Ep 30: Superhuman CEO Rahul Vohra on the Future of Email with AI and the Role Agents Will Play

I think if you were to go back a thousand years, when we were subsistence farming or we were hunter gatherers or Maybe even like 1500 years, those people would look at our jobs and go, that's not work. You're just playing.

I think a lot of people today would look at Pat and my job and say, that's not work, but it's kind of you to go back a thousand years. This is unsupervised learning. Redpoint's AI podcast. I'm Jacob Efron. I'm joined by my co host, Pat Chase. And today we had an awesome guest in Rahul Vora, the CEO of Superhuman. Superhuman is the lightning fast email tool that has taken the tech world by storm. And we hit on a lot of really interesting things, including all the AI features Superhuman's building, how they think about the future of email, and especially the role agents will play.

Journalists perennially claim the death of email and of course it never happens. And there's actually a pretty good reason that never happens.

And we talked about some really interesting product considerations when it comes to taking on incumbents and getting people to use new workflows. What areas are you like most excited about? What areas are you not as excited about as you do on the investing side?

Where I was bullish and continue to.

Be bullish is I think folks are really going to enjoy this one. In general, if you're enjoying the podcast, please make sure to like, subscribe, share with a friend. We're trying to get the word out and definitely stay around for Pat and my debrief. Well, thanks so much for, for being here. Really appreciate it.

Yeah, thanks for having me.

I've been excited for this one for a while. As, as Pat and I were saying, I think everybody in Redpoint is a superhuman user and we're all beneficiaries of these AI features you've done. So it's, it's exciting to get to dive into them. You know, I figured we'd start like in the most broad, futuristic way we could, which is just what do you think email is going to look like in five, 10 years?

I love this question because I'm, I mean, you've been in the business long enough. I've, I'm sure you've seen journalists perennially claim the death of email. And of course it never happens. And there's actually a pretty good reason that never happens. And it's to do less with email and more with email addresses. So think about the very first thing that happens when you join a company. You get an email address. And why is that? It's because the company needs a way to address you in a way that is owned by the company. Some people say, well, couldn't you use your cell number as your unique identifier? Well, the problem is that belongs to you. Unless the company is provisioning the number, you're going to take that with you. But if you think about an email address like rahuluperhuman.com there really isn't any more succinct, human readable way to talk about me as a specific company in a way that's owned by the company. So it's given to us to identify us. It's the unit of authentication. Even if you're going to use something like an okta or a rippling or some other SSO solution, guess what, you're probably logging into that thing with your email address. So it's like this underlying piece of infrastructure. So email addresses aren't going anywhere. But what email is is obviously going to change. And there are lots of themes that we can talk about, but I think two that are very top of mind today. Of course, this being an AI podcast, one is AI and we can get into all of the details on how email will change with respect to AI, but I think the, the short summary is we're heading towards an agentic AI future. So you can imagine AI agents in and around your email doing things on your behalf and then secondly, collaboration. So we were talking about this just before we started recording you all at Redpoint. Use Superhuman. You've probably started to see some of our team features. Team read statuses, team reply indicators, team snippets, team scheduling. These are some of the basic things and actually some of my favorite features, because you don't have to remember to use them. So long as all of you have turned on read statuses, you can actually see when a founder has opened your colleague's emails, which sometimes can make the difference in closing that deal. So really important stuff where we're getting to actually pretty shortly is going to be the ability to share a thread and then comment on that thread.

I mean, I'm intrigued by this, this notion of, you know, there'll be a bunch of agents running around our inbox helping us do things like maybe for our listeners, just paint the picture of what some of those might look like, you know, down the line.

Yeah. So let's actually zoom out and I think it might be helpful to share the context of AI in Superhuman. And there's a very particular phasing that we use that I think folks might find helpful and then we can get to the sort of the futurism stuff of agents. So there are three flagship AI features in Superhuman. There's actually a bunch of ways we leverage AI, but three flagship AI features. And the first was write with AI. So you simply jot down a few words and then we'll turn those into a fully written email. And we match the voice and tone in the emails you've already sent so it sounds like you. And once you've written it, you can then shorten it, you can lengthen it, you can improve your writing, you can fix all your mistakes, you can even change language. We built that I think last June or July. We then followed that up with our second flagship AI feature which was Auto Summarize, which we launched last year. Great picture. One of my personal favorites. I actually have it on my phone today that hasn't released yet.

Pat writes some really long emails and.

So it's very always need the auto, the summary.

It's honestly game changing. And so for folks that don't know about this above your conversation, imagine a one line summary that summarises the entire thing. It updates instantly when new emails arrive. If you want more information, you can just hit the shortcut and then you can see a bullet point summary. And as you guys have probably seen, this changes the way that people work because you almost always read the one line summary first and sometimes you don't need to read the email because the one line summary was enough. So you can just mark done or hit snooze and come back to it later.

Totally.

So that was our second flagship AI feature. And then our third that we released just a few weeks ago is Instant Reply. So imagine waking up to an inbox where every email already had a draft reply. You would just edit and then send and sometimes you wouldn't even need to edit. So that's the thing that we released a few weeks ago. It's blowing people away. Of course we consumer DNA in our company, everything is instrumented and we have metrics for everything. So we can actually measure how long it takes for people to write can send emails. And users who use instant reply are writing emails twice as fast, that is in half the time as before. That's real business impact. This actually blew us away. We were not expecting it to be that effective. Okay, so those are the things we built in terms of roadmap strategy. And I think this sort of applies to any founder who has an application, has some distribution, they're thinking, okay, what do we do? What's a sensible strategy for progressing through this phase one, what we started with was on demand AI features like Write with AI. These are features that the user has to remember to use. They're pretty easy to build, generally speaking, and they're pretty cheap to run because unless the user's mashing the same key all day long, they're not going to keep using them all day long. So it's a good way to dip your toe into the water and to see A, do you understand the technology? B, how good is it? And C, do users love it? For us, seeing how much users loved it, that gave us the confidence to move on to the next phase. And this is the phase of always on features. Features like Auto Summarize and Instant Reply. Now, these are a lot more ambitious. They're a lot more difficult to build, they are a lot more expensive to run because they're always on, always working for you. And they operate at massive scale. So to put this into context, when you train an LLM based on emails, the corpus most people use is the anbon corpus. That's 500,000 to 600,000 emails. If you've poked around with it since we've launched Auto Summarize and Instant Reply, we've already processed 4 billion emails. I don't know how many orders of magnitude more that is, but it's a lot. It's a lot.

I want to ask how many of those were from.

So that was the next phase. That was phase two and that's the phase that we're in right now. We have more always on AI features coming and it's paves the way and creates the foundation for phase three, which we alluded to at the beginning, which is our agentic AI future. So I believe the future of AI is we'll all have agents, probably multiple agents each, and you can totally imagine an inbox or an email agent. These agents will be mostly autonomous. You can give them goals, not tasks. They'll be able to plan and break those goals down into subtasks, work on those subtasks serially or in parallel. They'll be able to handle and resolve ambiguity by asking you questions and maybe interrogating other systems, Perhaps even internal APIs or your CRM. They'll even be able to interact with other agents. So that kind of an AI, we think, is where we're heading towards and that will free us all up to be more creative, strategic and impactful.

