Founder Mode: Emmett Shear, Founder, Softmax & Twitch

0:00Carolyn, I'm really excited. Today we are here in person with Emmett Shear, former founder and CEO of Twitch, but now founder and CEO of Softmax, which we want to hear all about. We are at the Y Combinator Founder Mode retreat. And so we're talking to people who know what they're talking about about Founder Mode. But first, Emmett, I just want to quickly hear about Softmax. Yeah. So after the whole OpenAI thing where I wound up doing that CEO gig for a record minimum time of like three days, I really feel like I did a good job as being a CEO. If it only takes you three days to be the CEO, you're really on top.

0:43How long you were the CEO for? I thought it was shorter. I thought it was like 48 hours. It was 58 hours or something. Love it. I love that. By the way, the day after that 58 hours, we ran your social radars interview that we had taped a few months before. It was perfect timing. That's so good. Okay. So after OpenAI. Basically, I felt like I'd gotten this message from the universe that was like, hey, Emma, you should pay attention to this AI thing. Because I kind of knew. I'd be like, oh, I should do some AI thing at some point. I was like, okay, okay, I guess I'm supposed to like do an AI. I guess I've message received.

1:17Fine. I'll go look into that. I'll figure out what to do. I didn't actually know what it was. I was just like, I guess I'm supposed to do this. and so I really carried out this question of like how would you make an AI system that was sort of robustly aligned with humanity what would it even what does it even mean what does it mean to build a system that stays in alignment with us because I had this sense that alignment was the right word that in some sense I wanted it to be on our side but I wanted to be aligned with it and I wound up sort of coming down to this idea that like what alignment is about is about this question of are if you have two you have two beings or you have many beings sometimes you have a bunch of individuals and sometimes there's like a we like we are here doing this we care about we have goals and i can i can sense myself like i have a goal but then sometimes we have goals and i have my own goals and i do the we goal instead of my goal and it's because i am at some level like my family is me my my community is me like there's we i am an individual but i'm also part of this thing that's bigger than me.

2:21And that's sort of like where this purpose and belonging comes from. And this comes from having agents, beings that can understand who they are, what they care about, and that can understand theory of mind style, who are the other agents around them, what do they care about, and can tell, like, are we part of the same thing? Because that's actually the question, right? Is there a we here? Like, are we a collective? What's the right collective to care about? Because there's lots of collectives you could care about. I feel like this here is a very comfortable and clear we. That's why we all like being together.

2:55Yeah, exactly. And 100 % that YC in general forms this incredibly powerful sense of we. And that sense of we are starting these companies together is one of the most important things that YC does, that YC is. And so like, how do you train an AI? How do you build or raise, grow an AI that knows itself, that has a model of itself that understands its goals, that understands, has a model of us, understands our goals? And if it is, if we are on its team and it is on our team, it can tell like, oh, I'm part of this. I'm part of this team. And it can feel that care that we feel because it correctly infers from reality that it is part of a team.

3:38Because that's not like, you're not hallucinating when you think you're part of a, like, you really know, it really is your family. It's not made up. That's real. And so that's what Softmax is dedicated to is how would you measure whether, how do you measure whether agents are capable of this or not? How do you measure if they are acting as a group? And then trying to make, build simulations, build reinforcement learning simulations, basically, where we can test our theories and see if we can build an agent. And the goal is to teach it to know a we, to be a we. Yeah, to raise it to know how to be a we.

4:10And it really is, once you end up with, eventually you end up having to parent the agents because ultimately no one can tell you who you are. That is something that everyone has to learn from themselves. You can't, I can tell you what you are to me, but I can't tell you what it's like to be you because I'm not you. If I was in there, I could tell you, but I'm not, then I would be you. So really, you're going to have to do it. And it's really funny. I have two kids right now. And like having kids, raising your kids, it's like, oh, like this is the same. Like you start and when you're born as a baby, you don't know who you are.

4:40And your job as a parent is to be a mirror and to give them a good mirror of themselves that helps them see themselves clearly. It's not your only job, but one of your most important jobs is to help your kids see themselves clearly so that they know who they are and to help them see you and help them see that they're, hey, we're all part of this family together. If they don't infer that, it's bad. And it's done all the time. It's a normal, ordinary thing that we're actually very good at. But it's also something we're like, how do you engineer a system that has the capacity to learn in that way?

