# From No-Code to AI-Powered Apps with Airtable’s Howie Liu

- **Show:** Gradient Dissent
- **Published:** 12 Sep 2024
- **Source:** https://wandb.ai/site/resources/podcast
- **Transcript:** https://podcasts.opennash.com/shows/gradient-dissent/from-no-code-to-ai-powered-apps-with-airtables-howie-liu-876c42/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Howie Liu argues that Airtable's lasting value comes from being a flexible platform where business people build their own database-centric apps, now extended with AI as familiar building blocks rather than a flashy add-on. He says durable products start from real recurring business problems, not AI novelty, and that most knowledge work needs humans and AI steps combined.

The episode moves from Airtable's patient origin and growth, through its push into enterprise workflows and AI features like Cobuilder, to a broader forecast. Liu is optimistic about AI's long-term economic value but expects a bubble-style correction when customers see no return, as happened after the dot-com era.

## Key insights

1. **Platform flexibility beat feature-rich vertical software** Liu says Salesforce dominated CRM because it was a customizable platform, not because of specific features. He built Airtable on that lesson so business users define their own data model.
2. **Slow, validated building protected Airtable's second startup** Airtable took about two and a half years to reach general availability, while his first company lasted one year. Liu says he tested wireframes, prototype backends and beta users throughout to confirm demand.
3. **AI works best as modular building blocks** Airtable exposes AI as fields and automations that people can review, edit and chain. Liu argues this fits creative, open-ended knowledge work better than end-to-end autonomous agents.
4. **Non-deterministic LLMs demand local experimentation** Liu says nobody knows in advance where models will excel across real business tasks. So companies should let people in each function test AI on their own data rather than rely on a top-down AI committee.
5. **Liu predicts an AI spending shakeout** He says some customers buy AI tools without measurable return, and will eventually cut spending, forcing investors to adjust. He compares it to the Web 1.0 collapse, which left durable companies like Amazon stronger.

## Questions this episode answers

### What does Airtable actually do?

Airtable is a platform for building internal business apps around a database, such as inventory, content production or CRM systems. Howie Liu says non-engineers can define their own data and workflows without writing code.

### What is Airtable's AI Cobuilder?

Cobuilder lets you describe the app you want in plain business language and it builds a first version for you. Liu says the underlying language models already understand industries like retail, so little explanation is needed.

### Will AI replace software developers or business workflows?

Liu doesn't think so. He expects developers to continue for 20 years because cheaper software creation means more software gets built, and he says most knowledge work needs AI steps inside human workflows.

### Why do many AI products have high interest but poor retention?

Liu calls it the 'AI tourism' effect: people try a flashy novelty once, but it solves no recurring business problem. He says Airtable focuses on lasting value instead.

### Does Liu think AI is a bubble?

He thinks the technology is real and worth trillions in value, but that current hype, funding and spending outrun delivered results. He expects a correction similar to the dot-com collapse.

