The Future of Email: Superhuman CTO on Your Inbox As the Real AI Agent (Not ChatGPT) — Loïc Houssier

We as a human species, like, we started to write because we didn't have, like, enough storage for stories that we were telling to each other, so we had to write to store those stories. Now, like all the content can be stored in YouTube, in TikTok or whatever, it's like, what's even the need to write?

What's the need? Because everything can be vocal. And I see kids now, they don't read article. They want a TikTok video talking about the article. Um, being a bit more grounded, what does that mean about, like, the future of the user experience for email and communication?

Will people still type or will they just talk to emails and they want to hear an email?

Mm-hmm.

And this is where it becomes interesting because Rahul, as a CEO, maybe next year he doesn't want to write to you, uh, with the new feature. Maybe he wants to talk to you. And then the way you will receive our marketing campaign about the new features, you in your car commuting, listening to Rahul talking about, uh, about that.

Hey everyone, welcome to the Late in Space Podcast. This is Alessio, founder of Kernel Labs, and I'm joined by Swyx, editor of Late in Space.

I just realized I have the tough job of always pronouncing names and-

I know, man. You gotta prep. Even on YouTube, namepronunciation.com.

Loïc Houssier, welcome.

Wow. I'm impressed.

It's, I- I... Did I get it right?

I know. You got it right.

Okay. Okay.

You got it right. I'm, I'm surprised. Usually I make the joke about like, "Yeah, you know what? You can say it, like, the way you want and everything," but, like, you nailed it, so I'm impressed. Thanks for having me, guys.

Yeah, of course.

Thanks for coming by. So you're CTO of Superhuman Mail, which is the new name for Superhuman. I've been using Superhuman for a long time. I think I was one of Rahul's personal onboarding things back in the day. And yeah, I think we're here to talk about all things AI engineering, but also you have a lot of history in products board, Firstbase, DocuSign, and n- nuclear submarines.

Yes, yes. That's kind of like the, the fun, like, icebreaker that I give to people sometimes. Like, like, two truth and a lie. Like, uh-

Yeah

... I went into a submarine, and the people are like, "Yeah, no way." But I did. I did. I spent one year on working around-

Yeah

... submarines and helped to carry across-

The, the trajectory is a bit weird. You were an engineer, and then you were sort of chief of staff on, on some submarines thing.

Yeah. So we-

And then you went back to engineering.

I, I started, like, uh, studying math. Uh, so I'm a math graduate. I was about, I was about to do, like, a PhD in math and applied math in cryptography. Uh, so, like, crypto before crypto-

Mm-hmm

... to some extent. It was cool for a moment, and then I was like, "No way I spend, like, three years of my life on the same topic." But in the same lab there was, like, a bunch of people doing, like, security, like offensive security type of stuff, and I was like, "That's what I wanna do."

So I was basically an engineer. I was a security researcher in that lab. But I did that in a pretty big corp. I saw one in Telco and then in the s- defense industry. And in the defense industry they have this nice kinda like career framework.

Like, you're young, high potential-ish, between quotes, so they want you to do, like, different type of jobs and kind of like have a spiral of career so that you at some point reach to the C level eventually. So they gave me the opportunity to be out of the tech industry for a year, and I went in the harbor, and I was there as, um, I mean, financial controller, process improvement type of person-

Yes. That's right

... and basically helping people do a better job, which was interesting because I had no clue. Torpedo system, radar systems, like, even, like, nuclear engine inside a submarine. Uh, but still, I had to help people take a step back from what they were doing and everything.

And that was really fun because I came from Paris, came with my tie and my suit and my ego. I was used to drive people through my technical legitimacy on the security space, and all of a sudden I didn't have any technical legitimacy at all, but I still had my ego.

So, like, it was a pretty fast ramp up and like whoop, put my ego in my pocket, and basically drive by questioning people, like, "How did that work? Like, help me. Like, I don't, I don't get it." And just by questioning, I kind of, like, build a new skill, which is, like, getting curious and understanding how people are working and being comfortable facing people that are way smarter than me, knowing better their field, but probably having a way to ask questions to help them, like, uh, identifying gaps or, like, uh, um, productivity gaps-

Mm-hmm

... uh, for example. So that was cool, but I missed the tech, so I moved back to the tech industry after, uh, basically two years.

Yeah. What are some of the other maybe highlights or stories you have been told about other experiences? I mean, DocuSign is another product that we all use. Um, yeah, any other untold-

No, DocuSign was cool. I mean, DocuSign was cool because it was an acquisition.

Opentrust and DocuSign, yeah.

Yeah. So I was a CTO of a small company in Paris, and, um, and we were like a typical, uh, I would say European company, Alessio. So, like, very focused on the tech, not very focused on the marketing, and we are trying to...

Like, we were one of the biggest signature company, uh, in Europe, but it's a very fragmented market. So we, we are winning France, starting to, uh, expand, and DocuSign is coming and like, "Hey, guys, we need to do a partnership and everything."

And pretty soon they understand that European market is tough and, like, the technology behind DocuSign is not sufficient. Lack of standards, lack of, uh, compliance and everything. So pretty soon they were, like, with us or against us.

Mm.

But the way they were explaining the value, I was like, "Holy cow." Like, we're not talking the same language. We're doing the same job. We're selling the same type of software. But, like, we are talking to CIOs from a technical standpoint.

They're talking to head of HR, head of, um, functions, and sell them the value. So pretty fast it was easy for us to understand, like, "Whoa, whoa, whoa. Uh, not the way to sell a product. Better to, uh, partner with them."

So they did an acquisition, but it's not a full acquisition. Uh, it was a security-oriented company. Two business line, one doing signature, which is, uh, I would say the one that, uh, and DocuSign was interested in. The other piece was doing strong authentication, so PKI stuff, SSL certificates, those type of, uh, of things.

And we were working for the Department of Defense in France. So we had the Ministry of Finance in France basically saying No way. No go.

Mm-hmm.

You cannot sell. So we had to do a carve-out, uh, which is, like, of the funniest acquisition type you can do. So you have your team. You need to divide everything in two: your team, your systems, your code source, and all of that.

Even your data center, you have to replicate and get rid of, like, all the shared systems and everything. So we did that for, like, something like six months to be able to sell the new carve-out company to DocuSign.

Crazy. Don't do that.

Are you, are you still involved at all with, like, the French, uh, startup ecosystem? I'm curious, like, how you've seen things evolve since then.

Yeah, it's pretty interesting. Like, I've seen a, I've seen a change. Uh, now that I'm getting some gray hair, and I have some experience, like, uh, I try to give back to some extent, so I spend more time, uh, helping, like, the ecosystem there.

But it's funny to see, like, the difference. Like, when you're here, you-- we live in a small bubble, and it's crazy to see how even, like, other tech scenes are different. So, like, the grit, like, just, like, the grit, like, to get shit done and, like, to, to, to move forward and everything.

They have great education. When I say they, sorry, like we, they, I don't know where am I now. But, um, so great education, great engineers, and all of that, but not the mindset of, like, creating things. So not-

Mm-hmm

... not a lot of entrepreneur that much. It's changing. We had, like, uh, some successes in Europe, uh, and especially in AI, like, there are some, some cool stuff happening. But still, like, the way to think about, uh, product-led growth, like, Superhuman nailed it.

But, uh, uh, ways to think about, like, the way to structure your organization to scale fast, uh, the level of ambition as well, how to, like, maybe not target France or target Italy to start with, target English and the world from the get-go, and, uh, that would be, uh, something to, to think about.

So I'm, I'm doing that, uh, quite some. Uh, highly rewarding, uh, but it's, um, yeah, it's, it's pretty cool.

There's a common question that people have about DocuSign that I'm just gonna indulge.

Sure.

What do all the people do at DocuSign?

I love it. I-

You, you know this is a meme, right? It's-

No, no, no. It's, it's-

I'm sure inside of DocuSign.

Why do you need, why do you need, like, so many people?

Why... When you have signing-

Yeah

... why do you need three thousand engineers?

It sounds crazy, but, like, you want to nail Europe. You need a different product. You need a different team to run your local data centers because of the compliance. You cannot just run your data centers from the US, so you need a local team there.

Oh, and by the way, the way to do digital signature in Europe, totally different. So, like, the stack itself is different.

Oh.

So, like, the way to make a digital signature is different, not the same standards and the same ways. So you need dedicated team to maintain that thing. The same way some people want to have DocuSign on-prem, so you need a team building appliance to basically plug and play and like, "Okay, you have your DocuSign appliance doing-"

There's a DocuSign box?

There's a DocuSign box.

Wow.