That's amazing. I love that. And it's been really cool to See, you guys ship instant replies and all of these new features and instant replies was interesting. I love the product and use it every day. Even if I wasn't going to use the replies, just having the three options of, oh, those are kind of the three ways that I could answer this email. I found that helpful in getting to a decision faster. So it's cool that you're kind of seeing that already have real impact on your users.

Yes, absolutely. And that was one of the things that doesn't come across in metrics, but obviously we qualitatively analyze every feature as well. One of the things we keep hearing is, wow, you know, I was sitting on this email because A, I didn't know how to reply or B, it was just like it was just below the threshold of doing, but it was still important, but just below the threshold. And the fact that there was an easy do it this way made the difference and people are actually now replying to things they otherwise wouldn't.

Totally, yeah. It's super cool to see. And I was thinking on mobile it would be amazing too, of like then being able to kind of draft or have the AI help you when you're on the road. It was also really cool.

I think it's the case that basically every AI feature you can imagine is 10 times more useful on mobile. That was the case with Write with AI, trust me, it's the case with Auto Summarize. Again, we only have it internally coming out probably in a few weeks and I believe it will be the case with Instant Reply as well.

Yeah, super cool. What's been the team structure that you guys have kind of put around AI, or how have you approached structuring the team to ship all of these AI features so quickly? It seems like you guys are really moving incredibly fast and have pushed a lot of things into production faster than almost any company out there.

Wow. Well, I'm glad it feels that way because to me it always feels way too slow and many other companies had a head start. But, you know, I think, I think one thing we do well, even if sometimes we start something late, is when we do something, we take it really seriously and like, it's full steam ahead. So we began our AI journey, specifically our generative AI journey, last February. We were doing AI things before then, obviously. We'd released Autocorrect in January of last year and that doesn't use generative AI, but it is still the best Autocorrect that you'll find on the planet.

Was it like immediately apparent to you when ChatGPT came out like, hey, this is going to be, you know, this is going to change the product or this is an all hands on deck effort. Or was there some moment where you were like, yeah, this is now time to dedicate a bunch of resources?

It wasn't immediately apparent. And you know, this is, I think this is one of those moments where I really had to step out of the office and just take some time out to zoom out and think, okay, this is a really big deal. And there's this conference that folks may be aware of, the lobby conference that David Hornick puts on. And it's beautiful. Conference takes place in Hawaii. Three or four days. It's not very big. Like 100 or so people like half founders, half VCs come and just basically hang out with each other and talk about the future. And obviously the topic last, I think it was February and March was AI. And being able to have that conversation in that environment without having to worry about running the business day to day was what gave me the perspective to realize how important this is. So I came back to the Bay Area and I was like, oh my gosh, we have to prioritize this. Everything is about to change very rapidly. But it was very hard to prioritize because we usually do prioritization in one of two ways. We either try and figure out how valuable is this feature going to be for the user or how valuable is this feature going to be for the business. And it was hard to know how valuable this feature would be or this collection of features would be for the user because we couldn't predict how well they would work or how they would feel. And it was hard to make a business case for them because we didn't know how much revenue they would bring in or how quickly they would help growth. And it's very rare that you have something like that. Normally you're able to, excuse me, reason about it on two dimensions. So what we did was instead we took a philosophical stance. We stated our core belief, which is that LLMs would change everything. It wasn't going to be obvious how and it wouldn't be overnight, but we knew that, or I knew that the way that we work was going to change. My job at that point is then to basically use founder fiat and say, this is it folks, we're going in this direction. Obviously it's not that easy. You have to tell the story many, many times. But that's essentially what we did. Very fortuitously in February. I had also just hired a president, Paul Tessier, Now, Paul is an amazing operator, an amazing leader, a chief product officer twice before working at Superhuman, and he's exactly the kind of person I brought in to make these kinds of changes possible. So he and I discussed and we came up with a new operating model for getting AI off the ground. So we have two operating models. One is alpha mode and the other is theta mode. You're probably wondering, why alpha mode? Why theta mode? You've probably heard of alpha brainwaves. Alpha brainwaves are when your brain is most coherent and in the present. And theta brainwaves are observed when your brain is thinking about the future. And so when a team is operating in alpha mode, which should be the default, they operate mostly autonomously. You give them goals and they're held accountable to the goals. That's normal business as usual for most Silicon Valley companies. When a team is operating in theta mode, it's quite different. It has an embedded executive, in which case was me. We all carry that goal together and the executive is doing individual contributor work on a day to day basis. And the expectation is this project can't fail because it's existential, right? And it moves really, really fast. So starting February, we went theta mode and AI and I knew basically nothing. I mean, which is kind of ironic because I dropped out of an AI PhD, but that was a very long time ago, so I had to catch up.

Well, I feel like everyone knew basically nothing because it was all so new, right? Even if you studied AI a couple years ago, then everything changed with LLMs.

Yeah, like, my stuff was 20 years old, so I was like, at least I had that background. And I very quickly got up to speed. We staffed a team with one designer, one marketer, maybe two or three engineers, and we just sat down and knuckled. And at that time, no one knew anything. No one even knew from a product design perspective what the user interfaces should be, right? Should we have a sidebar? Should it be conversational? Should it be built into the surface area of the product? Do you write the prompt, then select it, then say, turn this into an email? Or do you enter the AI mode first and then put your prompt there and say, turn that into an email? These things, maybe they sound small, but they're actually very big decisions in terms of the usability and how these things feel. So we knuckled through that for a few months and then we launched right, with AI, which at that time we called Superhuman AI last summer. And I mean, this thing went viral in a way that Superhuman hasn't been viral. For a few years and it was just like, oh, my gosh, AI has the ability to, you know, just make a product, make a company top of mind for everybody again. And that's something that's continued to be true for us since the summer of last year until now. I post a thirst trap graph on Twitter yesterday. You may have seen like the graph that goes, yeah, yeah, yeah, that's AI. Like, that is growth that happened because of AI.

Amazing.

Yeah.

Is that AI usage or. I guess.

No, no, it's even better than that. That is new, monetized retained seats per week.

Wow.

Amazing. Yeah, it was a good.

I liked your caption too. Just keep going, just keep going.

I mean, yeah, startups are hard. I've been doing this for seven or eight years and sometimes there are some months where you're like, oh, my God, this is so hard. We need a breakthrough. We need a breakthrough. But I am reminded of, I think Paul Graham puts it best. Just don't stop typing because what are the ways that startups die? You run out of cash. Okay, don't do that. Raise money, don't spend too much money. Make revenue. Other people talk better about that. Co founder fallouts don't fall prey to that. And I'm not the guy to talk to you about that. We've done fine, thankfully. But the thing that if you get through those two things and if you've lasted seven or eight years, you've got through those two things, the thing that then gets you is you just run out of steam. It's just too hard. And SaaS is especially prone to that because it takes a very long time to build a great SaaS business. You guys know this and so just keep going.

