Building Open World AI w/Andrew Mason

Cedric: [00:00:00] Join Nabeel, Fraser, and special guest Andrew Mason from Descript for an insightful discussion on the evolution of AI in product design. They explore how Airtable's innovative AI implementation is reshaping user interactions, dive into the challenges of open world agents, and examine how products are adapting to make AI more accessible and powerful.

 

Andrew: So I think it helps to have a buddy that is like constantly in the process, just like gently. And somewhat mockingly nudging you.

 

Nabeel: It's like Yahoo Pipes,

 

Andrew: but we have L Labs right now. It's such a fundamental mindset shift that it's not the kind of thing that you can just have like a corporate training for a day or something like that.

 

Introduction and Guest

 

Nabeel Hyatt (2): Hello everybody. Welcome. To Hallway Chat. I'm Nabeel. I'm Fraser, and we have a guest here today. Andrew, that's me. That is Andrew Mason from Descript. I was talking with Andrew maybe two [00:01:00] weeks ago.

 

Discussing Airtable's AI Implementation

 

Nabeel Hyatt (2): We were talking about As an established product that has been integrating AI really well . And you'd said something a little bit surprising, I guess, although once I thought about it, like, I use this product all the time, it makes sense, which is Airtable.

 

Andrew: Yeah. So context on that. Nabeel and I as a side project are opening, uh, physical board, game club, retail establishment, you know, and that's actually where we're, what we're using Airtable for.

 

I don't use it at Descript. We use other tools, but we start picked it up. You're adding notion. Heavy notion guy. Yep. And picked it up again for this project and got exposed to how they're agentifying the product. What I think is interesting about them and where we look to them for inspiration is that it feels like products are going in this direction.

 

"Cursor for X"

 

Andrew: Tools at least, of having. An editor that's great for humans to interact with. And then a chat sidebar for interacting with an agent that you can [00:02:00] use to delegate tasks. Um, and then other forms of AI based interaction interspersed with the, like, sprinkled on top of the editor as well. And you've seen this primarily in vibe coding tools so far, which is kind of the easiest domain because you're just generating text for the most part.

 

I mean, increasingly they are integrating more and more tools, whether it's interact terminal or or whatever else, but it's mostly just generating text. So even though it's. The most powerful use case. It's also one of the easiest to get started with and now everybody else is rushing to follow to agentify their tools.

 

Fraser: Cursor for x,

 

Andrew: cursor for X. So we kind of shamelessly talk about Descript is now cursor for video editing. Yep. Airtable is the best example I've seen so far of a non vibe coding tool that has agentified their product.

 

Nabeel: I'd probably use it even more heavily than you do now for sure. Inside of that. And, and I kind of agree, but I guess my question to you is like, [00:03:00] is that an execution thing on their part?

 

Is there something they're doing with the product that you feel is not intuitive ? Because for me, it's an easier problem to solve. I guess like the hard thing for me is not the agent running away and doing things inside of any product. So inside of Cursor, the hard thing is not, oh look, there's a chat bot.

 

I know where to put the chat bot. It's over on the right. We That's easy to find space for. Yeah. And then there's the articulation layer, like, oh, does it have tool calling? Can it do things inside of this product without me? And that is an execution problem. Did you do it smartly?

 

Andrew: The,

 

Nabeel: the problem feels to me like, can you surface the work?

 

The Challenge of Surfacing AI Work

 

Nabeel: Do you have proof of work? Can you surface the work the agent is doing in a way that the customer user can understand? And so the problem for an average user, if you can't recode and you're vibe coding, you don't really know what it's doing. And so it works really well for front end tasks where I can like, yes, I see something visual, like the box is in the wrong place and I don't have to look at the code.

 

I can see it's wrong when I can fix it. And it's hard [00:04:00] without being a coder for it to do like, I don't know, like backend or database work or things like that. 'cause you have no idea what it's doing, how many files it just worked. And often when people complain about it going off the rails, it's usually when that happens.

 

I say that now, like air table feels like it's serving in a way from just a more advantageous position. Because the entire product is a visualization layer for the data underneath. Like it's literally a database that you can look at visually. When I drag in, I drag in A PDF with like 50 pages in it.

 

And it did an incredibly good job taking a 50 page PDF and figuring out the data types spiral, importing it in properly. It was awesome, but also like I could have dragged that into clawed or something else and it would have an inherent problem because it's different product.

 

Do you know what I mean?

 

Andrew: Yeah, and like the dimensions upon which you audit the work are relatively limited. Like if you ask it to generate a bunch of new fields or something like that, then you can go in and you can audit the [00:05:00] formula Yeah. That it wrote or something like that. And then you can trust that it, that the data is right for all 20,000 rows or whatever it is that you have.

 

Review States and AI Adoption

 

Andrew: We talk about this in Descript to this problem of review states and how important is that even because. Cursor has had a lot of success despite and Replit tools like that, despite not really doing a super good job at solving the re review stake problem. Like, like you could be happily by coding away and have broken a part of your app that you're not refreshing over it over again without knowing about it until days later.

 

Yeah. If you're not being careful. But it's not, it's not clear how much it stands in the way of, it may not

 

Nabeel: matter. Yeah.

 

Andrew: Of adoption. Is that

 

Nabeel: because

 

Andrew: the

 

Nabeel: Like it is the bar different because a average person can't code, so you, or because

 

Andrew: they're people are using them for prototypes a lot of the time.

 

Uh, yeah. So it doesn't really have to, to be,

 

Nabeel: yeah. What, what

 

Andrew: do you think works so well at Airtable implementation? The most important thing for these products [00:06:00] is that they. Fulfill the promise that that they're making through the way that the product is presented. And Airtable kind of over time amped up the intensity of the promise that they were making.

 

So the first AI features that I remember using were kind of like AI fields and stuff like that. Like what was some of the stuff you were doing to autofill? Yeah. You would like autofill, basically.