Acquisition made in Tel Aviv, uh, at the time. Uh, wonderful people building, like, security appliance and where, like, you, you shake the box, poof, the keys disappear. Like, if someone is, like, stealing your box, no one can sign in your name.

Oh, you're kidding me. Oh my God.

No, I mean, some banks-

What if there's an earthquake?

And, uh-

Now they make more money

... that's a good question. They are mounted, like, on some, like, a-- So there's, like, earthquake mitigation, I would say, associated to this. So just that, but, like, apply to FedRAMP.

Yeah.

Dedicated teams-

Wow

... dedicated data centers. And like, oh, and we need, um, I would say, um, to have, like, uh, DocuSign run in Canada because data residency. Oh, we need the same in, uh, Australia. Okay, cool. And now you, you have, like, something even different.

We want Japan as a market. Oh, but Japan is not signature. It's hanko. It's kind of like a stamp. So you need a team to understand how Japanese market is thinking about even processing an agreement. Totally different. And then you have, like, verticalization, like, some different verticals and everything.

I mean, it's a good business. It's well-run, and, like, people are not coasting there. So there's a lot of work, and it's very interesting to see it from the inside because when you see those memes, you're like-

Right

... "Yeah, I know," but like, damn, I see the people. I see what they do.

I mean, you, you were the VP Eng, so you know, you know.

Stop.

You actually know. Yes.

Um, yeah, yeah.

I just wanted to get that. Obviously, this-

So I hope it's providing some-

This, this episode is not about DocuSign, but we have to ask

Oh, yeah, yeah.

No, no, of course, of course. Totally legit.

Yeah, let's talk about Superhuman. So you joined January 2025.

Yes.

Uh, just give people a lay of the land of, like, Superhuman AI. I think a lot of people that are listening are familiar with the email client.

Yeah.

I think the AI stuff is generally new.

Yes.

So just maybe give the canonical definition of what you wanna do with AI in Superhuman, and then we'll kind of dig through.

The main driver is, uh, how you can put AI in the product to accelerate the productivity of people. It's not to just, like, do AI things and, like, the sparkles and everything. We don't care that much about it.

Uh, our people are pretty high expectation oriented, and they don't want to slow down, so you cannot add latency. You cannot... So everything that we do is done in a way to, like, improve the productivity of people, so AI included.

First thing that we started to do was, like, auto-label, um, emails.

Mm-hmm.

Like, is it a pitch? Is it marketing? Is it... Kind of like typical classification that you could do. And so people can say, "Okay, everything that is a pitch, I will look at that, like, on a Friday." So, like, during my days, like, typical days, I don't look at it.

So, like, that, that was one of the first, uh, thing. Summaries, like, you have a long thread. What is this thread about that someone, like, shared with me? Okay, you have, like, a quick summary. So nothing, nothing that is-- that was very, like, br- groundbreaking, but, like, just well-thought, just, like, adding things that make sense at the right time.

Another example is now, like, we automatically detect if one of your, um, email requires an answer, and if no answer after two days, popping up, "Hey, this one needs to, to be like a," I would say, "You, you need to, to send another email to the, to the person because you, you didn't have an answer."

So that was the first step. Second step was like, "Hey, you know what? The draft is already ready. You can just hit Send." So it's very subtle. But it's like adding a, "Oh, damn, shoot, yes, I wanted to remind people to, to give me an answer," and the, the thread, the, the, the draft is already there.

Pretty cool. Sent. And now we have, like, more and more of that. Now it's detecting, "Oh, this is a request for you to ask your a- for your availability." Or you have an executive admin that is doing that for you.

Your draft is like, "Hey, let me CC the right person," and boom, so that it's ready, it's sent, it's done. And the typical chatbot, because more and more of the use case we see and people using AI inside Superhuman is to query your, your emails.

A good example as, uh, I would say tech people, we receive, like, a bunch of Substack, like a bunch of newsletter. Um, we say some are great, sometimes like, eh, the content is meh. I probably have like, I don't know, 30, 40 subscription, uh, because everyone has, like, something interesting to say at some point and everything.

Now I don't read them. I auto-archive those, and like every week on the Friday, I just, like, Ask AI, which is the name of the feature.

Mm-hmm.

I ask my email, "Hey, tell me about, like, the, the summary of all the Substack that I received this week. What should I pay attention to?" And then I can deep dive in, I would say the place where I-

Wow

... I want to, to pay attention to. So this is always thought in a, in a way to accelerate, I would say, the pace and try to not be in your way, hopefully. Feel free to ping me if-

Yeah

... that's not the case.

I, I would say I, I don't know if this is a recent change, but I feel like Ask AI, I've started using it a lot more. I've been a Superhuman user for many years, and you've had it a while, but somehow this year it kicked up a notch, and I don't know if it's because anything changed in the product, because I wasn't using it before, or is it just me trying it again?

Um, I... Now that's a good question.

Yeah.

That's a good question. I think people are more and more used to the muscle of querying things, uh, because ChatGPT-

Yeah, yeah

... because-

So the, the general consumer behavior is-

So... Yes, exactly. So the user experience-

Yeah

... people... I mean, now, like, every single product has a chatbot where you can ask questions, so it's becoming, like, more and more natural to ask questions compared to managing, like, a to-do list of emails.

And agentic search as well. Like, previously I was like, "Oh, you have to embed my documents, and then it's just gonna retrieve," and, like, I-- that's not what I want. But agentic search where you can actually figure out what do I mean when my question, when I ask, it's, like, half-formed.

You expand it, and then you actually answer it. Uh, it's, it's actually really good.

Yeah, and we spend a lot of time on the quality of the answers. Uh, so quality of the answers, and you've talked about the agentic framework, but, uh, one thing that is-

And this is a frame... It's not LangChain, right? It's, like, your own framework.

Uh, yes. I mean, we, we've done a lot of iteration-

Yeah

... and, like, there's a lot of subtleties and, like, multiple PCs are there, but, uh, and multiple different models be-

Mm-hmm

... based on where they're, like, really good at. Um, but where we spent quite some time lately is, like, around quality, uh, and making sure, uh, across different dimensions, but, like, making sure that we are generally good for typical queries and very optimizing for them.

And especially one thing we, we try to, to solve for is, uh, agent laziness. So through this chatbot, you can... But one of my use cases is I receive a Slack, and like, "Hey, Loic, can you review this, uh, document, please?"

Because whatever, it's a tech, uh, I would say tech strategy document, or I need to review the doc. I take the link, I go to Ask AI, and I basically pass and say, "Hey, find me 15 minute tomorrow.

I need to review this doc." And I don't need typically the agent to say, "Hey, I found this lot and this lot and this lot. Which one do you prefer?" I just asked for 15 minutes. "Find it, do it."

I have an admin when I was asking her, like, on Slack, "Find me 15 minutes," she's not asking me if, uh, I need, like, on the morning, on the afternoon. She's, uh, doing it. So working on this agent laziness because the handoff they were doing to the user is losing time.

So, like, working on, like, making things happen faster, uh, we sp- we spend a lot of time, uh, on this. So that's why you might have felt like, uh, that the, the overall quality-

Yeah

... is better.

Yeah. My, my, my old joke was because the way that you trigger it is you, you actually type it in the search bar, and when I was trying to normally do search, it would sometimes accidentally trigger the Ask AI, and I was like, my, my joke is, like, most of my AI usage is just accidental because I actually wanted to just search.

But then, then I started just using it more.

Yeah.

And, and then you-- the, the kind of questions that you ask changes.

Yeah. I use it to, like, find people's phone numbers-

Mm-hmm

... stuff like that. It's like, "Hey, what's-"

I use it to find my contracts, because I have so many contracts, right-

Yeah

... from all my sponsors and, like, venue things. Like, yeah.

Yeah.

Yeah. One of the use case that I, I would say that blew my mind, I, I was looking for like a... I, I was at a conference. They shared with me, like, a, a PowerPoint link, and it was, like, six months ago, and I couldn't find the deck because I wanted to reuse some of the content and everything.

Couldn't find it for whatever reason. Ask AI, "I'm pretty sure they shared with me, like, a PowerPoint link or something like this. Can you find it?" So fetching the context, there in the link. I couldn't... Yeah. I saved, like, probably 30 minutes, like, uh, searching, searching through my emails, so it's pretty cool.

It's to you, so it's, it's, it's-

Yes

... because there's no way you can fit all your email into a context window.

Yeah.

Right?

No.

Any- anything else that's more complicated that-

So, um, we had to do some pagination, uh, because, uh, if you do... Like, let's say I'm doing that. Like, oh, I'm pretty sure I had a conference I attended where they shared an, like a link with me.

In my case, I don't do, like, plenty of conference, but still someone like Rahul, my CEO, is basically doing a conference every three weeks or something. Um, not kidding. But de- but in the use case-

That, that is his job.