## Facts

- **Howie Liu** [0:09:32]: "my first company lasted exactly a year, um, almost to the day, from start to acquisition, right, by Salesforce."
- **Howie Liu** [0:09:32]: "Airtable, to your point, took probably 2 and a half years to even build the GA product that we could launch publicly."
- **Howie Liu** [0:07:19]: "I worked for this ad tech startup. And, like, the the, um, you know, I actually ended up realizing that the most valuable thing I could do during my summer there"
- **Howie Liu** [0:06:43]: "interning for your last company, CrowdFlower. It was kind of like early pioneer in in, um, uh, a lot of what would become this AI world and training data"
- **Howie Liu** [0:14:54]: "the Idaho Farm Bureau had a op-ed in it that referenced how Airtable was the new CattleMax."
- **Howie Liu** [0:15:43]: "we started around the same time as Figma, but I think what we both benefited from was both a shift in consumer"
- **Howie Liu** [0:16:12]: "using a lot of the benefits of, like, HTML5, WebGL, some of the, like, just modern standards and web"
- **Howie Liu** [0:20:11]: "I don't think we launched with even the ability to import spreadsheets, but that was a fast follow, right?"
- **Howie Liu** [0:21:53]: "we can compound our our biggest customers value by basically 2x every single year."
- **Howie Liu** [0:24:51]: "we see the best growth in retail, financial services, media, uh, and tech"
- **Howie Liu** [0:25:32]: "now we have a lot of customers who are pulling in data from, for instance, Snowflake or Databricks into Airtable"
- **Howie Liu** [0:42:20]: "we're actually building a new data store into Airtable that will allow you t uh, allow you to ingest multi-hundred million record data sets into Airtable directly"
- **Howie Liu** [0:31:42]: "we just launched, uh, our AI co-builder experience."
- **Howie Liu** [0:39:51]: "we did this reorg in our product group, um, about, call it 9 months ago now"
- **Howie Liu** [0:54:08]: "as the context windows have radically expanded from, like, you know, I think the best ones had a few thousand as of 2 years ago to now 200,000"
- **Howie Liu** [0:55:50]: "obviously, you know, very timely, uh, you know, to discuss this with the the Llama 400 billion parameter model coming out"
- **Howie Liu** [1:01:25]: "I'm pricing a lot of, like, upside into Adobe, right?"
- **Howie Liu** [1:01:56]: "Tome, as an example, started as more of, like, this AI, uh, slideshow tool, right?"
- **Howie Liu** [0:15:43]: "Web van predated Instacart by at least what, like, a decade, and that timing was wrong"
- **Howie Liu** [1:11:20]: "I signed up to be a beta tester for a smell printer company, right?"
- **Howie Liu** [1:11:20]: "you had really good companies like Amazon that actually were born in that era, right?"
- **Howie Liu** [1:09:23]: "somebody saying, like, I'm talking to a customer who, uh, you know, who is, you know, saying Microsoft Copilot is their main AI"

## Transcript

**[0:00:02] Lukas Biewald:** You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. In this interview, I'm talking with Howie Liu, an old friend of mine and the CEO of Airtable, a very successful app building company for businesses. I first met Howie when he was just out of college and he interned for me. And even at the time, it was obvious he was one of the smartest people thinking about product and business, and now he's become one of the smartest people I know thinking about the intersection of product and business and AI.

**[0:00:37] Lukas Biewald:** So, it was one of the easiest interviews I've ever prepped for. I had so many questions for him, and I hope you enjoy this interview as much as I enjoyed doing it. All right, Howie, good to see you. Awesome, thanks Thanks for doing you. the podcast. Um, you know, I feel like most of the people watching this or listening this probably know about Airtable and what it does. But, you know, maybe just to for the record, or maybe, um, for those that don't know, if you could just describe Airtable today and and the vision behind it.

**[0:01:05] Howie Liu:** Yeah, for the record, it's actually the name of our, um, blog, because, you know, Airtable records, for the record. Uh, yeah, so, you know, Airtable was founded with a very clear premise, which was, you know, I had worked at Salesforce for a year and a half before, um, had sold a small startup to them. And while at Salesforce, learned 2 really important things. One is, you know, the startup I had was kind of this small, disruptive CRM product, had a bunch of cool new features, like integrations with social, et cetera.

**[0:01:33] Howie Liu:** And, you know, obviously, Salesforce is this, you know, CRM, uh, CRM giant, right? Like, they they just dominate, and they, you know, even at that time, had multiple billions of revenue, they were already the clear winner. And so, first, I what I learned is that Salesforce is so dominant at CRM, not because they built all these fancy bells and whistles for CRM specifically, they have some of those. But the even more important premise is that they are a platform, right?

**[0:01:56] Howie Liu:** So they can be completely customized for any use case. The data model is custom to every customer, so it's not just always the same setup of contacts, accounts, et cetera. Um, so platform, you know, being powerful, that was one takeaway. And then I think the second is, you know, going after a really, really large market, in their case, CRM, and then later, you know, adjacent these to the sales CRM market, just lends all kinds of opportunities to continue to innovate and and kinda drive, sustain like this engine for, you know, being more ambitious over time.

