Fable 5 Power User Camp: A Frontier Model for Builders

0:00 All right, y'all. Let's get started. Um, I'm in a little bit of a different place today, so I don't have my normal setup. So, we're going to do some uh some improvising. Um, but I'm going to share my screen and the right button. Let's see.

0:33 All right. Welcome to Fable Power User Camp. I'm Dan Shipper. I'm the co-founder and CEO of Every. We're psyched to have you here. If you're here, you probably already know this, but just a reminder, every is the only subscription you need to stay at the edge of AI. And boy, did the Edge move this week. That's why we're all here talking about what we find there. Um you can think of us a little bit like a frontier lab for um future of knowledge work. Um and we do ideas, apps and training. So on the idea side we are constantly experimenting with uh with the models with ways to use them for our work and for our lives. Um and then we are constantly turning them into writing and podcasts and and all that kind of stuff to help articulate what we see at the edge. We also make apps. So, we build apps to um help us, you know, when we see something missing at the edge, we build something for ourselves to help us do better work. Stuff like Kora, which is an email agent and soon an email inbox. We'll be talking about that a little bit. Um Monologue, which is our speech to text app, Plus One, which is our um uh our agent product, which is coming soon again to people on the wait list, which we're psyched about, and many, many more things. We also do trainings, so we do camps like this. We do courses um and we also do uh lots of consulting so like executive leadership training all that kind of stuff. Um we are psyched to have you here. Um I want to start with talking about the moral of fable.

2:00 Um I I wrote a piece that is it should be coming out like a little bit later today, but I wanted to give you like a little bit of a preview. Um and I like I like this framing because all fables have morals. you know, if you're like if you're if you're thinking about uh if you're thinking about, you know, The Boy Who Cried Cried Wolf, Fable Moral, you know, it's it's there. Maybe Daario should should read that. Anyway, um uh I I think Fable is really interesting because there was such a divergence, and you'll actually probably hear some of this today. There's such a divergence in reaction internally when we tried it.

2:40 Um, it was anywhere from this is the best model ever and it has completely shifted what's possible to like I don't even know what to really do with this. Um, and that's new. That's that's new. And I think the reason is um, Fable is this like power tool that if you're you know we have this levels of AI scale um that that we that we do. Uh, let me just see if I can open that up. Can I No, I don't know how to Yeah. levels of AI. We have this levels of AI adoption scale.

3:16 Um, which you should to if you haven't done this, you should totally just like take this and paste it into your agent. Um, but uh uh what we found is that if you're anywhere from, you know, one to six on this scale, so one is I use chat sort of like Google search. anywhere from that to a six which is yeah it's like it's it's working as a little bit as an assistant but I'm not really delegating lots of stuff to it. um you're going to be like ah I don't know this isn't that much better because it's not it's like it's pretty incremental actually for any kind of task like a collab like for example collaborative writing not that much better um actually probably worse because it's slow uh but if you're in the seven or eight territory and you're doing multi- aent orchestration and you're writing dare I say it loops um you're going to be like holy [ __ ] this is crazy because I can now all the stuff where I was really trying to delegate. It's actually working and it's working overnight or for 3 or 4 hours and I can do amazing stuff. And if you saw the video I made, like I don't know how how many of you saw this, but like I one-shoted this.

4:28 This is not really a loop example, but I one shot at this and like it's crazy that it's possible to make a video game like this. This is a a uh version of the Bourhees story, the Library of Babel. Um so like it's it's an infinite library. has got all the books. You can go uh you know I can go from room to room. I can go up the stairs. Like this should be weeks of work. Um and now it's not. Uh and it's not really a particularly difficult prompt. I can even one of the cool things is I can even find like one of my pieces here. Um after automation um some of Katie's pieces, some of Kieran's pieces. Anyway, lots of cool stuff. Um, but I I think that that's there's this um the thing I want to point out is that divergence between the one and the eight on the scale and that the work what what has historically happened is that the workflows that developers and builders at the edge of AI edge of AI use have have leaked into knowledge work and that's like that's the historical trend.

5:32 So an example is uh if you started using cloud code where you started using cloud code about a year ago now if you're using codeex or co-work like all the same paradigm the whole paradigm the whole thing that we figured out how to do in cloud code is now a standard part of the knowledge work stack at the edge of AI um and so while fable may feel sort of weird and inaccessible to you right now um uh if you're a knowledge worker the stuff that uh the stuff that people who are like power users of Fable are figuring out how to do, you're going to be doing in the next couple months or a year as soon as it becomes more available. So, what I want to try to do with this with this camp is show you a little bit about why we're excited for it and what how how we are how we all think about it so that it might give you some inspiration for what to try and also get that back from you because we honestly feel like this is one of those things where we have not figured out um we've not figured out exactly how to um uh how to use this to its fullest extent. So, I want to give you a couple mental models for how I think about like why it's different and how it's different. Um, let me open up I wrote a piece like a couple years ago that I think is actually a really perfect one for this. Um, just open it up. Uh, it's called capability blindness in the future of creativity. Um, I wrote this in 2024 and I was trying to think about like what what will it be like when AI is very powerful and how will work change and the the way that I think about it thought about it then and I think it it's turning out to be to be pretty on point is the is the difference between being a sculptor and being a gardener.

7:16 So creative work, knowledge work, coding, it used to be being sort of like being like a sculptor. Like if you have a block of marble and you're turning it into the David, it's turned into the David because you every single thing you removed a piece of marble yourself. You you sculpted it. You you you shaped every little part of it to turn it into by hand to turn it into something. Um and that's like that's how coding used to work. That's how every kind of knowledge work used to work.

7:47 Um writing, design, all that kind of stuff. I think of the future that we're starting to peak into with Fable as being a bit more like gardening. Um, and a gardener is not directly making plants. A gardener is creating the conditions for plants to grow. Um, and so they're thinking about the soil and the sunlight and where to plant the seeds. They're watching the plants as they start to grow. and maybe you know they're growing they're growing crooked so you put up a trellis. I'm not a big gardener but like you know I know the basics. Um and uh and it's sort of this feedback loop where you set conditions you see what's happening and then as the as the plants start to grow you modify what you're doing to to help enhance the growth. And I think that that's exactly how you use this model well because you're not you're not you don't want to be collaborating with it. You don't want to be doing work for it. you want to be creating conditions for it to do the work and then evaluating the work that it's done to enhance the conditions.

8:53 Um, and that is exactly the buzzword that's going around right now, which is loops, which I think everyone's using. Few people could really understand to be honest. Like I feel like my understanding of it is like not even that incredible, but I can I can I can give you some ideas of of of how it works. Um, so if if you think about a loop as just being a gardener, I think that that hopefully should click a little bit. Um but there are there are there we've been talking about loops for a long time at every and the the like the biggest loop that we've been um we've been talking about is the competent engineering loop.

9:26 It's like you know you decide what you want to build you're you know it used to be the first step was plan but now the first step is ideulate and brainstorm on what you want to build with competent engineering and then the agent just goes and builds it. Um, so you're you're delegating the work, then you're reviewing the work and making sure it's like actually really good, and then you're compounding what what you learn from that process so that the next time you go build something, it gets easier.

9:50 Um, that's a combat engineering loop. And that's why people like Kieran, like Kieran has been, and I think he'll be on later, he's on an absolute tear right now. So, he's building uh an inbox, Kora inbox. Um uh and like this is how I do my email. I I don't use superhuman anymore. And um he is like he has gone from this working but it's like pretty slow to anytime someone reports an issue and everyone internally and every is using it and they just paste in Slack like hey can you fix this? he just has cursor with fable like go and just like churn through the backlog every day and everything gets solved within within like five hours something like that like on average about five hours it's it's truly crazy what what his development process looks like now um and it's because he has this loop set up where every time he has an idea I had like for example I had an idea yesterday for how the CLI could be better and I just had my agent make something and then he just kicked off his agent it built it and now it's live and I'm using it Um, and so once you have your loop set up, as data comes in for how can I make this better, it's just really easy to use something like Fable to delegate and have have it do really good work for you to to fix it. So you might be wondering how does this actually apply though? Um, like outside of how might it apply to knowledge work outside of coding.