So interesting. And I think, you know, I mean, one, it's interesting just the kind of the white space of design decisions you guys had early on. And you know, to your point, there's so many different ways you can do it. Obviously, you know, finding an intuitive way to have it seamlessly fit into the product feels like the North Star. But there's, you know, it's all kind of untrodden territory. Like, I don't even. I guess you probably just tried a bunch of the different designs and figured out which, like, felt the most intuitive to your end user or how do you, how do you even come to the design decisions? You guys reach out?

Well, why don't we take a specific feature like Instant Reply, which is the one that we just launched. Again, very non intuitive. No one has built anything remotely like this, actually, that's not entirely true. I think there were precursors to this. You could argue Gmail Smart Reply is a precursor. So this is the three little blue widgets it has into emails. All it does is it puts the text of that widget in the email. And I think Microsoft demonstrated in concept a copilot feature where there are some little widgets and you can click them and it writes a really long email. But there were many product decisions that we had to take to get instant reply feeling right. So, for example, the first one was, should they be visible before you start replying, or should they only be visible after you start replying? Or maybe they should be visible at both times. And my gut sense, for what it's worth, was they shouldn't be visible until you start replying. And I had a very logical argument for this. My logical argument was, you don't actually reply to most emails, so why would we clutter an interface that we've tried so, so hard to keep minimal over these last seven or eight years with this thing that you're not going to use all the time? Surely we should wait until you actually start replying. And then I realized, oh, wait a minute. Half the value of this feature, obviously, is to make your reply really quickly. But it goes back to what we were saying before, which is sometimes the activation energy to hit enter is the problem. Sometimes you need a bit of goddamn inspiration. You need to see the three ways to actually do the thing that wasn't obvious when we started building the feature. I think it wouldn't have been obvious to really any product design team. Like, you look at it and you go, okay, instant reply. What's it going to do? You're going to reply instantly, but half the value is inspiration to reply, which wasn't obvious. So that was the first thing. I think the second thing was, how do we make it from a visual design perspective, tucked away and minimal. We have this principle at Superhuman there when you want it and out of the way when you don't. And that is hard.

That's the dream for AI in general, right. In products.

I know, which actually makes me feel super lucky, because that is the razor that I've been honing, actually, since I was a kid. I've been a designer since forever. And constantly making software and user interface has been my obsession all day, every day. And so we were ready, we were ready to take on that challenge. And I wouldn't say we got it perfect, but I think it's pretty dialed in. We've had so, so many bits of feedback on this feature, no one has said, oh, I think these things are too intrusive, like I wish they weren't there until I started replying. So we had to dial that in. The next thing was actually the length of the replies. So write with AI, which is a thing that we launched last summer. If you've used it, you'll know it generates pretty medium length replies, right? It probably does a few sentences to a paragraph per points that you give it. And that's by design. That's actually what people want. We thought that's what people would have wanted with instant reply as well. Or I guess we didn't really question it. We kind of started with the prompt that we designed last summer for instant reply. But it felt weird. Like you'd hit tab a few times, you'd hit enter and it would put the thing in and it'd be like, oh, there was like an impedance mismatch here. I'm going really fast. And the thing that was put in is actually kind of medium sized. Something didn't connect. And then I realized when you're moving really fast, you actually also want the messages you create to match that and be really short. So instant reply by design only puts in one, maybe two sentences, right? And often doesn't even wrap the screen. It's just a really quick snappy response which matches the impedance of you moving, which means you can continue to move quickly because what you don't want is, is to be moving quickly. Hit enter and now you have to proofread a really long email and then maybe edit it, right? You want to hit enter, maybe change a word and then hit send and move on the interaction design. It's funny, I was at OpenAI yesterday and doing a talk for some A16Z founders and we talked, obviously we're at OpenAI, we talked a lot about prompt engineering and fine tuning and rag and model customization and all the other things that what startups are doing today. What I realized was the reason why, or one of the reasons why instant reply is so effective, isn't the prompt engineering and all of the stuff that we did on the AI side. Oh, that's really, really important. And that was very hard. It's actually the interface itself, right? The fact that you can mash tab and hit enter and that's it is key to making it work. And early versions of the feature didn't work that way. Early versions of the feature you had to sort of, I think you had to hit down, then you had to hit left or right. And this, by the way, this is one of the things that I spend a lot of time watching out for. The arrow key motion is a shoulder motion, it's an elbow motion. And if you measure the time it takes for people to do it versus tap enter, it's like 10 times as long. It's also quite painful. You do it too much, you get rsi. And for what it's worth, most people don't even bother doing it. So one of the changes we made was to switch from down and arrow keys to tab and enter, and it instantly became 10 times more usable and 10 times faster to use. So interesting. And obviously this is just the story of Superhuman, right? A lot of what we do is like really minor, seemingly minor UX tweaks like that, but you do 200 of them and you have this product that feels completely different to everything that came before it. And then speaking of prompt engineering, that was the last thing. That was really, really tough. If I were to show you the prompt for this thing, oh, my God, it would blow your minds. It is so big. In fact, our partners at OpenAI were like, we think you're the partner with the biggest prompt right now.

You were really pumped for the longer context windows.

You're like, all right, make Jacob sound more professional. Just do it. Do whatever you can.

It's actually funny. We specifically say avoid corporate jargon. Like, we apologize for the inconvenience.

I actually really like this about the Superhuman AI. And I don't know if that's what makes it sound. It sounds like something that I would have drafted or, you know, that it sounds like a human versus a bunch of the other things. You can kind of tell this was drafted with AI, and then that has the worst effect if you're receiving it right. If you're receiving one of those things that seems kind of extra verbose or whatever and, you know, was written by AI, so.

Yeah, yeah. And I think this goes back to what LLMs are trained on. We just don't have great training data when it comes to emails, which I think puts any email startup, including Superhuman, in a great position because our customers actually do want the AI to sound like them, and that's a thing that they're willing to let us help them achieve.

How do you actually make it sound like. I mean, does that end up being like, does that influence the prompt you do, or do you actually end up kind of fine tuning to some extent, like the output or, you know, how does. How do you get it so that, like, my emails Would would sound different than Pat's or how do you see that evolving over time? Hey guys, this is Jacob. Hope you're enjoying the episode. We're trying to get the word out as much as possible, so please share this episode if you think others will find it interesting. Now back to the show.

We avoid fine tuning as much as we possibly can because it takes time. It costs money. Actually, the money is the less important side. It takes time. You have to assemble a data set, you have to tag it, the training takes a while and most importantly, you don't really get to transfer that as the models update from one version to another, even successive versions of the same model or model generations. And so avoid fine tuning. That's sort of been our mantra. And push prompt engineering as far as it will go. And yeah, I think this is why OpenAI were like, wow, what you guys have done with prompt engineering and multi shot learning. Basically it's other emails that this person has written are like this. It's pretty interesting how far you can push that. I think you can push that really, really far.

Have you guys found any of the tools for prompt engineering or evaluation, any of those helpful? Or have you had to build those types of things in house or how do you think about developing these prompts and iterating on them? I'm sure they change inversion and there's a bunch of intelligence that goes into those.