**[0:02:28] Howie Liu:** So it's a self-fulfilling kind of prophecy. And I took both of those away in in the founding of Airtable, one because we wanted to build an app platform that was just radically easier to use, right? So Airtable is designed to enable people who are actually in the business, whether it's a marketing person or a product person, et cetera, to build a really useful app. But with this platform approach where, you know, we're not going out and prescribing, Hey, here's exactly what the solution for every single possible use case will look like.

**[0:02:55] Howie Liu:** We're enabling people to build their own apps, right, and to find their own data model and create their own workflow around that. Um, and then secondarily, you know, the the vision has always been about starting, you know, small, as is the case with with kind of the typical, like, you know, low-end disruption, but moving up and up and up and up into more and more high-scale, high-complexity, high-value use cases, which we now have done. So, you know, kind of the the condensed, you know, 10 year plus history of Airtable is we started with disruptive, you know, really, really simple, easy to use no-code app, uh, builder.

**[0:03:27] Howie Liu:** And since then have been scaling up to support really large, you know, kind of transformative end-to-end processes like content production at the world's largest media companies, which is a very multi-stage, you know, kind of almost ERP-like process. But instead of for physical assets, for, in their case, content or digital assets, right? And that's a very common type of pattern we've seen even in other industries like retail, tech, financial services, et cetera.

**[0:03:54] Lukas Biewald:** Interesting, that's different than what I thought you were you were going to say. I feel like maybe I, uh, I'm out of date or something. Um, I so so, let me repeat it back. So Airtable is like an app building platform. Is that is that fair? And is that all cause? I remember in the early days, kind of watching you build it and sort of thinking, this looks like, um, almost like a way to have a a database without having to administer a database was kind of my early experience of it.

**[0:04:23] Lukas Biewald:** Is that is that wrong? Or or

**[0:04:25] Howie Liu:** No, that that's correct. And I think what, um, you know, what we really centered on is that the types of apps that we're focused on are very database-centric, data model-centric, right? So, Airtable is for internal business apps, not for external, like, you wouldn't use us to build a game, you wouldn't use us to build, like, the next high-scale consumer app. You're not gonna build, like, Instagram using Airtable, right? But what you are gonna do is build internal use cases that look like CRM, that look like ERP, that look like inventory management.

**[0:04:59] Howie Liu:** And it turns out having a database-centric platform where now anyone, the business person who's closest to that, you know, use case and process, can actually define the data model they want. They can say, This is the schema I want. It's not projects, it's not tasks, it's really about representing this object. It could be, again, you know, kind of your unique kind of inventory, it could be your unique type of content assets. Um, you know, that's the starting point.

**[0:05:25] Howie Liu:** And I think, you know, if you look at, like, traditional programming, um, I always found there to be this big unlock when, you know, I first like shifted from early days of my programming career, you know, thinking about logic first and, like, writing C plus plus code and doing, like, you know, kind of Monte Carlo simulations and just logical stuff. To instead thinking about useful app development, um, and and usually the starting point of that was, like, first, you define the data model, right?

**[0:05:50] Howie Liu:** In, like, Ruby on Rails, often you start with, like, the ORM, the object, you know, model, and say, Here's the objects that I want to work with.

**[0:06:00] Lukas Biewald:** interesting? I I remember also when we were first talking about Airtable, but before I even saw it, you were talking about your love of of hypercard, which is something I remember hacking on as as a kid. Is is that is there any of that influence on still an Airtable product?

**[0:06:15] Howie Liu:** I mean, you're peeking you're revealing just how cool I am. You know, some people talk about, you know, sports, some people talk about, you know, their cool hobby. And here I was talking about hypercard, this, like, 30 year dead product that, like, was really cool and ahead of its time, right? But, um and by the way, like, I do wanna give a shout out to the the backstory, which is the way you and I met, uh, you know, my first job out of college, and I had done a lot of, like, tinkering, programming, knew I wanted to do a startup.