 

Nabeel: Kind of like the same things that the first, very first Claude. I dunno if anyone remembers this, but Claude used to have a Google Sheets plugin.

 

The same version as that, which is simply, I will run a query on this cell and I'll do that hundreds of times over and over and over again that was the first version, right?

 

Right.

 

Andrew: So they started small with something like that. It was useful. It did a pretty good job of fulfilling the promise. And then as they started rolling out the, the agent, the, the sidebar, it kind of started as a button in the header that would open a sidebar on the right. But it felt like a very kind of optional [00:07:00] thing, that you could very well get through using Airtable without even knowing that it existed.

 

Airtable's Refounding Moment

 

Andrew: Mm-hmm. And then they recently, I think after really getting their confidence levels pretty high. Have re redesigned, they've gone as far as to like brand it as a refounding moment of, of the company. Feels like it. Yeah. Where now they have like the starter, prompt, pat design pattern on the homepage, and then when you get in there it's big.

 

And on the left, on the left side. And this is what we call an open world agent, meaning the promise that it's making is that you can do anything. And so far in my experience, like it's pretty good. It's very difficult to know like what the bar is for, what percentage of times it needs to do what you ask it to do, or what failure is acceptable versus isn't.

 

Because people seem to have a pretty high tolerance for failure. Yeah. On a lot of these things. But it, it's enough that you keep going back and you believe in it enough to try again with your next ask. [00:08:00]

 

Nabeel: What is a open world agent that feels like a name We apply to things because we're trying to. Take back the word agentic from being overused and, uh, butchered and trying to define a specific behavior.

 

Open World Agents vs Agents on Rails

 

Nabeel: So what does that

 

Andrew: Yeah, so, so we, we, we should talk about agentifications of products that we think have not been successful by comparison, but a lot of the ones that we've seen end up feeling like hard-coded AI workflows. Yeah. Maybe with like a type ahead. Element that gives the illusion of choice and open worldness, but they really are just like, you can have any color party you want as long, as, long as it's blue kind of things.

 

Yep. Um, so those feel like agents on rails as opposed to something that can try to do whatever you ask.

 

Nabeel: Do you have an example of that? That, I mean, you can cut it if you too. All of

 

Andrew: our, all

 

Nabeel: of our competitors. All competitors not, yeah, that's [00:09:00] true. That's very true. Do you have, do you have a thought of one, like a, a company like that you feel like has launched agents or launched?

 

Oh, I open up a chat bot and I'm trying to do something, I'm trying to think of one right now too. Like, I open up a chat bot, I start doing something. I, I can do it as long as I, yeah, I, I say the five incantations,

 

Fraser: I think the entire first wave of tic like platforms were that where like you have a canvas and you say like.

 

When this happens, I'm going to draw a line over to this other function and I'm gonna draw a line over this. Like, Lindy, there's a whole bunch of them mm-hmm. That fit that bucket. I think that's what you're saying. Yeah. Similar. Yeah. Yeah, yeah. Mm-hmm. And, and like on Rams, like it, it can then go through a, a pipeline.

 

It's like you have pipes, but we have LLMs right now.

 

Nabeel: Yeah. It feels like less like coding, which is very open and almost like. The kind of phrasing of using game design would be like, it, it, there's an emergent property that can come out of the work, like you can surprise yourself with what it can do. Yeah.

 

The same way that like in diffusion models, you're suddenly trading around weird words and incantations [00:10:00] that lead to images you never thought of before, this kind of emergent property. And yeah, some of the early agents stuff, which just feels like you're kind of like natural language prompting into what is clearly a series of if then statements of backgrounds.

 

Yeah. Yeah. That, that feeling That's right. And so you're phrasing is to try and grab back, like all agent think is not the same. And open world is this feeling when you, uh, tell what you want and then like it'll go figure it out. And whether it's writing code on the back end or tool calling a bunch or who knows, but like great things come out the other side.

 

Is that the, is that the basic framing?

 

Andrew: Yeah. Where you don't feel like you're constrained by these discreet widgets necessarily. You change the way that you think about those projects and, and you approach 'em, like instead of having like a wizard with a bunch of predefined forks in the road, it's amazing to see how users just start asking so much more of their products and expecting so much more.

 

The Power of Chat as Primary Interface

 

Andrew: As soon as that switch flips in their mind, [00:11:00] to think of it as, oh, this is, this is truly open world. It's safe to ask for anything.

 

Fraser: Yeah. I, I have, um, something that builds on that. Before we sat down, you were saying that the difference historically between like a notion and a coda and an Airtable, is that some of them felt very doc centric and this felt very spreadsheet centric.

 

Mm-hmm. And I think that that's part of what makes this this big push into AI actually work. Uh, at least from my perspective in that. A spreadsheet is a late person's database, and you can use it for anything, right? But it has just enough structure that you can use it for, you know, planning your budget or planning your trip or tracking whatever your, your tabletop Yep.

 

Storefront, right? And if you think of it as a database, then you have chat as an interface to give it direction. Right. Chat is the lay person's way for interacting with software increasingly, it seems. And I think, I think for a very long time. [00:12:00] And then you, yeah, you're long on chat. I'm long on chat as the primary interface.

 

Yeah. But then, but then over the past, like the past 12 to 18 months, like things are so fluid and dynamic that there's been fundamental. New capabilities that have been delivered that we've all done in our spreadsheet anyway historically, like, oh, I want this cell to update with the most recent information from a search.

 

And historically you'd open up your Google sheets and you'd go do a search and maybe put in like the stock value or whatever, if you're tracking your portfolio or your fantasy baseball team and like a agentic like search with reasoning can just do that automatically now. Right. And they still have like the same wizard.