Yeah, that's actually been the hardest thing about working with LLMs as a technology. And the most surprising thing is how even successive versions of the same model break what we thought was a great prompt. So when we wrote the auto summarize prompt last November or so, I think there was a January version of GPT 3.5 Turbo which just completely broke it. And we were like, why? The prompt looks reasonable. We weren't doing anything too weird, but it just didn't work properly after that. So we've had to build a pretty robust regression testing and evals framework on top of it. We haven't built it in house. I fully believe it's a startup to do so. We use BrainTrust and we've been very happy with that. Every time we change a prompt or we change the model, we have a battery of tests that runs. We know did something break? Do we need to look at something? And the fun thing about these evals frameworks is you can of course use LLMs to assess themselves. So we did this with instant reply. Instant reply, by the way. Way, way harder to build than autosummarize, especially on really, really Long threads. A funny example is, let's say someone is joining the company or leaving the company. And these threads often get 20, 30, 40, 50 emails long. It's big, long thread. What the first few versions of Instant Reply would do is it'd be like, hey, Bob has joined the company. And then it now says, alice congratulates Bob on joining the company. And Instant Reply would be like, thank you Alice, for congratulating Bob on joining the company. That's bad. Instant Reply. I reminded of the bad intelligence from Team America. What else?

Great reference.

I know, right? Other examples would be not addressing the topic at hand. So the thread of the conversation will be about something else and it will go off on a side tangent. Another example would be not addressing the right person and replying to the wrong person. So in the Evals framework, we actually ask GPT to assess itself and answer the question, are you addressing the right topic? Are you addressing the right person? Is the response that you created as long as it should be, and a few other ones as well. So that's been the most surprising thing and we rely on braintrust to do that.

How have you thought about which models to use? It sounds like you're building on top of OpenAI and maybe some of their different models. Have you thought about going down the open source path? Or how did you, you know, land on choosing the OpenAI models?

When we started, I mean, it was tough. Like we peeled off, like I said, maybe three, four, five of us at most. And we didn't have the luxury that some larger companies have, which I think is to sort of do multiple things in parallel, like maybe run OpenAI models and anthropic models, or maybe even Mistral models as well, and just see and have that luxury of choice upfront. When you don't have the luxury of choice, I think you have to be even more precise in calculating in the decision that you make. But that said, it wasn't a particularly complex decision. My belief was that, and this is still true, that OpenAI is about six to nine months ahead when it comes to model quality. They are a little bit further behind. They have a lead, but not quite as far ahead, I should say, when it comes to model speed, and both of those are super important for our use cases, Speed is actually tremendously important and they've demonstrated their commitment to aggressively lowering costs. I mean, they don't have to do that, but they're doing it. And so every time we're thinking, well, this thing is too expensive, I'm able to point to the Fact that, well, it just keeps on coming down and I have every reason to believe that the costs will keep on coming down, going forwards. And if you run the clock all the way forwards, I don't think it's that long, maybe two, three years before we have Edge LLMs and it's actually running on device. So I mostly see this as relatively short term cost and the value is going to start accruing elsewhere in this entire ecosystem. So that was one of the reasons we picked OpenAI. The other is they've just been so incredible to work with. So we have a deep collaboration with them. Our executives are talking a whole bunch, our marketing teams are, our go to market teams are, our engineering teams are. When we've ran into difficulty with engineering our prompts, they're like, okay, try this, try this, try this. And making instant reply work really was a collaboration between the two companies that we've been very grateful for. I think that's because they're a startup and they want to win, they want to build AGI before anyone else does and for all the reasons that they want to do that. And they're very good at collaborating with startups to do that.

When you mentioned cost, I'm curious, how does it cost play into decisions today? Are there things you would do, you know, if OpenAI is good at, like you said, consistently dropping prices, like if prices came down like 100 times, is there stuff you would do differently or is it kind of like that's not, you know, as much of a consideration today because you're providing so much value for the stuff you are doing.

If prices came down 100 times, I think in some ways things would stay the same. You know, for example, we don't use GPT4 for instant reply because GPT4 reserved instances have a much lower tokens per minute limit than 3,5 instances. And actually the problem with something like instant reply is less the intelligence of the model and more the peak throughput. The problem and the challenge with instant reply and Auto Summarize is it's a very, very bursty feature and it needs to work in real time. And it turns out everyone queues up all their emails to Send out at 8:55 on Monday morning, which is that.

I was the only one.

Folks, please stop sending your emails at 8:55 Monday morning. Yeah, because our queue just gets insane, like it's nuts. And then we have to like so much of what we built is queue management and like saying, okay, we have now another billion emails. Okay, let's generate all the Summaries, let's generate all the instant replies. And it would be completely infeasible to do that with a model like GPT4 Turbo, but it is feasible to do it with 3 PT 3.5 Turbo. Now, if costs were to come down 10x or 100x, could we do it from a cost perspective? Maybe. But I think even from just a throughput and a latency perspective, 4turbo isn't going to cut it. So it's not just cost that has to change. I think it's also the advent of massively accelerated hardware, things like Grok, that will pave the way for new use cases and new applications.

It's interesting because obviously I imagine you're running the model to summarize and to draft an instant reply like I assume adds up at some point. But it seems that the bet you're making basically is, look, this stuff keeps coming down. Whatever the cost is today to do this is the highest it'll be for us.

That and Superhuman was designed this way from the ground up. We do charge $30 a month retail price, a little bit less if you're buying teams or a contracted account from us in order to do things like this. And in the early days of Superhuman, I'll use margin in air quotes because it's not actually margin, but the extra revenue, so to speak, would go into things like onboarding and into training. And this was years ago, right? Remember when Andreessen Horowitz led the Series B and you couldn't self service into the product, Whereas nowadays that ratio has flipped. So we went from 85, 15% concierge to self service to today it's flipped around 85% self service, 15% concierge. And what we've essentially done is we've shifted those dollars into AI. Right. It still costs money to serve the business, but we are deploying the revenue into what we believe is our duty to our customers, which is to build the fastest, most powerful AI email experience ever.

It was interesting you alluded to, obviously. I remember my Superhuman onboarding viscerally. I'm sure you do too.

I got the concierge treatment felt very important. Yeah.

But I think that one thing that I find so interesting about, you know, your story and relevant to a lot of AI startups today is you obviously have been so successful at introducing people in new workflows and getting people to do things differently. And, you know, I think this is something that a lot of AI companies are thinking about right now. How do we, you know, AI features look differently Workflows work differently. How do you kind of educate your users about that? And I'm wondering, one, you've obviously, you know, the always on ones, I guess maybe less relevant, but for some of the ones like the writing features, you've obviously had to teach people how to use them. And two, just any advice you'd have for, for these startups that are thinking about like, hey yeah, we've got something, a really cool net, new workflow if you use it. But how do we convince or how do we get people to to start using it?