**[0:06:43] Howie Liu:** Uh, even then was interning for your last company, CrowdFlower. It was kind of like early pioneer in in, um, uh, a lot of what would become this AI world and training data and and, you know, and so on. Um, but um, but I think even then, you know, something that really excited me was this idea of, like, solving the meta problem behind building apps, right? Like, I had personally been tinkering on lots of different apps. I think I even built maybe 1 or 2 in the context of of being, uh, your intern, um, you know, and then even, like, past work I had done while programming, like, a lot of it was like building some kind of app.

**[0:07:19] Howie Liu:** Like, act actually, at one, um, uh, previous internship, uh, while I was in school, uh, I worked for this ad tech startup. And, like, the the, um, you know, I actually ended up realizing that the most valuable thing I could do during my summer there was not the job they had intended for me, but, in fact, building them this like custom asset manager kind of like content management system for them, right? Like, for for basically creative assets as they were going out and pitching their their, uh, you know, their brand customers.

**[0:07:45] Howie Liu:** And what I ended up building was basically kind of this like database with a layer on top of it. And the point of all this is that, you know, I've always tried to step back and think about, you know, what what's the underlying abstraction that can save me the effort of having of having to make each and every one of these apps. And I think HyperCard was like a really early, um, and kind of like profound product in that, like, they really tried to make, you know, app development very accessible.

**[0:08:13] Howie Liu:** It was kind of maybe one of the first no-code products because while it felt like a spreadsheet or not not a spreadsheet, a, um, a slideshow product, like, you created different cards, you know, there were underlying data table concepts, and you could kinda create logic. And if you clicked this, it would go to, you know, this page. And so there was this, like, interface layer on top of a data layer and some, like, logic in between that, you know, as a kid, I remember, like, building some really interesting interactive apps with.

**[0:08:38] Howie Liu:** If you actually go back and look in the, um, like, historical, like, you know, archives, like, there's, like, videos of people who have talked about, like, you know, I'm a car dealership owner, and I built my entire inventory management and CRM system on HyperCard. So, like, just an early example of how, like, if you design a really good and flexible platform, people end up figuring out how to solve lots of different business problems with it.

**[0:09:01] Lukas Biewald:** Hm. I g- I guess 1 more question on the on the history of Airtable before I get into what you're doing now. I I one thing that I watched, you know, kinda from the sidelines was it took you a long time before you did a a public wreath, and it was it was kind of interesting cause it was your 2nd company. And it seemed like you actually slowed down the process before putting something onto the world, just so contrary to the, um, the advice that you normally get.

**[0:09:26] Lukas Biewald:** Can you talk about, like, what was the thinking behind that, and maybe, like, would you do it again the same way? Was that instrumental in your success?

**[0:09:32] Howie Liu:** Yeah. So, you know, the the so to to kind of reflect on your point, um, my first company lasted exactly a year, um, almost to the day, from start to acquisition, right, by Salesforce. And, uh, and Airtable, to your point, took probably 2 and a half years to even build the GA product that we could launch publicly. So, like, the entire span of my first company was, like, not even the time it took us to build and launch, you know, the the 2nd product, right, or the 2nd company, Airtable.

**[0:10:06] Howie Liu:** And I think there's a few reasons for this. 1 is, you know, when you don't have much conviction about the market and, you know, kind of your unique take on the market, I mean, ultimately, like, creating product-market fit is about having a differentiated point of view on how real world problems can be solved with a new offering, right? Um, if you're doing the same thing as everybody else, like, by definition, you're not gonna, like, really gain entry or be disruptive.

**[0:10:32] Howie Liu:** And so you have to have, like, a fresh take on, you know, the problem and maybe even what the problem is itself. And, you know, I think with my first startup, Etactz, I don't think we really had that much conviction about it. I mean, we we thought we had an inkling of it. You know, there were some interesting trends emerging around, like, you know, lighter weight CRM, CRM that had, you know, could incorporate, you know, kind of automatic intelligence around social media profiles, et cetera.