That is his job.

And he's fantastic at it.

Yeah, and it... Damn, I'm learning so much from him. Uh, but clearly this... I would say depending on the use case, I mean, of course, you have more than 40, 30, like even hundreds of emails that can semantically be close to your answer, so you need to go through that.

So we had to implement a pagination search. So like semantic search for like the first, I would say 40, and deep search. Not that one. Okay, next 40, next 40. So I'm kind of like using this agentic loop, and while you don't have find the answer, continue and even extending like the semantic search proximity until you find the right one, because it might be buried page two of the, of the, the, the search page, uh, technically.

Yeah. How did you design the tools to get to the agent? Just maybe give people overview of, like, the framework, what it looks like. Like, how are you structuring these interactions? Is there just one Superhuman agent that does everything or, like-

No

... do you have separate ones?

We have separated tools, uh, clearly. So, uh, even an agent, like I would call about t- I would say tools. So there's a bunch of tools. A tools to detect your availability, a tools to understand who are the people you interact with, a tool to write an email.

The tool to-- like, so every single action is very tool specific, so it's not a magic big tool that can do pretty much everything. It's a set of small tools that are used, uh, within the agentic framework. So like there's a first step that is like, "Hey, what is the best tool to do this?"

Kind of like building a plan-

Yeah

... like for each step, what is the tool, and then making the calls.

Yeah. I think now the tools versus skills that Anthropic talked about is like the hardest thing of how much you wanna put, and there's like the MCPs discussion. I'm curious how you evaluate the tools too. Like when you build them, it's like, how do you think about how to name them?

Like how to give the description. It's like, how much work have you had to do to nail it?

I don't think we spend that much time into... And again, like I will defer to my three engineers working on it, which is interesting. We can talk about like the, the amount of people you need to work on those stacks, uh, when you want to be serious.

And, uh, and I have fantastic people, so I feel blessed. Uh, and most of the time was trying like the different agentic framework, trying to understand the different models, the ones that are solving which type of problems, because every single model is good for something.

Sonnet was really great for like, uh, agent hando-handoff. Like the laziness was really great. OpenAI version of it was not that good. Uh, now we have Gemini coming in and like in the room like, uh, last week, like poof, okay, that one is, uh, is cool as well.

So I think we are-- I guess everyone has built like a way to, for one, switch easily, uh, from one model to the other.

Routers, like model routers.

Everyone has like an LLM proxy to some, to some extent, and like an agent proxy to, to, to implement different stuff. Uh, which is becoming interesting because the way to tweak them and tune them is different. So, uh, it's still easy to switch from one agentic framework to the other.

Uh, but at some point, I think it will be harder and harder, and like the stickiness of them will be tricky. But to answer your question, like, uh, we didn't, didn't spend that much time on the tools themselves, I believe.

How do you think about evals? Like are you s- evaling like one email draft at a time? Are you evaling a longer workflow? It's like, just run us through like, yeah, when you're testing Gemini, like how do you decide what it's good at, what it's not good at?

What's like the eval structure?

At first, we had a relatively naive approach, query, answer, query, answer, and having like a set of queries. We, over time, evolved into like, uh, thinking more about like the different dimensions that we want to target. Agent handoff is a very typical type of, uh, problem space, and that you want to, uh, make sure you select the right model for.

So typically getting a bunch of queries targeting hard handoff that we've i-identified by through dogfooding or whatever, but trying to target a set of what we call canonical, I would say, queries, uh, along that dimension of, uh, I would say that specific problem space of agent handoff.

But like there's more. Like there's the deep search, like shit ton of emails, and you want to find that needle in the haystack. That's a different type of, uh, category. So you need to have canonical queries that are like targeting that type of dimension.

Because every single user will have their own way to q- to question their own data set, and we cannot replicate every single data set, uh, of, uh, of people. The good thing is, uh, we have a bunch of users like Rahul or like myself, we receive like a shit ton of emails.

Pardon my French, by the way. I don't know if it's, uh, okay for the show.

No, you're good.

But he receives probably like five hundred to a thousand emails a day.

He's still part of the onboarding. It's like, I will send an email to Rahul, and he will reply. I'm sure it's not actually him.

Uh, sometimes it's him.

Yeah.

He's reading like pretty much everything. He's- I, I don't know how he's doing it. Uh, but he is really, really paying attention, uh, especially at the tone and why something is like going sideways and everything. He really associate the brand and tone of like the people talking the company with himself, which is kind of like bringing us to the, to the next level as well.

Yeah.

So thinking about all those dimensions, uh, is really key. And so like even if you have like an eval tool, like the way you structure the-- your different queries to target those dimensions is important. And then we have those specific queries, like the route queries, uh, typically.

The one we, we joke about, and the one that was one of the first, uh, that we used as a, as a way to calibrate our, uh, quality was, um, we have stories, but he did some, like five years ago, some refurbishing in his house, and he had this table, specific type of wood, and he was discussing this, uh, with the contractor.

And, um, he wanted to have Ask AI find that email and the type of wood that was discussed in the thread with that guy five years ago. And until we nailed that query, he was not satisfied with the deep search approach.

And this is where we're like, "Oh, damn." Okay. So that's a different set. Uh, but we're also talking about dates. Like another, I would say dimension, is dates. What is last quarter compared to today and everything? Large language, uh, models are not really good with dates.

So like how do you manage that? So these specific queries for that. So we're like, "Oh, okay, so there's dimensions that we need to care of." So now we structure the old evals, and as you were, um, asking end to end, what is the query?

Whatever happens there, there's like an answer. Was there like a good agent handoff? Ne- I would say dates, uh, were there-- were they nailed or not? And, uh, et cetera, et cetera, et cetera. So it's pretty intensive in terms of brainpower put in the quality.

Again, because, uh, Superhuman is a high perceived quality type of product, so we had to in-invest that amount of time there.

Yeah, high real quality. It's not just perceived.

No, but I think this is, this is important because, um, what is quality?

I don't know. Yeah.

The, the, the feeling.

Like if I buy a car like that is a Toyota, it's good quality.

Yeah.

And I get the quality for my bucks. If I buy an Audi or Porsche, I expect a different grade. So maybe it's grade. Uh, like the grade is different, but, uh... And it's high grade, but high expectations, so a high amount of time-

Yeah

... I spend on quality.

Yeah. In PM-ing, there's this concept of the high expectations user

And, and Rahul was one ex- one example of those. And I was just wondering, like, who are the most outlier, extreme people? How are they using-

Yeah

... AI in their email? You know, and just, just in general, like the most extreme examples that you've come across, obviously, 'cause that's how you work.

Oh, that's a good question.

Be- you-- for example, you had, "How much time do I spend in Waymo last month," right? Which is-

Yeah

... basically turns your email into an, like an accounting system 'cause it's kind of a source of truth.

Yeah.

I don't know if I would do that in Superhuman. Is it reliable?

It is reliable.

Wow.

And when you think about, like, the amount of work, and we're working right now with Anthropic to basically do kind of like a building on the fly, small kind of a keynote of Lambdas that will build the code to do the aggregation.

This is an easy example.

This is like a code execution thing.

It's... Yes, it's a code execution piece. But, like, this one is relatively simple because you just, just have to have, like, the agent extract from the email. So select the emails from Waymo, from the Waymo receipt, like extract the time, like the duration of the trip, and then do the aggregation.

But that's not easy. Like, data aggregation is not easy-

Yeah

... and LLMs are not good at math. Uh-

Yeah

... so, like, that-- there was some support about it, and right now we're discussing about, like, extending this approach to more. It's interesting to-

Is it-- are you operating on the email file itself or is there a fundament-- is it like a row in a database and you're just writing a SQL query?

No, the aggregation is-- so we don't extract that data on the-- So when we ingest the data-

Just email, you know, like a GPA

... so we ingest the data.

Yeah, okay.

Uh, we ingest the data. So we rely on, uh, Gmail and Outlook, uh, of course, because, uh, they do-- they are doing, like, some great stuff that we don't wanna do, spam detection, like-

And Superhuman will never do it.

And probably.

Probably never do it.

Probably. Like-

Which is being a IMAP server or-

Exactly.

Yeah.

Like, do I wanna do that? Probably not. Um, probably not. Uh, maybe-

You know, HEY, HEY, HEYmail did it.

Yeah. Um, they have an-- But, like, is it something where we want to spend time? Is it valuable for our end users, really? Not sure. They live in an ecosystem. They will live in a different company.

Yeah, you-- Outlook.

Yeah.

Yeah, yeah.

So, like, they have Outlook, and they have Gmail. It's already there. So, like, if we can just plug and make that better, I mean, it's, um, it's, it's, it's good there.

