or Too abbreviated. I cannot stand reading detailed accounts - and this could add to many professionals comprehending more.
If this could convert passive voice to active voice, or underline it as an error, this could be a great technical writing tool for inexperienced writers. Good technical writing is hard, a tool similar to this could be useful.