11:18 Um, one thing that I've been trying uh is uh what I'm now calling the compound editorial loop. So Kate, who I believe is on this, is our editor-in chief, and I have been trying to automate Kate's co copy edits for like four years. It's my it's my white whale. Um, and I think we're getting pretty close. Um, I'm not I'm not quite there yet, but it's fairly close. And what I've been doing, what I did yesterday is I I'm trying to make a loop for for her copy edits. And so what I did first is I downloaded a corpus of all of the all of the edits that Kate has made over the last like four years and it's like you know 600 700 documents and like 30,000 edits. So it's a lot of data which is actually really interesting to start. Um then I had um uh uh Fable try to use that data to copyedit the piece that I wrote for today.

12:17 Um, so I have a version of what it thinks Kate's gonna do. Kate is gonna copy it at the piece that comes out or if or if she she may already have done it. And then >> it's in process. >> It's in process. >> Yes. >> Um, and so we're going to look at the difference between what she does and what the system did and then compound it into new data and new rules that's going to make it better. And this is the kind of way that we all are starting to think internally at every I think we all need to be thinking now that we have access to this model. So Austin is also doing this too with on the growth side. It's like yeah any one growth experiment for every is actually not that interesting.

13:02 What's really interesting is creating a system or a loop to propose to anytime you propose a new growth experiment to be able to implement it immediately and see the results and maybe even to propose experiments to start with. Um, and once you have that architecture, it makes it so much easier to um do more valuable work because you're not it's not about any individual task. It's about building the system that can once you propose a task just get it done in a way that's really high quality. And I think there's so much potential for this and it's it's where we've all been pushing for a long time. But Fable is this step function change where a lot of it is going to start working in a way that it didn't before. So that was actually the point of this piece that I I showed earlier, the capability blindness piece from a couple years ago is um what what you don't want to get caught in with AI is capability blindness. It's like, um, I tried this thing a couple years ago or a couple months ago or a couple weeks ago and it didn't work. Therefore, AI can't do that. What what we found over the last couple years and it continues to be true, is you want to keep turning over those stones. You want to keep being like, okay, it didn't work last time, but there's a new model, maybe it will.

14:13 And I think Fable is one of those one of those times where you're like now is a time where I'm going to turn over a stone again and give it a shot because it it actually can do things that other models can't that that are kind of stunning. Um so uh with that that's a little bit of an introduction. Um and what I want to do now is Nateesh who um Nateesh please introduce yourself. Nate is a senior AI engineer at every he's been he's a huge First of all, huge Claude Stan um is the one that builds our internal agent Claudi, which is what our our consulting team uses to do all of their work and I think he is extremely excited about um about Fable and has a lot of deep insights about how to use it well um and is going to do some demos for us about what he's built. Natash.

15:03 >> Hey guys, Nesh here. So I as uh Dan said I'm a senior applied AI engineer working at the consulting team and a lot of the work that we do involves making decks. So um that that is like slide decks your PowerPoint presentations which we then present to the clients. So when when I got access to this model the first thing that uh came to my mind was why don't I try to uh have it make a deck for us. So we've been using a version of um the PPTX scale that comes built into uh into claude. So we've been using a version of that with our style guidelines. But um it it it it would be uh problematic to use uh you know the previous models would fail at that because uh on the good days sometimes you would just get your whole deck done 99% there and that's great. But then the on the bad days um the model would be you know uh it would just not be able to fix alignment issues or you know like some make really stupid um uh decisions about where things should be uh in a way that doesn't make sense visually. So when I got access to the model I asked the model to um you know make a deck for me um using uh using the PPTX skill. So it I was at work and it it wasn't that much better than the previous ones. So and I and then I even asked it to uh give me the token breakdown of what it uh how many tokens it used to make the six slide deck and it was astounding. It it had consumed like 30 million tokens input tokens um which if I if we were paying through the API it would have cost us like $130. So that's uh nuts. So that's that's [clears throat] like you know what uh that now it turns into an example of what to um how Fable uh how to use Fable to grow your ambition. So then I pointed fable and asked it to actually not just make the deck better but actually look at the reason why behind why this deck was so bad. So um so then uh this is the you know mostly the prompt that I gave it and by the way this deck that you're seeing is a result it's it's it was made by Fable. So I'll get into uh you know like how it was able to do that. But yeah so it I gave it this prompt and I asked it to analyze what the problem is where is it struggling. So it was able to do a lot of diagnosis and um it suggested that we build like what Claude doesn't have is a PowerPoint application. So um it suggested that we built a CLI tool uh which would serve uh as the tool which Claude uses like a human uses PowerPoint or you know Keynote or whatever and and then it went and built that and now with this CLI tool um it's it's uh here by the way uh it's it's open sourced and um with this CLI tool um all the decks that uh we have been making like I've been experimenting with you know over the last week uh including the ones made by the uh lower level models, they're all better because this is a fundamental structural gap that um Fable was able to identify and fix for us. So um it's uh you know uh there's there's uh there's like it it's able to give you're able to use like the agent is able to use this tool to do edits like we humans do. uh whereas the previous tool would have it uh do edits like uh a programmer. So uh this this was like the clearest example for me where you had to point fable to uh a more ambitious goal in order to see its power.

18:53 So um then uh yesterday another example this was uh you know uh this this deck as I said was created by uh by this new CLI tool. Um then yesterday one of our teammates Andre was asking about a tool for uh you know screen recording um cuz everybody here uh hates loom. Um so uh I I just uh I have been having that like you know brewing that idea in my head that you know um the the basics of a screen recorder are so cheap like you know you have uh Cloudflare R2 for storing all the data and um you have your browsers like users browsers which can do all the video stuff. So um I just pointed fable at it and asked it to make something for me and I was able to make this application one shot um and it was like you know not not one shot but in one chat and um it's a screen recorder which you know which works and does all the uh you know screen recording things and um it's uh it's exactly what Loom is not like all the you know problems that I have in Loom where with all the pop-ups and you know login screens. It's like a completely login free thing which it was able to build and um uh it's it's actually uh working and um this was done on the side on one using one prompt um where fable uh just went and uh you know uh launched sub agents uh created a dynamic workflow and built all of it. So um that uh like these like are the ambitious type of examples that I've been thinking of uh to giving Fable.

20:29 >> When would I want to use a dynamic workflow? Asking for a friend. >> Yeah. So dynamic workflow is basically um uh the the workflow like it's basically the tool that lets people be the orchestrator and um it's it's uh it's sort of like an alternative to loops. So uh you were just talking about loops where it's like do this task uh until this goal is satisfied and then loop until it is done. So it's uh so dynamic workflow is another way to get there but where uh you know claude is actually able to orchestrate uh all the phases between uh sub aents and um keep iterating. It's able to iterate using a dynamic workflow as well but to do a more processoriented task.

21:15 H fascinating. So I probably could use this for my editorial my my Kate uh automation. Yeah, it's a good idea. >> Yeah, I think I I would just uh ask uh so Fable is great at uh designing a workflow. So uh I would just uh try to ask that. >> I'm about to do that. Um >> yeah. >> Okay, I want to keep going. Uh Kieran, are you on the call? >> Yeah, >> here he is. Uh the man himself. So, uh, Karen just shared that he is on track, it seems, to spend, if we didn't have a lot of credits, $750,000 in AI this year because of Fable.

21:53 >> Only at Cursor only. >> Only at Cursor, which is >> Yeah. So, >> which is how many credits do we have? We'll we'll talk about this later. If if >> I don't know, it keeps working. Thank you. >> It's totally possible that we we we go bankrupt, in which case it will be because of Kieran, but we'll it will be worth it. Um, >> I'm doing this for you all so I can [laughter] report is it worth it or not?

22:16 >> Uh, and we're excited for sure. Yeah. >> So funny. Um, >> also Yeah. Yeah. >> Tell tell us tell us your tell us what you're doing. Tell us how you're feeling about Fable. Give us give us everything. >> Mhm. Uh there. So like this is the best week of my life. Uh maybe like this is how I feel like I'm so excited and we had early access but like like you do your benchmarks you try to work but like this week I felt this blick of like oh [ __ ] like it's actually doing it. So I'm building Kora. Uh this is a new version of Kora like this not the final design or form but uh it's a full email inbox.