Uh, I, I mean, in, in some case, I-- Superhuman was the original wrapper company. If you pe- if you think about- ... GPT wrappers, this is the Gmail wrapper, the Gmail wrapper. Fir- at first it was LinkedIn wrapper, and now Gmail wrapper.

Yeah.

I think more for it than Gmail itself, so.

It's very true. It's, it's very true. That said, you can question, like, what is like an SMTP server for real? Like, it's...

It's a server that conforms to a spec-

Yeah

... with some database. May- maybe not even.

Maybe not even.

Yeah.

Uh, maybe not even. I mean, uh, they're doing like Waymo stuff. Like, they have like crazy, like, especially Gmail, like the search capability is of course like, um-

Yeah

... I would say crazy good and all of that, but-

To do what you do, you need a server-side clone of my Gmail, and then you need also a local cache.

We didn't-- we need local cache. We work offline. That was one of the thing that we did as, uh, initially beside the UX, BI, and beside the, beside the, the speed. We have everything local. One of the reason is, uh, we want to be fast and under, like, every interaction should be under 100 milliseconds.

Yeah.

I mean, with network, you cannot. You just can't.

Yeah.

So everything needs to be local. So, uh, yes. So we have like a copy of emails local on device, uh, and works in the enterprise world because-

SQLite

... interestingly for mobile, it's, uh, used to be Realm. Um-

Yeah, RealmDB. Yeah, yeah, yeah.

Yeah.

Is it a Facebook, uh, tech?

Mongo.

Mongo.

Uh, it has been acquired by Mongo.

Yeah, yeah.

Uh, but, uh, and now it's like somewhat sunsetted, so we need to find a different way to do things now.

Oh.

Uh, might be, uh, SQLite. Uh, but yeah. So o- on device, uh, stored. But that was like the old search where we had basically like a database with rows of the emails. Uh, but everything that is AI, like we have, uh, all division embeddings and all of that, uh, so we have a hybrid search, and we use-- I don't know if we can name brands, but we use a TurboPuffer on the back end, uh, to store like, uh-

Yeah

... five years of history

... I think TurboPuffer is relatively public with their customer list, so I don't know.

No, yeah, and I, I think they're-

We'll, we'll let their PR department take it.

They talked, they talked about it, uh, anyway, but, uh, but it's a, I mean-

Yeah

... stable infrastructure. They do things pretty well. It's fast.

Yeah.

So.

I'll briefly comment that, uh, I know any number of local first database companies that would love to work with you. If you're saying that you're on the market for a Realm replacement, they will come and talk to you.

I mean, I'm more than happy. I'm more than happy. That's, uh, my AI, my, my, my mobile team, like, they're really looking for like something different.

They will love nothing more than to be Superhuman's database. Okay. I wanna just, like, focus on the AI side, right?

Sure.

So people want to know where is their inference running, what are you sending over, uh, what can, uh, what can the provider see, that kind of stuff.

It depends.

Yeah.

It depends. Depends on the use case, uh, depend on the type of model you, we wanna use. Uh, so there's some stuff we run on inference company, uh, with open models. We-- There's some stuff that, uh, uh, we run with, uh, OpenAI, with Anthropic.

So it's, it's pretty diverse.

Mm-hmm.

It changed because, um, also p- based on the quality of the models. Um, we're a GCP shop.

So lots of credits for Gemini?

Yes. So we have an incentive to for probably like, uh, spend some dollars there.

Yeah, I mean, it's nice that they're also a leading model anyway. So like you're not actually compromising-

And, and they are doing some pretty good, uh, good stuff there

... anything. Yeah.

Uh, but, uh, we use Baseten to run some, uh-

Yeah

... I would say some Llama, some, uh, BERT model for classify- classification. There's, uh-- We're doing probably some discovery discussion with like some YC companies about like a model on device, uh, as well, uh, because-

Yes, they work offline.

Yes. And interestingly, those companies, they started to do like, uh, on device mostly for cost reduction. Uh, that was their pitch, "We'll reduce your cost." I mean, uh, we don't care that much. Our people, our users, they want quality, and they are okay to pay for that quality.

But we want to solve for offline. Like we-- if you're offline, semantic search doesn't work as well. Uh, so, uh, so we are discussing with the companies.

What are your design constraints for offline inference? For example, right, like DeepSeek V 3.1 would be like 600 billion parameters. I don't think you want to take out six hundred gigs.

It's-- So we-- And, and people are somewhat complaining, uh, about like, uh, our footprint-

Yeah

... on, on the device.

It's probably like two, two gigs already.

Um, both in more memory-

Yeah

... and both like on device because we store local emails. Uh, like, uh, we store-- Like when you install Superhuman, we download the last thirty days of emails so that y- we can do search when you're offline at least for the last thirty, uh, thirty days.

But we keep that history, so it's starting at thirty days. And if you're like a customer for like two years, technically we optimize for two years of email in your device. So that's interesting.

On the local model, any thoughts on like every app is gonna have its own model versus you're gonna have a device model that people run, like?

Oof. I mean, uh, it's a lot of space.

What would you prefer? I'm curious. Like-

Um-

... would you rather have the user just take care of the inference and rely on that, or do you wanna own the whole experience?

I mean, Superhuman will want to own the full experience. Like, we're pretty picky in the way things are, I would say, happening. Um, so but at the same time, like if we talk about mobile, you want the mobile experience feel like your device.

So we are basically not doing React Native. We are doing Swift, we are doing Kotlin because we want the app to feel like the, uh, the, the user experience in generally in the, in on iOS or on Android.

So but for the models, um, that's a good question. Uh, I would love the device provider to be better.

Right. Hmm.

I mean, we can question like, uh, local devices-

Right

... uh, local, local, uh, like iOS has done some work there, but it was underwhelming so far. Uh, they're still working at it, and that's why we have like YC companies that are spending time there and doing some, some cool stuff.

Yeah. Amazing. Interesting question on Baseten. They're a very different cloud inference provider for open models compared to, let's say, the Fireworks and the Together AIs. The general pitch is that they don't charge by token. They charge by box, effectively.

Anything else that's interesting working with them versus the other inference providers that you buy?

Um, they're easy to work with.

Yeah.

I mean, that's when you're a startup, you want to move fast. They're really easy to work with. Uh, they know what they need.

So the prio-priority is like what? Cost? Speed?

For us-

Yeah

... it's quality.

But-

So it's quality and speed

... they all-- It's all open models. It's all the same quality

It's like we would always start with the highest and more expensive model-

Yeah

... to get the right quality, and when the quality is nailed, then we can spend time trying to optimize.

Right. But all these providers, Baseten, Fireworks-

Yeah

... Together, all these, they all have the same access to the same models.

Yeah.

So unless they quantize heavily, which all of them say they don't.

So in that case, like, um, the, the fact that it's a box, you control your cost way better.

Yeah, yeah.

So unless-

So it's like fixed capacity.

It's fixed capacity. So you know, when you are so-- when I discuss with my CFO, like-

Yeah

... uh, when it's token-based, it's like the exercise is way more trying to understand- ... like, uh, how we set the adoption and all of that. So the heuristic-

But that's serverless. That's serverless.

Sure.

Usage-based, serverless, it scales up, scales down.

Fair, but like the, the, the cost control is becoming like a thing.

Yeah.

Uh, it was a thing before the acquisition. Uh, now that we are part of a bigger umbrella, like understanding your cost structure and like being able to make projection that are closer to the reality is more important. Like all pre-IPO-ish companies, you want to really, uh, understand where you will be like in three months, six months from a cost standpoint.

So Baseten for that is pretty cool because you have more latitude to stay within the, the bracket of like, um, a box-

Yeah

... basically.

I was thinking about this. I-- You know, a lot of people think about cost in terms of dollars per million tokens-

Sure

... right? And I think that that is actually amateur thinking. That's only, only the kind of pricing you care about if you're a solo developer. But once you're in, in so large scale like you guys, uh, and so also something I learned about at Cognition, you should actually cost care about price per trillion tokens, 'cause we spend multiple trillions per month, and when you unlock that scale, you unlock different ways to spend.

That's not a serverless, uh, token-based pricing. And so basically I think Baseten makes a lot of sense on a price per trillion.

Yeah. I didn't look at it that way. It's, uh- ... it's pretty interesting. But no, no, that's fair. And I mean, we built like so many different models, uh, trying to understand like the cost per million of tokens, and then you have to infer like what is the average number of tokens because we treat every single email.

There's really short emails, very long emails. It's like you have to understand your data, like, uh, what is the median-

Yeah

... and all of that to, to make your projection. And it's always-- there's always some magic. The reality is like you don't have the time to-- I mean, I'm an advocate or like let's move fast, and if the-- if it's successful, it's great, even if it's expensive.

