Listen to this post:
Good morning,
This week’s Stratechery Interview is with Asana co-founder and Chairman Dustin Moskovitz. Moskovitz was a co-founder of Facebook with his Harvard roommate, Mark Zuckerberg, and was the company’s first CTO and Vice President of Engineering. Moskovitz left Facebook in 2008 to found Asana, a project management SaaS product that helped pioneer not only the SaaS business model, but also several web app technologies and product-led marketing. Asana went public in September 2020, and Moskovitz stepped down as CEO earlier this year.
In this interview, we talk about Moskovitz’ journey, including his time at Facebook, founding Asana, and whether or not SaaS companies will be disrupted by AI. As part of this, we discuss Moskovitz’s argument for a more modest approach to AI in the enterprise, and whether go-to-market and pricing models need to change. We end the interview by discussing why Moskovitz stepped down as CEO, including his focus on philanthropy and AI safety, and get to the edges of a debate about the best approaches to AI safety, and how that depends on your time horizon for AI take-off.
As an aside, I noted earlier that interviews would be more flexible in terms of timing now that I am in America; in this case, this interview was recorded over a week ago, on October 10; Moskovitz was gracious enough to allow me to delay this release so I could interview Dr. Gracelin Baskaran about rare earths last week. Expect normal Updates Tuesday through Thursday of this week (and an early interview next week as well).
As a reminder, all Stratechery content, including interviews, is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.
On to the Interview:
An Interview with Asana Founder Dustin Moskovitz about AI, SaaS, and Safety
This interview is lightly edited for content and clarity.
Topics:
Facebook | Starting Asana | SaaS and AI | AI Modesty | AI GTM | Focusing on AI Safety
Dustin Moskovitz, welcome to Stratechery.
Dustin Moskovitz: Yeah, thanks for having me, Ben.
So tell me about your background, which is not very long because I want background pre-Facebook, which obviously, famously, started in college. Where’d you grow up? Is it just total chance you ended up being roommates with one Mark Zuckerberg or what’s the story there?
DM: Yeah, I’ve got to go back 20 years now. Grew up in Central Florida in a pretty sleepy place in the middle of the state, and ended up going up to Harvard, and didn’t really have a lot of friends there. I don’t know if you know about Harvard, everyone’s just in the freshman dorms the first year, and then you group with six to eight other people, and you join a house together, and you live together for three years and I ended up joining a group with my girlfriend at the time that happened to include Mark. So I didn’t even know him, but through a hinge, ended up becoming roommates, basically. That was sophomore year, and we started Facebook in December, January of that year. Came out to Palo Alto that summer, and I’ve been in the Bay Area ever since.
Yeah, here you are. You were studying economics at the time, correct?
DM: Yeah, it was early in my college career, so I actually only took two econ courses, but that was my major at the time.
So did you have any experience with computers or technology? Or it was just suddenly, like, “This is an amazing thing, I’m jumping all in”?
DM: I was not a programmer, but I’d done a lot of web dev. I was like one of those kids that was using Dreamweaver. I don’t know if you’re familiar, it’s GUI-based.
Yeah, very much. I was sort of a half-generation before you, but everyone that I talked to, there’s a huge number that’s like, “I’m going to start a small business making web pages for businesses”, or something along those lines so I can relate to that category.
DM: That was a great job to have as a high schooler back then. You tended to have more skills than the adults and so I would do a little bit of contract web work. So that gave me a bit more — I sort of knew what it was, I knew what web hosting was and how to do a little DevOps, and then I was an Amazon referrer. I don’t know if you remember the referral program?
Yep.
DM: So I would review products, and put up a website, and earn a commission. It was quite lucrative in the 90s. I’d get like thousand dollar checks.
I was in college the Dot Com era, so I mean there’s all these schemes you could do. Number one, if you had your own domain, you could have infinite PayPal accounts, and then there was another virtual credit card, all which gave you cash bonuses, and then there was Value America or something like that, where you could buy computer parts, and you could have this circular scheme to buy computer parts and you could build computers for other people and sell them. That’s what I did my freshman year.
DM: That’s not bad!
Yeah. What were the early dynamics here? So you move in, your sophomore year, into this house, is it like, “Hey, I could use some help with the front page web dev”? Was this a thing where you were immediately intrigued or you got roped in? I’m just curious, what’s the dynamics in this house as this idea is forming?
DM: Yeah, path dependence is one of my favorite topics.
Right. That’s why I’m asking about it. It’s so fascinating.
DM: Yeah, very unintentional. I think the first thing I did to help was set up the email server, which was necessary to send out a confirmation email link.
Yep.
DM: I wasn’t like, “Oh, I work on this team”, at that point I was just, “Oh, I’m doing this thing for you”, then there was another thing. Mark was doing it himself for the first two weeks and then I really got much more heavily involved when we started deciding to go multi-school so I abstracted the code base to be able to do that. That involved a lot of DevOps, because they were almost like different applications at the time, isolated databases and stuff.
And you were certifying students by their email addresses, right?
DM: Yeah, that was how we did the access control. So if you had the college email, you were able to see the other people with college email.
As you zoom out 20 years on, what’s the biggest lesson you recall and learned from those early years?
DM: It’s hard to say what is totally applicable. I think a lot of what made us succeed was not knowing enough to doubt ourselves and just feeling like, “If you put out a good service, you could win”, but a lot of people think of Facebook as the first and the ever-dominant social network, but at the time there was MySpace and Friendster and they were making some critical mistakes, from my perspective, that were totally fixable but that left them vulnerable. I think seeing companies as genuinely vulnerable, even if you think they’re powerful, if they have some sort of systemic deficiencies, is a real opportunity. And if you move fast, then you can quickly get an edge.