I believe that the core trick, if there is a trick, is actually to start with the always on features like auto summarize like instant reply and then connect them to the on demand features in general, like in product management. My philosophy as a Product person is 100% Reach features are so vanishingly rare that when you have the opportunity to build them, you should build them. And sometimes it's non obvious. So for example, when we could have been starting with generative AI, so this was Q4, not of last year, but of the year before, we weren't. And the reason we weren't is because we were busy building autocorrect and the reason we were building Autocorrect is it's one of the very few opportunities to build a 100% reach feature. Whenever you have those build those, they are so, so rare because the main reason startups don't make progress, one of the main reasons is you just end up building stuff and people don't use it. Right. Like you probably see this all the time in your portfolio companies. Half the battle is getting people to even use the thing that you built. So we started building the always on features. A trick is to connect those to on demand features. Imagine you've just used instant reply and that's a much easier problem because they're always visible underneath qualifying emails. You now have an instant reply inserted. We haven't done this yet, but what we're planning to do is connect it to write with AI. So after you've inserted instant reply, you can imagine a few little buttons in the reply form like lengthen, which is going to be a likely one. Shorten is probably very unlikely, but what would they be? I mean, I'm just thinking about this on the fly. Let's just say shorten, lengthen and custom. And if you were to tab over to custom, you're now in the full write with AI prompt. When you enter that, you'll see something on the screen like to get here faster next time. Hit command J. And so I think that kind of connection between always on features and on demand features is the best way to teach people about on demand features.

Very cool. Yeah, I really love the UI that you guys have around, giving people options and buttons. And it seems like that is kind of one of the best ways of interacting with AI today versus at one point people were saying, hey, everything's going to be a prompt or chat or that sort of thing. And. And I think people want more control than that. Right. And I just, I think you guys have done a really great job of kind of introducing those to users. One thing I was curious about on the. On the pricing side, you talked about, you know, taking the budget that was for the onboarding and then moving that towards AI. And I was thinking about the cost when it came to things like the always on features. Right. Then you need to run that model for every single email that each user gets. But you haven't charged anymore for Superhuman. And is that by design? Like, there's been other companies out there that have increased their prices or, you know, I think notion they were charging more for notion AI. Have you guys thought about that? Or how do you think about pricing for AI features going forward?

I think at some point we will, but that's actually primarily the fact that it's been $30 a month for eight years. Yeah, it's a very, very different, much better product today than it was eight years ago. Obviously, it's just hard to do. We have a pricing project going on right now and we've talked primarily about AI today, but at the same time, we're building Superhuman for Sales, which is a version of Superhuman designed specifically for salespeople. And it has a HubSpot integration and a Salesforce integration and a recent Opens feed. We've also found that founders love Superhuman for sales.

I love Superhuman for sales. I love the recent open feed. I toggled it on and I love the recent open feed is great.

Yeah, it's super helpful. It's shocking. If you email someone within five minutes of them opening your email, assuming you don't make it weird, they're like 20x more likely to reply. Obviously, there's a limit to this. You don't want to sort of be creepy or whatever. And good salespeople and good recruiters and good founders and good investors know how to play this game properly. But it is a superpower and we want people to have these. So there's that line of business, there's the enterprise line of business where it's really Exciting. We've just closed a deal with a tier one strategy consulting firm. I can't say which one yet, but they bought superhuman for 1500 managing director partners. And the reason they're doing this is we've been able to show that the MDPs are replying to emails 8 to 13 hours faster. You're an MDP, you're running a book of business, you're selling projects, you're running a team. You live in email. This is real business impact.

Yeah. As a former lowly analyst, that would have been, that would have been nice when I was around.

You're waiting for your boss.

Yeah, exactly. Like, turn on your Superhuman, please.

But this is enterprise. This is, you know, do you have an intune integration? And I was like, what's intune? And I was like, oh my God, there's a whole bunch of other stuff. It's like, do you have data loss prevention? Can you please beef up your anti phishing features and all of that kind of stuff?

Yeah, not. I was the most. When I was at McKinsey, we were using Lotus Notes, so not always the most progressive.

You look way younger than that.

It was like, it was a decade ago.

So long story short, pricing is very, very difficult. And we have all of these streams of work that we're trying to figure out. How do we price Enterprise? How do we price Superhuman for sales? How do we price like a platinum edition? I think like every founder would just buy the give me all the tools edition. Right. How do we price AI? Hopefully in a few months we'll have an answer, but I don't know right now.

I mean, I guess, you know, shifting over. You obviously are quite an active investor. I'm sure you get your fair share of AI pitches as Pat and I do too. What areas are you like most excited about? What areas are you not as excited about as you do on the investing side?

So Todd and I were actually quite bearish on most de novo AI companies because most of the pitches that we got from scratch AI companies were, well, here's the app that we were going to build anyway. Now we're going to build it with AI. And at that point you almost have two companies to build, right? You have building the thing and then you have making it amazing with AI, and that's really, really hard. I would not want to be building an email client today and also have it try and be the best AI email client. That sounds, frankly impossible. Like you'd have to raise hundreds of millions of dollars and basically run two companies and it's just really, really hard. So I think I'm pretty bearish on those pitches. Where I was bullish and continue to be bullish is if you have some kind of distribution advantage. Perhaps it's contacts from your previous line of work or it's some distribution insights that you've figured out, some kind of viral loop, a viral mechanic that no one else has figured out, or you just have a track record. Like your Nikita Bay are just really, really good at launching consumer teenage apps, then sure, by all means. Because what you're actually investing in is the ability for the founder to create distribution and to retain users. AI is kind of incidental. AI might be a hook, but it's sort of incidental. Where I am really bullish is, as I mentioned, our agentic AI future and I don't know what that ecosystem looks like. You know, clearly we're going to need some kind of service. This could be chatgpt or it could be something else that federates all of the AI agents we have that has some kind of authentication handshake between them. You know, I imagine a world where I was going through. This is like a real example where my partner and I are doing family planning and we were going through the process or we want to plan for a potential pregnancy. And I was thinking, oh, when's my next open enrollment? Because that's like the last thing on my mind, right? I don't even know. And should I change my health insurance plan? The way I actually had to do that was I had to ping head of hr. Head of HR said, okay, go and talk to this person and go talk to that person. Like, we'll go talk to our benefits advisor firm, Benny, who's great. But then that's now another phone call. Let's schedule like a meeting. And I'm just like, please, just very simple questions. I imagine a world where I fire up my executive AI and I'm like, yo, we're doing family planning. When's next open enrollment? And it knows the roster of my other agent AIs. And it goes down the list and it goes, talk to the rippling one. And it's like, hey, rippling AI, I'm here on behalf of Rahul. And the rippling AI is like, okay, prove it. And they have some kind of cryptographic handshake. And it then says, okay, the question is, when's open enrollment? And it says May 24th. And then it says, okay, he's planning for pregnancy. These are the details. It probably goes, you know, what's the age of his partner what their health insurance details. Okay. Based on all of that, this is the plan that we recommend. And it comes back to me. And that whole thing is instantaneous. Right? It doesn't take. Which is actually crazy given that they'll probably be interacting in natural language. It's like a natural language conversation that happens instantaneously and it doesn't take meetings and days and phone calls.

And do you think all of those agents will live inside of Superhuman or do you think then it will be a Superhuman agent talking to the rippling agent? Or maybe people have some sort of AI API where they can connect into or I guess maybe it's very tbd.

