Andrew Lee

“If you're doing a startup, you want to ride, you know, whatever the trend is that's happening. I think the right time to launch a feature or launch a product is right before it seems possible. So in the case of AI Assistant, I think, you know, no one else had released something like what we do in email. I think a lot of people were like, we're not quite there yet. And that's what you want to get it out.”

“One of the core insights that we had early on when building this was that we couldn't get planning to work for the quality of models that we were working at the time. I think that's probably still true, where if you try to break it down into a series of steps where each step sort of feeds into the next step and each step does some piece of work, that there's going to be errors made by the models at each step that propagate through. So we changed it a little bit and we said, okay, what if the goal here was to end up with one prompt that had all of the information you need in context.”

“There's serially five calls, and the feature extract is actually like five different things in parallel. So every time you ask a question to that assistant, we're doing like 10 LLM calls. And I want to note that before we did that, we embedded all your emails, right? So there was a whole bunch of your processing done beforehand, and we had to pay on your millions, honestly, to set up the data to do that.”

“We did the thing where you take an email, you remove a section, and then you train it on the correct answer being the actual email you set in the first case. Your data set is emails with sections removed and then the correct output is the section completed. We did this in a bunch of cases. This taught it the formatting and generally how emails should work. That combined with the RAG approach that I talked about and some prompting was enough to get the voice right as well.”

“We chose Pinecone primarily for performance considerations. It has a feature that none of the other top-tier vector databases has, which is name-stacing, where we can, without a performance penalty, have a huge number of users on there together. So I think if you're in the process right now of picking your vector database, you should think, how many namespaces do I need? Is it one per user? Is it one per company? Is it one global one?”

“Another way is the economics don't make sense. And if you have confidence in the trends in the economics, you can afford to make that investment, and you can, you know, cover the gap with central capital. And then over time, it'll make sense. And the best example I have of this is YouTube. So YouTube was losing money like crazy because at the time, serving that amount of video infrastructure was really expensive for bandwidth and for storage and for, you know, re-encoding the videos and stuff. And that obviously worked out real well.”

“My experience is the better we are at doing our job of, like, helping you generate the emails, the more they are exactly like the emails that you had before, right? If we're doing a great job, the email that you write should be no different, whether we help you write it or not. We simply help you do it faster and we help you make fewer mistakes.”

“And then Google killed off Google Inbox, which to me was their next-gen email product. And I thought to myself, hey, if Google with its infinite resources, with the largest email user base in the world, that they're not willing to invest to try to figure out what the future of email is. Maybe I need to do something about this. And I have a long history with email. Actually, my dad and I ran an ISP in our basement in the 90s. So I call up a bunch of my Firebase buddies and I said, hey, you guys want to start another company? I'm thinking we should build an email app.”

“I look at it with LLMs and with the automation we would like to build if it was like auto triage and stuff, being more of a knowledge base. It is a corpus of information about everything that is going on at your business and everything you've ever sent and everyone you've ever talked to and all of your SaaS notifications, all of your meeting invites, everything. We can now mine that to do useful things for you. I think it's going to be a reframing from a tool to send and receive messages to a knowledge base that knows all about you that can help you get your job done.”

“I think what's going to happen here is, yeah, to some extent, the AI is going to help you triage them and things like that, but I think also the social network is going to start to bear a lot more, right? So, like, personally, I filter partly based on the content of emails, but a big part of my filter is, like, where I met that person. I think that sort of thing is going to become even more important of, like, who's connected. So, I see, like, higher importance for relationships in the social network and less importance on the actual content of the email because that's much more easy to engineer over time.”