Starting Asana
Well, there’s the path dependency lesson too, I think that’s probably a big one. You left Facebook in 2008 to start Asana — it’s actually, when I was preparing for this interview, it was kind of shocking to me that it’s been that long, made me feel very old.
DM: Totally.
Have you ever along the way in the 17 years had any regrets about leaving and doing this? Or it’s exactly what you’re supposed to do?
DM: I don’t think I have any regrets about leaving. You know, it was an interesting thing, because I was like 22, 23 at the time.
Right.
DM: From my perspective, I was working in a big corporate environment with like 800 people, and that was very foreign to me, and I didn’t particularly like it. I was Head of Engineering up until 2007 and I was like, “I don’t want to be the big manager”, and started basically making DevTools, which became the progenitor of Asana and I think that was a better fit to me personality wise.
But I think it also turned out, I still didn’t want to be a manager of people. We started Asana with Justin [Rosenstein] as the CEO, and then sort of switched by necessity at some point, so that was path dependent as well. Sometimes you’ve got to go the flow.
Well, tell me about the impetus for building Asana. You talked about you were building these DevTools internal to Facebook, I think that connection’s very interesting.
DM: I always like to say that we were democratizing what was the secret sauce of a lot of tech companies at the time, which is they all had, the larger ones had an internal, what we now call collaborative work management system. The one at Google was built by my co-founder, Justin, it’s called Tasks and it later became like the little task mole inside Gmail, where there’s a related path there. Apple has one called Radar and Amazon had one called Simple Issue Tracker. But they were these proprietary systems and they were helping these companies organize and move fast but you couldn’t really go get one off the shelf.
Yeah, what were the competitors at the time?
DM: There was very little, it was Basecamp, Trac was an open source project that was very popular, and then Jira was just getting started. So by the time we were starting Asana, Jira was like a thing, but it wasn’t the huge company it is now back then.
Right, and they were on premises at that time, I don’t think they had gone to the cloud yet.
DM: Oh, yeah.
I’m trying to recall back, 2008, this is still pretty early in this sort of concept of a SaaS application.
DM: Right, but at Facebook we were pioneering a lot of what enabled that later. I was part of redesigning Facebook from, back then, you would click and you’d do a whole page load and this is part of what made Friendster slow is it would be like a 20-second page load and then over time Facebook evolved to what I think of as a thick client. So you download a bunch of JavaScript, a bunch of CSS, and then it changes as little as possible the UI.
So one of the really big versions of this was the photo albums. You could just click right arrow, and go straight through it, and it would just swap out the image source and be extremely fast and that gave us a really native sense of like, “This is how web apps could work”, and you could design something that is like a desktop experience in the browser, but you couldn’t have done it five years prior to that. The JavaScript engines weren’t there.
Was that the key secret sauce for Asana that made Asana different? Was it the way you organized tasks and things along those lines? Particularly at that time, you think about Asana as competitors today, but back then, Asana, and Jira, Basecamp, so the traditional on-prem software, what do you think was your core differentiator?
DM: Yeah, definitely versus those, it was experience. But the thing we were mostly competing with was email and meetings.
Right, nothing.
DM: Yeah, nothing. And there, experience matters even more. So we used to say, “The design goal is to be faster than Notepad”, because that’s what people tended to do, it’s just to keep a little task list for themselves, and there are ways we paid that off. There are ways you can manipulate the task lists that are faster than Notepad, and it’s completely non-blocking, even if you don’t have an Internet connection, you can create new tasks by pressing enter, so it feels very fast. Then we also had innovative data model between the prototypes to be what we call a Work Graph model, it’s a much more object-oriented view of work than a document or a project-oriented view.
Right.
DM: It means tasks can live in as many places as you need them to, or nowhere at all and that is also similar to Notepad, where you’re just writing things in this file that you might not even save, and then later they may become more substantive, and you give them formatting and share them with people. Asana’s meant to give you that spectrum of lightweight.
How much were you thinking about go-to-market back then as a 22-year-old? Asana’s become well-known for being a pioneer in product-led growth, starting low in the enterprise and working its way up. Was that intentional all along, or you just wanted to make this cool product, and that ended up being the way that it could distribute thanks in part to being web-based and all those aspects that we talked about?
DM: Yeah, I think at the time it was a little less thought out from the beginning. This was when Dropbox was starting and Gmail was very popular and you have a sense of you could build a big consumer product, and cross over into business, and that’s just kind of fine, that’s where we target. Then as the field got more crowded, we needed to differentiate a bit and carve out what we were going to focus on and the benefits of the graph-based structure pay off more in a larger organization.
SaaS and AI
What’s interesting about the Asana story is it is a great origin story. You recounted in a bit about building a better task manager internally, you had these big companies that had one, you built one internally at Facebook. It’s like, “This is something that can be externalized shown to other people, Facebook doesn’t want to do it, I’ll go out to a startup and do it”.
DM: Yep.
And there’s this opportunity that you talked about the web, you couldn’t have done it five years prior and it’s completely new, you can start from ground zero. I guess the question I have, that was nearly 20 years ago, as we’re talking about a shockingly long time ago, is there a part of you that worries, like OpenAI published that article a few weeks ago, tools we use internally at OpenAI, and I don’t think they talked about task management. But is there a bit where you think, “Hm I wonder if companies starting right now, creating task managers from scratch might actually approach things totally differently just like we did 20 years ago”?
DM: Yeah, without naming customers, I’ll just say two of the top two AI labs are some customers and they use our AI scaffold on top of their own models because it’s adding something to their ability to get leverage out of them.
There’s always a worry about Innovator’s Dilemma. The thing I worry a little bit more about is whether the data model premises and the things you invested in as a mode over the past 10 years are still the relevant ones for what’s coming.