23:03 Um, it's going to come uh to beta users soon. If you want early access, just DM or like make some uh yeah, send us something. It's going to come to iOS, Android. And what we do now is Dan, Brandon, everyone internally is using this and they're sharing a lot of feedback uh to this Slack channel. And what I did is like I was like, "Okay, but there's just so much feedback and so much stuff like how can I leverage this new model to do all of it?" So I kind of had this idea of like maybe can I just make this promise of like if you drop in feedback here, I will fix it within 24 hours. Uh and I set this goal to myself and you can say, "Oh, why don't you use linear or like a combon board?" I I'm anti-combon boards. I think you should just ship things and skip the whole step together. So what I did was I uh used co-work actually again for a scheduled task that is like a alpha feedback pulse. It's scheduled in the morning in the afternoon and basically it runs through everything in Slack. It responds with uh I eyes. You can see here if it saw it. So it's tracked and responds with a a slash when it did everything.

24:25 And what it will do is it will create uh so this the alpha pulls it like has a state it loads the messages uh it stores it in uh like it downloads the images the videos all that and process it into uh like a brainstorm that I can use with LFG and there's a pull request. Why I like this is that it's not uh it's not continuous. So like if there are like 20 feedback items, there are not 20 pull requests. It's just too much to do. So I just say I just batched it together and I've not seen any model be able to do that and Fable can do it. So that is something that's very good batching stuff together and just say okay now uh it will kick off uh workflow. I use um work like uh Kora in the cloud on cursor and it will just call the API say pick up this um uh pull request you have LFG from compat engineering on everything and once or twice a day I will just go in and uh it's done and I see let me open here so I just uh run the polish command which is the last step.

25:48 So try here let me try again. So I use c polish uh on this pull request that comes out and c polish is showing me what to do. So, hey, how to get this feature? Do this, go here, and like this is what you should test. Because most of the time, this feedback that came in was not for me. It's from Dan or from Brandon. And I don't even know what it fixed. So, it's kind of like a a like a prime for me like, hey, these are all the things that is going to fix and how to test it. So, normally this works, but uh yeah. So what what I do is I have it on the right side here and check what it did and then if I like it I merge the PR and after the PR is merged uh my task will check GitHub and say yes this is done which which is cool like I feel like I'm in the factory it's kind of working I think Fable unlocked the batching part or like the bigger chunk and I think the bottleneck was always like having 20 pull requests, but now you can have one pull request with just more stuff fixed and also Fable is good at not messing up the LFG flow as well. So the quality is higher. You can really focus on just making it very well and really focus on like what should I build uh more the artistic or like the product decisions and these bugs is just automated. So this is something uh that's very cool.

27:18 And one one fun thing was when I realized this was cool is like I did this um I went to bed and said just merge it when it's green. I woke up and Dan and uh Brandon both says whoa this looks slick and I was like what looks slick and it just merged it. I was like oh yeah it looks slick. So like the agent was like there was the first time where I felt the factory and yeah that was very special. This is very hacky and probably I'll make it work better, but uh this works for me. Go try it out.

27:51 Slack integration with coowork is pretty good. I'm sure you can do it on any other, but Fable on Co is is great. You also get 2x usage. Uh so that's nice. >> How would you say this affects how we should use compound engineering and when we should use it? >> Um we should always use it because you you need it to have the factory. That's the thing like this works because you have everything automated like the goal is to automate yourself out of everything and I think compound engineering is a good way to do that since if something messes up you can say compound this knowledge uh you can make your own version of this obviously you can just say hey if I learn something store this to agents MD or cloth MD or whatever memory system you have but it's very important to connect all your sources the agent should have access to everything. It should uh yeah, you should set up the environment for the agent to uh succeed which is knowing what you want. uh so I think the philosophy of compound engineering you should yeah you need that always for everything and uh yeah even for little things you can like you say should you use fable for correcting copy on a copy no but if if that's one of the 20 things it does in one go it saves you so much time because you know it will do it and that is more important than uh yeah it feels like there's just an extra uh team member right now.

29:21 >> That's great. Um, okay. Uh, let's keep going because we've got we've got something really cool to show you about what you can build. Um, Jack, who is uh a senior editor at every um Jack, are you around? >> Yeah, I'm here. >> Do you want to introduce yourself and show the people what you made? >> Yeah. Um, so, hey, I'm I'm Jack. um senior editor editor here at every um uh uh just to kind of like preface this a little bit um I think like out of the editorial team I'm probably the one with the most like technical experience. I used to do some kind of like product and front-end work. Um and so I would I would be what Dan describes as like lightly technical. Um but I I I consider myself first and foremost a writer. Um and then so basically what um what I built here um you know like like so many others at every I was trying to use um you know I'm I'm usually in codecs and I was like oh I wish I had um you know I wish I had a codeex for cloud code. Um and then so once Fable came out I was like maybe I can build one. Um let me let me just like move Here, let me move my Okay. So, so, um, I'm calling this a kestrel, like the the hawk. So, because I'm picturing it as like this little, you know, bird companion that kind of sits on your shoulder and then goes and does things for you. Um so uh one I think like one main piece of this is so so so first of all you know uh you kind of have like built-in browser um one of the key ingredients of the codeex desktop app and you know I added things that were some like my favorite aspects of um of like apps like codeex and cursor which is like one of them is like the ability to annotate um on the web page and have it do the work. So, uh, what I have here open is basically we've been working on creating these benchmark pages for the site. Um, and this is one that is Katie's writing writing benchmark. And so, we have this page and you know, for instance, like I was looking at it and this text is kind of small. So, I can kind of like highlight it and say, uh, the labels are too small.

31:47 >> Can I just pause you? This is just nuts. So you just rebuilt like Codeex but using Fable so that you could use Fable and Codeex. >> Yeah. Yeah. Um, [laughter] >> so each >> So I think like it's like I'm I'm very like particular about like the way that my apps work. And I think like one of my pet theories is that like basically like AI turns everything into user experience like like user experience in the more general sense of the term rather than like specialized product focus because you know if if you're thinking about like okay this is how I want this thing to work these are how I want these interactions to work. Um, that's all it all comes down to like the feeling that you have when you're actually using uh using the app. Similar to uh similar to kind of the way that um uh like uh uh yeah, conductor does like basically this is running off of my existing um claude and codec subscriptions.

32:50 Um, and it has so therefore it has uh all my existing connectors that I have linked up to those accounts. Um, so you see here it bumped up the SVG label size and we can just kind of like refresh it's if it's going to work. Maybe it's not going to work. Um Um, so yeah. So, I can I can kind of like let that run. Um, and so the other thing I wanted to show you is uh one of the things that you know Dan's um been doing with his inbox sweep that um I don't know if he's showed here but is is working on is that basically like you know you want uh your agent monitoring um whatever you're working on whether that's like a web page or sometimes like in my case sometimes it's like a markdown document in a loop and and operating it on it in a loop. So what I can do is I can trigger this sweep here.

33:50 And what I have this sweep set up to do is basically like the first time I run it, I tell it what I want to do on this file. Um, and so what I've told it is that like anytime that I have like um um here, let me actually stop it here and and restart. So, so basically like um anytime that I have um uh TKS or things in brackets um it will then as part of its sweep pick up those indicators that uh you know when I'm writing I drop them in because I'm like okay I want you know I want like uh another like word here or I want like a different thing or I have to do research thing and what it's done is um on the sweep it's picked up that these are the things that I want wanted it to do and so what I can do is have it execute um and now it'll go about go and execute um those things so let's see if it writes us another good um verse about uh falcons okay So, um, yeah, I I think these the foreign forms >> this this is nuts. This [laughter] is so so crazy that you built this.

35:16 >> Kieran, are you watching this? Is this cool, Kier? >> So, so okay. So, so the part so how I built this >> how I built this was actually I got the basic project running um and and like like uh you know model dropped Tuesday. I didn't have early access to it like the the rest of the team. Um, I actually had the day off on Wednesday and so Tuesday night I was like, "What what can I do with this?" And I basically like built enough of this project to have the sweep and to have like an agent, you know, to to be able to use my my Claude um, and Codex accounts. Um, and then like Wednesday morning, I was working on it. You know, it was like there were little things that weren't working the way that I wanted it to work. Um, and so I was like, well, like, you know, why don't I just like try to use it to build itself? And so basically I have this this is the actual project repo for this app that you're looking at. And I have this document here that um, basically I'll, you know, here, right? I'll, yeah, I'll trigger another sweep. But this document is like the things that I want it to do. And what I've told it is every sweep if there's an item that feels like it's like u manageable and doesn't require too much work just like go ahead and do it and like make it a part of the sweep. Um, so what I can uh so yeah, we'll we'll let that run.