So rather than trying to optimize the cost too early, like just go with something that you control and fast, and you'll have time. I mean, it's a good problem to have.

Yeah.

Success is a good problem.

When do you think it's gonna break from like a cost perspective? Say you were to like draft every single email that I get, I'm sure you will lose money on the forty bucks a month.

Yes and no. I think that it's a matter of like how more productive we make you. Like, we have some customers that told us like in initially when you're talking about like the different models and everything, like take the better model.

Like, I'm ready to pay like two hundred bucks a month, but like get the m- the best model.

Yeah.

Like I don't want half crap because-

Right

... it's less expensive. So like always give me the best. And you know what?

'Cause these are all like high value CEOs and VCs.

Right. Yeah, yeah, yeah.

I mean, one hour of their time-

Yeah

... is worth like-

Way more

... ten times the, the, the amount of the subscription. So it's an easy-

Why, why isn't there a two hundred dollar a month subscription?

That's a good question.

Yeah.

I'm not in charge of the pricing and packaging.

Uh, wait, wait. Okay. Well, ma-maybe one sec. One-- An example would be like, well, what's one thing that you would like to do that you cannot do with today's models, even though you tried pushing quality? That peop-- your customers are telling you, "Actually, we really want this," or maybe Rahul's telling you that he really wants this.

I don't know.

Yeah.

I don't know. I think s- I think we have the means. We have the means to do, like, pretty much everything that we wanna do. Like, it's a matter of executing-

Yeah

... and doing it right.

The way I would put it is, like, if you can articulate what you cannot do today that you sh- think you should be able to do, and your customers would, would pay you for it, the model less will make it happen.

But the problem with-- that you have, and the problem that I have with Cog, is we cannot articulate what it is. We will know if it's better, but only once it exists.

No, that, that's, that's a good framing. And the, the other piece that I think it's, um, pretty tricky is that there's a transformation that is happening in the user experience. Like, even the way we are thinking about the, you know, user interface right now, it's totally switching.

Like, the way we think about emails right now, it's still like some sort of like a to-do list. It's a table to some extent with rows. What would it be like in a year? Because people will be more and more interacting with their systems through a conversational aspect.

Hmm.

Like I see my kids. My kids, they don't type on their phone, they talk.

Ah.

I mean, like, I-- all my kids. I have three kids, all, they talk on their phones.

What age?

Working, college, and, uh-

Okay

... middle school.

Okay. On WhatsApp?

WhatsApp, because they're European and they need to talk with the family. Uh, the reality is, like, Snapchat, uh, it's, uh, like TikTok, like whatever, like Instagram. Like, they communicate over Instagram. Like, I'm like, "That's not an image tool," like or something.

I feel like a boomer.

Yeah, I am. Uh, I'm definitely am. But, uh, what is interesting is that they-- and, and we can debate about like-- but like we as a human species, like we started to write because we didn't have like enough storage for stories that we were telling to each other, so we had to write to store those stories.

Now, like all the content can be stored in YouTube, in TikTok or whatever. It's like what's even the need to write? What's the need? Because everything can be vocal. And I see kids now, everything is vocal. They don't read article.

They want a TikTok video talking about the article. So coming back, and I'm sorry, like I'm getting like very high here, but, um, being a bit more grounded, what does that mean about like the future of the user experience for email and communication?

Will people still type or will they just talk to emails and they want to hear an email? And this is where it becomes interesting because Rahul as a CEO maybe next year he doesn't want to write to you, uh, with a new feature.

Maybe he wants to talk to you. And then the way you will receive our marketing campaign about the new features will be, well, he said, discussed to you or talk to you with his voice. Uh, not just voice and tone in terms of like, uh, uh, writing, but like, uh, really like, uh, you in your car commuting, listening to Rahul talking about, uh, about that.

So coming back to the what cannot be done right now, I think like the, the main problem is like nailing the un- new user experience. I mean, OpenAI now you can do stuff with emails. They're trying to do some stuff there.

Um, like all those chatbot, they try to be like this basically the new OS to some extent. So how do you interact with those new apps? So what is an app even in this new world? So that's what is like really, uh, interesting, and that's why I'm glad to work with Rahul because the guy is so freaking visionary.

And, uh, if there's one company to nail it, there's not a lot, and I believe like, uh, Superhuman is one of them.

Yeah. I think the inbox is like the ultimate private data source. I feel like even when I see all these companies that are like, you know, talk to like your AI clone to get advice or like, you know, things like that, I feel like so many times, man, I'm just writing the same thing over and over.

Like, you know, how many founders email me asking about help for XYZ task?

Yep.

And like the answer is almost always the same, you know? And like there should be a way almost for Superhuman to like be the advisor on my behalf in a way. It's like you should be able to predict what I will respond-

Yeah

... to this email.

It's called auto draft for respond. Uh, we're, we're still, we're testing it internally, uh, because like there's... If... Especially, sorry to, to cut you up, but like, um, same, same for me. Like how many companies are reaching out to me to pitch whatever-

Right

... like AI frameworks or like AI tooling or like whatever. And my answer is like, although I don't answer because I receive like hundreds of them, uh, honestly like, "Thank you, don't have the time," and everything that's so it's cool.

But like because I want to be polite, like right now, like, like it's automatically generated for me because they learn that I'm usually don't care.

Yeah.

Uh, and that's my answer. Or if it's someone that is pitching me for, uh, like, "Hey, I want to work with you guys," and everything, like, uh, like someone that is applying, my answer is usually, "Oh, please, uh, reach out to HR.

I'm CCing HR and everything."

Yeah.

So now we are now able to understand how you reply typically, but it's always like if it's covering only eighty percent of your use cases and you need to discard twenty percent, where is like the cost-benefit value? Is it annoying to have like twenty percent where you like, uh, discard, I want to write it, uh, myself?

Is it good? Like what is the limit? Ninety ten, eighty twenty?

I think it's like AI plus the snippets that you have. I think that's kind of like... Like I have snippets for a bunch of things like vendors. I have this like super long snippet. "Thank you so much for reaching out about your company."

Yeah.

"Sounds like a great product. We're not currently in the mar..." Blah, blah, blah, blah, blah.

Yeah.

It goes on, and then the response is like, "Thank you so much for your thoughtful response." And I'm like, "Great, get it out of the way." But I feel like if you could use that-

Yes

... plus AI to do the small-

Yes

... kinda like last smile-

Yeah

... thing, I think that would be enough.

Yeah.

You don't really need a GI. I'm excited for it.

Q-Q1, Q1, Q2, something like this.

I mean, dude, I pay two hundred bucks a month to OpenAI, to Anthropic. Like I'll, I'll give you two hundred bucks a month- ... if you like make me not write the same thing over and over. Deal. I think more generally what he's trying to get at and what Superhuman is starting from a very good basis, but not there yet, is kind of like AI EA.

I don't know if this comes up a lot. Where I have people I work with who do read my emails and respond for me.

Yeah.

And they have memory, and they, they know my normal preferences. They have human judgment, which, um, LLMs don't have. Is that something that you would want to build, or do you think what you wanna leave to others?

That, that's, that's the goal. When we kick off really like the revamp of our AI- In world and what AI means for like Superhuman. Uh, Rahul did a pretty good like pitch on it, and there was like a pretty nice video, I think it was in March, uh, for the launch of like the new AI.

That's the vision. Like the vision is like you have an EA, and most of the people who are using Superhuman, C-suite, founders, and all of that. So pretty fast they need someone to help them with their emails, and we want to do like most of that job.

So we're getting there. We're getting there, but that's, that's the goal. That's the goal. Like the, the first thing, like, um, answering your availability. Right now we can do it. I mean, right now it's in beta.

Mm-hmm.

Uh, but right now my emails, like internally when someone is asking, "Hello, can we meet next week for lunch?" Automatically, I, I will have like three slots proposed in a draft, and I can just like send the draft that is, uh, prepared for me.

Yeah.

It's still up to you to decide whether or not you want to send the draft.

That's the thing. I, I don't want to be involved.

And this is where your EA will always be better than an LLM-

Yeah

... because she knows the type of people you are okay to have lunch with, or maybe they have the context because, um-

Yeah, sometimes you're busy, but you're like, "Oh, VIP, I will move this."

Exactly.

You know what I mean? Like an...

And your-

And your calendar is not gonna know.

I mean, we're getting closer because w- we know how much time you interacted with that person, but like how much time you interacted, it doesn't mean that maybe last week you had like a bad discussion with them, and now, now you're not friends anymore for whatever reason.