What’s an example of one that you might be worried about?
DM: Well, I’ll just give you something sort of adjacent, which is slide decks. Right now a lot of people are trying to make AI generate keynotes and PowerPoint files and I think those pieces of software exist that humans can manipulate formatted slides, and that’s just tricky. You’ve got to align things and show them stuff and I think two years from now, humans won’t be doing it at all. Then why do we still have that format and service of this UI that’s meant to be used by humans?
It’s like Google Docs being stuck on a print layout for ages and ages, which I think they finally changed the default there, but yes.
DM: Yeah, and so the path dependence leaves you hobbled in a lot of cases and it might be better to just do a fresh start. I think we’ll see some SaaS apps do that, they’ll just be like, “Here’s a totally new version of the product”, and then that’ll be weird, because that’s not the premise of amortized software returns.
But where I’m going with this is generally I think that the most important thing for SaaS is to adapt and treat this as something that justifies working differently and then I think there are just a lot of advantages for whoever could adapt well and move fast. So yeah, I don’t think any particular SaaS company is vulnerable unless they’re too stuck in like status quo bias.
Well, I guess that is a big question for everyone, for all the SaaS companies, it’s a big question for you. I’ll give you the opening statement, why is it that if investors are looking around and saying, “All these SaaS apps are vulnerable from AI, there’s going to be new approaches”, why would you say, “Well, I can’t speak for all SaaS but Asana, we’re actually well placed in going to win in this regard”. What’s the thesis there?
DM: Well, it depends on the timeline. The thing I can defend better is Asana’s well-positioned for the next few years to be part of this transition, and the reason I’d give for that is that what most enterprises try and do right now with AI is they buy chat licenses for their people and they’re like, “Here, go use these chat licenses”.
That’s successful for some of the people, but you end up with a lot of unused seats and shelf wear and what Asana gives you instead is basically a project workload that is automating the AI for you. A human doesn’t need to drive it every day because it’s already part of your workflow automation and that is where I expect most inference to come from over the next two years.
I always like to say, people are always like, “Oh, I’ve never used gen AI at all”, and I’m like, “Well, have you ever been to an Amazon product page?”, because I think that little review summary is the most prolific version of it or just the most visible, but nobody’s engaging in chatbot.
Give me an example of this, because I know you’ve written about the proper way to develop AI right now is to start with these simple workflows and expand from there. What’s an example of how AI is impacting the Asana product right now through your workers or wherever else it might be?
DM: We get a lot of customer feedback and also feedback from the sales team and our support team about things that’s changed in the roadmap and for many years now, we’ve done this process to try and stack rank, merge them and come up with the highest priority things to work on. Historically that has taken hundreds of hours to classify these things and interpret them and now we can just, as something comes in, automatically process it, keep the tallies and also create backlinks. So if a human does want to deep dive, “How did this thing become number three?”, they can go back and see all of the tickets that led up to it, and that’ll let us close the loop at the end too. When we actually ship that thing, we can go back and write to everyone who asked us about it and a lot of those things you would just wouldn’t bother doing. You would sample them or something instead to make it a manageable amount of work or you wouldn’t bother following up with them. But once you’ve set this up in Asana once it’s just running all the time and doing inference for you and you can go look at the result when you want.
Is this an example because it’s a great example of how AI is beneficial, or I could be cynical and say, “Well, it’s great he picked an example of AI doing something that wouldn’t be done otherwise?”, so there’s no sort of follow on questions about seat elimination or things along those lines.
DM: Well, again, the product manager part we would’ve done, and so this is a way to give you more leverage on your UXR [user experience research] and PM teams. We might not have done the follow through part and so I do think there’s going to be some displacement of work and that’s just what all the companies are doing. Currently we use third party products through AI support, and again, I think it has this property of that is useful because it is in this context of a workflow of, “Receive a ticket and respond to it”, and the system has more situational awareness and can narrow the distribution drifts and all sorts of things that can make that more successful.
The larger lesson I think is just that the structure makes you more successful, there’s less ways for the user to cut themselves and they’re not figuring out what’s possible. It’s just somebody else has sort of written pseudo-software for you.
Is this pseudo-software, is this written by Asana or you’re enabling these companies internally to write it for their specific situations?
DM: All of the above and more. We have some out-of-the-box templates, that’s when you create a new project, you can just use some AI rules we wrote, then you can edit those, you can make them yourselves, you can work with a service partner. We also have an AI assistant bot so you can just chat to an AI and be like, “This is the workflow I want”, and we’ll build it for you, but we will build it out as we have a canvas that does step-by-step workflow creation you’ve probably seen before.
You mentioned the empty seat problem, people buy chat licenses for everyone, then how many are actually being used? I do think this is a real question as far as AI penetration in the enterprise, which is right now, particularly the chatbot interface, requires volition on the side of the user to actually go and use it, and one of my thesis about SaaS in general is SaaS wasn’t just a function of developing technology, it was also like a demographic change. You had a whole generation of workers coming up that were used to doing stuff in the browser or used to using apps on their computer and so it was just very natural to do that. I could see a similar story with AI where the younger generation comes up, of course they use AI for everything, to get the older generation to is more difficult. So one of my questions, thesis, theories, is some of the most valuable AI penetration is actually going to be pretty explicitly job eliminating, kind of like how mainframes eliminated bookkeeping back in the day, just because there’s no change management, it’s top down, this is going to do new things. Is that a theory that you think makes sense or would you push back on that?