I think it's so tbd. This is the trillion dollar question, which is how does this ecosystem of AI agents come together? I mean, do they all live inside a superhuman? I think from the perspective of Pane of Glass, yeah. One of the strategic advantages we have as Superhuman, and I think this is why it was such an easy investment decision for IVP for A6 and Z for first round. And everyone else is. It's one of the only three hour a day products. Right. 50% or more of our users are in superhuman for three hours a day because it's email. What else does that? WhatsApp, iMessage, email, Slack. It's basically chat, messaging and email that our three hour a day products, nothing else even comes close. And so. Well, I guess maybe operating systems, but apart from that, you've got this Pane of Glass opportunity to be where people do work. Now, do we have to own the agents? I don't think so. And it probably isn't even good strategy for us to try and own every single thing that you might do. So I think the way that this probably proceeds is every company builds an agent that makes sense within their domain. So you could imagine an email agent that triages on your behalf, that drafts emails on your behalf, that schedules on your behalf, that maybe even sends emails on your behalf and is interacting with all the other agents and you're interacting with that agent ecosystem in a sidebar inside of Superhuman or maybe even one day with your voice.

I love, I love that vision. And we'll know to send the emails at 8:54 on Monday to beat the beat the 855 rush.

Well, I want my sales start messing with Jacob's agents and then, you know, send them on wild goose chases to do crazy things.

And yeah, it's like your, your open air bill from last month was massive because Pat's agent had you Checking all sorts of things. You know, I think another, again, like I said, I just think about how relevant the lessons are from Superhuman to a lot of these AI startups we meet. You know, another thing is everyone says, oh, the incumbents have such an advantage, you know, in AI, right? They have, you know, they're where you already are, they have all your existing data, you know, and I imagine this was, you know, and it still is something you probably deal with with like people moving over from Gmail or other email clients where, I mean, obviously you can port your emails, but as I understand it, you know, you got filters, you have other things set up and it's like there's some. It would have been easy at the time to be like, oh, that's, that's a space where incumbents are too advantaged. And so I'm sure a lot of our listeners are hearing from folks, yeah, the incumbents are so well positioned. I mean, any thoughts on, on that general statement of like incumbent positioning in AI and then any lessons you've learned or would think about for these companies on how to take on the incumbents?

I think it is generally more true than in previous cycles that incumbents have a greater advantage, but I don't think it's such a great advantage that you shouldn't try. They're still incumbents, right? They're still encumbered. They move really, really slowly and be limited by the fact that they have to build one size fits all solutions. I was on a podcast recently, I found a podcast and the host said, you know, one of the craziest things about Superhuman is you're one of the few companies in recent memory that has specifically gone after a space where really the only competition is an incumbent or two incumbents, namely Outlook and Gmail. Do you have advice for founders on. It's scary. Do you have advice for founders on that whole strategy? And I said, as an investor, I specifically look for that. I specifically look for founders going after incumbents because in many ways it's a really great strategy. First of all, incumbents have to do one size fits all solutions. So Microsoft outlook or Office 365 more generally has something like 400 million plus paid seats. Gmail has a north of a billion users. Between the two of them, that's basically everyone who does email. And by definition that's a one size fits all solution. It's really easy, relatively speaking, for a founder to go into that and go, okay, what is a underserved yet economically powerful subsegment of this market? That I could simply build a better product for. And at that point you've turned a problem from something you don't have control over into something you do have control over, which is build a better product. I'm basically saying get good, but you can actually get good and build a better product as opposed to somehow figure out some unfair distribution advantage. So that is very doable. The other thing is, and this ties in directly to the previous tactic, incumbents specifically struggle with with things like product speed. So you'll recall that Superhuman launched on the platform of speed, the fastest email experience of all time. We say that you get through your remarks twice as fast, you reply to emails once, two days sooner that you say four hours or more every single week. And the features that were the foundation for that at the very beginning was simply how fast the app was. Instantaneous Response Instantaneous search Sub 100 milliseconds User interactions It is very, very hard for an incumbent to make that change. Superhuman looks like a web app if you use our web version, but it's actually not. It was the first email client to be entirely written in JavaScript. Now imagine how difficult that would be for an incumbent to make that switch after 20 years of being a client server application. The third thing is design, right? Like familiar with Conway's Law, where the structure of a product actually tells you more about the structure of the organization. You're looking at how the. Org is structured, then whatever is logical for the user. We had a funny moment a few days ago in Superhuman where we were in Gmail and we were looking down the left hand side of Google Workplace and I think it's, if I remember correctly, it's mail and then chat or spaces and then there's one other icon, I don't remember the icon. And I was like, why isn't Calendar there? That's a really weird choice. And the PM I was working with was like, oh well, Calendar's up on the top right hand side where you have like the little grid of nine squares and you click that and there's a whole bunch of other things there.

I honestly don't even know. I'm usually in Superhuman.

I feel like you were trying to bait us in that memory.

The only correct place to be. And I was like, holy crap, this is just the organization of Google. There's a products manager who runs this, there's a products manager who runs that. This is not in the benefit of the user in any way, or at least I don't think it is. And it's understandable. Look, people have careers, they come and do these tours of duty where they run a product for two or three years and it gets handed on. And at least historically in Google, the big criticism has been you just get rewarded for launching some stuff. No matter how well or badly it does, launching gets rewarded. So people keep on launching things. This is why we've had six or seven different versions of Google Meet Hangout. I don't know, whatever they've been called. That's the other thing that incumbents really struggle with. So there's actually so many reasons to go up against an incumbent.

Yeah, it's super interesting. I feel like for a lot of these AI companies, the question is like, do you go, do you take off a vertical from the incumbent or go horizontal? And I think we had Arvind from perplexity on and I think he was saying this tension early on of some folks were like, well, just build the best search for X thing. But he's obviously tried to build a broad product there to compete. And I think that's, that's something a lot of folks are thinking about right now. Do you start with one specific user archetype or try and just go for glory?

Well, I've been a very big fan of the specific user archetype. Right. And initially it was founders and VCs, and then it became leaders and managers. And now it's really what at LinkedIn we had internal lingo for this. It's really any outbound professional, right? You could be a founder, you could be a VC manager, but you could be a salesperson, you could be a recruiter, you could be a BD person, a consultant, a community evangelist. There are so, so many different people that now use Superhuman in every industry, in every vertical, in every possible way. So I think there is a path from one Persona to multiple. But I had a very clear sense of what the progression would be. And as an investor, I think one of the key questions to ask if this is what the founder is pitching is can you just articulate the series of Personas that it's going to be?

Yeah, it's interesting too, because I think VC funds and even people within VC funds have different biases on. They like to invest in companies where there's a big incumbent or there's not a big incumbent or those different things. And I think we always love markets where there is a big incumbent because it kind of shows that the market is enormous. And then even by building a product for the superheroes of whatever that market is, or for the even a small segment of that, then you can build a really big product. And I've never thought of speed as really a vector to compete with some of those folks on of. You know, if you're competing against the Microsoft or Googles of the world, then it is hard for them to make their app faster because they've been building it over, I don't know, 20 years or that sort of thing. And yeah, I think it's really interesting.