 

Experience in Airtable. Mm-hmm. Where you say, like, for this cell, if this trigger happens, then I want these three other things to occur. Yeah. But the, the nice thing here is that you can, you can set all of that up by going underneath the hood and fiddling around with the pipes if you want. If you can also do it through check and you can be like, when this column, [00:13:00] uh, is updated, yep.

 

Uh, do this thing and this thing can be a web search or it can integrate into Gmail or this, that, or the other thing can just be a trigger. I think it's far more approachable for, for most people with that type of experience.

 

Nabeel: Well, the workable metaphor, like has some start ahead of everybody else, right?

 

If you've been working with spreadsheets for years, you may not even understand. I don't even how even understand that your table is a database, uh, versus like a spreadsheet that for sure, you know, like, like for most people. But that metaphor is, is powerful because you probably bumped up into it a bunch of times and so you're trying to get ag and you're trying to be open world, allow anybody to type anything, you know, they have a really.

 

A broad set of desires, right? Like, you know, that they've tried to, I don't know, do a very complicated formula at one time, or they just wish that every single time I have to this combo would update the other way. Like there's just a lot of embedded wishes, and I don't know that that's true everywhere.

 

Like I don't know that open world, a agentic cursor for X is actually like universal approach. It should be [00:14:00] applied everywhere. I think it works similarly for video and you know, you don't have to your own board, but like I think you've done. I personally, I use all these products. You've done the best job of actually giving something that's emergent in open world.

 

A lot of your competitors are kind of on rails, but the thing that I felt when I first used it is like, I, unlike Fraser complain some about chat interfaces. It's like, and I'm an old school design guy, and I'm like, no. Like people are dumb. I, I put an Alexa in my house and nobody knows what it's good for and they don't know what commands are there.

 

Like you need. Interfaces and buttons. So I know that I can do this and do that. But the thing about video that is very similar to spreadsheets is just think you know what you want. You may not know that script is good enough to do it yet, but you have a sense of like, Hey, this thing feels choppy. Can you smooth it out?

 

I stay out too much. I'm so angry. Can you fix that? Like you do have these like innate natural language desires the same way that you might have for a spreadsheet. 'cause the metaphor is so fixed. Um, so you're like other words like, and these categories are more alike than frankly, like, you know [00:15:00] for sure.

 

I don't know certain parts of finance software or sales CRMs or lots of the other areas where they're trying to get ag agent tech. Yeah.

 

Andrew: There's another thing that is interesting where, um, if you look at the vibe coding tools, like the lovable and the bolts, when you go to their homepage, they'll have like a big open text field and then they have some starter.

 

Kind of templates down below that are like, build me a app that I can use for tracking my next home purchase or, or something like that. It's almost certainly not something, not what you want, that you want, but it's designed to just give you inspiration about the kinds of things just to get your mind going.

 

But in tools like Airtable or like. Descript, there's like a workflow discovery aspect to it that you have to do with with these suggestions that you're having. So when you open the Airtable chat agent, it will have these kind of like [00:16:00] personalized suggestions on types of analysis that you might do or interfaces that you might build on top of your database because there is this like discovery problem of what is possible.

 

Like, what is this thing capable of doing with my data? Mm-hmm. That probably reflects the fact that these things aren't truly open world. Like you can't literally do anything with your video or your, your data set in here. It has like a discreet set of tools. Yeah. And even once you have all the tools in the app hooked up, there's still other tools that you haven't built yet.

 

Nabeel: What does it suggest right now if you click on, uh, so Fraser, just so you know, this is. You're looking at the backend for the, for the tabletop library over here. So it has things like all the reservations in the system, a list of every game we carry, all the members, stuff like that. It's, it's, it's all that stuff.

 

Andrew: So it has three sections of suggested prompts, ask, analyze, and build. Under Ask [00:17:00] Find reservations hosted by David Chen on January 7th. I wonder if there's a reservation

 

Nabeel: on January 7th that David Chen has

 

Andrew: and then build. It has like actual apps you can build on top of the data. Yeah, data. I mean, ask, analyze,

 

Nabeel: build is, uh, is basically taking the three things you want to do genetically right?

 

Ask is very simple looped questions. Look at the data and give me an answer. Yeah. Analyze is basically a way to say structured SQL pairs. Yeah. That's kind of the version of that, right? It is like, we don't wanna call this a SQL generator because that's gonna freak people out when they don't know what that is.

 

But it's basically they do some analysis on all of the information that I have. And then build is some version of talk to a backend agent and fill out a bunch of fields, or does it do more than that? I think it does,

 

Andrew: yeah. 'cause they have these interfaces that you build in, so it'll just like automatically create views of your data that have a limited number of fields for.

 

Whatever for whatever you using. Yeah, yeah,

 

Nabeel: yeah. Go build a login flow [00:18:00] page on interfaces. Mm-hmm. And somebody with the database or something like that.

 

Fraser: Yeah. A design decision that I love is that that chat ui is on the left. It basically says like, this is the primary way. You should be interacting. Here's your interaction point.

 

And then here on the right is the, the, I'm hesitant to say artifact 'cause it's quite different from what that word means now, but like here

 

Nabeel: is the No, it is. That's the adobe art. Sorry. Sorry. It is, that's, that's the anthropic Yeah. Artifact. That is the thing I'm manipulating is on the right. The thing I'm manipulating.

 

Fraser: But here the thing that you're manipulating is, is familiar, consistent, ever present? Like is the spreadsheet?

 

Nabeel: Yeah.

 

Andrew: I think actually a good contrast is notions, implementation of ai, which I think is quite bad, even though I like notion a lot as a product. And I'm sure this is going to evolve, but what they have is they have a little, like, almost like the typical help bubble design pattern that exists in the bottom corner of the app.

 

They've turned that [00:19:00] into ai. It's really cool. As like a enterprise search kind of gle, is that what it's called? Yeah. That they're trying to, they're turning it into that kind of a competitor. Yeah, but as like a co-editor, it's really. A poor experience. It's not designed in a way that suggests that, and when you actually do it, it's often like rewriting the document in chat and then forcing you to bring it in and merge it, or it puts it in the document as a suggestion in a way that just probably seems like a really good idea to not getting back to review states.