**[0:10:59] Howie Liu:** But if you, like, said to me, Hey, you gotta bet, like, your entire future livelihood, like, your entire next 30 years of your career on this product idea being the one that's gonna, like, you know, just be right and make it. I would've said, like, Okay, well, I mean, you know, like, jigs up, like, I I give up, right? And to some extent, that's kind of what happened, right? Like, a year in, we had made some traction.

**[0:11:21] Howie Liu:** Um, you know, we had gotten, you know, gotten so far and, like, I think, done some interesting things and actually started to monetize that product. But, like, also kind of realized there was a lot we didn't know, right? And so, like, having the humility to even admit, you know, to yourself, Hey, like, I don't know what I don't know. Um, and so going into Salesforce, I both learned about, you know, and and and kind of formalized my thinking around this platform opportunity.

**[0:11:42] Howie Liu:** But that also kinda got a a glimpse into a company that was a, you know, the first, like, big company I ever worked for, right? Um, company that had really kind of reached, um, you know, takeoff trajectory, right? Um, and, you know, probably at a scale of, like, a couple billion in revenue, but still, you know, early and scrappy relative to, like, the the huge, huge giant they are now. Um, and I think that also gave me some conviction of seeing what is a company that has made it and has figured out the market and now has this massive flywheel of momentum look like.

**[0:12:14] Howie Liu:** And so going into Airtable, I think my my approach was look, if I can, like, be very deliberate and thoughtful about the product bat that I'm making and, you know, really invested in creating a differentiated and, um, you know, hard to replicate product, then, you know, sort of similar to, like, building a rocket, like, if you're SpaceX, you know, I know that once we do it, there is a market for it. I understand the use cases and the pain point.

**[0:12:42] Howie Liu:** And I also believe that we can create this flywheel so it gets easier over time and we get more leverage where right versus I think there are some products where it's faster to build and launch, but then you has have less durable advantages over time, right? Um, and you could be a flash in the pan taking advantage of kind of, like, a near-term, you know, especially in AI right now, I think there's a lot of, like, you know, cool products that ultimately become short-lived because it's more of, like, you know, kind of a cool, fatty, like, AI tourist experience, um, like all the avatar products where, I mean, they're fun and cool, but it's hard to imagine kind of a

**[0:13:15] Howie Liu:** durable flywheel, uh, compounding from there.

**[0:13:20] Lukas Biewald:** I guess what y- you've defined your product very broadly. So, I guess, you know, like, in some sense, some kind of app building tool surely would be successful. But what gave you, like, the what what gave you, like, the confidence, um, to to, like, what gave you that that confidence? Was it sort of seeing Salesforce customers? Like, I I do remember talking to you about, like, wow, like, you know, Salesforce has all this, like, missed opportunity here by not, like, making things like more abstract, um, so that customers could build these types of apps.

**[0:13:54] Lukas Biewald:** Is that where it was coming from?

**[0:13:56] Howie Liu:** I think there are a number of different insights that, you know, kind of added together gave me that that overall conviction, right? Like, you know, it's kind of like your Bayesian priors. If you add enough of them up, like, you you kind of get a a, you know, a total confidence that's that's higher. And so, you know, in my case, it was partly seeing some of the insights from, you know, like, the ways in which Salesforce was really strong, and also the ways in which it was kind of underserving use cases that were more away from CRM, right, traditional CRM.

**[0:14:26] Howie Liu:** Like, the further away from that you got, the harder it was to kind of, you know, there there weren't many good solutions to build, like, a t to stand up an app, right? Um, and aside, like, you know, as I explored some of these app ideas, like inventory management for SMBs, let's say, like, there were very crappy, kind of not very flexible or scalable, um, uh, custom solutions or vertical solutions, right? Like, there were entire companies that just focus on that one use case, or it could be, like, donor fundraising, right?

**[0:14:54] Howie Liu:** Um, there were actual products that that just focus on this. Um, and by the way, we would later encounter a lot of these and, like, kind of subsume, I think, the micro markets for a lot of these after we launch. Like, one of my favorite pieces of press to this date was, um, the Idaho Farm Bureau had a op-ed in it that referenced how Airtable was the new CattleMax. And apparently, like, if you Google CattleMax, it actually was this, like, verticalized software for just cattle tracking.

