Join us as we chat with Shir, the engineering lead of Notion AI, about her journey into AI, from her early fascination with Apple's speech-to-text technology to her impactful contributions at Google and Waymo. Shir shares insights into her current role at Notion, the innovative AI projects she's leading, and the unique culture of experimentation and failure that drives product development. Discover how AI drafts performance reviews in minutes, celebrates code deletion in team demos, and integrates seamlessly with Slack to provide precise answers. Plus, learn about innovative tools like the AI writer that helps overcome writer’s block and enables effortless content creation. Get a behind-the-scenes look at how Notion's unique culture fosters innovation and continuous improvement. Learn how Notion leverages AI to enhance productivity, creativity, and collaboration, and get inspired by Shir's perspectives on the future of AI.
Website: NotionAi
00:00 Introduction to Shir and AI Journey
00:47 Early Fascination with AI
01:36 Career at Google and Voice Search
03:43 Transition to Waymo and Challenges
05:11 Joining Notion and Initial Impressions
06:09 Adopting AI at Notion
11:55 Internal Use and Impact of AI
15:12 Challenges and Innovations in AI Integration
20:05 Monitoring and Feedback Mechanisms
22:39 Strength and Flexibility in Product Development
23:44 Exploration vs. Exploitation Phases
24:47 Speed Enables Quality
27:50 Overcoming Sunk Cost Fallacy
30:36 Celebrating Failures and Learning
33:11 The Future of AI and Creativity
34:44 Practical Uses of AI in Daily Work
38:04 The Evolution of Human-Computer Interaction
41:02 Final Thoughts and Reflections
📜 Read the transcript for this episode: Transcript of How to Increase Workplace Productivity. Notion's Lead AI Engineer Shir Yehoshua on High-Risk, High-Reward AI Projects |
[00:00:00] Shir Yehoshua: Hello, I'm Shir. I love AI products. Um, and I currently lead Notion AI engineering.
[00:00:08] Jeremy Utley: Thanks for making time to chat with us.
I don't know if you've had the chance to hear any of our past episodes, but we're just a couple of nerds who are eager to learn from world experts like yourself. ,
[00:00:17] Shir Yehoshua: yeah, I'm, honored and humbled that that's, you considered me an expert because I fell into it.
[00:00:22] Henrik Werdelin: It's so fascinating. What we learned is that Most of the people who have on the podcast are like the experts kind of like stumbled into it. and because the technology is so new, very few have, really been working on it for more than a few years, obviously they, from the engineering background.
How did you stumble into AI? Like, do you remember the first time you played with playground or what was kind of like your experience? Like
[00:00:46] Shir Yehoshua: Yeah. so I've actually kind of been interested in the space because I thought it was cool for a really long time.
I grew up in the Bay Area and my first experience, I guess with, what is now called AI was I went to a take your kid to work day at Apple. My parents didn't work there, but I was just the nerdy kid. My friend didn't want to go, so I went instead. and they had a demo of, um, uh, speech to text.
And, I thought it was like the coolest thing absolutely in the world. and I was like, we're going to have science fiction. It's the future. Like I'm going to be talking to my computer really soon. and of course it wasn't soon. and I think it was at least maybe 10 more years before Siri actually came out, or was good.
but I've always been fascinated by it. Cause I just think it's really cool. I guess that would be my start, and then after I went to school, I studied computer science, I joined Google to, work with voice search at the time, which eventually became the Google Assistant, before we didn't have, the technology to do speech recognition properly, but we're finally, here now, so it's going to work.
and then it kind of worked. you could set timers and alarms really well with your phone. And those are some of the, initial, actions that I built at Google. but we didn't have LLMs yet. That's, kind of the next wave, and now I'm finally excited that we can make it to that supercomputer and that computer that you can talk to, and do, all the mundane stuff.
You can just have something else do, and then you can focus your brain on the creative stuff instead.
[00:02:21] Henrik Werdelin: you've obviously been following it.
[00:02:22] Shir Yehoshua: Yeah.
[00:02:23] Henrik Werdelin: I remember back, I think the first time I tried to like to talk to somebody was Alice. Oh, yeah, I think the bot was, was called that.
And it was like these very scripted kind of conversation I'd go through. But over the last few years, obviously, I think. , the technology has moved so fast. Was there kind of like a point where you were like, okay, it might actually finally have arrived all this stuff that we've been dreaming about for a while?
[00:02:48] Shir Yehoshua: Yeah. it was a bittersweet. moment for me. one of the things that I built at Google Paint, it took almost a year, was an action that very few people use, which was the, okay, Google add a calendar event. figuring out what time, figuring out like, what do you want it to be called?
What's the description? Who do you want to invite? It was very formulaic, and it was the first time we were introducing follow up, questions. With a single prompt and LLM can do all of that. and can fill in, like you tell it what fields you need and then it fills everything in.
[00:03:19] Jeremy Utley: Just so you're aware, Shir, you've now done such a good job with your past products, you actually triggered my Google assistant
[00:03:28] Henrik Werdelin: my Google assistant also fired off.
[00:03:32] Jeremy Utley: I'm like rapidly muting over here. It's just going, what's the title of the event? I'm like, no, stop.
