0:00Emanuele Melis of AI One says you don't need handovers between people: agents can cut through a good portion of those scenarios, and the human only comes in at the last mile, without a large data transformation program.
2:17Emanuele Melis says Fable rebuilt a contraption puzzle game he remembered from childhood in minutes from a short description. Jordan Metzner says the impressive part is the accurate real-time physics on a first attempt.
4:47Jordan Metzner says Fable is slow and seems expensive, and its highest setting burned through his tokens, but it found security bugs, implemented features predictably and produced world-class designs when told to 100x the design.
8:14Emanuele Melis says agentic workflows start with the autonomy slider at zero, where the agent only drafts and researches; as knowledge builds in Context One, the slider moves right, and 30% to 40% autonomy changes daily work.
10:38Emanuele Melis says AI One targets high-complexity, low-judgment use cases. Instead of piling up rules over years, he says agents handle nuance and edge cases much more quickly than a set of rules.
12:05Emanuele Melis says separate agents check transactional databases, documents and monitoring logs, turning what used to be a week of investigation into a report within a couple of hours so a person can reply to the customer.
14:41Emanuele Melis says when his kids ask to color something like a dinosaur on a space shuttle, he asks ChatGPT for a PDF of it, which makes him the best parent of the year.
15:36Jordan Metzner says as tokens get cheaper, consumption goes up, a classic Jevons paradox: he builds more and crazier apps. Emanuele Melis says GitHub had to limit project creation because of the flood of new projects.
19:41Jordan Metzner says SpaceX has monopolies, with the Falcon 9 flying about 250 times a year, and Starlink could make cell service much cheaper. Sam Nadler says he is bullish but not a buyer on day one.
22:54Emanuele Melis says LinkedIn is the best place to find him and his views on AI, data and enterprise.
Emanuele Melis [0:00]: Well, what we found is that, actually, you don't need all this hands handover between people. You could have software. You could have agents that help you and cut through a good portion of all those scenarios where effectively just saying, well, let's let the agents do as much as possible. And then the human only comes in at the very last mile.
[0:19] Theme song: Built this week, breaking it down. Built this week, we show you how. A fresh idea, a clever tweak you locked in. You built this week.
Sam Nadler [0:29]: Hey, everyone, and welcome to Built This Week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam Nadler, cofounder here at Ryz Labs, and each and every week, I'm joined by my friend, business partner, cohost, Jordan Metzner. How are doing today, Jordan?
Jordan Metzner [0:46]: Yo, Sam. How's it going? Happy to be back. Another big week in AI news. Obviously, Anthropic dropping Fable five, which has been a bombshell in the AI industry. And then as we're recording this Friday, SpaceX IPO, obviously doing pretty well in the market. So, yeah, huge week in AI and technology and super excited to get this week's episode started because we got a great guest this week.
Sam Nadler [1:10]: Yeah. And I'm super excited to introduce our guest. But really quick, before we do, don't forget to like and subscribe. We're just about to hit the 28,000 subscriber mark on YouTube. New episodes out every Friday. And with that, I'd like to welcome Emanuele, who leads field engineering at AI One. Emanuele, if you don't mind, just give us a quick intro about you and your company, and then we'll jump into the demo we have planned for today.
Emanuele Melis [1:36]: Hey, Sam. Happy Friday, Jordan. Great to be here today. So Emanuele Melis is, as you said, a lead field engineer at AI One. We're a New York City based team that has worked across high performance regulated trading and data platform in the city and in Europe. And two years ago, we've started with this belief that enterprises don't need to go through large data transformation programs to start benefiting from AI. Right? So we decided to build Context One, which is a context control layer that sits between the enterprise systems where LLMs and agents can basically use it to give accurate governed context. And it can be used to automate complex workflows.
Sam Nadler [2:17]: Very cool. So, I think we're gonna dive pretty deep into AI One in a few minutes. But in the meantime, I think you built us a little demo. If I'm not mistaken, you used Fable, which is pretty topical. Let's jump in.
Emanuele Melis [2:28]: Yes. You guys remember there was early in the nineties, this little game where you had to build contraptions to effectively solve a puzzle. So I was like, I can't find it anymore. I don't remember. I used to play as a kid. So when Fable came out, that was the first thing that I tried. I have to try this. Can you put this together for me? And in fact, it did. And it doesn't look quite like it, but the idea was that you had this page where you were presented with, puzzle that you have to solve and you would place things in there. And in this case, you have to bounce the ball till the basket's right. And I was so surprised when it literally took Fable a couple minutes to put this together. Right? And you could play it and you could work it and it will just effectively, you put in the your funnier, you have the ball bounce and put the trampoline in and if you're lucky, it gets all the way up to your little basket. Well, maybe after a few tries, you will.