 

Right. But it turns out that it feels better, I think for most people, when the agent is pretty aggressive about just making changes and then users understand that there's an undo button, there's.

 

Nabeel: Yeah. Did you ever play around with G Script at a little helper like that?

 

Andrew: Um, I think we had the blessing of seeing that it didn't like I'm trying Yeah.

 

From how Notion did it and kind of bypass that. We have it as a right sidebar right now, [00:20:00] which reflects where we are in our ability to fulfill the promise of the agent. Hmm. And it will eventually make its way to the left, which just feels like you were saying Fraser, like that. That's primary. Yeah. If you're serious about this thing being the way that you want people to interface with your tool, then that's where it goes.

 

Yeah.

 

Fraser: The other thing that I thought was, was really important was that you can interact with it through chat. You can go deep into the bowels and, and like mess around with the settings to set things up if you want. You can also just treat it as a spreadsheet. Like you can go in and edit a row. Yeah. And you can, you can make things the way that you want.

 

And back to a couple of things that we've discussed for over many months now, is that engenders trust. Your work product is there. You can QC it, you can qa like you, you can, you can gain confidence that it's doing what you want it to do when you walk away. Yep. By looking at it in the familiar UI with which you expect it to be.[00:21:00]

 

If it's not right, you can edit it. You haven't played around with it enough. But I also imagine that you could, another thing that we've discussed is that, um, sometimes you just want to show the AI the way that it should be, and I imagine that if you're building A-A-C-R-M or some other thing that you use a spreadsheet for, you could probably set up a row.

 

And be like, when new things come in, I want you to complete the,

 

Nabeel: you can't do that. And that is my, my number one thought. Yeah. But I'm so glad, I'm so glad you walked into that. 'cause if you think about Claude as an interaction, yeah. This happens all the time. Like I'm trying to teach Claude to write a certain way and I'll give it a couple of samples and I'll say like, like this.

 

Or you know, can you summarize a sentence down to this? And then your four more sentences do the same thing. Like, it it, so it's like, give it a month. Right? Think it's gonna be a month. It's going to be run. I, I hope so. Yeah. I hope so, because right now. You're still back in when I start, when I do that kind of work, you're back in Formula and I am opening up the Formula tab and then in the Formula tab and [00:22:00] Airtable, there's a way to fill it out with ai.

 

Mm-hmm. And you're typing in something, you're hitting go, you're looking at 20 samples, you're like, that's not it. You're trying to change one word in the prompt bit. 20 samples. That's a lot of fun tokens for Air Table to charge me for. Look at the outputs versus this kind of like. A habit or a user pattern, which seems much more valuable for how LMS and Transformers work, which is like, just show me three or four or five examples and I will intuit, instead of you trying to figure out the English language, fun way of explaining to me the work I'm supposed to doing, just show me outputs and let me figure out how to get from here to there.

 

You,

 

Fraser: in this case, being the ai, not not unit, user. User, yes. Right, but like I want to fill out two, two rows as a user. Yep. And be like, I'm going to automate the, the new names in column A. Yep. And whenever a new name shows up in column A, go and fill out the spreadsheet. Like Rose. Rose wanted to. Yep.

 

Right. Just

 

Nabeel: that'd be great.

 

Fraser: I

 

Nabeel: wish it was that easy. Yeah. I do want to go back real quick to [00:23:00] the primary interface being on the left. The three modes of, there's a very interesting decision and doing the three modes that I think is actually really prescient, which is. You, I'm going to end up using the word open world a lot from here and forward, so thanks for that.

 

The interesting feeling of an open world experience is you feel like you've pushed all of the power to the user to think about how they really wanna work with this product. So nevermind. What are the four buttons we put in the top left, which is how you probably use this product or whatever it is. It's like now I get to actually ask questions about what do I want my workflow actually to be?

 

Mm-hmm. And can I like have a couple of prompts that I copy and paste and share with my other teammates and we just. Run every Airtable through these sets of prompts or every Descript or whatever. That's the beauty of what the world is to ask you this question, how you really wanna work versus a workflow that's forced upon you.

 

Awesome. The emerging property that is like, what are the choices that you have? And the problem of course is like you have a lot of choices. Mm-hmm. In Airtable, there's like thousands, and finding a way to bucket into three basic [00:24:00] modes I think is something that it's enough that a user can keep it in their head.

 

Right. Hey, there are three things you're really supposed to be thinking about AI as roles. Yep. Inside of Airtable in a way that like for instance, like if you just look at Chatt, PT, Claude Perplexity, I don't think you've seen enough of that. Probably. Arguably the only thing you've seen that is doing that is like they took a whole bunch of post training and other like choices they made and they boxed that all together and they called that deep research and we now know what that is and it's across all the apps and it's a different mode of interaction, which is great.

 

Surprising to me that we're like two to three years in and like there's basically one thing

 

Andrew: Totally. And one of our kind of. Goals for the second half of this year at Dscr is a internal facing goal that we call AI gene therapy. And it's, it's really about, um, rewiring everybody's brains to be more AI first and incorporate mm-hmm.

 

These AI tools into their workflows. And [00:25:00] I think every company is struggling with this transition. It's uneven. Yeah. Where some people take to it very naturally. And other people just have a harder time thinking of how to make use of these tools. And I think that comes down to what you're, you're talking about is.

 

We haven't yet created like a PAT framework uhhuh for what these things do. We haven't taken this alien power and packaged it into little widgets, which is what people are used to, is having these products handed to them. It's just unbridled power and the people who are doing well with it are the ones who can have the sensitivity to figure out like, oh, I could use it for this.

 

I could use it for this. And probably once that. You know, framework comes along, we'll see more people making the shift.