**[0:15:18] Howie Liu:** It was like a cattle tracking database powered app, right? And and so, like, you know, there was just a lot of these, like, very brittle, kind of narrow niche markets that I think were being poorly served because there wasn't, like, this flexible app platform that had, you know, our case, also, like, a lot of modern innovations, like real-time collaboration, a mobile-friendly layout, and just, like, this high degree of UX ease of use that was reflective of, like, a new generation of software.

**[0:15:43] Howie Liu:** And I will note also that I think the timing was right. I think a lot of, you know, business ideas, you know, are sometimes about the timing, right? Like, web van predated Instacart by at least what, like, a decade, and that timing was wrong, but Instacart's timing clearly was correct, right? And so I think with Airtable, uh, we we started around the same time as Figma, but I think what we both benefited from was both a shift in consumer or, you know, even, like, business end user expectations for their software.

**[0:16:12] Howie Liu:** They wanted it to be more flexible, better UX, um, web-based, and also real-time collaborative. And then, second of all, I think the technology to build that kind of software became more, um, accessible and also potent, right? Like, we were able to go and build a web-native product that felt like a desktop-grade experience, using a lot of the benefits of, like, HTML5, WebGL, some of the, like, just modern standards and web that, you know, in my prior company, even just a few years prior, like, were not possible, right?

**[0:16:42] Howie Liu:** Uh, were really, really hard. Even, like, the real-time, you know, backend element and creating our own proprietary kind of database engine would have been much harder, I think, 3 or 4 or 5 years prior. Um, all that being said, we also, you know, we didn't just lock ourselves in a room for 2 and a half years and build this product hoping it would work. I think we were doing a lot of market validation even without a fully launched product in between.

**[0:17:04] Howie Liu:** So, early on, it would be like taking, like, a wireframe prototype to some target, you know, customers users, often in our, you know, kind of friends and family network, um, and, like, showing, you know, product mockups to them and saying, like, Hey, do you think you could see yourself using this for X, Y, Z use case? And so we were kinda doing use case discovery and trying to find, like, the jobs to be done specifically for for this product even early on.

**[0:17:29] Howie Liu:** And we would basically do that at every phase, right? So, at some point, we built back built a backend. It was kind of a, like, a very, um, like, flimsy backend or, like, an intentionally just kind of a prototype backend. But it was enough to start testing it for even, like, small, solo use cases, right? Um, so I think we had a few customers, you know, if you could even call them that, like, beta testing this early, early version.

**[0:17:51] Howie Liu:** Um, and then eventually, we'd get, like, more, uh, robust use cases even before launch on our beta. So, I think at every point, we were able to validate a, like, people needed this kind of an app platform, and that, like, the UX that we had designed around it made sense to them. B that, you know, like, they could actually, you know, fully implement a valuable use case once once we had built it. Cause, like, the theory versus, like, the reality of these things sometimes, like, you know, is not the same.

**[0:18:17] Howie Liu:** Um, and then C, like, once we launched, I think we we figured out, like, there actually was enough stickiness and virality to this product that you could create this compounding growth flywheel from it.

**[0:18:28] Lukas Biewald:** Hm. And I I get, I I promise we'll get to the AI soon, but I I I do have some other questions. I I remember, like, I hadn't seen you in a couple years, maybe, like, 5 years, 6 years, I'm not sure. And I was at an investor conference and I watched you, like, present this chart of your, like, user numbers that looked like it looked like a mathematical function or, like, a caricature of, like it was just, like, this beautiful, like, smooth exponential curve that I I was just, like, so, like, deeply jealous of that.

**[0:18:57] Lukas Biewald:** Like, was there any, like, uh, like, along the way, has it been this sort of, like, smooth, like, kind of rocket ship to success, like, it's looked like from the outside? Or have there been sort of, like, you know, kind of pivotal moments or or or, like, doubt in your in your strategy?