36:44 And what it should do is it should kind of like pick up several of these things um that are like kind of the lowhanging fruit. Um and and like once it's completed with the item, it moves it into the done section. So basically like this is like everything that I've done on this app like all the tasks I've thrown it. Um oh it says nothing new. Okay. So Oh yeah. So I I think I think what it does is it actually like um looks for kind of like compares it against the previous version. So I have to actually ask add something.

37:28 So like like Wednesday morning I was just like doing doing my laundry and like in between folding clothes I would think of things to add and I would just add it and I had the sweep running. Um so here let me let me start over because it'll um it'll uh be a shorter timer. But basically like I had it running like it would sweep and then automatically do those improvements and then like refresh and I would like fold some clothes, put some clothes away and then come back to it and like see some other things that I realized I wanted change, go do those things like like add those things to this list. Um and then like go off and do do other errands. So um we can uh let's see we can kind of like come back to Yeah. So you see you see it also like you know when it did that it replaced those. So um let let's see if I let's see if I I can actually like write a haiku verse to to complete this. So, Falconer Wits.

38:40 So, all right. And then we'll go back to to Kestrel. And it still has nothing new because um let's see. Uh let me ask you a question like um uh or there's nothing new times in a row. Can you collapse the messages?

39:26 Don't do it yet. Just answer. Um, okay. So, let's go back to our our camp. So, okay. So, it's says complete the fourth hyon. I can also have it just like auto execute. So, every sweep it'll just go and execute. Um, and there you go. So, two talons clushed, a light landing on the worn leather glove. The meadow goes still. That That's a pretty good like I feel like that's a pretty good completion.

39:58 >> That's amazing. um to to that start. >> I've never been more had more wrapped attention to watch someone write a haiku. Like that is [clears throat] that was incredible. So cool. >> Um let me let me just see if I can I can actually get this get the sweep to work cuz uh um but yeah, like I think like thinking about, you know, like thinking about the Codeex app, right? like the Codeex app, it's meant for primarily for developers and we're seeing, you know, seeing folks that are like, you know, starting to use it for uh other types of knowledge work and, you know, we at every are starting to use it for other types of knowledge work. But then I think there's still certain aspects of it that are primarily for developers. So with this, I'm like, what if, you know, is there a way for me to actually um like create my own version of it that is more um supportive of the wide range of tasks tasks that that I do. Um, so what it's going to do here now, it's going to like hopefully answer this question um about um about collapsing the messages um Yeah. Do do you guys have any questions about this? So, this is so new that like most of the team hasn't even seen it cuz like it's literally been 48 hours and I'm just like putting it together into like an app that the rest of the team that I can actually like share with the rest of the team.

41:51 >> Uh I I I'm sure there's a ton of questions. One one thing that I would love to know um and maybe we'll do one or two questions, but one thing I would love to know. So for sweep, I would actually prefer it to allow me to if I type something in there to then steer it. So it like sends a message. So I could like, you know, if I do a bracket or TK, it just steers the chat. Does that make sense? Yeah. Rather than the chat having to pull that kind of >> So So why don't we why don't we do something like that? Um >> that's what I was thinking.

42:22 Um, so, okay. So, Dan, uh, here, let's, um, what what what what do you want to sweep? Like, do we want to do a file or just like a text file or like >> maybe a proof doc or um or text file? Text file is fine. Yeah, that >> Yeah, let's do a text. I think there's still like some connection issues that I have to um so start a new text document called dan.md and open it.

42:54 Um and then I'll like really quickly jump back here. So you know what it's done is it's it's gone through that sweep and it's actually answered my question. Um where oh well it opened it here. Wait. All right. Let me open that here. Um. All right. So, so now I'm going to like initialize like start the sweep. And if you're doing it for a new document or a new browser, new surface, what it does is the first time around, it actually says, I need to know what I what you want me to do. So, like Dan, what how do you want the steering to happen? Uh whenever I uh whenever I type TK, I want it to then activate the chat um to take action on whatever I said TK about.

43:50 >> Okay. but in in the chat rather than in the document itself. >> Or >> uh well I want basically I just want the AI to run whenever I do TK >> and do whatever is it seems indicated by what I type TK on. >> Okay. Um yeah I need to Okay.

44:21 Let me >> you can just do it on XI probably and then we will share the result later. Yeah. >> Um >> yeah, I want basically what I'm asking is I want it to not be on a sweep, I want it to be on demand. So like instead of >> Oh yeah, >> instead of it polling, I want it to uh be activated. >> Yeah. >> Um so so yeah. So right now it doesn't um I don't have that built in yet.

44:48 partially because it's, you know, it's like it's uh it's very costly with with Fable to to do that. >> Yeah. >> Um but what I do have it do is um Okay. So So see it says the brief saved. Um and now it'll execute. Um so what I do have it do. So So let let's let's do an example like TK. What what do we want it to do? Uh uh TK title for my next piece about uh why like about how Jack completely like blew my mind with his Kestrel demo.

45:34 >> Um how about title and intro? >> Great. Um, >> all right. So, >> um, >> um, we're getting some some some comments about like people are maybe a bit lost. Um, >> okay. >> And I think I think we're getting we're getting pretty far into Kestrel and I want to like I want to pull us out for a second. Um, >> and for someone who's a bit lost, can you explain how they should think about using um using Fable from from what you've learned or observed so far over the past 38 48 hours?

46:19 Yeah, I think from what I've observed so far is one it's that like Fable even at low um low effort level is like really really capable in that like I can trust it to do things do like smaller tasks um accurately in a way that I couldn't trust previous models. Um, and so that's one piece of it. And then the other thing is like to think about like, you know, uh, I I think like it took me a while to wrap my head around this idea of loops. Um, and I think the, you know, the the the thing I was using, um, the the basically the the text file that I was using to actually like work on is an example of like it's like I don't want to have to like prompt it every single time. uh I have like an idea for something, right? It's like I just want to go about my day, capture my ideas in like the same note. Um or or you know monologue like monologue it into monologue notes. Um and then just have it automatically like like act you know like like act on that note or like check that note on a regular interval and then like perform actions based on what's >> he said I said thank you again cuz I talked to his dad yesterday. Okay.

47:42 >> Okay. [clears throat] >> Keep going. >> We I think like transferring more model to the agency. I see uh Andrew's question. Um I think it's like or trans transferring more like agency to the model. Like I I I think it it's that like I I did want previous models to take like to have more agency, but I couldn't always reliably trust them to do the right thing. Um and this is the first one that let's like it's right often enough that like I trust um being able to kind of like hand things off to it.

48:27 Um, I think the other thing is like, you know, it's like it's like I I'm like I I I really don't like multitasking if I can like, you know, if if I don't have to. Um, and so, for instance, like I we also have like a young child. Um, we have an almost two-year-old at home and like it's like when I'm watching him, I don't want to be like, you know, having a a chat with, you know, with Claude or with whatever bot, right? It's it's sort of like I want to like, you know, be able to like pay attention to him, but at the same time, I do want to like capture my thoughts. And so before this, what I would do is, you know, I have a notes app that like, you know, basically like every time I had an random idea, I'd save it as a note and it would put it into my like daily note in Obsidian.

49:20 Um and I think like how that workflow how that flow has evolved over the year over not the not even the years over like the weeks. um is that basically like like two weeks ago I was having um uh I was having Claude basically scan that note file every night and then like you know take ideas and stuff and like organize them into other notes that I have that are more like about projects and ideas. Um, and like now it's like I feel like now like there there's kind of like two things like one is like I can actually do you know um I I can do what like what Dan's you know uh Dan's demo example here is doing is like I can not only like just like save ideas and notes for it to organize. I can actually like save tasks for it to do. Um, and I think the other thing is that like um like with you know with writing the haik coup is that like like it's also reliable enough that like I can simultaneously have it run live in the same document. And so you know it's like I'll go about and I'll like write a first draft of a piece and usually I'll kind of like stub out, you know, stuff that I need to do research on or like like yeah I need like a transition sentence here, right?