 

Nabeel: Have you found hints for how to nudge people from, uh, what was the phrasing you used?

 

Andrew: [00:26:00] Well, we have a, a spectrum that we, of like a. People's adoption, proclivity of any of any piece of technology where it goes from on one end, end, end of the spectrum, you have technology that you're like actively hostile towards.

 

Mm-hmm. Then you have people who are skeptical of something and then you have converted and then rewired and where like I would consider myself with AI as someone who's converted but not rewired. Being converted is like when you've used Amazon forever to buy books. And, but they sell paper towels. But you still go to Costco to buy your paper towels and then one day you wake up and you're like, why am I doing this?

 

I can just buy my paper towels. Everything on Amazon that's being rewired is when you really like Yep. Cognitively change your neural pathways to a. Adopted this new technology and I think we're all like somewhere along that journey. Yeah. Like for me the biggest thing was actually like starting this, you know, side project with you where we were forcing ourselves to use [00:27:00] AI tools just like using them.

 

Yep. And getting outside of your normal, um, habits. And that

 

Nabeel: was partially 'cause starting a whole bunch. One, we don't wanna hire a bunch of people and it's a side project and we don't have a lot of time. Two, we're both very curious about AI already. And three, and this is where it's hard to do this inside of an org that already exists, or a pattern that already exists is like, we were doing all this for the first time, so it's like, oh, how are we gonna come up with a mission statement?

 

It's like, how can AI help us do that? Right? How are we gonna come up with a website? How is AI gonna help us do that? How are we gonna build a database? How is it like it? So you get to litigate this all from scratch, which does feel like a much harder thing than I'm at year. Nine at Proctor and Gamble, you know, making my internal spreadsheets or something like that, and like, now I'm gonna try to do it an entirely new way next week.

 

Andrew: Yeah. And it's, it's such a fundamental mindset shift that it's not the kind of thing that you can just have like a corporate training for a day or something like that. And, and be done with it. We did a hackathon for the entire company and that was [00:28:00] a step in the right direction. But if it stops with that, then there's too much inertia to just fall back to your old ways.

 

Fraser: Yeah. Yeah. Uh, um, to bring this conversation back to the Airtable conversation, the one thing that Nabeel and I have, have disagreed with a lot is on these broad welcome to RB on the broad horizontal assistance like Che team Claude. I think that solving that rewiring problem at the product level is a fool's errand in the short term.

 

And people are going to have to just suffer through a, a, a cathartic rewiring process. And that product teams and the research teams have much more fruitful things to focus on there. I think in these open world cases that at your definition, DS script, Airtable, that are open world agents, but still kind of constrained to a, a domain or a Senate tasks.

 

I think, I think I, you lean all the way into product that accelerates the

 

Nabeel: rewiring. Why do you make that, why do you make that distinction that it's a fool's errand [00:29:00] for claw and, or, or, or like where in the spectrum would you be? Well, 'cause

 

Fraser: there's, there's, you are coming to Descript, like you are coming there to edit.

 

Yeah. Uh, audio or video. For a very broad, diverse set of reasons. But you're still there to do like one, one job. Yeah. You know your goal. Yeah. And I think that you can have not, rails is not the right way. You can help nudge people in the open world problem to figure out what it is that they can do in a much more graceful way because of that natural constraint where it's like, where the hell do you start with rewiring somebody in clawed.

 

Everything that you can do. I mean, I use it as my health coach, my fitness coach, my financial advisor, my tax account. Like, uh, I use it for work. I use it for all these things. It would be a cruddy experience that would fail. I have high conviction would just fail. It'd be messy. And there's, there's higher order things that they can do in the short term that try to solve that problem and then that problem gets [00:30:00] solved for them by having people.

 

Is this your conception of why something like GBTs was just too early? Sure it was just too early and I also think it was maybe not architected the way.

 

Nabeel: Yeah, yeah, yeah, yeah, yeah. Sure. But, but your point is, it's partly cultural. Like you haven't, you could obviously product market it better or structure it better, but part of it is like, it's funny that to say that in front of me because obviously like a lot of the conversations when you have about these group is like how hard it is to make a product.

 

That is horizontal versus vertical in nature. So nevermind looking up at cloud, you're looking, you know, at even more verticalized software and you're looking at the, you know, the product management challenge of building something like descriptor notion or you know, or Airtable, these things that are very horizontal.

 

And it's an interesting point I hadn't thought about the before, which is like, how much more different would it be to navigate AI when you're at the fully horizontal? It can be everything to everybody, uh, standpoint. And I don't know the answer to what it leads you to it, it certainly should think into different decisions.

 

I [00:31:00] think I,

 

Fraser: I'll just keep going on that while I'm on a rant. And your, your deep thought is, um, I think the equivalent would be insane in to Google in 2004 is there's a lot of things that you can search for and you should, you should add scaffolding for the end users to help 'em figure out what it is they want to search for or like how to better search.

 

Mm-hmm. And I think that. History has shown that the winning bet was, uh, over time solved that implicitly automatically on the backend for the end user. Yeah. But that wasn't like a day two problem for them to solve.

 

Nabeel: Yeah. And it also was emergent from using a horizontal

 

Fraser: not interface. And like, imagine, imagine if you like typed into a search in 2004 and then they were like, it looks like you're trying to do this type of search.

 

Would you like to actually end it with this type of thing? And yeah, like you'd mean cram. Yeah.

 

Andrew: Yeah. There's an equivalent to the, let me Google that for you. Um, thing happening with ai, I think. Yes. Which is like, like that meme is like basically reflective of people who have not yet fully been rewired for [00:32:00] search, right?

 

Yes.

 

Nabeel: And yes. Um, right. They're asking the question 'cause they haven't rewired for search and so it's like, well, fine, I have, I will go use Google for you and I'll get the answer. Yeah. Right.