But your EA would know. So there will be like a always limitation, uh, to this, but, um... And that's why we want two people to always be in the loop. And maybe it's your EA that is in the loop.

Yeah. It's so helpful when it's, when I'm not in the loop.

Yeah.

We can batch it and like I have my once a day call with, with the EA. But yeah, ob- obviously that will happen. You know what? A lot-- Some, some ways that other people are acquire- uh, pursuing this, like Notion's trying to go after it, right?

Yeah.

They have Notion mail, Notion calendar, then obviously they really care about AI. Some other people are doing this interesting thing where they buy an EA company, like a, a company that already does virtual assistance, and then just monitor what they do and then just...

First of all, Superhuman can provide me an EA that is a human, and then it slowly replace parts of it with AI.

Hmm.

I'm curious what you think about that. That's a more aggressive approach if you really wanna-

I mean, that's probably the best way- ... to understand how an EA is working and like the type of work that they're doing and everything and-

Yeah. Make your own data.

Yeah. I mean, um, I mean, that's, that's intense. Uh, that's intense, but like sure.

You have the money.

And you, you pretty fast understand what are the type of workflows you want to automate first. So like having that data would be like, uh-

Yeah

... obviously pretty, uh, pretty interesting.

One of his portfolio companies, they bought a sort of-

Law firm

... legal form. Yeah. Yeah. Uh, do you think, do you think that's an accurate description or am I glorifying it too much?

Uh, no, it's an accurate description. It's like it just behaves as a law firm though.

Right. J- just treat it as a law firm, and then internally start to optimize.

I mean, you have now so many customers that- ... there might be... You might need a lot of EAs too to do it for, for everybody. But I'm curious, I think like the, yeah, the memory is kinda like the killer feature of the EA.

It's like-

Yes

... understanding in real time. I'm curious like now that you're like within Superhuman, the company now, Superhuman Mail.

Yep.

Do you feel like there's like a lot of advantages of being email plus documents plus being embedded in everything? Like, do you feel like that helps closing some of these gaps?

Yeah. Um, so for example, like Coda is an interesting, uh, I would say piece of software. So Coda is like a Notion equivalent. Um-

Yeah, we used it at, uh, Amazon.

Yep. Uh, it's a pretty good one, and a lot of like enterprise companies start to like use Coda more and more because of the flexibility and everything. And Coda has this concept of like Coda packs, uh, which is integrations, uh, glorified integrations, if I would, uh, I, I can say this in this way, but they're ingesting the data.

So like the data is there. So like every time you have Coda. So we have technically an ingestion pipeline that, uh, can aggregate all the knowledge about you in the company, uh, which is great. And now if you add Grammarly, Grammarly is ubiquitous.

Our, I would say the users of Grammarly, Grammarly knows that you're in Google Doc. Grammarly knows that you're, I would say, crafting a, like a, a post on LinkedIn. Grammarly know- knows... Technically they can know. Doesn't mean that they, they use the data-

Right. Yeah, of course

... but like they're everywhere. Uh, so like when you have this, I'm everywhere, oh, you're getting into your email, but I know that you are currently like on Jira with that context. So all of a sudden I can, poof, pop up like some of the context.

I know that you're writing to that person. Oh, it's about this. I can expand and like augment your email because I know where you are coming from. So the data will be there through Coda. Grammarly knows basically where you, I say, you're switching from Google Doc to Salesforce to LinkedIn, and now you're writing an email.

So we have this augmented context even more so like much more precise compared to something like, um, ChatGPT, for example. They don't know where you are-

Yeah

... because you're switching windows. You're coming from to, uh, I would say to GPT from Salesforce to ChatGPT. They don't know where you were. They wait for you to pass the content to get the context. If you're Grammarly, I know where you're coming from.

So like when everything will be converged, and we've been acquired only like three months ago.

Yeah.

But when everything will be converged from a contextualization standpoint and knowledge standpoint, we know way more. So we'll be like way more accurate in the way to help you.

Maybe predicting a fourth acquisition, but wouldn't it make sense to have your own browser?

That's a good question. I think there's much more to be done on the productivity space before like, uh, I would say solving a browser and everyone is trying to do a browser.

Yeah. Atlassian and Perplexity, OpenAI.

I'm still sad that Arc, uh, is not like getting into development anymore because of Dia, but Dia has been stopped.

They're, they're rebuilding Arc in Dia.

Yeah. But like it's, it feels like, it feels very unstable now. So like more and more people are basically saying like, "Okay, let's go back to Firefox."

What?

I mean, more and more people are doing that because like there's so many browser. Like, you're like, you want to wait for the war to be done and to have like the clear winner.

No, no, no, no, no, no, no, no.

You don't feel?

I disagree. I disagree. You should-

I'm in Alice

... you should go all in. What, what are you using?

I use Alice. Yeah.

Alice? Yeah. I'm, I'm also Alice now.

Oh.

Yeah, yeah.

Interesting. I'm still on Arc because I'm-

It doesn't have profiles still.

Yeah.

That's the biggest issue.

So, so I, I- Based on the different emails I have, or logins I have, I switch between Atlas and Chrome and-

Yeah

... Arc.

Interesting.

Yeah.

Yeah, my personal one, it's on Chrome.

But I'm just saying, like, well, okay, if that context matters to you, right, you- you've Coda and all those things, then, then Grammarly, all those, you might as well have your browser. It's the season of... No one, no one will-

Yeah

... get upset at you for saying, "Oh, we have a browser." Like, it'll be like, "Yeah, makes sense."

Or it will be like, "Oh, no. One more?"

But, but it's the Superhuman one, and that's a good brand.

That's interesting. I, I foresee, like, uh, like, browser to disappear completely. Like, I, um-

Hmm.

Like-

Ooh, okay, that's the title.

I mean- ... uh, my main, like, central, I would say, piece of software that I use in my productivity tool is Raycast.

Ah.

Yeah.

I'm a... I mean, I'm a Mac user, uh-

Raycast

... so I use Raycast. For the people that don't know Raycast, uh, it's basically like a, a way better spotlight, uh, on Mac. And I don't need bookmarks in my browser anymore. What i- what is doing a browser beside providing you a view on a website?

Nothing. So it's just like... So even, like, to some extent, Raycast should be like just a web view because what I do with, uh, Raycast is like-

Then you're turning Raycast into a browser.

Ah. Is that a browser if it's just rendering HTML?

Yeah.

Okay, so like-

Right

... if-

E- everything is browser.

So yeah, if it's o- only, like, a rendering HTML-

What else you want? You want JavaScript? You want-

I don't know

... local storage? You want what?

Yeah. Like, local storage is one.

Extensions.

Like, y- you need a browser, like, uh, to, to have, like, your local extension.

Mm.

But, uh, to, to have, like, your local storage that is, like, pretty massive, like Superhuman. But, I mean, what's left? If you... Like, everything that was making a, a browser a browser before, which was, like, bookmarks, like, uh, basically the, the, the last, um, history that you had, maybe, like, uh, cookies and every- Like, what's...

If you get rid of that, it's just a view, a web view to some extent.

Yeah. I- it's a clean application platform, uh, where- with that open app store, you know, that there's a... There's a Marc Andreessen line of, well, the operating system is just a poorly debugged set of devi- of, uh, device drivers for the browser.

He said the browser is the actual application interface.

Uh, it's interesting.

From the person that made the browser.

Yeah, yeah. Of course.

That makes sense.

Yeah. I, I think the browser will be, like, more and more thin. I believe they will be, like, thinner and thinner, but, um, they will, they will disappear, or they would be, like, just-

Yeah

... embedded in the OS eventually.

Yeah.

So.

One more technical sort of thing-

Sure

... and then we can go to sort of organizational things. You mentioned understanding the person. You know, one part of memory is just, like, the knowledge graph, and one part of the knowledge graph that really matters is the entities that I deal with, right?

Like, I deal with him for, for four years, and, and we have that context. And basically, what exists today in Superhuman, and maybe what is possible in the future, right?

Hmm.

Do you, do you-

That's-

For example, do you use a graph database or something like that?

Not yet. And it's interesting because you are mentioning, like, what's missing right now. I think that these knowledge graph, like, you know, oriented, uh, database, I'm not there yet to some extent.

But have you actually tried, or are you just saying that-

No, we, we didn't try. Not right now.

Yeah, that's the thing. Like, it's, it's not fair to say they're not there yet if you haven't tried.

Correct.

Yeah.

Correct. Uh, but, uh, even from a taxonomy standpoint, uh, when you think about those entities-

Yeah

... what are those?

Yeah.

Uh, if you're a verticalized size-

People, companies.

Yes. Um, but, like, then you start in... You'll start talking like, um, about projects. But is the project, is it a task? Is it an initiative? Is it a hierarchical aspect to those? Um, how deep is-

Yeah, yeah

... the tree?