DM: I don’t know that the only way to cause inference is behavior change. We haven’t done this yet, but one of the things I’ve really wanted us to do is when we do a performance review cycle, Asana should just run a dossier on every employee and give it to the reviewer. Right now what we do is we give them an AI tool where they’re like, “You can type in these commands and it’ll give you a dossier”, but if we just gave you that, you don’t have to use AI, but you were getting a benefit from it. We have a lot of built-ins in Asana too, project summaries and portfolio summaries where it’s like you just have to hit a toggle and we’re going to send it to you every week, you don’t have to go to the chat app. That stuff I think is viable. Also, Asana doesn’t charge by the seat in order to try and align value better with what you’re actually using.
How do you charge?
DM: That has its own problems too. So we basically have a fixed platform fee that gives you a certain amount of tokens and then you pay a la carte after that, so it’s meant to be consumption-based in practice. A lot of people anchor on what that first token allotment is, so we still have some friction there, but they’re not arguing with us about like, “Oh, only 20 people are getting value and the other 50 aren’t”, it’s just like, “This is how many tokens you use out of your budget”, so it ends up feeling more like a platform fee.
But the core Asana product is still on a per-seat basis, right?
DM: Yeah.
So how do you think about that tension of right now, sure, if you can do work that’s not done, that’s great, but actually if you’re really effective, some companies might not need as many people and that could a problem for you?
DM: I think that’s real. And I think [Asana CEO] Dan [Rogers] thinks that’s real too. One of the reasons, I don’t know if you knew, but we switched CEOs at Asana.
I was going to ask you about that, we just jumped right into the AI stuff, but yes.
DM: One of the reasons I wanted to hire him is he was thinking about this at LaunchDarkly and moving them more towards usage-based pricing models, so I do think that’s part of the future, but you’ll need to get there in steps. Have some SKUs that are not seat-based and that’ll displace revenue and you’ll do some swaps at renewal time and over time, I do hope to lower seat list price and maybe it even goes to zero because the other thing about this platform fee is we have people paying tens of thousands of dollars, but they only have 10 seats and so the platform fee is just totally swapping that seat revenue already. Whenever I saw those deals, I was just like, “Whatever, comp them on the seats, we just want them, this is the new business model for that particular customer”.
How do you ensure internally though that you’re not slow rolling that transition, waiting for the revenue to come up — if it actually is the right thing to do, you push that forward?
DM: That’s just the challenge of leadership, but it is a public company and it’s a tricky story to tell and the customers are only at a certain point, I think our products haven’t been in market that long. We’ve learned a lot, but the scale that will affect the revenue numbers is sort of like the next step I think will happen in the next year.
But when you start a new revenue line and it’s single digit millions, we’ve got a $700 million business, it’s just not going to show up and so you can’t go too fast. But we are thinking a lot about how are new customers coming in? Let’s set them up for success for the long run, and generally if we can get them using the paid AI features early in their life cycle, make it more part of the onboarding experience, that’ll help us move faster.
Is there a part of you that, especially with this AI transition, that wishes you were a private company?
DM: Yeah. I mean, public company in the past three years I think was just a nightmare in general.
Well, I was going to ask you about that. I mean, so you did the direct listing, you had this massive run up in 2021 along with all the other SaaS companies, and then I would imagine a very painful come down since then and been scrounging along ever since.
DM: Loss aversion is a big deal. If we had just been a private company, maybe it just would’ve felt like modest growth the whole time or something but having the expectation set so high and then every SaaS company is down from that I think is not been great in a lot of ways and was certainly not fun as a CEO just constantly on the earnings call being like, “Well, there’s a big new tariff push and a war and all these things”.
What did you think of the dynamics of that? Obviously there was that COVID bubble stuff that happened and that got punctured. Does it feel like the AI narrative has made it impossible to recover from that? That came on right afterwards and now everyone has this doubt about SaaS and you’re just sort of stuck.
DM: The timing is not great, but also it’s just a very uncertain environment and uncertainty is bad for business. The customers on the other end are uncertain about their businesses and they’re uncertain about where the software will go and which AI platform they want to adopt and so I think it’s a tough time to be a SaaS company on the public market, but it would’ve been tough in different ways to be private. But I think our timing was particularly bad and when COVID happened, we should have just hit pause, but it felt like, “Oh, we’ve got to follow through on this thing at that point”.
So where do you think in the long run as this shift in AI happens, do you see this as an opportunity to build out your business with new customers or is this a matter of expanding your business internally with the customers you already have?
DM: Well, one and then the other, I think as we become known more as a good AI platform, which should be good for tracking business, and also the best thing that could possibly happen is the market learns that Asana is the best way to adopt AI because of these automated workflows and the fact that you don’t need to retrain your entire employee base.
AI Modesty
Tell me about the thesis — you wrote this memo for Kleiner Perkins basically making the case that a lot of people are trying to get AI to do everything, and I don’t think that’s the right approach. What in your estimation is the right approach? Give me that big picture overview, we’ve been touching on it on the sides, but I know you’ve written this down very explicitly.
DM: I think most of people who are tracking AI are waiting for this moment where it’s fully autonomous either for a particular role or for all roles at once, it’s how they define ASI [artificial superintelligence], but there’s this intermediate step where it can do 98% of the work for you and maybe be confident that it correctly did 90% of that, and you could give the human a lot of leverage if you could focus them on the correct 2%-to-10% of what’s remaining and a lot of roles don’t lead to be defined that way.
So when talking about customer support, I’m always like, “I think neither the Sierra model or the cost-per-ticket model make any sense because I actually have a very wide variety of tickets”, I have free users that are signing up and kicking the tires and they’re going to plan a wedding, they’re not enterprise companies at all. And I’ve got the CFO of a Fortune 500 company writing in, and they’re both using the same email queue and I want to route that second one to a human, or I want to pay $1,000 for an AI to really think about it, but I don’t want to pay the flat $1 price to handle both those situations.
So that’s an argument for usage over outcome-based.