Emanuele Melis [3:31]: Yeah. Yeah. Almost. That's it. This way. Almost.
Sam Nadler [3:38]: But you guys remember this. Yeah. I remember this too. Right?
Emanuele Melis [3:42]: Yeah. Yeah. Yeah. So fun. I used to play all the time as a kid. But now yeah. There you go. It works.
Sam Nadler [3:47]: It works. That's awesome.
Emanuele Melis [3:49]: What's crazy to think about it is that a year ago, this would not have been possible, but today with Fable, literally took five minutes to put this together. I just told it, hey. I had this game as a kid. I don't remember the name but I loved it. Can you put it together for me on like in a web UI? And ten minutes later, boom. I was wasting my time in the office playing this.
Jordan Metzner [4:10]: Yeah. I think what this is what makes this so cool is not just that it's a simple puzzle game, is that the physics are very accurate to how the game expects to play. So that's a trampoline and the ball bounces expectingly like a trampoline or flat to the wood or the wind, etcetera. And so, I think those are always impressive when you see a model be able to implement, real time physics essentially, especially in a first go. So and obviously, yeah, this looks super fun. I'm sure you could theme it to a bunch of different ideas or whatnot, but make different levels, turn into a mobile game. Anyway, that's your new mobile career. Oh, there you go.
Emanuele Melis [4:45]: Great idea.
Sam Nadler [4:46]: Yeah. I know we're gonna dive into the Fable launch. But Jordan, you've playing around with Fable, high level thoughts and the thirty six hours it's been live?
Jordan Metzner [4:55]: Yeah. So, to give you kind of my daily driver as I usually prefer to use Codex. I use I have Claude and Codex and other tools as well, Cursor as well. My daily driver is usually Codex 5.5 extra high, kinda like turn it on there. So Fable came out. First, I put on like the highest. And then that starts like sub agents and a bunch of other stuff. My spend went really high, so I had to like kinda turn it down a little bit. It's a little hard to understand. There's like I mean, maybe like 10 different degrees of like highness that was like max and super high and then regular high and like, what am I supposed to know? I mean, like, even if they put number values on it, like, how am I supposed to understand? Just like, tell me what the difference is in the sense of like, this one costs a little, this one costs a lot or, just tell me the difference in trade off of speed and time. So I think like, there's still some like way too much ambiguity in understanding like this like, slider of how high you can pull this model forward. Right? Because like, if cost is an issue, let's pull it all the way forward.
Jordan Metzner [5:46]: But then what you also understand is that the speed goes down. I mean, it just takes so long to get things done. So one thing I did notice with Fable is I would just run stuff and then just leave. I mean, it's honestly Fable coming out during the World Cup is perfect because just start it right before the match starts and you don't even have to come back till half time. It's perfect. I'm serious. Sometimes I would check at the water break, they got those new water breaks in there and it's like, nope, still working. So Yeah. Even the FIFA water breaks don't kill Fable. So yeah. I mean, slow in the Sense that it just takes a Long time to run. Mean but I found it to be pretty accurate. I found it found some security errors and bugs in my applications. It was able to implement features and libraries successfully and predictably. I saw some feedback on Twitter and you know, one guy said, just told it a 100 x the design.
Jordan Metzner [6:40]: And I think he said I don't know if he said loop it or not loop it, but I've been playing with the loop and not loop. But you know, I just said literally a 100 x the design and boom, a 100 x the design. Like literally a 100 x the design. So you know, I've the designs are incredible. Like I don't think they can get any better. I tried a 100 x the second time and it got like marginally better because like there's not that much more to go. Like, we're in world class like web app web design, app design at this point. It's incredible. It is very slow. It is seems expensive. I haven't run it in parallels. Like, it has like sub workers if you put on the extra high, but I haven't like run it. I ran it once like that and like my they like immediately told me like I ran out of tokens. I tried to step back a second. So again, I don't think and like, it seems like this is the common place and I'm sure many of us are the same like, I don't think anyone understands how many tokens they're using.
Jordan Metzner [7:25]: Like, know, you send a command because you send a command, like, it doesn't come back and be like, that's gonna cost you like a $100. Are you sure you want me to do that? Because like, could do this for $5 instead. Right? That's kinda like probably a better direction of where we should kinda get feedback before these things go and like chew through your token consumption. But anyway, Fable five, super impressive. Obviously, I'm expecting OpenAI to come back with something equally impressive relatively soon. Can't count out Google and Microsoft as well, and obviously, Cursor and Grok and SpaceX with the IPO. I mean, you gotta see Composer three coming down soon. So anyway, that was my quick fable and a fable and five for you.