5:14And then what would it mean to parent an AI? I don't think anyone knows. So that's the kind of what we're working on in the abstract. I'm curious. What's the end goal? Which I hesitate to use that expression with AI. But let's say you're successful in measuring and understanding that you created a we. Yeah. What do you want it to do once it's a we, once it recognizes the we? I want it. I want the same thing for it that I want for my children, which is I want it to live a life of virtue and nobility where it both cares for itself and grows to be the happiest, most flourishing life it can have. And also.

5:50cares about the greater holes of which it's a part and acts with nobility, tries to make a difference, cares. Because what other thing could you ever wish for another being that you brought into this world other than it has a good life and also purpose and belonging? I don't want it to have some solipsistic good life on its own. That's not actually a good life. Nobody wants that. I want it to see the truth that we need it and it needs us and we're part of something together. That'd be the ideal. That is, in my mind, the bar for success. If what we produce are alignment benchmarks and ways of training agents that someone else uses to then train, I view that as a success.

6:31So we're going to be launching things like benchmarks to help people see, are they succeeding at this? Because ultimately for society, somebody has to. But we sort of view it as a holistic system. You don't get to just pick pieces of it because if the thing lacks the capacity, it won't learn it. If you don't have the right training environment, it won't have the capacity. If it doesn't have experience with a high variety of, you don't just raise your kid as parents, you need peers. It needs to know how to help other agents and help teach. It needs to teach as well as to learn. So you need this very, actually quite sophisticated environment.

7:14So we spend almost all of our time thinking about what's the learning environment, not what's the learning architecture. And obviously at some level, they all go hand in hand, they connect to each other. But I'd say we're almost like an inverted AI company that cares about the learning environment more and trying to make the agent into something specific less. And it's been the most interesting work I've ever done in my life. I'm sure. This sounds so interesting. and I bet you're having fun. Oh my God, it's so good. When did you start it and how big are you now? So we started it, geez, like a year and a half ago, me and Adam Goldstein started basically with the question and like two employees and like trying to figure out like all the things I'm saying now, we had a proto version of them, but I would not have been able to articulate them clearly.

8:09we were groping in the darkness. And we spent the next year, roughly, groping in the darkness and made a lot of progress. And I got to the point where I had this day where I was like, we were doing the next architecture design. We're like, we need to reset our architecture. And we're like, okay, we have the plan now. We know what we need to go do. Here's the open questions. Here's the things we need to go measure. And then we invited David, David Blumen, by the office to give us a talk on something just sort of random. And he comes in and he lays out this architecture he's been building for the past year and a half, two years, that is, note for note, the thing that I had just articulated to the team that we had to go build.

8:46And it is the weirdest, synchronous thing I've ever had in my life. I cannot explain it. I cannot explain it. I mean, I can't explain it. I'm out talking about this shit all the time, so of course he heard about it. It's not a coincidence at all, but it was weird. Who's David Blumen? David Blumen is my second co-founder, and he's an engineer who's at Google Asana. and just, you know, one of the best engineers I've ever worked with, but also like saw the same thing from a less theory point of view at the same time and just sort of like in parallel, almost at the same time we did, sort of building in parallel, but where we took a theory and then had come to the practice.

9:19He started the practice and had sort of come to the theory and we kind of came together and like this was then, so then in January this year, he joined the team. And so it's been seven months basically of like, we have the full, really we started in some ways we really started seven months ago. And so that's where we are now. So you're seasoned now, Emmett. What have you learned about founder mode? And tell us if you can highlight any big mistakes or times when you went into founder mode. I'd be curious. Yeah. I think founder mode, when I think about what that means to me, it's like taking personal responsibility for the decisions the company makes and sort of treating the company as a full extension of you.

10:04And that is not always a good idea because the company both is and isn't that. There's a way in which that is 100 % true. Companies are reflections of their founders, reflections of their CEOs, and shying away from your responsibility to make decisions, to set direction, to take risks is like betraying the company and your customers and yourself. And the company's made of a bunch of different people who are not you. and I think that there's this danger in treating it that way where you it becomes a company with just one brain one brain many hands yeah and one brain many hands is not it's just not a like you you don't have enough intelligence you're you're just not smart enough I guarantee it you cannot pay attention to the whole company at once like even if you're really really good there's just a limit and so the my experience at founder mode has been what has been successful as I was talking about like at Twitch, when I shifted into leading Twitch from Justin TV and it was sort of this new thing, I went through this first phase of like doing a lot of the work myself.