 

Andrew: And Nabeel was ahead of me on getting rewired for ai. So we would be in these meetings and he would often be the one.

 

That would do the equivalent action of the, let me Google that for you. Where we'd be like pondering, pondering over a, a mission statement or like how do we get a permit or something like that. Right. And he would remind us of this alien power, right? That we all have access to. It is to solve the problem for us.

 

So I think it helps to have a buddy that is like constantly. In the process of just like gently and somewhat mockingly nudging you.

 

Fraser: Well then at your company you, you, you have a lot of different levers of your control and, and a lot of incentive to try to, to politely nudge people to, to adopt this new alien technology.

 

Nabeel: Yeah. Right. Yeah.

 

Fraser: And, and so like, I think that, that I would be, be [00:33:00] in on that,

 

Nabeel: huh? This makes me think the left panel, which sounds as simple as like, you know, moving UX over to the left. In a way, it's that buddy. Do you know what I mean? Uhhuh, like this is a software company who is telling you that it's time to think differently about this product.

 

It is sending a very different message to a user than, you know, the two years ago version of this was, or a year ago version of this was everybody asking, you know, adding the little ask AI button in the upper right hand corner, the little floating. Thing that you can now ask eye questions or notions case, the bottom right corner anchoring it to the left feels like the person inside of your org who is kindly telling you Uhhuh, that if you're just interacting with the artifact on the right, you're probably under utilizing this product.

 

I don't think I'd really ever thought about that until right now, but it's actually. A kind of a profound statement [00:34:00] that the piece of software is making if you're trying to get your users to change their behavior and think about you differently For sure. Because even if you don't open it that often, um, or use it that often in the beginning, it's like trying to tell you you should start from the left and move to the right.

 

So obviously your brain should be moving in this direction. That that's why legitimately feels

 

Fraser: like we're refounding on the company. Like they're, they're the primary way that they're telling their users to interface with their product has changed dramatically. And my guess is, uh. Gemini products. Google products have all sorts of challenges.

 

Yeah. It'd be hard. Google Sheet already has like, it, it like slides in from the right and you can like in theory be chatting with your Google sheet. Yep. Much like this, except it feels profoundly different. Yep. Um, the other thing along that same line that I absolutely loved, and it took me a moment to realize the brilliance of an kudos to whichever PM and or engineering team at at Airtable came up with this is.

 

Your your little nested, uh, history. Yes. When you're working on it. I was like, I wanna go back and [00:35:00] fiddle around with what I was doing previously, and it's anchored on chat where I first, I brought my mouse up and it said like, okay, you can go back to chats.

 

Andrew: Hmm.

 

Fraser: I'm like, well, wait, I want the previous re iteration of the spreadsheet.

 

And it's like, no, it, it is telling you again that your, your primary interface here is, is chat.

 

Andrew: Oh, wow. I see what you're looking at. I wouldn't be surprised if the PM. That, or the designer that designed that is hearing this and saying, oh, we didn't do that on purpose, not design, but it's so cool that

 

Fraser: half

 

Nabeel: half of the great features Yeah.

 

Come about that way. The implicit decision from made. Right, because

 

Andrew: you're saying that this feels like a, like a version history nav or like back to your projects or something like that. It

 

Fraser: just reiterates the idea that chat is the primary way with which you're interacting with this product. It, it is.

 

You're not going back to your sheets. You're not going back to your, your database or whatever the Airtable, uh, proper noun is for those. You're going back to your chats.

 

Nabeel: And then I, and then ironically, of course, like much like cursor, you could have two or three [00:36:00] or four chats. That are ostensibly different projects, different areas of this Airtable I'm working on.

 

Yeah. And by, by, uh, categorizing them that way, it's saying, Hey, I've got three or four workflows that are going on different subjects that is obviously just a few steps away from what we're now seeing in Cursor and Rep and elsewhere. We're just like, oh, that's a background agent. Right. That is an ongoing task that just keeps doing it in the background and so on and so forth.

 

Right? Yeah. Instead of just doing this one time, can you just keep doing that?

 

Andrew: Right. You can imagine. This starting to feel like an inbox where stuff is active or archived or starred or whatever. Sure.

 

Fraser: And Nabeel was showing me some of the, the TIC aspects that you all have on your air tables for your project.

 

And it like, it runs almost like CR jog, ay, and it's like this thing's been updated and then it's forcing these other five rows to go and update. And so you can absolutely imagine it feeling like an inbox where it's It's done work. Yeah. The work is [00:37:00] outputted into a spreadsheet that, that, I don't know how many white collar people live in spreadsheets but aren't actually using it for like Excel type formulas.

 

Right. It's just to like organize the thoughts.

 

Nabeel: Well, it's not just Excel type formulas. Right. The beauty of this is that most people who use this probably uses this spreadsheet, not a database. A hundred percent. And so Matt, I'm, I'm sure you know, executive of the Airtable. Over the years, you know, continually bangs their head up against, why isn't we using like a database that's so much more powerful?

 

Yeah. And the good news about having an agent built for you versus you building it yourself is that it can build a lot of those lookup tables and a lot of that like interconnection between sheets that you wouldn't build yourself. In other words, it can build it like a database, which makes 10 more powerful.

 

So on the backend you feel like, oh wow, we're just gonna end up with. People using Airtable in ways they can never use Google sheets.

 

Fraser: Yeah, for sure. And then, and then it takes like information from previous cells. It kicks off like web agent searches off of that and then fills up the rest of your spreadsheet.

 

Feature Discovery Through AI

 

Fraser: It's really remarkable.

 

Andrew: One kinda like side note to all of this that I feel like I should [00:38:00] call out is we're talking a lot about product design and I think like for us at at DS script, we've been very explicit about saying that all of that is. 10th of. The importance of what happens inside of that chat window, wherever hell it is.