[00:03:37] Shir Yehoshua: Oh, yeah.
So my, my husband works on that. I'll tell him that he's done it. He's done a great job.
[00:03:43] Henrik Werdelin: And then you moved on to Waymo. And so, I mean, like, which I also tried, which is also mind blowing.
So what made you kind of go from the AI cause to notion, which obviously, is a different technology problem.
And now I guess from a customer founder fit kind of perspective, you now have to understand, a completely different world, which is not the physical world, but the mental model world instead.
[00:04:11] Shir Yehoshua: Yeah. I kind of just chased things that I thought were interesting.
I was never like, I'm going to do AI or I'm going to do voice or I'm going to do whatever. I just was talking to somebody who was working at Waymo, and I thought it was really cool. It wasn't Waymo at the time, it was just the Google, chauffeur project. I was like, I know how to write code, just throw me in and I'll figure out how to help.
it just seemed super cool. And I think the thing that I learned at Waymo that was super, super challenging is that it's incredibly easy to get a demo that works 80 percent of the time, 90 percent of the time, even 95 percent of the time, but to actually drive a car. on the road safely, it needs to work like basically 100 percent of the time.
And you
[00:04:54] Jeremy Utley: don't, you don't want to arrive safely your destination 95 percent of the time.
[00:04:57] Shir Yehoshua: Exactly.
Exactly. Um, and that's what took so long is going because you need actually a completely different stack, to go from, it works most of the time to it works reliably all the time. but that was a really cool experience.
And then, I joined Notion because I just started using it and I thought it was really cool. I actually, so even though I'm a Silicon Valley baby, I never used any kind of software to take notes, keep to do lists, do task management. I had like a composition notebook, white paper, because everything else was too constraining, too limiting.
And then I started using Notion, and I was like, Oh, I can let go of this because it's so flexible. It's this innovative product that I, a bunch of building blocks, my brain can finally be translated, into a digital product and actually, even better for me than a pen and paper. and so Notion was my evolution into finally arriving at the digital space.
and that's why I joined. I thought it was a great product and I was like, I want to work on this. I want to, I don't know anything about, user, like UX, and UIs and building really, really great products. I've been on the technology side, but I want to learn. So that's why I joined.
[00:06:09] Jeremy Utley: can you talk about where, where was Gen AI at that time?
Where were LLNs and what was Notion's kind of, you know, position relative to large language models then?
[00:06:20] Shir Yehoshua: Yeah. So I joined Notion before LLMs were a big deal. I actually started working on the platform side. I was working on our public APIs and our third party integrations. then, there's this legendary story of our founders got access to GPT 4, were wowed by it.
I talked to them at the time and I was like, well, you know, Can you do these four things? And then it was like, well, actually, yes.
[00:06:46] Jeremy Utley: what were the things?
[00:06:47] Shir Yehoshua: Oh, so like the, my example of building a calendar, like calendars, scheduler of the thing that was like so hard for me to do when I was at Google, that was a single prompt, with GPT 4.
Like to me, that was the, that was the difference. and the fact that you could, I was worried about, you know, security and, like saying the wrong thing, saying something insensitive, but like you can actually prompt an LLM, keep the LLM from saying something insensitive. That's like wild to me and so cool.
and so there's still a lot of problems, that need to be figured out, but. I just felt very, very different from the technology from before.
[00:07:23] Jeremy Utley: So you had this conversation with the founders and then it's, or do you get deployed to share, go figure this out or like what, because you're working on the platform side, like how does your career transition within notions?
Yeah.
[00:07:35] Shir Yehoshua: Yeah. So, I was working on the platform side and we did this very fast redistribution, of the work and I just created a new team. So I worked really closely with Simon, who's our co founder. He kind of built the very first prototypes, and then started hiring a team, both from within, Notion and from the outside, kind of trying to get a mix of the, AI expertise, of which there was very little of it at that time, because this was, in January of, last year.
and then also like the internal notion folks who knew the craft, of product development. And so when you brought those two groups together, that's sort of where the magic happens. but yeah, it was a really hard pivot of like, okay, this is brand new. let's go.
[00:08:19] Jeremy Utley: And what's it like when you shine the bat signal in the sky at Notion, so to speak? Is it like, is this a project people want to avoid? Does everybody pile on? Because, I mean, it's easy with hindsight's 20 20, you go, this is a really big deal. January 23 are people going, dude, sheer, let me on your team. Or are they thinking career limiting move?
[00:08:42] Shir Yehoshua: I think it was mixed. I think there was a really strong conviction from our CEO, Ivan. And I think that was contagious. And that helped bring the whole company along on like, okay, let's see what this goes. And it was always framed as this is a bet that we're making.
A bet is very different from like a critical investment. a bet is something that you invest a lot in, and there's a large chance that it doesn't work out. but it's like a high risk, high reward. Versus a critical project, which may be like, investments in infrastructure or reliability.
Like, that has to work. and I think it was always framed as this bet. It wasn't framed as like, we're, Gonna do everything we can on AI and if it fails, our company fails. It was like, we're gonna do a lot of investment and give it a proper go. and then if it works, it's this high reward.