50:49 and and like you know I can either have it like suggest those things or I can like have it you know like actually like write an example of those drafts but but it's it's sort of like simultane like both of the two things that you know I think like we we talk about at every of like having one agent that you kind of like delegate everything to and just like don't care about being in a document live and then have the other that you're like working with simultaneously while you're in the same file. Um, but like but I think like both of those things, right? What you need for both of those things is actually like you need an agent that is reliable enough to to like you know do all those things um like somewhat accurately. Um, and I think like you also need uh, you know, for the real-time collaboration, you need an agent that is like fast enough to actually be able to like respond um, in real time. Um, and so, so I think like like I found that with like Fable um, Fable set to like low effort is that like that feels like it's like quick enough for me to actually like um, do Yeah. to to do. Um, >> hey Jack, I want to Yeah.

52:07 >> I want to um I want to keep keep moving a little bit. I mean, this is incredible. Uh, and also I love that it just wrote the article and I actually think um it's it's not a bad first draft. Um, but I want I want to keep I want to keep going because I want to get we have about uh an hourish left of the of the camp and um uh I think Austin I think we're going to move into breakouts. Austin, did you want to tee that up? Yeah, I um we're gonna move into breakouts next, but I think what might be helpful is to like I so I lead growth here I would definitely identify as like a a knowledge worker type. And so before we do breakouts, I'm going to show and explain a prompt that I actually started running this morning that might give some inspiration for like how how to use Fable if you haven't yet in a way I've really liked. So, I'm going to share uh my screen here um if I can. All right. I have a new computer, which is uh unfortunate because I need to reset to share my screen. So, I'll I'll talk over the prompts real quick. Um so, what I did, I've been using Fable inside of uh Cloud Code and the warp CLI. And what I did this morning is I uh my team on the on the grow side is trying to revamp our subscribe page to drive more paid subscribers.

53:29 >> We can't see your screen if you're if you're sharing. >> I have to like reset Zoom because I have a new computer to share it. So I'm just going to talk over it to just show the prompt, which I think will still be helpful. So I um I went into Cloud Code. I have Fable as the model on a high setting. I point it at all of the um sources it might need, which is a Slack thread where we're coming up with this problem where we want to design a bunch of um experiments on our payment page to have more conversion. Um all of our meeting notes where we've had conversations about what that could look like. Uh Figma designs from our designers for like what the ecosystem could could look at. And then um I asked it I did a monologue where I was like hey here's a thing we want to set up the the problem I want you to solve um doing a LFG compound engineering flow so that I can it can brainstorm plan execute and review um go on that path to one identify what's up with our current subscribe flow how it's doing look at the post hog analytics for it look at all of the events see all of that Um, go do an audit of all of our competitors.

54:41 Um, I'm going to give you like a ramp card. Actually, go pay for competitors that are interesting. See what they're doing that's working. Audit all of that. And then, um, come up with a stacked ranked list of 10 things we should ship to improve conversion on our subscribe flow. And it's been running for a few hours this morning. Interestingly, the first thing it came back with was was that we don't have the infrastructure yet to automate experiments on that page where we should be able someone has an idea in Slack kind of what Kieran was talking about someone has an idea in Slack and um you can just go chip it and that was the first thing it came up with which is very cool because I think a thing that I found on the knowledge work side that might be holding us back from using it the most is that you need the architecture in place to come up with loops and so what it's doing right now is building that experimental architecture before it starts shipping all of the experiments for the subscribe page. And so I think about it that way has been really helpful for me of what's the actual problem, does it have access to it to all the sources for it? And then have I been really clear about the the final result I want? Um and so as you all as we think about like in breakout rooms like what kind of thing might you want want to use Fable for?

55:57 Like that has been incredibly helpful for me. Sick. I love it. Um, okay. Do you wanna Do you want to um kick off the the breakouts, Austin? >> Yes. You want me to try? Yeah. >> Uh why don't we why don't we both try if you >> I know how to I I I can do it. Let's just um do you have a vision for how we want to do how you want to do the breakouts?

56:18 >> Yeah. So, um we're going to give like 20ish minutes uh for breakout rooms. While you're in there, please uh introduce yourself when you when you show up, say hi. And then I do think it'd be good to frame it this way, right? of talking about like say your experience with this stuff uh how much you've used it. But I think either as a as a group or maybe like individually talk about like what you would want to use Fable for and how and then when we come back I think it'd be great for a few people to say like here's the you know the system kind of similar to what I described that like either I want to try or we decided we want to try and we can all jam on it together. Dan and Katie, Natasha, Kieran and I can hop in and say like this is great. Here's how we think it would work. here's some things we might want to improve on it.

57:00 Um because I think coming out of this to get a sense of I actually want to go ship this on a personal project or a work project would be really helpful. >> That's great. Yeah. And so we're going to make it'll be about 45 breakouts room breakout rooms. We're going to aim for like five or six people per room. And uh I think also so like introducing yourselves, talking about what you want to use Fable for, getting feedback from people, but also just know this is just a very um it's a very special group of people. Everyone here is like really trying to use these tools in new and interesting ways. And so this is your opportunity to get to know people in the community that are trying to do the same kinds of stuff as you are. It's also the first time we've done this, so we definitely want feedback about how to do this better. But just want to give you guys a chance to all meet each other.

57:45 We'll be in the breakout rooms, too. We'll be hopping around. Um, so we're going to take about 20 minutes and we'll come back at around 20 after the hour. I think everyone's backish. So, what I'd love to do is if you were in the if you were in a breakout um and you want to share and and tell us about what you talked about and um uh you know any any takeaways or questions that came up like please uh raise your hand and we will call on you and um yeah, we'll take it from there. Uh Lauren, do you want to kick it off?

58:26 >> Sure. Uh I'll be quick because there were just two of us. Olaf uh from England and myself. Olaf is a photographer. Uh I'm a consultant uh early in AI uh here in Northeast Oregon. Uh we we just shared our experiences. We're we're not uh probably as caught up to some things as the rest of you all uh actually are. So, we're just enjoying these sessions. Uh trying to read small text on a screen. uh trying to keep up with you and um uh definitely here to here to learn and here to help wherever we can. Uh Olaf, if you want to add anything to that. I don't have much more to say than that for our conversation at this point.

59:12 Uh I don't really have anything to add to it. It was very interesting to meet Lawrence, see we're in the same position, but basically um a lot of experience with AI in terms of image generation, but not at the level that we're talking about here. So my presence here is to learn from a lot of the knowledgeable people that seem to be on the call. >> Amazing. Well, we love to have you here. Thank you both for for coming and for for participating. And any way that we can help you get more out of this, like please let us know in the chat or or send us an email afterwards. Um, all right. Let's keep going. Uh, let's say Isaac.

59:53 >> Hey, uh, Isaac Parker. There's another Isaac here. >> Oh, Isaac Rodans. I was thinking Isaac. Yes. >> Okay. Go ahead. >> Uh, >> his chair is empty right now, though. >> Is it? Okay. Uh, was gonna answer for him. >> Yeah. Uh, Isaac Parker, did you have something you wanted to say? Uh not not so much. Um I'm just kind of enjoying. Thanks. >> No, no worries. Okay, let's keep going. Uh Kelton.

60:24 >> Hey, thanks Dan. Uh so our group was uh Rob uh uh Brian and myself. Uh Rob Rob and I are I would say tech I speak for Rob technically adjacent. We're not developers. Um uh come from knowledge work. uh uh were users, put it that way, where uh Brian um was a software developer um had spent a lot of time in management and we we all discussed on how how this is going from the uh way of old software development from where the the tool is several stages removed from the actual user and by the time you actually get to the user it's you know the tool is nice and it looks good but it might not work to where we're all talking about tool tools that we're making for our teams, whether it be, you know, supply chain or in accounting and finance. Uh, and it's just it's revolutionizing how we work. Uh, also, you know, a lot of, uh, praise for monologue and the way it helps, uh, expand how you can convey with, um, your AI agent.