DM: Well, it’s an argument for narrowing the scope of the problem you’re trying to have the AI address. Lower stakes tickets, maybe they’re much more common type of stuff, and FAQ might be answering, but the user didn’t look in FAQ and then really monitoring like CSATs [customer satisfaction scores] on those tickets and monitoring false positives where the system did something it wasn’t supposed to, and only giving it more rope as you get confidence. Whereas what people do now is they’re just like, “We’re going to fire this customer support rep and throw in a bot”, I’m like, “Oh, shoot, it’s only working 60% of the time, now we got to go invest more”, and I think just a lot of people are cutting themselves trying to go straight to that 100% outcome.
So I use Claude Code a lot, and I think the magic of Claude Code is actually how the agent talks to the user. It does a lot, it can go for two hours of human time, and then it gets stuck at some point, it needs you to do some DevOps stuff or it didn’t correctly identify how to solve a bug and the UI, literally it color codes the text that the user needs to reach. The problems I’ve made the most mistakes where I’m moving too fast and not actually paying attention to what it’s trying to tell me but when we work together really well, we go farther together because I’m able to solve the stuff it can’t and it can do a lot of shit that I don’t know how to do it all, and that is just different to me than trying to have it fully do a pull request from start to finish and then I just look on whether it succeeded at the end.
I think what you’re saying makes a lot of sense, and it fits with my understanding of where things are and I think you appropriately talked about the fact that you feel good about the next couple of years. But insinuated in that is in the long term, “Yeah, maybe AI can do it all”. I do have to ask though, if I was talking to Dustin Moskovitz in the late 2000s who is praising the virtues of just moving fast and other people are stuck in the past paradigm, how would he respond to what you just said?
DM: Wait, how does that relate to moving fast?
Well, I think the alternative argument would be just YOLO it, and, “Yeah, we’re going to put that customer service bot in there, it’s going to screw up 10%, but we’re going to get a lot of iterations, it’s going to get better, it’s going to be painful, but we’re going to get to a truly functional version that much more quickly than trying to be very careful and conservative about it”.
DM: I see, yeah. Well, first of all, it’s different in enterprise software, the stakes are higher.
For sure. That’s a tough thing.
DM: And the stakes of what these things do are higher. There are those reports of giving away a refund on a full truck or something, it’s a costly way to break things and generally I think the technologies are quite dangerous.
When I look at the ecosystem, the thing that actually has me most spooked is all the demos work without access control, and basically the most powerful systems, the ones that look like they’re doing the coolest things, have the least controls just from a normal data security perspective.
Is that intrinsic to why they’re powerful?
DM: Yeah, it’s just the easiest way to make a demo is not have permissions and not have it stop. But you also have things like prompt injection out there, there’s a lot of novel security risks and so I think one of the arguments we make in AI safety in general is you should move a little slower in order to move fast because if you make really big mistakes, then you have to really stop, or you have a huge brand damage issue and so I think it’s more like that.
I think in AI, inside companies, the way this manifests is disillusionment. You tried AI, you’re like, “It’s shit, AI is never going to work”, and then you learn the wrong lesson, you let the first cut scar. Whereas I think if you do what I was suggesting with support, you can still move really fast, you could displace 50% of your tickets, it’ll just be the easiest tickets, and you’ll build from success and you could still move fast, you can do daily iterations.
Somebody asked me at that Kleiner thing, they were like, “How often do you think we should look at QA for our support tickets? Should we do it by quarterly, monthly?”, and I was like, “What do you mean? Have a real-time dashboard, look at it right away, ook at it the next hour”. There’s plenty of information, do the tickets get resolved? And how many turns? You can know.
I sort of stole the promise of this question on the fly, but I did have written down here along the same lines, “You focused a lot on the importance of security and governance, can the Dustin Moskovitz of 2008 imagine caring so much about these enterprise realities?”, but there’s a bit where these enterprise realities, is that actually a bit of a moat for you?
DM: Well, I hope so. I think the risk is that customers YOLO it because that’s what they think they need to do to succeed, but not all of them will be like that. But yeah, that’s the pitch we make is this is an enterprise-grade AI platform, and it’s not even just that it’s secure, it also has the enterprise functionality. When you use ChatGPT Enterprise or Claude Enterprise, you can’t share a conversation, it’s like very basic collaboration things are not there. Whereas when you do AI in Asana, it’s all existing task threads and comments, you can add whoever you want, they can participate with the bot, you can have two bots in the conversation together. It’s just a more built-out platform that now has that stuff added in it.
AI GTM
We talked about the product-led growth, and you had a very compelling freemium model. I tried Asana ages ago, didn’t quite fit with the way I worked, but that was I think a “me” problem rather than a “you” problem, but I am curious how you think about your go-to-market with AI. Take freemium for example, if AI usage has these significant marginal costs, can you sustainably still have a compelling freemium market? Or if you’re trying to make the pitch on things like governance and access controls, is it much more of a top-down sales motion, you have to go in and get the CIO and the CTO on board and then go in, which is sort of an inverse of the way things operated for you at the beginning?
DM: Yeah, we’re trying to do both. We launched a self-serve version of Studio actually three weeks after I left, so I haven’t seen the data, but I think both stories make sense and I hope you can kind of close it from both ends because the CIO sale is pretty hard right now, partially because of this trough of disillusionment, they’ve been burned by the other vendors.
We’ve got to deal with all of that, all the over-hype, so that’s pretty complicated, but it seems like it was going to be easy at first, and so we doubled down on that. Then ran into some more friction and we’re like, “Well, let’s at least allow people to self-serve in the SMB land if they want”. We have talked a lot about, “How does this show up in freemium?”, so we do have a lot of AI features in the free product now, and we have a freemium version of Studio itself. If you have Asana Premium, you can use the workflows with rate limits basically and we debate a lot about how much are we willing to lose and all this, I think being-
Is this bit of a new muscle? Because usually you think about software and you just ignore the cost of serving. Yes, there’s some marginal cost of serving, but it’s very small, and now it’s front-and-center.