[00:09:33] Jeremy Utley: So, so Henry and I, you may or may not know this about us, but we're super nerds when it comes to innovation and organizations. Yeah. We're unrequited lovers of innovation in organizations just to say. We keep trying despite the evidence to the contrary. Right. can you talk for a second about, governance mechanisms, et cetera.
When you frame it as a bet, and you said high risk, high reward, how is that framed internally and how is it different than maybe any other innovation projects? what are the practical implications in terms of. Expectations on you and the team, your expectations of success and failure when it's framed as a bet in that way.
[00:10:08] Shir Yehoshua: Yeah. So, I think you have to have space to fail, in order for a bet to succeed. And I think maybe that's the primary difference between, a bet and something that's more critical. I actually, have a bunch of dance training and I've learned a lot from that world. and if you're trying to learn a new trick, you can't ever get it until you fall, a gazillion times and you have to learn how to fall first before you can actually hit the turn that you want, the leap that you want, the whatever move that you want.
And it's the same thing. You have to be able to safely, like with mats and protections and knee pads and whatever, you have to be able to safely fail. in order to succeed. And I think that is what you want for a bet. it was something that's more, maybe still hard, but definitely possible to do.
you maybe don't need that same pressure and that same safety net. maybe you want to be motivated by the fact that, okay, this can't fail. but I think for innovation, you really need That safety net.
[00:11:07] Henrik Werdelin: Do you feel that you guys, good, let's say a little bit of the organization and then we'll kind of go into, what have you discovered that people can use this for?
But in terms of organization, we often look at how does people internally using it? What are the processes that can be optimized and how can you change your product? Obviously I'm a very, very big user of Notion. Jeremy is not. so you'll get the nerdy questions from here.
But obviously over the last few releases, you can see how like AI is getting, you know, more and more of a prominent kind of position in the UI, which often means that that is clearly a priority. do you feel that with kind of like the excitement from your founders on AI, you've also actually started to use it as part of the day to day work?
[00:11:55] Shir Yehoshua: yes, a hundred percent, we had, so, we're kind of building our next generation of the AI writer, which helps you draft, and edit content in your workspace. in preparation for our performance review season, which happened maybe a month or two ago, someone on my team.
Made a, two minute video, about how she was writing her performance reviews, asking AI to, find all the relevant, like, what did I work on? What are my strengths? What are my weaknesses? can you frame this feedback into this framework, which was the framework that we were told to use?
And she demonstrated how you can write your performance reviews in like a couple minutes instead of agonizingly over a couple hours because you're stumbling over writer's block and you're procrastinating because you don't want to do it. and I think that was like a big moment inside of Notion.
Publish a video and we looked at our metrics and our usage internally of Notion AI just like spiked, that entire week. and then for myself personally, it was kind of the first time that as I was writing my manager reviews and manager assessments with people on the team. It took me way less time because I wasn't doing the manual work of, figuring out what everybody worked on, AI could help me summarize that, and then I can focus all of my attention on giving really, really good feedback, and I got the most number of compliments on my reviews this cycle, even though I spent the least amount of time, but it's because I got to put my creative energy towards that.
[00:13:24] Jeremy Utley: So can you tell us like, what, what is, what was the workflow before? How did you go find stuff before and then leveraging notion or whatever other tools you might use? How did the tool practically save you time?
[00:13:37] Shir Yehoshua: Yeah, so in the alternate previous universe, let's say I was writing a review for somebody, I would go read what they wrote, copy and paste, the sort of bullets of what they said they did, then I would go read every single bit of peer feedback, copy and paste All of the additional things that their peers said that they did.
So I have a summary of their projects. and then I'd probably also look through either my notes with them or, other meeting notes or project pages to make sure that I'm not missing something. That's like an hour, maybe two hours already, just right there. and with AI, I just say Per
[00:14:15] Jeremy Utley: Per person. It's an hour per person.
[00:14:16] Shir Yehoshua: and then with AI, I did have to, like, because, we use a couple different tools, I, put it all in one place, which takes maybe, a minute or two, and then I just ask AI to pull out what did this person work on, And it comes out with a list of bullets.
now I'm still reading everything, so it's not like I'm skipping that part, but I don't have to do that, mental mapping, copying and pasting, finessing of information. And so that's one example of a huge time saved.
[00:14:42] Henrik Werdelin: Do you have a, we love these cases, right? Because obviously for somebody, for most people who hear about AI, they hear that there's a lot of cool stuff.
That you can do and then they get hit by the day to day of, you know, everyday kind of minutiae and then they, they don't get to use it. And so do you have another one of these where like suddenly you did something internally and you saw the usage spike or you were like, Hey, this is actually. Incredibly useful.
[00:15:12] Shir Yehoshua: I think the other one is a little bit more, it's not as dramatic, but it's a lot more pervasive, which is just Q& A. So, I think we've seen, or like we've seen, so we have an internal AI tips and tricks channel, where people will post screenshots of just cool things that they did.
and so much of the time it's AI, just like our Q and A product, discovering some bit of information that you didn't even realize was there. my toy example is I was trying to get the model to hallucinate stuff about celebrities, but because of our AI connector product, which we just launched into Slack, I was asking questions about Beyonce and because there was a whole lively discussion in Slack about Beyonce, I couldn't get the model.
to hallucinate or make anything up about Beyonce because it kept pulling, stuff.