Yeah. So two thoughts on that. First, the market size thing, it's interesting because as an investor, that's usually the main thing that would cause Todd and I not to make an investment. It's often not the founder isn't good enough, the product isn't good enough, or it doesn't solve a problem. It's like, oh, the market here isn't good enough. And I think about all of my experience pitching Superhuman. And it was so different. No one ever questioned the number of people that do email. And it's funny because in many ways the pitches could be boiled down to a few really simple facts, which is there's a billion professionals in the world on average we spend three hours a day reading and writing email. That's 3 billion hours every single day. That's north of a trillion hours every single year. It is basically impossible to find something that people spend more time on. The only other thing people spend an equivalent amount of time on is team chat. Right. Slack Microsoft Teams. And there's great businesses to be built in that area as well. And if people ever wanted a bottoms up analysis, I'd be like, okay, how many subscribers do you need at $30 a month to be a hundred million dollar run rent company? I assume this was Arvind's pitch as well. And you do the math and it's like, okay, you need 300,000. That's it. There are a billion people who do email professionally. And I actually remember being in the A16Z. This was series B pitch for 2019 in the boardroom and I asked the GPS, I was like, does anyone not believe that we'll get to 300,000 subscribers at some point? And everyone's like, yeah, you'll get to 300,000 subscribers. Like this is okay. So it's like, okay, so we've agreed this is at least a several billion dollar company. Let's now talk about $10 billion plus and how we get there. So that's my general thought on markets.

We always like to end our interviews with a quick fire round where we get your thoughts on probably overly broad questions. That shouldn't be in quickfire, but they're standard so we keep them there anyway, I guess. First, what's overhyped and underhyped in AI today?

So my overhyped and underhyped is actually going to be the same thing for both, which is this idea that AI is coming for our jobs. So I think it's underhyped in the sense that the vast majority of people, and by people I just mean like people in general. I think about my family who don't work in tech don't yet see how quickly it's going to happen for certain specific roles. So I'm thinking about, for example, entry level customer support roles or maybe entry level sales roles where I think Jason Lemkin said this best. It doesn't have to be as good as a salesperson. It has to be only 80% as good. Because if it's 10% of the cost or less, that's a trade off that many companies will happily, happily make. And if anything, it's going to be a better consumer experience. Right. I think we all know how terrible today's robotic voice customer support things are. When those are good, companies are going to adopt them en masse. So that is underhyped right now, I think what is overhyped? Sorry, that's overhyped. What's underhyped is exactly the same thing, which is there is nothing unique about customer service and sales roles except for perhaps by headcount, there's a lot of them. But it's coming for my job and it's coming for your jobs and it's coming for our friends, his job as well, and that's fine. At some point AI will be able to do most of my CEO job better than I do it. And yeah, bring it on. I think if you were to go back a thousand years when we were subsistence farming or we were hunter gatherers or Maybe even like 1500 years, those people would look at our jobs and go, that's not work. You're just playing like you're just having fun most days.

I think a lot of people today would look at Patton my job and say, that's not work. But it's kind of you to go back a thousand years, right?

Yeah, well, you know, make it easy for everyone. But I think that will happen again and I think it will be a lot faster than a thousand years this time. We'll probably look 50 years into the future and see people primarily doing what we perceive to be as leisure. And I think Bertrand Russell said this really well. A while ago. To us it will look like leisure. But you know what the funny thing is? To them it won't feel like it. Because humans are just hardwired to work and to get stressed about stuff and to play silly games of status with each other and to be competitive and to try and amass power and wealth and this, that the other, but the actual stuff that they're doing is going to seem silly. That's super interesting.

Super interesting. Besides Superhuman, what other company do you think would be most interesting to be building generative AI features at?

I go back to this idea that Superhuman is one of the few products in the world where it's a three hour a day product. So what are the products like that? It's basically chat and messaging and I've built my career on doing things which are valuable in the workplace and for professionals because it's really fun. And I'm obsessed with productivity and helping unlock human potential. And so it would probably be Slack or Microsoft Teams and I actually wouldn't start with AI. I have a very clear vision for how I would build either of those products way better than they are today. I would do that and it would be AI enabled, but that wouldn't be the point of the product.

Very cool.

What's the top thing you do?

Oh gosh. Okay. Q. Slack Rant. You guys use Slack?

Slack and Zoom Human. Yeah.

Cool. Okay, so there are some things that email does shockingly well and this is actually the main reason why it has pervaded for so long. In email, conversations aren't ordered like they are in Slack by when people happened to start them. They're ordered by when people last replied. In email you have an at a glance view of what a conversation is about. We call them subject lines. In Slack you don't have those things unless you manually write them in. In email you can mark things as done. You can snooze things. Soon, in next generation email clients like Superhuman, you'll be able to assign those threads and share those threads and leave comments and so on. In Slack you can't do any of those things. And so if I were to show you my Slack this morning, which I haven't looked at yet, and I'm dreading the thought. Just as I say would be without exaggeration, probably 100 to 150 unread channels in each of which are multiple conversations. Some people have thread disciplines, some people don't. Some conversations meander in and out of threads. They're all inter jumbled with each other. I wouldn't be able to tell where to start or where to finish. I'd probably look at it and go, fuck that, I'm going to go do something else. Which is actually how I structure my day. I open it, I close it, then I look at my calendar. I'm like, calendar, what should I be doing? Review this document? Yes. This feels like a much better use of my time. Which is why when people want to get work on my plate, they're like, put it on Rahul's calendar. Because that's the only way it will happen. Slack is just so overwhelming now. There are some problems with email, right? Email is immutable. You can't change a subject line once you've set it. Messages are immutable. You can't change it once you've set it. It's fully decentralized, which is an insane architecture. Everyone stores every copy of every email, which, oh boy, Building shared threads and comments on top of that is hard. It's very complicated. But we figured it out. It's going to be amazing. So what I would do is I would just combine the best of both worlds. I would build a workplace experience of the future that is designed for thoughtful, asynchronous discussion where you can. Where you can have subject lines, where you can archive things, where you can snooze things, where you can assign things, where you have an inbox where you can hit inbox. Zero chat is also there as well. But you have the choice of modality and it doesn't become a free for all. And you can actually, you can legitimately run a company and there's a reason again why email hasn't died. And it's because Slack didn't capture, and neither has teams for that matter. These products didn't capture the thing that email does so well.

The whole time you're talking about your Slack, I just pictured you need a Slack agent, something that can just do this all for you.

Yeah, I mean it does beg the question, I need a Slack agent. I think there's like a future of work agents, right? And I don't know whether it's a Slack agent or not, but there is an agent that maybe somehow knows what's important to you and what's going on in your Slack and what's going on in your email and where the free spots are in your calendar and just tries to organize the stuff that you should do next. One of the biggest life changing moments for me as a founder has been when I switched from having anxiety over what to do next to simply Saying, I'm not going to worry about that. I'm going to worry about whatever my calendar tells me to do next. And I'm going to tell my executive assistant, who is phenomenal at what she does, and she tracks all the priorities in the company at all time, what kinds of things I want to work on. And she has a knowledge of my preferences, what I like to do on a plane, what I like to do in the mornings, what I like to do in the evenings, and in greater detail than that, too. And it's been amazing. I don't have any anxiety or stress. Yes, I have an infinite amount of work. Yes, I'm always behind everything. That's okay because I'm still making consistent progress and I still feel good about the number of hours I work. I'm like, okay, I worked a solid day. Yes, I could work more, but sleep's also really important. And I'm eight years in. There's another 10 years here at least. Like, let's keep going, right? Like, this is a marathon. And I think having a healthy attitude to work is actually more important than anything else, which is ultimately the vision of Superhuman. The mission, I should say it's not about email. Our mission is to help professionals end every day feeling happier, more productive, and closer to achieving their potential. So that's what we spend all day thinking about.