DM: Yeah, COGS are going to change, for sure, but I think we’re being too cautious. I think, again, people have learned a lot of lessons from the past 20 years to be scared of COGS, whereas I think spending $10 on every new user would definitely get ROI based on conversion, but then you also have other options, like you could not put the most powerful thinking models in, and then you’ll get more runway on a faster cheaper model.
Right, but how do you tell the user, “Actually this could be so much better if you just pay a little bit more”?
DM: Yeah, that’s the trick, you want them to have the best possible experience. The product we have coming out is a really perverse version of this, Teammates. It’s sort of having a deep research tool inside Asana, but the variance on how much we spend is from 10 cents to $100 or something, and then try and put a margin on top of that. I kind of think you need to reverse it.
I think what’s happening in practice with the models is the user now says, they hit the extended thinking button or they say, “Think hard” in Claude Code, so they’re telling the agent how much to spend, and that’s where I kind of suspect things will go. You’ll be like, “This is a one-point task, a two-point task”, or maybe the system will do a quick eval and check in with you, “Do you want me to write a page here or 10 pages?”.
Right. That’s a pretty big cognitive burden to put on the user though.
DM: It is, but it may be worth it if it’s the difference between spending $10 and $100 and frankly it’s time too. I’m having trouble with the new Sonnet 4.5, I’m like, “Hey, help me find a backpack”, and it’s like, “Here’s 25 pages on backpacks”, I was like, “No, I wanted you to do less”.
So, it is a cognitive burden, but you can have defaults and I think the power can be worth it and generally, I have a view that the rate limits distort AI just incredibly, and it could be way bigger than it is. My favorite example is Claude Code. I don’t know if you are on Reddit or you see the users complaining about the rate limits and trying to game them, it’s like their whole life is like, “How do I avoid paying $1 more on my subscription fee?”.
Talk about cognitive burdens!
DM: Yeah, talk about cognitive burdens. I just use my API key, I’m using the same software with them and I’m paying a la carte, and I never think about it at all, and I do 10 times as much. I’m fully willing to automate things and I have different willingness to pay, but I don’t think I’m being irrational. I’m not spending $100,000 — instead of $200, maybe I’m spending a $1,000 or $2,000, I think I’m getting well more than that back.
Isn’t that going to accentuate the need to do a top-down sales motion though? Because you can talk to a CIO or a CTO and convince them of the value such that they should have an expansive budget for their users because they’re actually getting so much more productivity. It’s a lot harder to, on the bottoms-up, have a guy saying, “Manager, can I have $5,000 worth of credits?”, or whatever it might be, “I swear this is replacing an employee”.
DM: What we’re trying to do is basically build you a resource management system for this so that you can sort of fractally delegate it. So, you’d buy $1 million at the CEO level and you’d give $100,000 each of your department heads, and they would use their judgment to distribute it, we already do this for time. We have time tracking and time sheets, and so this would just be your compute budget in addition to your time budget and then the managers about to express judgment about whether that’s a good use of budget.
Do you think this problem will be solved, like models will become good enough that you don’t need to be on the absolute bleeding edge, the trade-off for speed and cost will be worth it, that you can get back to the zero marginal cost world that SaaS came up in?
DM: I think that it’s sort of like a little column A little column B. The way Claude Code works right now is there’s a lead agent that is like Sonnet 4.5, and then it delegates to these agents to run simple tools, and it makes a judgment like, “Haiku can do this”, I’m just looking, I’m going to run date on the command line or something, and I think that’s a lot more where things are going. That’s certainly how Atlassian Rovo is built and are built in into Asana. For the planning agent, you want the best one, I think and I think that will likely be true for the foreseeable future.
So, just thinking about that question, I will say that is the line the labs used for a long time, I assume you know they were negative gross margin on inference for a long time, and I was like, “This is crazy, raise your prices, why would you be negative gross margin?”, and they would always say, “Well, the models will get cheaper in the future, and so we’ll just lock in these token prices”.
But then they did release better models and charged more for them and I think the market responded well to that. I think there’s a large number of people that just want to use the best and one of the places where you don’t see this subscription rate-limiting issue, by the way, is the financial firms. They just write custom software on the API, and they’re willing to spend quite large amounts of money because they can see the ROI. So that’s what they think, but I think empirically it’s not what’s happened and what they’ve done instead is introduce higher price subscription tiers, and then they still lose money on that because now they’re bifurcating their customer base into the power users.
I don’t see a future there, but I respect that this is what the users think they want, is real price certainty, but I think another thing that could happen is you might get to a world where you’re reserving a GPU and you might just pay for the GPU rental time, and the system is trying to use it as much as possible because a lot of the stuff I described for Asana, it’s not synchronous. We could do it in the off time, we could do it overnight, which also means we get the duck curve benefits on their infrastructure, they could charge us less.
There’s a lot of ways to make inference cheaper and just be more targeted about what you’re trying to do but ultimately I think more power is more power, and a lot of people are going to want the best. I can’t imagine a financial firm not using the most powerful model, for example.
Well, you mentioned there was an allergy to COGS. Do you think there was — not misplaced, but appropriate at the time — but now increasingly misplaced allergy to services, in that you need to go into firms and really help them establish these workflows? Particularly if you have the thesis that, look, AI works really well if you constrain it properly, so we need to go and actually help you do that.