[00:16:01] Jeremy Utley: context.
[00:16:01] Shir Yehoshua: you'd be surprised at just how, exhaustive Q and A can be at finding you your answer. And you get to, stumble upon things that you wouldn't have stumbled on before.
[00:16:13] Henrik Werdelin: Do you mind just explaining the connector with Slack to somebody who hasn't tried it?
[00:16:19] Shir Yehoshua: Yeah, so
[00:16:20] Jeremy Utley: Like
[00:16:20] Henrik Werdelin: this guy.
[00:16:20] Jeremy Utley: you just explain it to me.
[00:16:22] Shir Yehoshua: well, I don't blame you because we launched it two days ago. So you have an excuse there. yeah, so we have a Q& A product, which, if you've heard of RAG, it's a RAG system. Basically, we index your entire workspace, and then you can ask the LLM a question, the LLM will effectively search the whole workspace, find content that's related, see if it answers your question, and then answer your question, if it found the answer, and it's using that reasoning capability.
[00:16:53] Henrik Werdelin: the workspace is, can include all Slack channel,
[00:16:56] Shir Yehoshua: so as of two days ago, we now also include Slack. So you can also connect Slack up to the same Q and A, system. And we won't just search over your Notion workspace, we'll also search over public Slack channels. a lot of times you're not sure where the answer to your question is.
did I write a doc about this? Did I write a post about this? Maybe I did both. I found it really helpful in being able to search across books.
[00:17:22] Jeremy Utley: And what's like, what, okay. This is useful, not just for teams, obviously, but I can't remember where I put stuff, right? How do you make sure that you, like I had an exam, I had, I had an experience yesterday.
I've been blogging for years, so I have like, you know, thousands of blog posts. I have only been able to write about a fraction of the things that I know or find interesting in books.
Right? So. As an example, I'm on a call with someone and we're talking about the virtues of team versus idea and talking about the, that if you have to choose between one, obviously you pick the better team with a bad idea rather than the better idea with the worst team, right? Because a good team's gonna pin it.
I know for sure. Ed Catmull said that. I know for sure, but I've interviewed him on my, you know, other podcast. Did he say it there? did he say it in an article I read on HBR? Did he say it in his book? You know, which, which I don't have, you know, and so I asked Notebook LM cause I've uploaded a bunch of stuff there and it's given me nothing.
And I'm still, it becomes a user frustration of like, what else do I need to upload to this system before I can reliably query it? Because like, now I'm, you know, I asked the system if it gave me something great. If it doesn't, then I'm basically left at. Well, and by the way, I scraped my own website, so I don't even know if I got all my blog posts, right?
There's like, I got like 85%, right? So it's possible it's actually there. It's just telling me it's not there, right? So all that, I don't know what the exact question is, but I think part of it to me is, how do I know what of my world to add if I really want to be able to say, Where did I hear that thing, right?
[00:19:06] Shir Yehoshua: Yeah. I think the short answer is you probably need to somehow connect everything. but it's like, what percent of everything is everything? but I also think that RAG as a technology stack is still quite early. just last two weeks ago, we added the capability to search for two things at a time.
when you ask a question in the LLM, maybe actually it was a little ambiguous, or maybe there's two variants, like, do you mean, one definition of this word or a different definition of this word? only very recently did we add the capability to do two searches at once, get two sets of results, and look at them both.
and the whole functionality worked really well even without that. so maybe the missing percent of everything, you can't find. might be on the indexing side of making sure that content is there, but it also might just be on the.
quality of retrieval side of making, the stack much better at actually finding the needle in the haystack.
[00:20:04] Jeremy Utley: Yeah. I love the Data point that you shared with us a second ago about how, after this two minute video, you skyrocketed, tell us about monitoring use internally and how that I don't monitor sounds, I don't mean it bad.
I mean, in a wonderful way, I think many organizations don't even have the call it tooling to be able to know that. How did you integrate that capability into your process? And then how do you use it to inform?
[00:20:34] Shir Yehoshua: so we're, in an interesting place. I think a lot of, SAS tools are like this, but we heavily and aggressively use our own tool for everything.
we are our biggest, source of feedback. anything that we build, we actually test out internally first. all of our AI features are tested internally first, and we use the product analytics internally to vet, whether or not we actually should proceed and launch to a set of beta customers, alpha customers, and then more publicly,
Of both AI features to make our own employees more efficient, productive, kind of unleash their creativity. and as a, like, our poor employees who have to suffer through all of the different variations of our features that didn't actually work. long story short, we're not monitoring it necessarily from a, our company, like all employees must use AI.
We're thinking about it in the reverse way, or at least I'm thinking about it in the reverse way of, I'm building this AI feature, if it's Useful and people are using it. Then I've done a good job, right?
[00:21:40] Jeremy Utley: customers. It's not their early customers.