I love that. And I feel like that's the perfect note to wrap on. I mean, I guess we'd just love to leave the final word to you. Where can people go to learn more about you about Superhuman? This is your chance to plug anything you would like for our listeners.

Okay? Okay. So here come the plugs. If you want more of my crazy ramblings, I'm on Twitter. Just my first name, last name, R A H U L V O H R A. Superhuman is just Superhuman as well. If you don't use Superhuman and you're a founder, I mean, what are you doing? Come on, let's be real. If you're a VC and you don't use Superhuman, this actually, you're just bad at your job. What's happening? So go to superhuman.com sign up our customers legitimately, get through their inbox twice as fast as before. They reply one to two business days sooner. They save four hours or more every single week. And we have team versions of Superhuman now. We talked about team read statuses, team reply indicators, team snippets. Soon you'll be able to share threads and comment on them. In addition to all of the amazing AI stuff that We've just built. I think that's it.

Amazing. Well, thanks. This has been an amazing conversation. Really, really appreciate it.

Thank you.

Well, Pat, I know we've both been big users for a while, but it was fun to meet the man who saved us a few hours or two in person.

Really, really cool to hear how they think about building AI products. And they've already shipped so many features, so it was, it was a lot of fun.

Yeah, I mean I loved his, his point that like the average, you know, data set that folks train email features on is 500 or 600,000 emails and he's got 4 billion. That combined with the fact that he's one of the few applications that folks spend, you know, three hours a day on. What a cool seat to have to build some of these features.

You know, how he was talking about how they designed the AI features and the distinction between the always on AI features and the kind of user activated AI features and how they, they started with the user activated, then moved to always on. But it seems like in hindsight he probably would have done the, the opposite and started with the always on because those are, those bring people in and they're cool.

You don't need to teach users how to use them. Yeah, it's interesting. I would have thought that would have been, you know, from a cost perspective. It's funny, I remember when we were talking with the intercom, you know, co founder Des, he was like, look, we'd love to summarize all the conversations all the times, but that, that adds up pretty fast. But it seems like the, you know, obviously with the price point they have, they've been able to, you know, take every email and still run those.

I liked when he talked about the thirst trap graph that he posted.

I'm glad he told us what it was. I was wondering if.

Yeah, he, he does seem like, he does seem good at building hype for his, his company and you could even see he's a really good marketer even in how he talks about, you know, what they're building in the future and, and all of that. So it was, it was funny that he, I don't know, self aware enough to call it a thirst draft draft.

Well, I love that also, I mean, I think another really interesting thing he said was he was talking about how like all the AI features they build are like 10 times more relevant for mobile. Which makes, I mean, mobile is like you're kind of half paying attention, you're on the go and it makes sense that that's, you know, those features are even more valuable. You know, I find often, even with email, it's like I'm like, well, do I try to do this on my phone? Okay, I'll just wait till I get home and like, have a little more, you know, focus time. And it's amazing. It's to think about the context in which some of these features are most valuable.

Totally. There's that meme that's like, I'm a millennial. I can't write any important email from my phone. And that really resonates with me. If it's important, I'm getting in front of a computer. But if AI is writing it, then maybe I'll write it at, from, from my phone.

For the end, we got a little philosophical there. But I loved his, his, his quoting of Bertrand Russell that like, no matter how much progress we make, we'll still find a way to work a lot and, and, and you know, play status games and you know, that, yeah, that, that I thought the point was really well taken that, you know, a thousand years ago folks would think what we do now is, is not work. And I'm sure we'll think the same of, of whatever comes post AGI.

Yeah, we'll, we'll find a way to stress our. The. I thought it was really interesting how much he's thinking about agents and building agents into Superhuman. And that seemed like something that was really on their horizon for, for how they think about the world is orchestrating these agents from Superhuman and interfacing to different products via the agent that you launched from Superhuman. And I thought that was really interesting. I think it's really TBD on who runs the agents and how they authenticate and how they talk to one another. And I think email is a really good place to start for a lot of these workflows. And I thought that was, that was fascinating to hear them talk about.

I think there's a real battle brewing for, you know, what's the homepage for all your agents? Like what and who has context on you and memory of the things you've done. And I think, you know, it's the three hour day apps that are probably in the best position to do that. And I think it's kind of still tbd. Like are they going to need some third party to help with a lot of these agent interactions because they're so hard to build and they need to be, you know, tailored perfectly. Like, we still haven't seen that many agent demos that work, so maybe each individual one will be really hard to Build or do we get to a point where you actually do have more of a horizontal agent, which was always the dream during the agent hype cycle of six, nine months ago that they can kind of just plug into Superhuman. So it'll be interesting to see how it plays out. But yeah, I agree with you. Like clearly he's, he's thinking, you know, a few steps ahead and, and I feel like we've been, we've been through the full agent hype cycle and now we're, we're back to the, to the real. I forget what they call the end of the hype cycle, but wherever it.

Is, trough of dissolution.

We were past that.

I feel like now we're on the something to enlightenment. Oh, that sounds right, that sounds, I don't know.

Enlightenment sounds right.

The trough of disillusionment that I think.

I feel like we only remember the trough of disillusionment, which is pretty negative on our part. But there's definitely a good part after that.

Yeah, good part. Coming soon.

You know, obviously the Mistral partnership with Microsoft was really exciting and I think, you know, it really does feel like it's coalescing around the set of, you know, six or so model providers that are going to be.

What about the Mistral partnership with Snowflake?

We had to get the Snowflake mention as well. Yeah, of course, two Redpoint portfolio companies working together. But yes, I think, you know, it really does feel like it's cementing, you know, who the folks are. I feel like in the last, you know, few weeks, you know, there's been a bunch of buzzy fundraises. There's a new robotics company that came out that thrive and a bunch of other funds funded. And you know, I just think the application of foundation models and scaling laws to spaces like robotics and biology is going to be super interesting to see. I mean, I think the big question in those spaces is you don't have the entire Internet to train on. So what is the, what's the data source in biology? A lot of these folks have to do wet labs and actually create their own data sets. But if people can find ways to, you know, to create large data sets in those spaces, would be incredible to see the emerging capabilities that happen from those models.

Yeah, I haven't looked at any. Anywhere near the number of bio or health care companies that you have. Are they full stack where they're building their own lab to get a lot of them?

A lot of them have to because there's just not enough existing data to throw into like a large model. And so there's different approaches, and some are, you know, doing simulations, and there's lots, you know, I think there's still an ongoing debate about how much you can do outside of actual, what they call wet lab data. But a lot of them, you know, a lot of the big ones do have their own labs.

Yeah. Wow. Fascinating.

Well, Pat, that was a fun one. And I think we've got a we've got a great lineup coming up. We've got Arthur from Astral, Ito from Pinecone, some mystery guests to be revealed later. But it really, it's a lot of fun.

As always, it was a really fun episode. I hope you enjoyed it. If you found it helpful or interesting, please, like comment, Share, subscribe. It would really help the channel. Until next time.

Sam.