These are all valid questions.

I think it's very-

And, and, like, you know, Superhuman's history is reportive, where, like, the person is, like, the core of the-

Correct

... the universe.

No, no, but there's some obvious entities.

Yeah.

But, like, if you think, uh... If you want things to be really personalized, these entities are, like, very, very subjective. Like, I'm a user of Obsidian.

Yeah.

So I'm a note-taking nerd, and for the people that use... know Obsidian, it's, uh-

Another local-first app? Yeah

... it's another, like, local-first app in which you build your own workflows and where you will basically, through templates, define your own entities that make sense for you. And there's no two, like, graph, uh, that is similar, even if you're using the, the note app, uh, say, for the same thing.

So trying to, uh, infer, like, a, a generic knowledge graph that can be u- reused with, like, dedicated entities, people, task, a project, and everything, it's harder than it seems.

Oh, yeah.

Interestingly, like, uh, we were thinking about it when I was at Productboard. At Productboard we have, like, the roadmaps of, like, so many tools. Based on that, you can probably infer some taxonomy about what is a SaaS product.

But even trying to generalize this into, like, a, an, a tree that can be repeatable for people, it's hard. There's some common stuff, auth- authentication, authorization, billing, uh, user management, dashboards, whatever. Every SaaS company has this. But then when you come...

you, you enter, like, the domain of the, the company, totally different because their features, their, like, um, uh, surface area is, is very different. So, like, even there, trying to form the knowledge that you have, abstract the entities that will be the same for everyone-

Yeah

... is not easy. So it means that then for each user, you need to have a, an unoptimized graph that is, like, subjective and dependent on the people. So you need to build the graph based on the, like, just the data, and you don't, you don't have, like, a real way to, to optimize for it.

But you're fair. Like, you're right. We didn't try. Uh, but also because-

Many people have failed. It's fine

... and, and I don't even foresee a path where that can be surfaced into, like, more productivity gain. At the end of the day, what is the problem you're trying to solve? It's super nice from a technology standpoint and, like, a, even, like, a thinking process standpoint, like, what is, like, the ultimate data model for productivity nerd and all that?

But what are you improving from a, an experience standpoint? Is it like the accuracy of your draft that I'm framing you?

I, I want my AI EA to remember everything I've talked, everything I've done, everything I talked to everyone, every conversation I've had, you know?

Yeah, but then it's Jarvis, and it's like almost AGI-

Yeah

... to some extent.

Yeah.

So-

You have the context that no one else has.

Yeah, but like the amount of compute and the amount... Because you need to recompute like your graph every time you receive new stuff and new things. So it's, uh, it's an interesting space. I think, so to your point, uh, we probably, as an endpoint solution, we probably won't be the one solving for that.

I think that there's like companies that should focus on this-

Yeah

... and be like, "Hey, I'm the engine that will ingest everything that you're doing, and we build a graph, and the graph will be like the best graph ever, and it will be like for each account or each, uh, uh, tenant, uh, will build a graph for you."

Yeah.

That would be great. But is it something for Chopper Buffer? Is it something for like those vector, uh, database companies, uh, to, to solve for? Maybe.

Yeah.

I don't know.

And so for what it's worth, uh, I'm actually dating someone who's doing Upside, and they're mining emails for the CR- basically like CRM population and building a knowledge graph from emails.

Interesting.

So ba- basically, they're happy that you're not doing it. Because-

I'd love to have an intro.

Because obviously if you do it, then you, you, you're a very serious competitor.

No, but it's, um... I think it's not easy.

Yeah.

So I would love to discuss, uh, uh-

Sure

... but like I think we would be probably more a consumer of the outcome-

Yeah

... uh, rather than the builder, uh, of that, uh, of that layer.

Yeah. I think the other big consumer obviously would be OpenAI.

Of course.

They, they clearly want to eat everything inside of ChatGPT.

I mean, this is a cool exit strategy for the- such a company.

For, for them, yeah.

Of course.

I mean, like, do you want to build an Superhuman app inside of ChatGPT? Or I, I feel like that the answer is no, right?

Oh, the, the answer is like ChatGPT or like OpenAI and Superhuman are competitors.

Okay.

Like this is what we fight against-

Yeah

... uh, to some extent. We have a different approach, I think. But, uh, and especially this like ubiquitous Grammarly presence, we are everywhere and everything. I think we'll be-- we want to be more proactive because we are where you work, we can be more proactive compared to ChatGPT that is waiting for you to do things to help you do the thing.

Mm-hmm.

So there's like reactive versus proactive. I think we're more on the proactive side. But so that's the competition. Like just, uh, I would say for notes, but like when Rahul is questioning the quality of our, uh, say AI, uh, queries on Superhuman, he's comparing us to Gemini, he's comparing us to, uh, OpenAI.

So that's the competition we are fighting against.

Yeah. I mean, and, and speaking with Gemini, the chat app obviously has privileged access to all of Google. So they can also-

Privileged access and like the, the search engine is crazy good.

Break them up. Rahul, break them up.

All right. Yeah, yeah.

Awesome. On a more broader side, so you mentioned you only have three people working on AI. What's kind of like the coding AI adoption at Superhuman on the engineering team?

Yeah. Interestingly, like, uh, our path was... So we, we started to really think about it like in Q1, like bunch of like people using some stuff and everything. We didn't have any data, uh, just anecdotal feedback and all of that.

The first thing we've done is like cut the, the red tape.

Mm-hmm.

Like, "Hey, folks, free for all. I will approve the budget like in one hour. You can try anything you want, and deal with the security team, 24 hours turnaround to get, uh, things approved from a security standpoint because you don't want to do some-

Right

... crazy things." Um, huge Q1 was like everyone was trying everything. It was very interesting to see like how things were like working super well on the front end, a bit less on the back end. We are, we are Go shop, uh, on the back end.

And, um, everyone like working on iOS and Swift, and we're like, "Eh, not that good at the time." But like a huge adoption in terms of, uh, in terms of tooling. Also, like on the product side, uh, a lot of like, um, v0, I would say, um-

For Next.js?

No, v0. v0 is, uh, is kind of like your boat, uh, boat, um-

Yeah

... in those like, um-

Because they, they build Next.js sites, right? Or apps.

Uh, yes, but like it's-

They just use it for anything.

We just use it for like a prototyping.

Ah.

To be like as close because, uh, we have a, a founder that is very picky and wants to review the design, and like a design on Figma is great, but like when you can click and do like real stuff, it's so much better, and Figma is not there just yet.

Yeah, Figma has Figma Make.

Yeah.

We interviewed Dylan.

Yeah. Sure.

It's, uh-

Sure.

It, it, it's, it's getting better, I would say.

It's getting better.

Yeah, yeah, yeah.

But like as a PM, they use v0 or whatever, like a tooling like this because it's like-

No livable?

Ah, Superhuman is like v0. v0 is the standard. And, and again, it was like free for all.

Yeah, yeah.

Try whatever you want and everything.

Free market, right?

So free market. Uh, and free market-

v0 won

... v0 won. Uh, always winning. It's still a free market. Q2 was more about, okay, let's try to understand where this is working, where this is not working.

Yeah.

So compile a, a huge list of wins and an area where like, ah, to do this, not good. Wow, to onboard a new, I would say, code area, amazing. I used to spend like a full day to understand all the entry point, the dependencies on the code stack that I didn't know.

Now I need like 30 minutes with, uh, Cloud Code, and I understand how things are working. Even for me, like I'm not in the code anymore, but like instead of like asking my engineers like, "How are we managing like the refresh tokens with Gmail?"

Like now I just like Cloud Code, and I'm using Warp.

Yeah.

I'm a-

Warp?

Warp.

Yeah.

Warp is good. Uh, but anyway, uh, Warp, Cloud Code, like how, how this shit is working, and boom, boom, boom, boom, boom. I'm providing like the links to the right files, explaining you like the, the high-level concept and everything, and I don't waste my engineer's time, uh, to just answer a question.

So pretty cool. So that was Q2, and we started measuring. Uh, so every PR, we have to put a label, "I used AI," uh, or, "I didn't use AI." And if I used AI, it was productive or it was not.

So trying to understand, uh, the lay of the land. Um, Rafi said, "I think we have like 80% of people that are really flagging the PR." Out of that, uh, 80%, um, probably

Eight, ninety percent, uh, I would say of AI usage. So it's all declarative. We're n- we're not, like, plugging any tool to measure, like, the real number of tokens and everything. And out of those ninety percent, again, ninety percent of positive impact.

But it's not always in the code. It might be, like, just the discovery, understanding, like, the lay of the land or, like, uh, stuff like this.