[00:21:43] Shir Yehoshua: Exactly. and then I'll also like the team will sometimes get discouraged because we'll launch something and then get like a fire hose of feedback.
Where we're getting so much feedback about everything that's broken and everything that needs to be fixed and like this pixel is in the wrong place. and that to me is the biggest signal that we've actually hit something right. because if you're not getting any feedback, it means nobody's using your product.
and so those are my favorite moments is when we're getting this insight on how we're supposed to improve.
[00:22:14] Henrik Werdelin: Can you tell a little bit more about. The let's call it the ideation process, obviously with the technology moving so fast and your ability to build so fast, because. in many ways, where do you go to, to find the opportunity space?
[00:22:33] Jeremy Utley: and also just adding to that question, how often do you go there? There's where do you go, but then what's the cadence?
[00:22:38] Shir Yehoshua: Yeah. okay, so I'm gonna steal from my old dance teacher again. she used to say all the time that, to be a graceful, artistic, beautiful dancer, you can't just be flexible and you can't just be strong.
You have to be both. you have to be strong in order to move, but then you also have to be nimble. and so I think the same is true, when you're developing product, you have to be really strong in the convictions of what problem are you actually trying to solve?
If you build, one product feature for problem A, it doesn't work. And then you build a different thing for problem B and it doesn't work. And then a different thing for problem C and it doesn't work. you don't necessarily learn very much and you're not like incubating Very much. and so I think it's important to stay strong in, well, what's the person we're trying to solve this for?
What's the problem we're trying to solve? And really marinate in that. but then be really, really flexible in maybe the what and the how. and so just because You have this one problem, AI might not be the right way to solve it. it might be, but it might not be. and so having, the courage to change the what drastically, but keep that conviction the same.
in terms of how often, I see it as like a pendulum swinging. there's an explore phase and an exploit phase. we just had a launch. So we were in hardcore exploit mode for the last month or two. we kind of started thinking about this in January and in January we were in explore mode.
And so for, we were trying to, the problem we were trying to solve was let's be able to answer questions from Slack. and that like, we didn't really falter on that. Um, uh, but like how, uh, we, we were open to a lot of different possibilities. Um, if in the end we would have decided actually integrating it into the Notion Q& A product is not the best way to do this, um, we would have built something completely different.
That openness to that flexibility, but then the like conviction of the problem, um, does that make sense?
[00:24:39] Jeremy Utley: Totally. Yeah. What to work on versus how to deliver it. Yeah. And knowing when you're in which mode is huge.
[00:24:47] Henrik Werdelin: Um, I was reading some of your blog posts, uh, leading into this and you had this one headline, which obviously as a product of fellow product developer is intriguing, which is speed enables quality.
[00:24:59] Jeremy Utley: Uh, do
[00:25:00] Henrik Werdelin: you mind talking a little bit to that?
[00:25:03] Shir Yehoshua: Um, yeah, so, um, I think if you spend all of your time Um, and all of your time being rigorous about, uh, like all the different edge cases and all the different things you are constrained only by where your mind can go. Um, but if you can build a quick prototype, learn where it fails, fix those failures, build another quick prototype, learn where it fails, fix those failures.
It's a much more efficient path to finding all the problems, first of all. , and you're also not constrained by the imagination of your brain of, like, where this could go wrong. You're, you're a lot more grounded in reality. So, If you're not iterating really, really quickly, you're building a high quality product for the universe of your brain, not a high quality product for like the universe of reality.
[00:26:02] Henrik Werdelin: Amen. Why do you think that is so difficult for so many product teams?
[00:26:10] Jeremy Utley: I think so. Yeah, I don't know if you I don't know if you ever read that will never work by Mark Randolph about Netflix. There's this great part where he basically talks about the fact that in the early days, they take two weeks to build like this perfect thing all the links worked everything and then nobody used it and they realize it took us two weeks to learn we had a bad idea.
They said, okay, how can we learn if this idea is bad in a week, you know, and that means some of the links are broken, but it doesn't matter. And to your point earlier, if people are adamant about this link is broken, it actually means we should figure out how to fix that. Right. But if in a week we send like a kind of crappy version and we don't get any feedback, then we wasted a week.
Right. And he said, I love the quote. It's something like it became, it was not about having good ideas. It was about a system for having and testing lots of bad ideas quickly. Yes. Which is just a totally different, you know, process. What, what role, what, one thing that I'm kind of, uh, a theme that I feel like is emerging a little bit, Henrik, in some of these conversations, Shir, we talked to David Akhniev, who's the co founder of Typeform, and he was able to build this really cool product, Formless, like AI first product, because he's the founder.
You look at Mike Newp has done at Zapier. He's the founder. Then you mentioned this mythical moment where the founders get access to the golden scrolls, you know. Um, can you talk for a second about the, is that, am I, am I just cherry picking data there, which is totally possible, right? Or is there something to a founder's ability to call it see and seize an opportunity that may be unique?
[00:27:46] Shir Yehoshua: Yeah. Um, I think it's probably easier for founders because, um, I think a big thing you need to not play into is sunk cost fallacy in order to innovate. So just because you spent the last year launching a thing, um, If the next thing to do is completely throw that away and try something new, um, you need a significant amount of courage to do that.