Sam Nadler [8:00]: Planning a team off-site sounds fun until you actually have to plan it. Flights, hotels, schedules, coordination, Offsiteio handles all of it. Fable in five. Okay. I love it. Why? I wanna dive into AI One. Emanuele, and maybe it's me, but if you don't mind, I would love like a kind of an ELI5 or an ELI10 version of what AI One does. I feel like I have an understanding, but, I think I need to hear from you.
Emanuele Melis [8:31]: Yeah. Absolutely. So imagine you're a large enterprise. You have processes that touch multiple teams and the process could be anything. Let's imagine I am onboarding as a customer for this larger enterprise and my onboarding process, I have to fill in a questionnaire, the questionnaire goes to team a, then team a passes on to team b for some other checks and teams b comes on to team c and this is all money. Right? There's teams that go across. And now team d finds an error. They can fix it themselves. They'll have to send it back to team a. Now team a can fix the error. They have to send it back to the sales team sends it back to me. It takes a whole month to do something very simple. Well, what we found is that, actually, you don't need all this hands over hand over between people. You could have software. You could have agents that help you and cut through a good portion of all those scenarios where effectively just saying, well, let's send the agents to as much as possible. And then the human only comes in at the very last mile. We have this concept we call the autonomy slider. Right?
Emanuele Melis [9:38]: You start every agentive workflow with the autonomy slider set to zero. So the agent really has no autonomy. It would draft things. Would do research. It would present you something but then you as a person acting there, you will take the next action. But slowly but surely, the agent will build confidence, you build confidence. Build knowledge, the knowledge lives in Context One and the autonomous slider starts to move a little bit to the right. So maybe it's 5%, maybe it's 10%. It may never get to 100% and that's fine. Right? But what does it mean for a large organization if the autonomous slider moves to 30%, 40%? Then that starts to make a difference in the everyday life of the people and the enterprise. And what we do really is help these large companies go through that transition and through those processes. My team specifically is the one that will come to you and help you install it, get onboarded, and develop these first use cases on top of Context One.
Sam Nadler [10:38]: Perfect. So are there two questions for you. Are there a particular I think you mentioned client onboarding, but is there a particular like process or workflow that as of now is just seems to be a sweet spot? It's just working incredibly well. And in addition to that, has it is there a particular industry where you've gained a lot of traction or is it just it's useful for any in enterprise of a certain size or who's who has certain similar processes?
Emanuele Melis [11:03]: So I would say the size of the enterprise is is usually where when we see that need. But what we target is mostly high complexity, low judgment use cases, which basically means something that is very difficult but does not require the human judgment to be completed. So that's when you keep that is when you can put an agent. Now if you think of the evolution about, computer science and how processes were completed. Right? You have computers. Computers failed, and then you put a rule to address the failure. But now that rule itself would fail, then you put another rule, And that one rule will fail. And then ten years down the line, you have 150,000 rules all to fall back on the process. Right? Which is where we're that actually with GenAI, Procedure specifically, you don't need to do that cascade of scenarios and we're finding that the agent is much quicker at dealing with the Nuance and all those edge cases than a set of retail rules would do.
Jordan Metzner [12:04]: Okay. So Emanuele, like walk me through a use case like obviously like, you don't have to tell me the company like Coca Cola per se, but just walk me through kind of a pharmaceutical company or kind of, what kind of enterprise was coming in here and what kind of tools are they popping out that have really impressed you guys.
Emanuele Melis [12:17]: I'll tell you my favorite project because I never thought that was I've never I never thought in my mind I would see that type of project being deployed in a regulated environment. I just never thought in my lifetime we will see AI doing that or a version of AI, Gen AI, however you want to call it, which was effectively user request comes in. So somebody receives an email. Now usually that email means a person has to spend a week investigating something that happened on the back end. And usually the email is about money so people are very nervous and you know whomever sent the email, they want to have a response right away, they wanna make sure that doesn't happen again And the person investigating, they have somebody on their shoulder that drives them to go fast, get response, let's go back to the client, find the root cause, who needs to fix it, what happened, when, why, all those things.
Emanuele Melis [13:13]: And so what I'm really seeing is when you have when you can build a sort of agency and scenes to help the person do that. Now what used to be a two days investigation now compresses to one day. And you have an agent that goes on the bucket system transactional database and get all the transaction. You have another agent that goes in and checks the documents. Is there an edge case that should have been taken into account or was it? And then you have another agent that goes into the logs and says, is there a failure in the monitoring system that was supposed to be caught or was missed because of x y z reasons? And now you end up after a couple hours of what used to be a week of work, you end up with a report that tells you exactly all the data points that a person needs to then draft an email to the customer. The agent could draft the email. Sure. But the idea is that you don't need a person going into and keying at the keyboard onto that system.