So eighty-one percent. Like ninety times-

So technically, like, uh, yes, it's like, uh- ... ninety, ninety of eighty. Uh, but by, by inference, I, I would like if I caricature, I would say eighty percent of usage and, uh, happy usage.

So, like, roughly eighty percent of lines of code written in Superhuman. But probably-

It's not always line of code because-

Probably more than that. Yes. Yeah, because of PRs.

Yeah, because it's, well, it's the, the, the discovery. Like, most of the time you spend is not writing code. It's, like, trying to understand what you need to solve for, and this is the part that has been reduced.

Uh, in terms of real KPI, and it's-- AI is not only the, the only reason why we have accelerated, but in Q1, we were roughly said at four PR per engineer per week. Q-- So that was in Q1.

Q2, we were, like, closer to five PR per engineer per week, and Q3, we're closer to six. So the global throughput, and again, PR per engineer per week, we can debate.

Right. Yeah, yeah, yeah.

Uh, but that's a throughput measure, and it increased quite a lot. But again, AI is only a piece of it. Technical strategy, clarity of what you n- wanna do, organization. There's a lot associated to that. So we feel pretty good.

One question that a lot of, like, the AI leadership people I talk to have is like, "Am I supposed to, like, ask more of my engineering team now? Like, am I supposed to, like, you know, hire less people?

Should we ship more as a company?" I think most-- The, the thing about AI is like you can do a lot more, but most companies are not built to do a lot more. Like, you know, especially, like, if you ship 100 more features, you don't really have marketing to market 100 more features.

You don't have support to learn 100 more features. Like, how do you think about structuring teams and, like, the expectations of it?

That's interesting because, um, Superhuman historically was, um, very lean in terms of organization. So, like, Superhuman may, like... We have fifty-

Fifty. Crazy

... fifty engineers.

And your user base is roughly-

Uh-

... how many million?

Yeah. Like, less than that.

Okay.

Like, uh, paying users, probably one hundred thousand.

Okay.

Something like this. So it's still, like, a relatively small-

You're still supporting a lot.

But it's-- Yeah. So it's, I would say, small team, uh, pretty senior, and the average tenure is probably four years. So, like, long tenure.

Mm-hmm.

Fully remote, uh, as well, uh, which is interesting. So my AI team is distributed between Patagonia and Canada. Uh, so, uh, access to a different pool of, uh-

Mm-hmm

... I would say, bright people, not trying to compete in the Bay, uh, because people want to go to Anthropic. They want to go to OpenAI.

Right.

And, like, those guys, like, they have, like, the-

They pay too much money.

Yeah. I mean, it's, it's not the same, obviously, competition.

Yeah.

So we, we find the, the people where they are and people that don't want to move to the Bay and all of that, and there are some great people there. Anyway, long story short, um, relatively small teams, um, and we increase the capacity.

We try to not move too fast because we're qualitative. Like, because there's-

Mm-hmm

... it's kind of like a vicious circle. "Oh, we can do more. Let's do more." But, like, all of a sudden, like, the number of, like, bugs coming in-

Right. Yeah

... is also growing and everything, so we try to be conscious. Now we're working on the Grammarly/the new Superhuman, so there's, uh, also, like, an incentive to be... like, to invest a bit more-

Mm-hmm

... because it's a product that is working. And Shishir is, uh, really willing to, uh, implement a model that is called, like, the compound startup. We're still a startup within, uh, Grammarly, so we have our own P&L. We have still, like, Rahul as a founder.

Uh, the only difference between now and before is, like, our board is Shishir and the exec team-

Right

... uh, at Grammarly or/Superhuman. But we need-- We want more people. We want, obviously, Superhuman to, to have, like, more reach and, like, uh, to, to do a bit more. So now we are kind of, like, scaling that, and we are adding, like, uh, more capacity.

Yeah.

Uh, so AI is helping, of course, but it's also helping, like, for the onboarding. It's helping-

Yeah

... for, like, a, a lot of that. But we are, we are adding some capacity.

Yeah. Yeah. I, I think, like, you know, the mainstream maybe pushback on it is like, "Hey, like, you used to pay me X to do four PRs a week, so am I getting paid fifty percent more then I ship six PRs a week?"

I think that's the thing that... That's why there's a lot of pushback around AI as well from people is like: Hey, look, you know, I'm using this and you're getting more out of it, but I'm not getting more out of it.

I-

I think it's, like, the usual, like, you know, new tech-

I so strongly disagree. I would say disagree with that statement.

I would disagree too.

Yeah. I, I, I disagree too.

I mean-

But I, I, I'm saying, like, when you listen to people outside of our bubble-

Yeah

... there's, like, a lot of, like, this discussion around-

Oh

... you know, where the value is accruing and, like, you know-

So, so basically, if you only look at it as you're paying for output, was the previous payment wrong or was the current payment wrong?

Right.

One of them is wrong.

Exactly.

No, no, no. That, that's, that's, uh, that's an interesting point. The way I see it is, like, engineers are well paid. Like, we are, like, a very fortunate, I would say, part of the population. Our salaries are probably, I would say, pretty good and part of, like, the top, whatever, like, five percent in the country or, like, even in the world.

I think that when it talks about, like-- when we talk about, like, the Maslow pyramid, like, engineers, at some point when they're, like, pretty senior, they don't rush, like, for, like, ten more K or twenty more K or, like...

I mean, if we talk about, like, millions and everything, sure, but that's, like, the one percent of the one percent.

Mm-hmm.

For the rest of the population like us, I think that just, like, the joy and the dopamine is coming from what you ship. So, like, having this ability to ship more value and have more customers, like, being happy with what you do, like, you f- end your day and you feel like, "Damn, that was a good day."

So I think that the, the discussion is not about, like, the, the money itself. It's like, "Oh, damn, I'm in an environment where I ship fast. I can have, like, all the tools they request, uh, within twenty-four hours.

I can basically be, like, the best version of myself, and I have fun in a good team." You don't have a lot of attrition when-

Yeah

... I would say you have an environment like this. So like sure money, so you need to pay people like fair, I would say a fair amount. But if you're like just like fair, people tend to stay if you have the right environment.

And like-

Mm

... helping them to go from four PR week to six, they're like, "Shoot, like I'm so much better than beginning of the year. That's so cool." And you don't have that everywhere.

Yeah. I, I'm with you. I, I'm curious to see more of the, the scores evolve. Um, awesome. Any parting thoughts?

Just generally, uh, your take on AI on the software industry. You've been in this for two decades. Do you think that people should still learn to code? Do you think the junior developer is screwed? You know, any, any of those opinions that are common topics.

Yes, of course. Of course, uh, you need to learn to code. Uh, like I, I see this about like, uh, kind of like the switch from like assembly to C. Like-

Yeah, it's a higher level

... it's just another level of abstraction. Uh, but at the end of the day, you still need to understand how a computer is working. You need to understand how memory is working, like, like swaps and all of these things happening on a, on a, on a server, like how a server is working, like serverless between quotes.

It's always a server of someone. Uh, you need to understand the fundamentals, uh, to be good with AI. I do believe that AI will do only one thing. It will separate faster the good engineers from the bad engineers.

If you're good engineer, and you're using AI well, you will be an amazing engineer. Uh, if you're a poor, lazy engineer, and you don't want to understand things that you're doing, AI will make you even worse, uh, because you will have the feeling that you get it, but you're not being...

You're going like behind the magic, behind the curtain, behind things and how they work. So I think AI is a, is a blessing for our job.

Awesome.

Great.

Any, uh, final call to actions, hiring people, things you want people to do in trying the product and give you feedback on?

Of course, try the product. Of course, complain to me if things are not doing, I would say, great, uh, and, and they are not great. Yes, we're hiring. So we're hiring product engineers, so people that have a strong appetite for like the user experience, uh, because I do believe that in the world where the technical moat is not that a moat anymore because like startups, in two weeks, they can build something that is close to what you're building.

The difference is like the how you think about the user, the flow, and all of that. So people that have this appetite for nice interface, beautiful product that people love, this is the type of engineers we want. Good engineers, that's a baseline, of course, but like with this like spike into like the user experience, even if you're a backend engineer.

Backend engineer, but you care about like the latency because it's, uh, having an impact on the end user and all of that, this is the type of engineers we're looking for, and we don't care where you are. So you can be in Patagonia, as I said- ...

or you can be like up north in Canada. We try to, uh, limit, uh, limit things to-

Time zones, yeah

... like Americas basically, and um... But, um, yeah, just looking for like, uh, bright, gritty people that want to have fun. Uh, we're seriously fun.

Cool. Thanks for joining us, man. This was fun.

That was cool.

Thank you.

Thanks for having me.