Um, and I think the, in a weird way, the safety of being a founder, gives you the flexibility to do that. Um, but I don't think you need to be a founder. I think I, um, when I was at Waymo, There were plenty of folks who spearheaded, like, complete re architectures and revamps of the system, um, that were not founders, but they just had this, like, really strong conviction, and they very, very importantly did not pay into the sunk cost fallacy.
Um, and so I think that's the key, the key thing, but it might be easier for founders.
[00:28:45] Jeremy Utley: Is there any kind of hack to overcoming sunk cost? Like, how do you, maybe, maybe to say differently, how do you know when you're falling prey to it and how do you overcome that kind of like the gravitational pull of the sunk cost fallacy?
[00:28:58] Shir Yehoshua: Um, I think it's cultural, to be honest, a little bit of, if a culture of, Throwing away a lot of things and a lot of ideas and being okay with that and celebrating that. I think that, uh, that helps people like almost brag about it. like my proudest moments are like the, is the code that I've deleted.
and so, uh, and so I, I think that's part of it. Um, but I don't know, I'll get back to you once I figure it out.
[00:29:26] Henrik Werdelin: I can definitely echo being a founder of a few companies myself, that there's nothing better when people come with conviction with something that is a bit audacious, right? Like where you go, like, wow, like just go for it.
I think any founder thing in these exponential kind of like leaps, right? You know, that's how. We probably got to start our first company. And so I very much agree a that it's easier for founder, but also I would just like, I guess, echo you that most founders love, love, love when somebody from the team come with these kinds of things.
I remember we, we recently, um, for bark launched an airline for dogs. And, uh, I remember kind of like pitching that and kind of realizing that. It was, it would probably have been difficult for anybody else in the team to pitch that because that it was, you know, such an audacious idea. Um, but I think to your other point is that now that we increasingly have a culture of that, we're definitely getting more and more of those kind of like, Um, not linear ideas kind of coming at us, which is wonderful
[00:30:30] Jeremy Utley: to get, to get super, uh, practical.
How do you celebrate throwing stuff away?
[00:30:36] Shir Yehoshua: Uh, um, so we have, we have an extremely lightweight process. I try, I like hate process and I try and have as little of it as possible. Um, our team, um, structure right now, we have a daily standup. That's it. No more meetings. Um, and then we have demos. No stand ups on Fridays.
Just demos. Um, and anytime somebody has learned something or failed at something, or like, idea failed, Um, uh, they succeeded because they learned something. Um, they presented at demos. So, like, some of the Some of the demos that get the most amount of applause are like, uh, a couple months ago, somebody demoed their PR, their pull requests, where they just showed all of the code that they deleted.
Um, and they were like roaring cheers from the audience. Um, and so, um, I think that, uh, It's like, once you get it started, it's a little bit easier to keep up. Um, but just demoing what you learned and demoing the fact that you discovered that this was a dead end. Um, and then celebrating that.
[00:31:45] Jeremy Utley: That's so cool.
That's so cool. I mean, I, I, there is. a metaphor of innovation is exploration has always stuck with me and if you think about kind of dividing your crew you don't know where the water is for example let's all go in a bunch of different directions and then come back and report on there's there's a stream down there right or there's a whatever right um but i think if unless you have those mechanisms for celebrating dude it's a desert that way You know, if the person who goes to the desert gets punished, you know, it's like they didn't know there was a desert over there and nobody did until you sent the crew.
Right. But that's it's really a different way of that's not what we learn in school. You know, we don't learn that in school. We don't learn that in our degree programs, maybe a little bit like in PhDs. Do people really treasure dead ends and, you know, as kind of a means of expanding knowledge, but It's something that in a lot of organizations and then the and then what gets rewarded is succeeding and doing well and what gets brushed under the rug is the stuff that doesn't work and so people become kind of institutionalized.
I think having mechanisms like that and practices like that actually are part of the. You call it the Rehabilitation Program, perhaps.
[00:32:53] Shir Yehoshua: yeah. And it's, it's that, like, that same thing I learned in my dance training. You have to fall so many times in order to, like, actually succeed, at the end. And I think there's some cultures that really value that, some that need to value it more.
[00:33:08] Henrik Werdelin: In the Gen AI space, what's the next kind of area that you are fascinated by? You, obviously the rack models are kind of all the rack methods kind of like come in as a multimodal, like where's your head going or is it just kind of attacking more of these use cases in an, in an effective way?
[00:33:28] Shir Yehoshua: Yeah. Um, to be honest, I have no idea.
My whole, my whole career, I've just sort of. I've never tried to predict the future and go there. I just stumble across something and I'm like, Oh, that seems really cool. Let me go. Let me go jump in there. Um, so that's kind of my attitude right now. Um, I think, um, I think where Where AI is probably headed is more on like, I, I get really excited about that unlocking of creativity.
If you get to spread, like, spend all of your brain energy on the creative, artistic side, um, that All right. Like, we need more of that in the world, because so many people are doing like mundane, boring, terrible work, um, and AI I think can help them and kind of unlock more. And so that's, that's maybe where I'm most, most excited, um, for the, for the creativity that we can unlock.