 

Oh yeah, yeah, yeah. Like DS script historically is primarily a design idea. It is like, what if you could take a video editor and make it work like a doc? And from there it's just like solving that. Systems design puzzle propped up by really cool technology and as we've started to focus on the agent, we've had to make a break from that and say like, that remains true on the editor, but for the agent, and there's no better example of this than cursor.

 

Which is terribly designed product like, and if you remember when, you know, chat or the composer first came out, it was almost impossible to find it, right? It's still like doesn't open by default, it's in the left sidebar. It's [00:39:00] really easy to miss it, and they've been successful in spite of all of that.

 

And so, so I think like as. All this stuff is really cool, but it's, it's still just like a, kind of like a side note into the importance of the, does the agent to fulfill the promise of, of what you ask it to do.

 

Fraser: Yep. So we, we were, you were showing me the agent in Descript. Yeah. And it was like fulfilling the promise.

 

You, you were

 

Nabeel: blown away. Like, if I had to list a handful of pro, first of all, I don't think very many companies that actually for all the like age agentic last year, plus Fraser, like I actually don't think. Many companies have actually fulfill on the promise of making things. If we just went through the like cursor for X and we talk about this on a regular basis 'cause we're looking for companies that are like actually rewiring their whole companies to be this way and actually aren't very many.

 

I think part of that is not 'cause those people are not competent or blah blah. I actually think it's an opportunity more for companies that are horizontal in nature. [00:40:00] A a lot of sas. Software companies, especially the last five years, have been trending, trending more towards verticalized SaaS, which means like inherently, nevermind how good the product people are, the CEO, they are a single threaded workflow tool.

 

Yeah. Like that is what they are. And so the opportunity afforded to them to be quote unquote agent is just less. Mm-hmm. Like the same thing that was mm-hmm. That was a frustrating product exercise a few years ago for the script, which is like, oh my God, we have this product promise. Of doing this thing.

 

Can I add a word Doc is like, can I edit a video? Like a word doc is also like an unbelievable challenge. The same way that Airtable like, can I make a, A database as easy as a spreadsheet is an unbelievable challenge that takes years and years and years. That difficulty is exactly the opportunity when you get to agent ai.

 

If you can execute well, I'm really impressed by it, but I don't give Andrew and his team all the credit. Like I think part of it is like you guys are in a problem space. That is perfect for this type of [00:41:00] problem. It is a situation where I want to see the word doc on the right, that shows me what an agent is doing.

 

I need a display layer. That's great, but ultimately. Do I wanna be hunting and pecking around a bunch of interfaces to find a little button that says, remove us and ums. Or I just wanna talk to a thing on the left, tell it what I wanna do with the product and say like, Hey, every time I drop a file in here, it's like I tend to ramble in the beginning.

 

So make sure you heavily cut the beginning and then remove my As and um, like just give it a list of instructions the same way you would talk to Cursor and then know it's gonna do that reliably on a regular basis. Like I think you have a, like an interesting problem space. And there's just like not a lot of 'em, like I actually think we couldn't do, we couldn't do this monthly for 20 other products that are, that are behaving in this way that's actually transformative to the business.

 

Andrew: Yeah. So another example of that, getting back to this whole like feature discovery thing and how. The agent almost unexpectedly, we think, will improve feature [00:42:00] discovery by being open worlds. And like before the agent came along, we had this problem of, of building new features. Mm-hmm. And not having. Like intuitive language to describe it.

 

Hmm. For example, like we had a ai when we had remove filler words as an action, like first of all, when we release that, nobody knew what a filler word was. Right? Right. That's like a technical concept, and so what do you call that? That's the best we could come up with. We have another AI action called Edit for Clarity.

 

You know, what does that do? Is anyone gonna go and look for a feature called Edit for Clarity? No way. But now you can just go into an agent and say, Hey, can you. However you would express that. Yeah.

 

Nabeel: I'm rambling a little bit. Can you fix that?

 

Andrew: Yeah. However you'd say. Same thing with like, one of the cool things you can do with the agent is you can take a recording, you can split it into different, what we call scenes.

 

Yeah. And then apply different layouts. Apply like B roll and like what we call internally window dressing a video. But that's not a word [00:43:00] anybody knows. There's no word for that. Like I've asked. Everybody I know that, you know, works in, in, in video, there's no word for that concept of like adding a visual, doing a visual pass, everybody calls it something a little bit different, right?

 

And it's really hard to like wrap that up into a product and expose it at the right moment, but for sure. But people will type it into the agent quite naturally using whatever language they want

 

Fraser: in chat. You fine, you're fine. See, so, um, I agree like none of this stuff matters if the, if the actual product that gets delivered ultimately is, is good.

 

In your case, it's these edits, it's the stylistic changes and whatnot. Have you benefited from investing with that technology and now the agents are making it more like usable, more like discoverable. Or are you designing things that the agents can, that, uh, consume Yeah.

 

Andrew: Are designing [00:44:00] things that the agent can consume?

 

Nabeel: Right. Right. Now, your first pass, like, are you making. Tools for the agent, or are you giving the agent access to all the tools you had previously used?

 

Andrew: Oh, uh, it, well, it started with giving agent, like wiring up all the tools that we already had and now it's building new agent first tools where we're, we're, uh, we have to remind ourselves not to kill ourselves.

 

Thinking about how we would let people. Design a, an avatar, like an AI avatar or something like that, and how we expose that through the editor interface, because most people very quickly are probably just going to be doing that through the agent, right? So our expectation is that within a year, you know, three quarters of.

 

Interactions and, uh, in the app and actions that people take, were gonna be agent first. We added like a bit on every analytics event on whether a human did this thing or the agent. [00:45:00] Whoa, and where's that at now? I mean, we're, we're rolling. I don't know where it is now. We're still in the process of Okay.

