Hacker News new | ask | show | jobs
by ankit219 296 days ago
I like the idea. I also think you may need a mechanism to detect adverse actions before they are executed. This becomes important because an email cannot be unsent, and if I dont review the text, the 1 in 100 chance of the email sounding weird would freak me out. There is also the basic corner cases against AI prompt injection and all those spam and phishing emails that are rampant and more and more plausible sounding. Wonder if you have ideas around how to deal with those?
1 comments

Absolutely. We will be adding a safe mode where April can perform only read, summarize emails & calendar schedule, create drafts but cannot send or delete emails.
Maybe reading out the draft to confirm before sending could be a nice middle ground? Could help with peace of mind about exactly what's being sent, and you could go back and forth on the message if it's not quite right. Once people use it for awhile and trust it more, they could then enable auto-send.

Congrats on the launch btw—cool stuff!

That happens now - before Send/Delete it will ask for confirmation.
Oh nice. Does it read you the text of the draft? Didn’t think I saw that in the demo, but maybe I missed it.
Yeah, It does. We mentioned only few of the things in the demo or else it would have been very long. Apart from hard delete, create /edit labels - all operations can be done through April.

Create/edit/delete draft are very common use cases as people would like to read before sending.