[00:34:22] Jeremy Utley: To borrow a phrase from my friend, Frederick, who's until recently the chief, innovation evangelist at Google. He just wrote a book this week. So shout out to Frederick called what's next is now. Henrik asked you what's next. and to borrow his phrase, if you accept the premise that what's next for most people is now for you.
What are your now present kind of top uses of AI as a personal productivity or team unlocking. What, what are your regular uses of AI?
[00:34:53] Shir Yehoshua: Yeah, so, okay, this, uh, this morning, uh, when, like right before I talked to you, I used, um, we have a prototype right now of dictation, um, uh, that some of our users have access to, and I tried to answer one of the questions that you sent me, and I it.
Talked it out. It was way too long. Um, I highlighted it, and I asked AI to simplify it, and it was a much better answer. Um, and, um, and, like, then I got to, like, practice saying it again. Um, and that was just this morning, and I do something like that. Every single day of, because my, my whole work is words and text.
Um, and so AI does all the easy stuff for me so that I can do the hard stuff.
[00:35:43] Jeremy Utley: Well, I mean, you, you know, there's a future coming here where like, this is just V1 of the conversation. And then there's an AI that just takes everything that we say, like Jeremy's question, his random riff on like the, you know, Ed Catmull thing was way too long.
We're going to now. produce the question he should have asked, and we won't even be involved. They will just look way smart right now. It's just
[00:36:06] Henrik Werdelin: me doing the edits right
[00:36:08] Jeremy Utley: now. Henrik is the AI. He's like, how do I shave the meat or the fat off of this question? Yeah,
[00:36:15] Henrik Werdelin: you have, um, for somebody who is. just kind of like started to do stuff or somebody who has to recommend somebody just to start doing this stuff and they're a Notion user.
What would you say is kind of like the, the first little thing with a little potty trick that you could either show or ask people to do that will make them go like, wow.
[00:36:33] Shir Yehoshua: Um, yeah, I think so for me, the like two biggest aha moments were solving writer's block. So if you have to like sit down. Write a blog post, write an email, write like a, like a, if I have a really difficult piece of feedback to give, I'll have AI draft a version for me first, um, and then I rewrite it.
Um, but it like, it just helps you get over that writer's block hump. Um, so that's one thing, one side of it on like the, the very first input. Um, and then, uh, on the output side, once you've actually written everything and messed with it and massaged it, um, you can also use Notion AI to improve the writing.
So any, like, the simple things of fixed spelling grammar to also, like, simplify, make it longer, make it shorter. Um, and, uh, I'll often not actually accept Like all of the suggestions that AI makes, but it, it like helps you see the space of what's possible and be like, Oh yeah, that's, that is a bad sentence.
Like, I didn't like your correction, but let me come up with a better one.
[00:37:40] Henrik Werdelin: You also, I would add, like, you have a feature where it says just continue writing.
[00:37:44] Jeremy Utley: Yes.
[00:37:45] Henrik Werdelin: And I would say often, I don't necessarily agree with what it is, but it always prompts me to kind of like go like, yeah, I didn't like that, but that actually triggered an idea and then you kind of can go back and correct.
Right.
[00:37:57] Shir Yehoshua: Yeah. Yeah. It helps
[00:37:58] Henrik Werdelin: you know
[00:37:58] Jeremy Utley: what you are looking for. Yeah. That's cool.
[00:38:00] Shir Yehoshua: Yeah. And it's also like, um, if you think about humans, like we, we have evolved to talk to each other and converse with each other and collaborate with each other. Storytelling is like conversational, usually in nature. Um, uh, all of my favorite podcasts, uh, are always like people just talking to each other, uh, riffing on ideas.
And I think that's the best structure. And so it makes a lot of sense that if you have an AI. riffing with you, um, that that would also help unlock your creativity.
[00:38:33] Jeremy Utley: I've often thought that the, um, fingers, you know, when we take it, talk about natural language, fingers and thumbs are not our natural, they're actually the bottleneck.
And insofar as I can communicate with it back going like full circle, maybe in a way back to that Apple demo where bring your not kid, bring your kid's friend to work day. Right. We're able to access our actual thinking Um, and actual processing by bypassing our fingers and our thumbs. And it's pretty exciting.
I mean, you talk about riffing, I think. What needs to be done now is humans actually need to be retrained to stop interacting through our bottlenecks and to start interacting in natural language. To me, that's like, that's a behavior change. Like, technology has trained us now on one mode of interaction that is going to be archaic.
I don't think that the best users of computers are going to be using their hands, you know? And yet, most people right now, that's the way that they interact.
[00:39:33] Shir Yehoshua: I actually think I would take it one step further. Um, computers right now Like we just have like buttons and text boxes and fingers and typing. Um, and then you've got natural language with just like pure words, but often like if brainstorming with someone, you still want a whiteboard.
You still want to visualize it. You still need like props. Like we talk with our hands too. Um, and I think what an LLM, um, or it's like all of these advancements, especially in like vision and stuff can do is they can talk to you, but they can also show you a picture or. Like give you a little, a little micro app for the conversation that you are having or the idea that you are having to explore the space.