61:29 >> That's awesome. Uh, Kell, I'm curious, what are are you building anything in particular um, like for your teams? And if so, how are you using how are you thinking about you you might use Fable for it? >> The what I told the uh the guys on our call is in my mind right now I'm only about two months into the whole Agentic um space longtime user of uh Generative but um my capstone uh that I would call is I want to create a command center for our finance operations. Um our organization we're a nonprofit. We've grown grown 50% over the last several years and um we have outgrown using QuickBooks online from a reporting and workflow standpoint. However, the to graduate up, you're dropping 2550. I mean, a lot of thousand dollars for a, you know, not so large organization. Um, and it gets a lot more technically uh involved.

62:23 Whereas, I could build something to fit exactly what our team needs from advanced workflows, segregation of duties, controls, approvals, uh, reporting, uh, dashboarding, things like that. um and use that as our finance command center uh and utilize APIs with the different providers we use like bill.com, Expensify, uh QBO, whatever else we utilize. Um and it be auditable. That's the big that's the big deal is when it comes to our annual audit, being able to turn around and essentially provide our auditors with, hey, here's the receipts. Not just the receipts that you're looking for test, but the receipts of how the workflow went.

63:05 >> That's really cool. I love it. >> I'm I'm I'm a ways out from doing that, but that's where what's in my mind right now. >> Well, keep us posted uh as you make it. Would love to see what you what you make. Um let's go with Cartik. >> Hi everyone. Yeah, so we had a group with Diana and Marcus and couple people dropped by and then it disappeared. But we mainly talked about the budgetary constraints that you know us uh single users have because it's like I'm already in a max plan. I had a you know it was a $100 max plan per month and then once fable hit it's like I had to change it to a $200 [laughter] max plan within like two hours. So and then also the feedback we have was there are so many different tiers of users here right in using experience. So it's like some of the demos were a little bit over our heads.

64:01 So we probably need to kind of figure out, you know, how to demo certain features. But again, you do have it as an advanced, you know, camp for fable, but still. And then we were really interested in what Austin had to say about the growth of uh growth hacking or go to market and we wanted to kind of get more details on that. >> Really cool. H how are you think I think this is a this is a question that we should probably be able to figure out how to how to help people with and it's coming up is like uh thinking about how to budget tokens when to you know have a max plan how many tokens to use like do you have any personal feelings about how you're going to deal with this going forward? Well, because it's >> the way I'm thinking about it is I have till June 22nd, right? And I'm going to use X high till June 22nd and see what I can get with it. You know, that's the way I'm thinking about it. And once I get to that limit, right, then I'm going to probably drop down and see, you know, because everybody talks about factory to do like a harness so that you can use different agents because I'm still not there. It's like I'm like what are we like if I'm going to get the best model do the best thing why am I you know budgeting it to me if I can keep going you know wa watch a soccer match for like four hours come back and reset my limit that's the way I'm budgeting but >> this girl is like >> uh that that makes total sense Bridget I'm I'm gonna mute you um let's just keep going yeah I think that makes total sense we're gonna keep we're gonna keep rolling on on answering that question, but I I yes, >> I feel you on using using max plans uh uh while while we can.

65:47 >> While we can. Yeah, because once it gets to, you know, because I already have like $200 on the extra usage because they give you a hundred, so I put another 100. I'm like, then, you know, we'll hedge our bets, see what happens where [laughter] >> Great. Uh let's keep going. Matt, >> hey everybody. Um yeah, no I appreciate you guys and I just want to say Dan like ever since I came across Avery um it completely has changed the way I see AI 100%. Um executive vice president for an agency and um run our strategic consulting division. My challenge and I think what we talked about um is you know how do I build an AI operating system that I can help clients build and I work in the tourism and destination marketing industry that is able to do the things that I see you guys accomplishing but doesn't rely on you know a very expensive model to accomplish like how do you build a system that's just complex enough and capable enough but can be done in an organization where a lot people are using $20 accounts. And I think that's my biggest challenge. Um, and I have I feel like a lot of times been myself around a a poll trying to figure out like how do I build these systems so that they're complex enough to be capable, simple enough for like people who don't subscribe to Avery or don't like live in this world all the time and you know to to get use out of it. And I and I think that's that's I live in that kind of adoption challenge world even in my own organization. And so we talked about that a lot. Um and I also want to say that monologue for the record has changed my entire life. Um I now go out every day for a 30-minute walk. And you know, my first wow moment with Fable was like having it improve the pipeline so that when I go out for my 30-minute walks in the morning, automatically taking that uh monologue transcript and filtering it out into, you know, here's my task, book that, here's my knowledge, you know, for the day, my alpha knowledge that I'm like wanting to store. And so, um, Monologue has made it so easy. And when I say it's changed my life, it really has in that um yeah, I walk about 10 miles a day now. And uh yeah, that's entirely because of you guys' app. So my girlfriend just wants to say thank you. Down five pounds and um yeah, just sincerely appreciate all the work that you guys do, >> man. I'm gonna be hitting you up for some uh testimonial stuff for every that's incredible. Thank you.

68:25 >> Yeah. Yeah. Absolutely. Lifechanging for for somebody with ADHD for the record. Uh and I have a really nasty case of it. Um, having something like Monologue where I can go out and be a divergent thinker, um, and it and it just works still is, yeah, life-changing. >> So cool. Uh, Katie, did you have something you wanted to add? >> Yeah, just speaking to the whole thing about like how to get the most value out of these systems when you're working on those lower plans. Something that I've really um and just talking to some of the folks at the labs, they've really underscored this for me is just the importance of with these more sophisticated models as they get smarter and smarter and smarter. Um just having the right context in the right places and really thinking about how you organize your context in a way that's going to make it really easy for the agent to find what it's looking for. So like agents.mmd like your markdown file with like that's the first thing in that an agent knows to look for and it's going to like you know treating that as a directory where it like points it out to like you know like I have a context MD folder that has a lot of my preferences about like me personally in terms of like even like as granular as like don't like use contractions cuz like codeex never wants to use contractions like on its like default models. Um, and so like I have like information about that in like the folders. And so like like like this is like something that OpenClaw honestly playing around with OpenClaw really like underscored for me is like the way that you structure your information really with these models that are so smart and are built to like make use of context it really makes a huge difference in how much how far you're able to go and how quickly.

70:10 My other one piece of advice is that I I think um for the kind of situation you're talking about, Matt, I I think Notion is just getting really good at this, especially with their ability to auto select into cheaper models, its integration with codeex is is quite good. And that I think one of the best things we we do to level up is that we have an all meetings database so that all public meetings sit in one place and anyone can query it however they want.

70:36 It sits in notion. Granola has this with an MCP. That's actually like not that costly um to do. It doesn't require a a like high-end model. We're working on integrating it with Monologue, too. And I find that like getting that kind of early win with with everyone. Um even for maybe a set of meetings everyone's down to have recorded and shared is a real way to level up and build momentum without needing like a high level of cost.

71:04 >> Sweet. All right. Um, let's do at least one more. Austin, I just want to check on time. So, we've got 25 minutes left off. How much more time do you want to spend on Q&A on uh, you know, sharing versus other >> Let's do two more questions. >> Okay, cool. Anna, hey. Um, we were a group of four. Uh, there was Arlo, Rajie, Am from New York to Vancouver and to Warso in Poland myself. Um really useful insight for me is I work in health tech and there's lots of uh compliance rules and anything you should uh be able to uh not to use AI for everything but to anonymize data and stuff and we thought that table could be used for building a workflow that includes open source models um like on device to anonymize data for health tech uh require requirements and second part is tokens limit is a pain even if you have max with new models like we discussed previously and thank you for the real life events in New York City uh they say it's really cool and and um I would love to have one in Warso in Poland so I will contact you later I don't know if you have the community here but I'd love to host one cuz I think that it is a really special community and uh Europe also has to have a chance you know to attend. So thank you.