Emanuele Melis [14:12]: What happened for customer one two three on that day, at that time? And the agent will be much quicker because they also can effectively brute force their way into the systems and just run multiple hypotheses at a much higher speed than a person. Right? So that is the one use case that I never thought I would see happen really, especially in highly regulated environment, but I'm seeing a lot of traction because it actually make a diff it does actually make a difference to Baba to the everyday life of people.
Jordan Metzner [14:41]: Yeah. That's awesome. Okay. Just maybe before we wrap up on your app, let me can you just give us kind of one other fun use case, maybe like maybe less technical and maybe a little more fun. Maybe like you've seen some kids make some cool things or even just like maybe around the office of some cool projects that you guys have built that could be fun.
Emanuele Melis [14:59]: Yeah. The one thing that I'm having a lot of fun with is coloring books with my kids. So they're like, daddy, I want to color this. I'm like, okay. How am I gonna get a dinosaur on a space shuttle right now? And the Internet is what? You can find it. But now you have ChatGPT. You're going to ChatGPT, and you can ask it, hey. Can you prepare a PDF of a dinosaur on a space shuttle? And now I'm the best parent of the year. Right? And you can do that every single day. So that's that's definitely much more entertaining than to the kids than talking about my daily job today. Yeah.
Jordan Metzner [15:35]: Yeah. For sure. Okay. Cool. Well, again, as it was a big week in tech and in news and AI and everything else, should we jump into the news really quickly, Sam, and kinda roll into that?
Sam Nadler [15:45]: Claude announced Fable. There's been tons of articles on it. We shared one before the call. We touched upon, it's been, I think, thirty six hours, maybe forty eight hours. I barely used it, so I don't have a strong opinion. But, I you know, every week there seems to be significant step level upgrades in the model. And, what's it what are your all your thoughts on, what's to come and what this means for the space?
Jordan Metzner [16:12]: Oh, I'll go first. Was an email from the economist at a leading alternative asset manager today saying that, basically as tokens per token getting cheaper, token consumption's going up, and classic Jevons paradox here, and I think that's what's going on, you know. I mean, as I get more tokens, I build more apps, I build crazier apps, I have more ideas for apps to build. I don't know. I know my token consumption over time is going 100 x
Sam Nadler [16:32]: Make better.
Jordan Metzner [16:34]: 100 x my design. Yeah. You got it. You got it. I mean, I should try a thousand x or 10,000 x. But yeah, I think and it gets more and more fun. Know, I remember using cursor. I mean, it maybe was a year ago or so, and it would make these like markdown notes and then every markdown note had a random number to it and then it wouldn't even be in sequential numbers. And you'd be like, oh man, it has no context. And now I just I go watch half a soccer game and I'm back. So, it's only getting better and better. Hopefully, it'll keep getting cheaper and cheaper. And, yeah, it's getting more and more fun and more productive, at least from my perspective. And Emanuele, how about for you? I mean, as an engineering team.
Emanuele Melis [17:16]: It's it's crazy for me to see how quick and how much we have evolved in the last eighteen months. And I made a bet, that was May last year, that I would try to go my career engineering career, not looking at a line of code anymore. It didn't quite turn out
Sam Nadler [17:32]: How's going?
Emanuele Melis [17:35]: I still have to make a code for the most part. However, if you said cursor a year ago or the tools that we had a year ago, I think Claude launched a year and some months ago as well. And if you Claude Code. Right? And so, like, how wild is it that today you have Fable when you can give it a backslash command, you need give it go and just go do something and you keep iterating until you do that. Right? And Codex has the browser view. So if you're doing web apps, it will just build a code, go and click on the UI, and keep going and doing until it does what it needs. It's it's to me personally, an engineer, it's mind blowing. And, like, to your earlier point, we'll just use more tokens. Right? People will have more ideas, more things to play with. The amount of throw away projects that get generated. There was a statistics on GitHub, which is the world's repository for all the code. Right? That they actually had to put a limit on how many projects can be created because they were just being overwhelmed by all this by the code where there were new projects that were coming out that just didn't have capacity. Right? Yeah. Well, you would Which like definitely wasn't the case.
Jordan Metzner [18:42]: A repo to GitHub has like a marginal cost of zero, but then, people were pushing so many changes to every repo times multiples. Actually, it's not zero. It's nominal, but not zero. Right? Yeah.