I think that like, in a lot of ways, natural language is also limiting. Um, but if you get to have even more power from that and you put everything together, um, that's kind of the ultimate unlock.
[00:40:28] Jeremy Utley: So to riff on Ilya, you know, saying that. The hottest new programming language is English. I think part of what, part of what we're getting at perhaps is that the hottest new programming language is going to be communication, right?
It's transcending even English and transcending even the spoken word, but the way we communicate is the way we will program. Yeah, that's cool.
. We're
[00:40:49] Henrik Werdelin: incredibly grateful for you spending time. I mean, it is so cool. And I love the way you introduced like, uh, the thinking from the dance world and stuff like that. So
[00:40:56] Jeremy Utley: totally, totally. That's brilliant.
[00:40:59] Henrik Werdelin: Okay, give us, uh, give us your kind of initial thoughts.
[00:41:04] Jeremy Utley: I mean, there's so much there, right? We're talking to an AI product leader, but wow, what an innovator. I loved her references to her lessons from dance. I thought, you know, the thing about strength and flexibility is amazing. I love how they frame their experimentation and AI as a bet versus as a critical project.
I think that's actually a huge. Paradigm shift because right now in so many organizations, it's mission critical and it's the future, you know, capital F all this stuff, and framing it as a bet where there's high risk and high reward, I think is, it was a refreshing framing. I've got other stuff, but I want, I wanna riff with you.
What, what, what are you thinking?
[00:41:40] Henrik Werdelin: Yeah, I mean, like, I, I, obviously, it's so inspiring talking to somebody like her who, who. I think is kind of also this next generation of product people who is not like very rigid or very methodical, but very playful in the way that they think about like what products to build.
[00:41:59] Jeremy Utley: And
[00:41:59] Henrik Werdelin: I think you can feel that in the, uh, in the product itself, you kind of get features out there and some of them are clearly not like. Thought through to the nth degree, but it doesn't matter because obviously it's software so they can, can change it. And so I find the topic and notion, but also how she works and how she thinks to be just, uh, incredibly inspiring.
I wish I was a little bit more like that.
[00:42:26] Jeremy Utley: I think it's an interesting question. You know, they're same with Google, right? Dog fooding. They've long had the practice of using their stuff on themselves. There's a real advantage when you're. Where you're designing for your own employees are also early customers.
I feel that that your learning can accelerate. I wonder, this is just an open question. I haven't thought about like this before. Curious to get your thoughts. There's, you know, there's, there's a negative view of that, which is me search versus research, right? And we're solving our own problem versus solving a real customer problem, which, you know, may be problematic approach innovation.
But that being said, clearly designing for employees who are also first customers and getting feedback from them as first customers is important. Is an enormous unlock for Notion, Google, et cetera. I wonder how non software companies could treat employees more like customers who give them early feedback.
I don't know if you have thoughts. I mean,
[00:43:18] Henrik Werdelin: like, I, I think one of the, you know, I'm writing with, you know, my colleague Nicholas, uh, Thorn, this new book, which is basically on this topic, that is the move from customer. Uh, kind of like product market fit to customer founder fit. Basically the thesis that in this world of AI, where everything moves so fast and where a lot of stuff get democratized, you need to have this very intuitive understanding on the customer you serve by either being insanely close to the infinity group that you're building for.
Or honestly be a part of it and part of it being easier so that you don't have to go out and ask customers for feedback all the time. You can basically run that feedback loop internally if you And so I think obviously Notion is an incredible example of it, but increasingly we'll see more and more companies and entrepreneurs building these incredible companies where they build them because they're just very, very close to the customer, sometimes because the customer of themselves.
[00:44:16] Jeremy Utley: It's funny, I've seen, I would say there's two, if it's a narrow way, so to speak, there's two cliffs you can fall off of, you know, in terms of product development. For example, with a car manufacturer, one thing I saw was the executives never had to go through the car buying experience. They just got new cars and they never realized how painful the financing office was, right?
That's an example of somebody who's so, you know, insulated from their customer pain that they can't be trusted to make an innovation. On the other end of the spectrum, I work with a restaurant, for example, who is, there's, they order all of their meals from their own restaurant. They do. And they all know the little hacks and, oh yeah, this doesn't work.
So you always have to do this. And I think in a sense, they're almost inoculated. So you're either insulated by non use or you become inoculated through super use. And I don't know how to thread that needle to where You're not oblivious, but you're not such a super user that you overlook the stuff that like that, that new people or new users are going to get hung up on.
[00:45:18] Henrik Werdelin: I think that's a great point. I also promise you to have a heart stop because something else you have to do. So I will throw you off this podcast and then just beg our audience to and subscribe and share this to people. would have folks,
[00:45:29] Jeremy Utley: folks, we would have had a deeper debrief today, but I've got to go.
Okay. So I'm going to leave you in the capable hands of my cohost, Henry Quirtleman to wrap this episode.
[00:45:39] Henrik Werdelin: I was just leaving That's so fun.
[00:45:41] Jeremy Utley: so good. Have a good one. This week. So take care friends. Bye.