DM: Well, in the very short run, I think that’s true, but again, we have templates, so we try and take what we learn, package it up, we have the AI bot to help you, we can program that with a bunch of best practices and examples. Shouldn’t require handholding every time, but right now just nobody knows what’s possible and doesn’t want to get burned, and so a lot of them want that for sure. But we often find in a customer, you’ll end up with some kind of maverick who’s just like, “Oh, I know how this works”.
The in-house evangelist that sets it up for everyone.
DM: Yeah, exactly, and they can just do so much. They can write a workflow by themselves that uses the whole platform budget if they’re innovative enough.
Focusing on AI Safety
I’ve asked you a lot of questions about Asana as if you’re CEO. As you mentioned and alluded to, you’re not CEO anymore, you took a step back to Chairman. Why that shift now? The stock been punished enough, it took another punishment when you stepped down, the market thought well of you. What was the impetus there, and how are you feeling six months on?
DM: Yeah, a lot of it’s pretty personal for me, I was there a long time.
Yeah, 2008, I still can’t believe it.
DM: Yeah, 2008.
I tell you, I was researching this, I’m like, “I remember when Asana came out, and I can’t believe it was that long ago”.
DM: Yeah. And I also mentioned by personality, I don’t like to manage teams, and it wasn’t my intention when we started Asana. I’d intended to be more of a independent or Head of Engineering or something again. Then one thing led to another and I was CEO for 13 years and I just found it quite exhausting.
I’m an introvert, I had to just kind of put on this face day after day and then in the beginning I was like, “Oh, it’s going to get easier, the company will get more mature”, and then the world just kept getting more chaotic — the first Trump presidency and the pandemic and all the race stuff, it made it just a lot less of the company building, being a CEO is a lot more reacting to problems and doing this sort of thing.
Then the other big thing that’s happened at the same time is just we’ve been working on our philanthropy since 2012, that has scaled quite a bit and has become more complex and then higher stakes in particularly timely ways. So we’re working a lot on biosecurity and pandemic risk and we work on AI safety and those things meet in a very dangerous world. I’m pretty much a short timelines person, so I think these problems are now and I want to be more focused on them and I was just finding it increasingly difficult to serve both of those as well as they should be.
Do you see a tension in, “On one hand my company needs to be super aggressive in adopting AI so we don’t get disrupted”, versus the, “I am a well-known and the biggest donator to AI safety causes”, and things along those lines? Do you see a tension between those two positions?
DM: Superficially, yes. But like we were talking about earlier, I actually think that the better way to go fast is to go slow, or at least slower than we are. I think these technologies are really dangerous, I think we’re going to have something pretty bad happen and I think that will cause a chilling effect and I think that will be bad for Asana and bad for the labs, and I think it’s unnecessary, and so I think that’s the best way to resolve the tension.
But the harder thing has been like I’m a board observer at Anthropic and we are early donors to OpenAI and there’s always this tension on if you’re creating capabilities, are you also creating dangerous capabilities? Even when you do direct safety research like interpretability, there is still that risk, it’s basically dual-use in bio. So that is super real tension, something we talk about a lot, I don’t feel super resolved on.
Everyone is certainly aggressive, investing very aggressively. The founders of Anthropic back when OpenAI released GPT-2 think was a real driving force in not releasing that publicly at the time, and in 2025, that seems fairly quaint given the development of just the models generally, open source models both in the US, in China and elsewhere. Has this sort of changed your position at all? Or it’s just, do you worry about there being sort of a boy that cried wolf effect that has diminished the concerns about this? Or, where’s your thinking evolved since say the 2020 timeline?
DM: Well, I think mostly the boy who cried wolf is a misunderstanding where I don’t think you’re going to find any researchers who thought the GPT-2 release was a danger for takeover reasons, they would say it’s a danger because it accelerated the race, it was the starting gun. So I don’t know that it’s quaint. I think it’s just like we are in this world. We might’ve been in a different world where AI was developed differently, but now we’re in this one, we deal with this reality.
I do think a lot of potential solutions I now see as quaint and naive. My history on this is I was dismissive, I was investing in AI since 2012, this company Vicarious, and we were starting to donate near the EA [effective altruism] community and this was a big deal for them. At the time I was like, “That’s sci-fi, it’s not going to turn out like this”, and had these stories like, “Oh, we’ll develop it in an air-gapped way, it’ll be secure, we’ll turn it off if there’s a problem”, and now I see how it’s actually developing. I’m like, “That is so dumb”. The very first thing we do is connect it to the Internet and like I said do the demo without any controls at all because that’s the best demo.
Then people literally do things like write ChaosGPT, where it’s like your goal is to destroy the world. Ten years ago we’d be like, “Well, we just would never give it that goal”, it’s like, “Well, other people will”. And then the buyer risk stuff I became super convinced on is these are tools—
It sounds like one of the lessons, like I asked you about lessons that you’ve learned over the years, is that the world is maybe a lot more complicated than it first seems and there’s a lot more competing incentives than is maybe obvious at first glance. Is that sort of a fair summary?
DM: Yeah, I think it’d be hard to grow up without coming to that point of view where it’s more complex, but I think a more specific version of that is I’ve really grown to appreciate Moloch and the underlying capitalist forces that shape a lot of this, and I’ve been a little surprised — I know you’re kind of a safety skeptic yourself, but I’m surprised by some of the rhetoric where I’m like, “It’s really not clear to me whether you believe this or if this is just like a tactic”.
Well, I wouldn’t say safety skeptic. I think that the approach of, and I’m fairly critical of EA specifically, I think doesn’t appreciate the complication and competing incentives of the world generally. And to that end, to your point, especially the early proposals around AI safety I didn’t consider feasible because of, to your point, the capitalist impulse and competition and all those sorts of things. And the idea, I find it much more plausible, maybe not desirable but plausible that we get to a world of competing AIs and the best safety measures to have the best AI and AIs will keep each other in check more so than we’re not going to have any AI at all, or there’s going to be God’s gift to regulators who has the wisdom to actually lay down regulation that functions well and works well.