Emanuele Melis [18:55]: So many. I think it was around, like, 25,000 that they had a limit
Sam Nadler [18:58]: That was like, how do you even get to 25,000 feet?
Emanuele Melis [19:02]: That that's just so much. But then, I guess, as you said, it's so quick to iterate now and so fast to get a prototype out. And half of those prototypes are more than a prototype. Something very cool that a CRM project manager do recently, they had Claude Fable that effectively listened to the transcript of their conversation as they were live. So effectively, they were discussing with the engineers and Claude Fable Deployed built in live just by listening them talking to the com into the conversation, which is that would never be possible. Right? So definitely fun times for all the builders out there. There are a lot of possibilities to build.
Sam Nadler [19:41]: We have a couple minutes left. SpaceX IPO ed a couple hours ago. I'm bullish on SpaceX. I think, it's it's it's, if anyone can drive the valuation to new heights, it'll be Elon, the data centers in space, xAI. Yeah. I forget the third prong.
Jordan Metzner [20:00]: But Space logistics.
Sam Nadler [20:02]: I'm not a buyer today. Space logistics. Oh, yeah. I forgot. Yeah. I'm not a buyer today though. Jordan is. I don't know if you are, Emanuele. But what do you guys
Emanuele Melis [20:11]: No. I haven't. It's it's going to be interesting. I'll I'll wait it out to see for a couple days and see what
Jordan Metzner [20:17]: Okay, guys. First off first of all, don't bet against Elon. Okay? So that's first mistake number one. Second, you guys have to understand that, SpaceX has a, complete monopoly in various businesses. So it's the SpaceX goes to space more often than all other countries' space programs combined, including The United States' because The US doesn't have a space program outside of SpaceX, as NASA shut down their space program. The Falcon nine rocket goes to space something like 250 times a year. It's like the number one mechanism to get satellites into space. It's the thing that funds almost every other space startup that has satellites. So everything from like Hubble Networks to Varda Space and even Starlink itself. Starlink network has a mirrored network for the US government as its own private, controllable network. If you've seen how powerful Starlink is, which I know you have Sam, because you have a satellite Starlink Mini with you, which gives you internet anywhere in the world. And you're gonna see those little satellite dishes get smaller and smaller, and they're gonna move into your Mac, they're gonna move into your phone, they're gonna move into your car, they're gonna move on to your bicycle. Right? And they're gonna offer cell network service anywhere you are in the world.
Jordan Metzner [21:25]: So instead of paying your expensive fees being international, you'll be able to connect to Starlink and your fees will go down significantly because all they'll need is some local ground area satellites to connect up through. So you're gonna see probably a cell phone service come in really really cheap, which yesterday, I don't know if you saw the news from Cash App, they announced a $40 a month plan. So I mean, if that becomes the new norm, $40 times what I'm paying for AT and T, think SpaceX could, make some real damage in the space. And then obviously we haven't talked about their business in regards to going to the moon and going to space and everything else. And of course, Cursor, the acquisition there, the models there, they're oh, anyway, talk about the AWS, the Elons of the world and how they took the data center and rented it out to Google and to Anthropic and to Cursor. So I mean, guys, this is multi pronged approach. It seems obvious they're gonna acquire Tesla into this thing. I mean, there's just it has so much ambiguity. How could you not be excited about it? But anyway, alright.
Jordan Metzner [22:17]: I'll be a dreamer. I'll be over here. And I think I told you guys this, but I think I was telling you guys this. But, yeah, space Tesla from when it IPO till today from a dollar 11 base cost, this is like after the price cuts with this price splits whatever, to $3.86 today. So basically about 380 x since IPO on Tesla. I mean, guys, 1 point something x trillion dollars, $1,700,000,000,000. You don't see how Tesla becomes a 100 as test SpaceX becomes a $100,000,000,000,000 company? I see it. Alright, guys.
Sam Nadler [22:51]: Alright. I love it. I we're excited. Trust me. I'm excited. At least I'm sure Emanuele is is as well. I just I want a little perspective, guess, a few days to play out before I get involved. But Emanuele, where can people find you?
Emanuele Melis [23:06]: You can find me on LinkedIn. It's probably the best place where you'll see me, and you'll see my opinions disseminated on AI and data and enterprise on the web.
Sam Nadler [23:17]: Thanks for joining. It was great to hear about what you're building. Thanks for building the little game. And anything to wrap up, Jordan?
Jordan Metzner [23:24]: No. That was awesome episode. Yeah. Thanks for showing us a little bit of your app and the game you built. And, yeah, it's great to talk shop and see you guys all next week. Thanks so much. And bye.
Sam Nadler [23:34]: Thanks everyone.
Emanuele Melis [23:35]: Thank you.