72:38 >> We would love to do that. If you if you want to host something please reach out. If anyone here who wants to host something in their own city like you're totally welcome and we will absolutely support you in doing that. Um that would be amazing. >> Um uh last one Tessa. >> Oh that gave me a shock. Uh hi guys. I'm Tessa. I'm the CMO of a B2B fintech startup called Seintoint in London. I was in a group with Carol who's in El Salvador which with much nicer weather than me. So that was upsetting immediately. Um we shared some really interesting ideas around like using this for design and also some of the knowledge work examples that Austin shared I found really useful like the sort of subscription experimentation guide. So like I actually had my agent listening to him when he was describing the prompt and come up with ideas for this breakout. So I totally cheated and one of the ideas was like have Fable work on a stackranked list of experiments that we should run for Coint to increase the registration rate based on just pointing at all the context that we have. like all of Notion, all of Slack, all of our customer interviews, everything. And like have it run all of that like multiple source work over several hours very quickly and come back with stackranked experiments rather than just like a kind of vanilla list. So yeah, hope that's helpful for anybody else.

74:04 >> That's awesome. Um, okay, Austin, take it away. And who if people want to reach out to schedule one of these events, who should how should they do that? Let's we'll include it in the follow-up email. We'll like get organized in this. Everyone who RSVPd will get a follow-up email. We definitely want to do more of these. So, we'll find either the right inbox for people um because we really want to make that happen. Uh I think next I'd love for um Katie to to come on with with Dan and with me because I think the a thing that's really interesting with Fable and I think we want to get into is like how it both is and isn't good and helpful for the kind of like writing taste stuff that Katie digs into. So, well, and then also for the last part of this, I think when we're reaching for fable versus when we're reaching for codeex or co-work, Kieran was getting into a lot of that.

74:53 So, uh yeah, Katie, take it away. >> Yeah. Hi guys. Um I'm Katie for anyone who doesn't know. I'm a staff writer here at every squarely in the non-technical camp. Um I am the normie. I speak or the Lorax. I speak for the normies. um don't have a lot of um technical background and really when I'm talking to a model I literally just talk. Um and I'm a very sort of impulsive prompter. Um I I like to like thank goodness for steering in codeex because half the time I forget half of my prompt and then I need to go in and say oh wait there are like three more pieces of information you need in order to be able to execute this well.

75:33 Uh and sort of like to get into fable that's not something that works with fable super well. So to sort of like to to start with something that I've had to learn over the course of working with Fable for the last year and a half is uh or year and a half, it feels like that week and a half [laughter] um is that you kind you you have to invest a little bit more time upfront in sort of specifying um the like the outcome that you want um the sort of like like what your vision is for what you're trying to accomplish.

76:03 Um, and once once Fable gets going, it's going to show really good judgment of like finding the resources that it's that are available to it to execute on that ask. But if you're not investing the time up front in um in telling it what the outcome, what good looks like, it's going to come up with its own ideas and and Fable's ideas may not be your ideas. And then you kind of like I think the way I think of Fable is that it's a it's a steam ship. And like if you've ever seen the Titanic, it takes a long long time for that ship to turn and sometimes you hit an iceberg and then you sink. And we're not going to hit icebergs with our models, but we are going to like, you know, spend through our tokens as we've discussed on this call. And um and so that's something that that's just like a tip off the top of my um my head. But let me share my screen and just show you some of the things that I did with um Fable. If I can find the right uh Okay. Um side by side.

77:09 Can you see my codeex or No. >> Oh, wait. >> Not yet. >> Oh, well, here's >> now now we can now we see something. Yeah, >> we see >> both your We see you and part of Codex. >> Okay. Okay. A part of Codex. Okay. Here. Or here's Okay. Do you see cloud code? >> I think you're only sharing the the browser. >> Oh, this is my nightmare. Okay. Um, let me try again.

77:54 They moved everything around. Oh, here it is. Okay, here's Claude. Thank you. Okay, so like and so like just to like give you some >> Does everybody I'm seeing Katie's face in the screen share. Is that the everyone else's experience too or >> No, we can see her cloud now. >> Yeah, I'm seeing I'm seeing >> you're seeing it fine. Okay, then >> Oh, >> yeah. Yeah, she's sharing both her face and I don't know how she's doing it though. I've never seen anybody pull that off.

78:24 >> I don't know. >> I don't know why, but I can't. [laughter] Yeah, just try sharing your your whole screen. >> Share the entire screen. Share. Okay. >> There we go. >> The same thing. >> No. [laughter] Okay. I don't >> You have a Google monitor, Katie. Is that what's happening or >> No, I'm on my little MacBook Air in a call booth in the general the GP office in San Francisco across the street from Dan.

78:54 >> Okay. Try turning your camera off and try turn your camera off for a sec and try that. Try that. And now let's see what happens. [laughter] I've never seen this before. I truly >> Oh, someone shared presenter. >> I'm screen. >> I can also see screen. >> Yeah, >> I only see part of it. >> Part of it. No. The font is so small.

79:25 Cannot read it. >> It's got to be a setting. >> She's setting it side by side and she should do content only. I think >> is Katie pinned as a in zoom. >> It's a zoom setting for sure. >> It's a presenter. Someone said presenter layout content only. >> Not side by side. >> This is how we know AGI has not arrived is Zoom is [laughter] still acting this way.

79:55 Katie, are you able to change your presenter setting? >> Yeah, I think I did it. >> Okay. >> Sorry. Is this better? >> Nope. [laughter] >> Okay. Why don't you just talk us through it? Um and turn your video on. Talk us through it and then uh and we'll figure this out. >> For the people who can see what uh her screen was showing like because we could some some of us could see. >> Oh, some people can. So, so so talk us through it. And I'm sorry if she can't see her see her screen. She's going to try to >> uh >> I'm going to think about my life. Okay. So, um yeah.

80:31 So, some of what I did um so again like um a lot of the my breakthrough in getting the most out of not just Fable but also um codeex is just like file architecture and making everything really clean and organized and able for the agent to find it. So, for example, this is my list. I'm working here in my working overtime folder. Um, where I've actually been workshopping a a bunch of different ideas in my idea farm. Shout out to people who know about idea farms.

81:01 If you know, you know. You want to constantly be you never want to like run out of ideas because if you run out of ideas, you're in for a bad time when it's your turn to publish and Kate is relying on you. So, I kind of have this ongoing list of ideas that um that I'm kind of noodling over. And one of them was this idea about um flying a little too close to the sun with my vibe coding. And so I um I built an MCP into my app Tastemaker um with with Opus 48 when I was doing the Opus 48 vibe check and was feeling on top of the world. As Dan says, never make any major life decisions within 30 days of a meditation retreat, a psychedelic experience, or contact with a Frontier model. I made a a life decision to add an MCP to my app and that was a mistake and then Codeex caught the mistake. And so this was the idea for this piece is like I flew too close to the sun and AI caught me. And so this is where I started with Fable. I have all this context from interviews I've done. It has access to my code bases so it actually knows exactly what happened better than I do. And I just said I'd like to return to the idea from my idea farm about flying too close to the sun and AI catching me. there's a SQL brewing that might be part of the same piece because I've gone back to that code base since with Fable and I've added the MCP connector back in. Can you look around at what I've been doing with the code and see what might be interesting to talk about? And so that's kind of like like kicking off an interview process. But what you'll see when we get down here, this is a massive wall of text. And so this is a lot to sort of deal with and like I don't this isn't necessarily useful to me as an like as a interview process. So that's So for some context, if you didn't read the vibe check, I'm the yellow. I was not green or gold on this model. I was situational because I I have a hard time imagining when I would use it. And stuff like this is why because like it just doesn't have a good sense of how much information I'm going to be able to reasonably consume in my human brain um in order to like progress this process of moving with the forward with the idea. So, I went ahead and told it, "Okay, that's a lot of words. Can you break it down more simply?" And then I got a more I would call this an opus-shaped response. Um, and it gave me a couple of things to It just gave me a status report on the codebase. And so, and like how I could tell the story and then I go ahead and go into outline mode and like review and see how the how the outline is coming together. And then it wrote a draft for me. And well, this is the outline. Where's the draft? Um, so yeah, and you just go back and forth with like I asked for different options on the hook and picked a hook. Um, this is um yeah um I wanted to [laughter and gasps] make sure that this specific story moment came in because there was the moment where I was like, "Oops, I already shipped the code. Lol.

84:01 It's live." And I wanted to make sure that that made it into the piece. So, part of what is going on here is I'm just being opinionated about what I want it to do and like pushing it in the direction that I want it to go in. And at a certain point, it asks you questions because Claude loves to ask you questions. And then this is what I wound up with um on a um just off the top of its head with u writing the intro. So, uh I'm just going to read aloud, which is always mortifying, but here we go. There's a line of code in my app that's been sitting there for almost two weeks, and I'm the only person on Earth who's allowed to change it.