I think a critique I would have is I think the previous administration bought a little bit too deeply into some of the AI safety rhetoric in a way that really messed up some of the long-term relationship with China, for example, and headed down a road of, “This is going to happen very quickly, so we should fire off a bullet that we can fire once” — well, what if it doesn’t happen quickly and we have to think about 10 years or 20 years or 30 years? So it’s not that I don’t acknowledge the risks, it’s that I think a lot of people trumpeting the risks, there’s a lot of people who are good at diagnosing problems and I can say that, but might not agree with their solutions, if that’s a fair characterization.
DM: Yeah, just the one thing I would point out though is people tend to view the EA as a monolith, they’re presenting these solutions together. But the reality is they actually just all disagree with each other along these same lines.
But if you give them millions, I mean you’re sort of enabling the whole lot of them.
DM: Well, but again, there is no EA org, this is like saying you’re giving Antifa money. We give particular orgs money to do particular things that we think are justified. We certainly were involved in the chip exports, so that may be a credible or just legitimate disagreement have. Like I said, I do have short timelines, so I think those are going to be relevant.
Right. Which by the way, I acknowledge it might be an area I’m totally wrong. My timeline is more medium-term than short-term, but I think I’ve said pretty consistently we could take off much sooner than I think and that would change the conversation.
DM: That’s where I come from a lot is just my timelines are inclusive of short timelines and therefore I think it’s worth acting, but I don’t know what’s going to happen. But yeah, some of these other things, some EA’s want to pause, I really think that that is just a particular subgroup, I’m not on that letter, Eliezer [Yudkowsky] is out there with a lot of very confident statements. But when you look at the EA forum, you see these people fighting with each other and there’s a lot of internecine rivalry so yeah, maybe some of them don’t understand complexity.
But, I’m a chair of at Open Philanthropy and I’m a business guy and I’m also in the boardroom in at Anthropic and I know all the players. I’m sure I’m naive in some ways, but I think the collection of people is not quite that naive and I think Open Philanthropy has been more measured certainly on the rhetoric side and then in terms of what we’ve actually done, most of it is just gotten nowhere, the whole patchwork of regulations, I’m like, “Can we have one first?”.
Is this actually, are we circling back to why you stepped down as CEO? Because there’s a bit where just contributing money here, “I actually need to step in and have a more direct hand on what’s going on”?
DM: Yeah, I think that’s part of it, and I would get drawn into these conversations where I disagree with the program officers at OP, but I wouldn’t have time to really pursue my version of the argument and like I said, I’m working with Claude Code right now, I’m prototyping some things, I’m getting a much more native understanding and spending a lot more time seeing what’s out there and what people are talking about. So yeah, I think that was probably the prime driver is over the past five years I’ve been seeing AI play out and be like, “This is it, I’ve got to focus”.
How has it played out differently than you expected or has it played out basically exactly as you expected with let’s say from 2020 on? I know you’ve been in the space for a long time, but let’s say from GPT-2 on.
DM: I was pretty pessimistic even in the GPT-2 era in that timeframe, I was like, “They can’t do math, this will never scale, these are stochastic parrots”, and part of my journey on safety is basically I made statements like that and then I tried to review them later and I was totally wrong about that and basically set up a bunch of falsifiable assumptions that got invalidated. So I think I’ve come a long way on thinking that the elements of the future, I don’t think you just scale a bigger model and that becomes an ASI, I think the scaffold is a really important part of architecture. But, I do think they are on the path and we will get there.
Do you still believe in one is going to get there first and is going to win it all or do you see this end up being a multipolar world?
DM: I think it will be multipolar if we survive. This is also a weird thing, I’ve been trying to find the source of this, but a lot of people project on the EA is this idea of an AI hegemony, “You want the government to nationalize an AI and that AI just puts down all the others”, I don’t think very many EAs really believe that, but I certainly don’t.
I emphasize really the engineering part and the scaffold and so I think there’s going to be a lot of innovation there over the next few years that we’ll be able to turn what we see as weaker models now into more powerful models. Just generally, I don’t think there’s much of a technical moat for any of the labs. The one exception to that may be the coding agents, if that were to happen, I think would be because, I don’t know, Anthropic decided to keep the next generation for themselves and use it to build something. Right now what happens is all the labs using Anthropic to code, which is really crazy, but in terms of the model scale and the size and the researchers are, I don’t think there’s enough to really result in that uniform world.
You think back to the last 15 years of building Asana and even go back to Facebook and its growth, which is maybe a little bit of a contrast there, we have the bottoms-up bit, you have the enterprise bit, all the go-to-market things we’ve talked about. Do you think it ends up being that we do get ASI, however that’s defined, but it’s not widely distributed just because people don’t appreciate it and don’t use it? Or, is it sort an overwhelming force that’s sort of irresistible?
DM: I’m on the overwhelming force side. It’s also, it sort of doesn’t matter if some people resist because it’s going to be like a software eating the world-type situation and so I do expect very disruptive AI companies that will change the way the world works right now and again, you go far enough out, I’m like, I don’t know if there’s still capitalism, I’m quite into things will get very, very different.
Very good. Well, Dustin, it was great to talk to you. Very interesting conversation and we’ll look forward to following what you do next.
DM: Okay, thanks Ben. Look forward to it.
This Daily Update Interview is also available as a podcast. To receive it in your podcast player, visit Stratechery.
The Daily Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.
Thanks for being a supporter, and have a great day!
Source: Stratechery by Ben Thompson.
Leave a Reply