11:07And that was kind of founder mode. I was like really just in it every day, talking to customers, doing stuff. But then as the company started to grow, founder mode was the new thing. It wasn't. I had to completely stop programming, completely stop really doing that much product management. And I spent a lot of time on the weekly presentation every week. I spent a lot of time. I made the keynote and I presented to the company every week. Yeah. And because I needed everyone to be smart. I need everyone. I needed to I couldn't I couldn't make all the decisions. And so I needed to like I needed them to get it.

11:37And so I put a lot of effort into communicating. And this was when we were like 100 people. Yeah. OK. Communicating to everyone in the company. And when it's 100 people, 150 people, you can actually with enough effort. I synchronized everyone like I I didn't have to be everywhere because they would make the decision I would have made. because I have conveyed my mental model into their head and they could just use how I see it. But then plus also how they see it. They had a good enough Emmett in their head that they knew what I would say and they could figure out how to integrate that in that local situation.

12:11I think this is really good and interesting because this is like founder mode in terms of empowering others versus in terms of like, I have to be the one that takes all this on. This is another iteration of how people view Founder Mode. The thing is, it's not less egotistical. I got feedback from people because basically I was just gauging how well am I getting my mental... I had the answer and I was putting it into everyone else's head through repetition. The things I cared about, I just had a dissection for it. Every week I would talk about which streamers switched to Switch this week and I'd talk about why.

12:45And I would talk about why it was important that we get every streamer on every competing platform to switch to us and like how to think about when it was worth doing and when it wasn't and how to target them. And like, it was just every week over and over because I need everyone thinking about that that way. Or when we launched a product, I would talk about whether it was successful or not. But it was really about me pushing out. And then we grew and we grew and we got to the point where you just can't, most people can't hold all the, the context got too big. And the CEO's job is to be the holder of the context.

13:19The CEO's job fundamentally at the company, well, you have many jobs, but one of the most important ones is everyone else has something they're responsible for. Even the CFO is responsible for the CFO's area and not the other things. And they might hold the rest of the context, but you have the job of holding everything, being the context for the company. And the problem is what I learned is that I spend all day loading the context into my head. I just cannot – at some point, the content got too big. I couldn't even with effort jam the important parts of it into anybody's head. It's not possible.

13:49It was hard for me to keep up with all the context. And then I started to start – so then you have to start breaking things up and delegating and, like, releasing responsibility. And at first, that was good. At first, I delegated a lot. And it was actually, I think, really critical for the company that I had someone else who was a CTO. And I didn't really understand our architecture direction. But I worked with him enough that I – well, we went through a number of CEOs. and they didn't work, they didn't work, they're bad. And then we got someone really good. We got the right person. And like every, he was just a really good leader and I could actually delegate.

14:22And then that delegation turned into like, I had got this idea that I think Brian's talked about this, but like somehow I wasn't supposed to override my people. I'd hired experts. I was supposed to let them do their jobs. Let them do their jobs well. And like when I didn't like how something felt or where I was taking the company, that I should crush down that concern and not act on it. And that was a mistake. Like, I allowed people who were, I think, well-meaning, but who weren't me and who didn't really understand what we were doing because actually no one but the CEO ever really understands what's going on because it's not their job.

14:57That's my job. And I abrogate, I did not, I was failing in my responsibility to the company as the holder of the context to override people when they were wrong, to make sure that the things were correct in the context. Because one of your jobs is to grow people and to delegate and all that stuff. And one of your jobs is to make sure and to know, to have the discernment to know, is this the kind of decision that we have to get right? Or is this the kind of decision where it's actually okay if we screw it up and it can be a learning experience for somebody? Because there are decisions that can be learning experiences for other people.

15:30That's how I learned all this stuff is I made a lot of mistakes and I got to learn from them. But there's a lot of decisions that are not, you actually do need to make this decision correctly. Yeah. There's more of those than you think. It is important to make decisions right. And I just at some point I kind of realized like, oh, I have to I can't have a head of product. Really, I have to be the head of product. I can't expect anyone who doesn't have all the context to make to make these decisions. Although I did learn something that was like really interesting. There's this incredible law, Conway's Law.