84:36 Emergency connector shutdown equals true. One word, flip true to false, and a feature I built. The one that lets an AI assistant write your write in your voice by reading the passages you saved comes back to life. The security hole that got got it shut down is fixed. The fix has been audited, tested, and rebuilt with the kind of paranoia I didn't know to have the first time. That's pretty good. like that's that's something that I might actually say. Um, every technical reason to flip it has been satisfied for days, but I haven't flipped it yet. Um, so that's the kind of thing that it's doing. And like this is like this is like a decent I would say like something that I'm always looking at with a model is like how does it do at sequencing the information in a way that's going to be accessible to someone who doesn't know what I'm talking about. And I'm not 100% sure that Claude ha that Fable really has that judgment about um what information to sequence in what order. Uh and we kind of see that in the um the benchmarks which Jack briefly showed is there's not a ton of variation um between the different models in terms of what they're able to do in terms of like writing a hook, writing an email.

85:49 And so given like that, like for example, oh now I'm going to do something scary and try to switch over my window. So if I'm about to ruin everything. Um Oh wait, no. I have my whole screen. So um proof. This is a version of the of a draft that um >> we can't see proof. >> Oh, do you see middle picks? >> I think you might only be sharing, Claude. >> Okay. All right. Well, I um I can't I I I had GPT right to do a path on the same piece and it was just much more linear in its sort of progression of information that I think that there's just something about this is something that I've noticed about the characteristic of like claude models versus GPT uh open AI models generally is that open AI models tend to be very linear and like they do they like they go XYZ and they're going to execute and whereas like OP model or uh Claude models and I found this with Fable in particular are very kind of exploratory in their writing and interviewing in a way that I find really productive actually and I think it pushes me to really interesting places. But if you're trying to just like shoot off a marketing email, not that that's not easy, like not that that's like easy compared to editorial. That's not what I'm saying. But like we found that like Codeex is really good at sort of like linearly like executing a like a piece of writing whereas um Claude is better and Fable is a little bit better for more exploratory work.

87:27 >> Yeah. >> Sweet. Um all right, Austin, what what else? Where should we uh where should we take this? We've got 10 minutes left. I think I can um that I heard a little interest in it and now that I I feel somewhat confident I can share my screen. Uh I can actually show um how uh I've been using it this morning to run a a go to market experiment using LFG and the compound engineering flow and uh and a loop um as someone who never ran LFG until uh we got Fable and also never ran a loop. So, uh, it might be interesting for people to see that and then please hop in with questions. Uh, everyone can see my full screen. We good? It's working.

88:15 >> Amazing. Love to see it. Okay. So, what you're looking at is I use um Cloud Code and the Warp CLI mostly because it's just like insanely fast. I find it to be faster than the desktop experience and Codeex's speed has like ruined any other app for me. And what you're seeing is this like this wall of text is a big monologue brain dump of that example I was giving giving earlier around basically I want the I want Fable to audit our subscribe flow and then um make these recommendations for uh ways it can be better. Um and then at the end it just have I basically like dumped a bunch of stuff down of like the meeting notes, the GitHub repo. Um, you can see I just see like I just say like LFG loop Chrome like let me just make sure you have what you need to run in a loop hit the Chrome browser. Um, one thing Mike Kger from Anthropic said on on his pod with Dan is like try dynamic workflows.

89:12 And the way it works in cloud is if if you just like type dynamic workflows, cloud will like factor that in potentially. And I I've I've been finding that especially with the way that Kieran and Trevan have set up LFG, it's this like really great organizer and rapper to a big problem like this. Um so that I can just walk away and I told it I was going to do this camp and I said basically just go do your thing.

89:35 It identifies a big like autonomous uh research and plan operation. It goes and does its thing and what it spit out was this big proof doc. The part I want to show in particular here where it has its like this is a format I like where it's like 10 insights um the sprint it can actually go operate and then like what it's going to do next. Um and then on the insight site I'll show this which is like the thing that was really interesting me. It's like here's what's broken. Um the page the subscribe page can't autonomously run experiments and that's going to hold us back. And so while we were in this camp, I said like, "Okay, well, you should definitely fix that. Just go off and fix that." And I I do find it to be somewhat collaborative that way of here. I was like, "Oh, go go go prioritize this thing." And so, while we've been in the background, it has shipped um these four PRs. I can tell you I have no idea about like the quality of these PRs, but I do have actually like a growing confidence in the model's ability to ship them. as someone who like my my skill I feel like is on the strategic side of of growth not in like what is actual the code behind this but um we have these like architects as Dan would describe them who can review the code before we ship it we have like a go to market engineer on our team Yosh who's amazing who can like look at this before we ship it um we can test it in in staging or whatever and so like this is the kind of flow that I found incredibly helpful and that Like really the only prompting I gave it was describing the problem, describing the situation I want and giving it all the all the context and then just wring letting it run over and over again.

91:29 >> Uh Cart, >> so Austin, do you review the PRs before it submits it? Like because you said you're the strategy guy. So, how do you >> we make sure someone uh uh technical on our team who who like there's there's like an owner on the on the website who will review it in some capacity. But I know like >> not the technical aspect of it, more of the strategy like you already gave it all the prompts, right? Then it gives you these recommendations, but do you review the recommendations or does it just dump into a PR?

91:59 >> Yeah. So, it depends, right? like there's certain things that are um that I I find the best way to review them is just to see them out in the wild. I wouldn't do that for something live and prod on the um every site. But for instance, a thing I've been trying to crack is actually using uh an an agent a lot of times cloud code and romotion this like codebased video production tool to oneshot product demos in um uh that we can post on social and my way of reviewing that is like oh just go make the videos right like let me try like make the video uh one review it yourself. When I when I do these, the agent actually goes and takes screenshots of every frame and reviews it against uh the persona it's imagining. Is someone scrolling Twitter?

92:46 Like will it get them to stop? And then will the video make sense to them? And then once the agent once the reviewer agents are confident that it's a passable quality, then the review stage to me is like, okay, then I go watch the video and give give feedback because that's low stakes as opposed to pushing something to the site. Um I do for this kind of thing I um uh before like shipping that two week sprint plan I would go look at it and usually my flow is like oh it suggests 10 things ranked 1 through 10 and I'll say great go ship 1 four five and seven I I like 1 145 and seven we're not doing three for this reason but just do it and then I'll review your output.

93:25 >> Awesome. >> Can I ask a follow-up to that Austin? >> Yeah please. In the spirit of compound engineering, what is the equivalent that you do on your growth team? You said we're not going to do three for this reason. Does that all get documented to your notion so that it learns over time or what what's your flow there? >> So my flow there is that in in any agentic session uh basically I only really I work in codeex like as the daily driver. I go to task there every day. And then now there's always a fable session running for at least like a three or four hour sprint. Um, but both of those uh tools have a built-in rule that whenever it gets done with a job or whenever I'm seemingly done with a task, it it uh prompts me, do you want to compound this session? And then it has two um two types of compounding. one is like a full compound of like let me actually go do the whole thing and then save it to um either the the GitHub repo powering the growth work we do it's called like every growth OS and updates that um or do you want me to like suggest what to compound and I can like um query like okay you don't need to compound this thing that's actually not a useful learning just compound two or three and I I find that that built-in rule um is really helpful so that I can both like make sure it happens and then also make sure we're compounding the right stuff.

94:51 >> Super helpful. Thank you. >> Yeah, of course. >> Sweet. All right, everybody. We are pretty much out of time, but this was fantastic. So fun. We are definitely going to do these more with breakouts. I love getting to see all your faces and having you all meet each other. Um, please join the Discord. Um, please interact with each other. If you want to set up more meetups so you can meet up in person, please do that.

95:19 Share every with your friends. Um, tell one person about the only subscription you need to stay at the edge of AI today. And we will be back on the 26th with uh a Codeex native camp. And between now and then, we'll be publishing more stuff. We've got a lot of new product updates coming. Um, and enjoy Fable. >> Thanks y'all. We'll we'll recap this in a follow-up email, too, for all the prompts and stuff we talked about.