 

Of, of launching the agent. Right. Um, so it's pretty low. We

 

Fraser: might be here to talk about Airtable and so we can just say, we don't wanna talk about this, but like, I'm so interested around if you think the primary way in a year's time is going to be interacting with the agent, are you going to have a refounding of Ds script?

 

The Future of Product Design with AI

 

Fraser: And what I think by that is, is how are you going to navigate that? You have a large important user base who is familiar with the product today. Hmm. Versus your excitement and enthusiasm for what will likely feel like a dramatic new product?

 

Andrew: Hmm. I think the question that, so it does feel, it's the first time I've ever used the word pivot uhhuh to talk about DS script is like the focus on the agent.

 

Up until then, pretty much everything we've done is something that we thought we were gonna do at some point. Okay. And the, the agent has really been the thing to take us by surprise. And whether you call it like a, [00:46:00] a refounding moment, that's just like a branding exercise. Yes. But I think it's an interesting question that I've thought about, like N nabeel, you should.

 

Talk about like replic, like the question is whether, whether you as a company need to make really hard choices. Yes. That will alienate your existing users. Yes. In order to bring about the agent. Yes. And

 

Nabeel: yeah. How much of your current user base, and current knowledge and current success. Is a platform of which to stand on to get to the next thing.

 

Or an albatross that will That's right. Actually keep you from the next thing you're trying to build and yeah. The rep example is that that company, which was obviously not an agent company or even a code gen company initially, it was, you know, the way your 16-year-old son or daughter first made their first code online.

 

It's like, you know, it was basically like a. A coding environment in the cloud. Your first cloud little thing to make your, usually to make your Discord bought, which is how it first came on my radar. That [00:47:00] group has a low propensity to pay, um, and a really high engagement cycle. And that's wildly different from if you think you're now an agent to cloud company that is gonna compete with Cursor and is for the everybody on the planet, as they would phrase it.

 

For everybody on the planet who has never coded this is what they want to be their first experience talking to. Someone who's coding and now you're using Claude or their internal models, you're charging for token. You can't be free to use it can't be structured like a social network where you're sharing all your things with everybody else.

 

'cause everybody else is gonna talk to Claude as well, and that'll charge you money like that. That whole structure means that rep went through a process that they. Mostly ended up having to move on from a lot of the users that frankly got them up to a billion dollar valuation like that they were sitting on and like that's a lot of courage.

 

Frankly, there's a lot of CEOs and leadership teams that would not have been able to do that. And you know, kudos to them for navigating that. This does not feel like the same thing for dscr. I'm sure there are users [00:48:00] that, for you guys. Have come along that have certain sets of expectations and keep asking you for some esoteric little thing that they use in Adobe After Effects or something like that, if they're pros.

 

I'm sure you've gotten that forever, and I'm sure with Airtable similarly, they've navigated this like, well, if I'm a real database nerd, why does an Airtable have this database nerd? But that doesn't mean you have to leave them behind. I don't think.

 

Andrew: I don't think so either, but it's something I ask myself a lot.

 

Right. It's there's, there's not just, not just the pricing change that they made, but also, I mean, if you had used Repla a year ago or whatever it was, it was just like an IDE in the cloud and it felt like an IDE where you're like a code editor. Yeah. And now if that. Part of the app exists anymore. It's incredibly buried.

 

Mm-hmm. That's true. Yeah. They're trying to go in and, and write capture. You can still go in and like open a file and write code, but

 

Nabeel: someone we're buried, this app will be twice as buried in another six months.

 

Andrew: Yeah. Yeah. So there's a question of like, is there an equivalent move to that? For us, I think [00:49:00] like what, what Repli has done seems so smart to me.

 

Cursor by comparison. You know, I, I think there's like a 0% chance that a year or two from now you're gonna see like, product managers, vibe, coding, and a fork of VS. Code. Yeah. It's just like not the right product. Yeah. Agree. Or abstraction level. Even the, the cursor, you know, founders and other people talk about how.

 

There's gonna be this new abstraction layer that becomes the source of truth for the stuff that the software that we're writing that looks more like a spec. Yep, yep. Uh, a natural language eng English thing that, um, yeah,

 

Nabeel: you should be talking to a PRD, not the document.

 

Andrew: Right. Yeah. So like what Replit has done has said like, code is not the right layer to focus on.

 

We're actually gonna start building out a human editor, and they're putting effort into like a wissy wig style editor, so that mm-hmm. You know, sometimes you don't wanna tell the agents to change the hex code on, uh, a button [00:50:00] or something like that, right. You can just click on it and change it yourself, and that's more efficient and precise, so.

 

We feel like we've, that's what we've been up to all along, is creating that new human editor abstraction layer that's higher level than words instead of like timelines. And so we're pretty well suited to it, but, but trying to keep asking ourselves like if there's something more aggressive and disruptive that we're missing.

 

Nabeel: I mean, arguably the one, the only one, and this is true for all the question that everybody asks. 'cause everything in AI asks questions, and every layer of the abstraction is like, do humans have to be involved at all? There's a bet in Airtable, there's a bet in Descript. There's a bet in in bed, in cursor and rept that humans are involved, that a human has to look at the PRD, um, even if you're, even if you're upping the abstraction layer, there's a human involved and I don't know.

 

I'm happy to make that bet that the creativity of a few minutes is still somewhere in the loop, even if that loop moves from every 10 seconds to every minute, to every five [00:51:00] minutes, or whatever it is. Yeah. Cool. I I think we should, uh, we should be done. Stuart. It, thank you so much for coming and hanging and, and chatting.

 

Fraser: I think you were like, for a period of time, the epitome of the Silicon Valley promise of founders. It felt like it was very pure, like you wanted to go do this thing. And we wanted to build and, you know, the, the outcome was the outcome. But he even handled that in a way that just continued to build on the promise of like Silicon Valley as a, as a concept.

 

And so like, thank you.

 

Andrew: Yeah. My pleasure. It was fun. Thank you. Thanks guys.