16:02Conway's Law says organizations, products, always reflect the information flows within those organizations. The structure of that product reflects the information flows in the organizations. If you have a communities team and a videos team, I will go to your website and you will find a communities page and a videos page. I guarantee you every time. Okay. And that's because if you're both shipping to the same thing, you have to be coordinated, which means you have to have information flows between your teams. And if you don't, you'll keep stepping on each other's toes. Right. Yeah. And so you can have two teams if they're like deeply interconnected, but then you don't really have two teams.

16:36You only have one where you've labeled half the people, video or community. But like, but if you have two teams that are actually separate in any way, they have to have separate things. And so the only way to fix that as a CEO, you have two tools. One is you can be, you are the join point. If you personally take accountability, you become the information flow thing that merges the two pages. You basically have to design the thing. Because you're not going to look at two pages. You'd much rather look at one. And because you are integrating all the information from the video team and from the community team.

17:10And therefore, you can make the correct contextual decision. Or you have a second tool at your disposal. I think people in founder mode are way under use this. If you want to merge two pages of the website, you can just re-reg your team. like the easiest way to change the website if you think that oh we have these two pages i want them to be one or if these one page i want it to be two just just reorg your company give them different titles yeah just make it if you have two if you have one team you would have one page if you had two teams you had two pages and it's so funny to watch it happen because like they'll come to the conclusion themselves and they'll be right that's the logical thing for them to do this is conway's law conway's law yeah this guy this business i love it this uh you know organization theorist from like the 70s the 80s so the cybernetics people which we're really into at Softmax now.

17:52Okay. I knew about this at Twitch and I tested it a couple times. Not that much. I really learned it like later on in my Twitch career so I didn't get to use it as much as I'd like but I tested it. It definitely works. And you'll use it at Softmax? Oh, you'll use it 100%. We're using it very deeply at Softmax. I have ideas about how to use it too. It is. That's a great, this is great. And the thing is people in founder mode are really, people talk about being scared of taking product risk stuff and overriding people on product but what you're really scared of is overriding people on, I think, I tend to think about things in a company

18:22But I'm a systems person whose primary medium I like to think is people. The systems I care most about have always been people systems. How is it that people systematically relate? And so for me, founder mode is mostly about the willingness to be like, we're going to have a team focused on this. I'm going to decide, no, we're putting these things together. We're splitting them apart. and like if you set the expectations right, people are okay with that because they understand that the goal is to have the teams that reflect the work that needs to be done. Not like, it's not because anyone's succeeding or failing.

19:02It's about what, the cool thing on Combi's law is it works even if you explain it. It doesn't matter. It's a law. Knowing about it doesn't change anything. You can know what's happening and it still works and so like, and there's no clear, in some way you can also see it normatively. there's no clearer commitment to we're going to merge these two products into the same thing than to merge the teams how else can you say to the company so clearly i'm serious and so that's that i think that that for me that's that is the i think people in founder mode there's often this really emphasis on like i'm going to make a lot of the detailed decisions and that can be right there's a there are products or situations where that is where leadership lies that's what you do need to do and i did that at twitch also but i think maybe the thing i've heard less that i would encourage founders to think about is like, what if you treated your organization's structure, your communication to the company, your people, your leadership with the people with as much seriousness as you treated the product?

20:04And I find that like that takes, it's a little bit of a weird feeling at first. It's like, I'm going to spend, I'm going to make a keynote. I'm going to spend three hours every week building a keynote. But like, no, because if I do that, it's going to be good and it's going to be clear. And I have way more amplification of my work there. Sometimes only if the thing you, if the thing you need is to be at Twitch, the big thing we needed was to be smart all over. I needed every individual negotiation to be smart. And so that's, I can't be in all the negotiations. So I have to make the team smart.

20:32If what you're doing is building a product with like a two, two surfaces and those services have to be right. Like I think actually Airbnb is a good example. It's like the checkout experience and like the, like those experiences have to be right. And if they're not right, it doesn't matter if you're smart all over and the team that's pitching the people is good. You're just not going to succeed. Whereas on Twitch, the core thing was kind of was working. We need it. But I think that that's like a it's something people should reach for as a tool more often. Yeah. OK. I love it. Emmett, it's such a treat to have you.

21:03I'm not kidding. It's so nice to hear about Softmax and to hear your thoughts on Founder Mode. And I hope someday you can come back on the social radars and talk about softmax. Yeah. I'd love to. That would be fantastic. Whenever the timing is right. Yeah, absolutely. Thank you for having me. Best of luck. Thank you, Emmett. Thanks, Emmett. Bye. Bye.