**[0:19:14] Howie Liu:** Yeah, I mean, so there's definitely been hiccups for sure, but even, like, you know, the early days growth, I would say, first of all, like, you know, when you zoom out enough, everything looks like a smoother distribution or, like, a curve than if you zoom in, right? Like, you know, you zoom out and, like, there's normal distributions and the you know, like, there there's, like, growth curves and so on. But, like, the truth is, like, it was probably very chunky and, like, there were kind of different inflection rates throughout time, even from the earliest days, right?

**[0:19:40] Howie Liu:** And I remember even

**[0:19:41] Lukas Biewald:** Oh, no, I'm gonna ask, uh, the investor to send me that graph. It would be funny

**[0:19:45] Howie Liu:** So you stare at it, like, you know, te- 10 times, right? And, like, look at this section, that section. Um, but like, it's it's it's always fascinated me this, like, the statistical concept of, like, you know, when you zoom out enough, like, things start to, like, smooth out. Um, but, uh, but yeah, I mean, like, living the experience of, like, the day to day and the the kind of, like, month to month, like, you know, challenges we had as a, you know, as a company, like, trying to gain that momentum.

**[0:20:11] Howie Liu:** I mean, you know, early on, we we got some early traction, right, when we launched, and there were a lot of people who were enthusiastic about the product, but then we quickly realized we had to, you know, like, build easier ways to bring data into Airtable, right? So, like, I don't think we launched with even the ability to import spreadsheets, but that was a fast follow, right? Because because it was obvious that people needed to easily be able to bring in their data from existing spreadsheets or tools into Airtable.

**[0:20:34] Howie Liu:** Or, for instance, like, quickly building out templates, because as we started to see people come in, some people figured out, like, here's how I build the use case I have in my mind in into Airtable. But other people needed to kinda see a starting point to react to or to edit. Uh, so there were definitely a lot of, like, kinda moments where I think had we not done those things, like, the growth probably would have linearized.

**[0:20:55] Howie Liu:** And I think, I mean, probably a lot of companies whose zoomed out growth looks very smooth, actually sort of in the moment had to execute in ways that unblocked the the smooth continuum of growth, right? It's almost like, you know, if you have, like, an organically, uh, multiplying, you know, organism, like single celled, like bacteria or something. Like, if you don't give it enough of its necessary inputs and, like, nutrients, like, it's gonna level out at some point.

**[0:21:22] Howie Liu:** So I think it's ca- it feels like you're constantly trying to find, like, the level out point and figure out, like, what is gonna cause that, um, and then unblock it, right? Um, and so that could be, like, these activation techniques, these ways of, like, inspiring people with the specific use cases. At some point, we made this transition into more of an enterprise sold, uh, growth model that layered on top of and supports, uh, you know, um, the the original, like, bottoms-up, uh, product-led growth engine, um, you know.

**[0:21:53] Howie Liu:** And, you know, I think, like, even now, as we transition into, you know, becoming this enterprise-grade platform that can solve really large spanning use cases, that means that, like, from a revenue curve standpoint, we can continue to, like, really compound our revenue. Because, you know, if we win these large, enterprise spanning use cases that are very valuable, they they're worth multiple millions, right, each. Um, and that's that's something we've really been focused on, like, we can compound our our biggest customers value by basically 2x every single year.

**[0:22:22] Howie Liu:** So this is a a, you know, one of the most important, more recent stats that has helped us continue to, like, you know, compound the the growth. But, um, but, you know, you switch from just getting lots and lots and lots of small customers at the top of the funnel, you know, to keep streaming in to, instead, you still get them, but then you can also, like, double down on and really compound the biggest appointments. So, I think it's always been about, like, getting ahead of the plateau point, cause every I think every system, every type of, like, product or organism, and in a way, like, a company is sort of like an organism, like, it's trying to grow and

**[0:22:58] Howie Liu:** consume more to, like, then, like, sustain itself to keep getting bigger. And, you know, but at some point, you're gonna, like, run out of, like, the space it needs to grow unless you go and, like, figure out how to, like, unblock. So it's definitely felt like that. And, you know, um, certainly each of the the transitions that I described in the moment was, like, actually more complicated or required more, like, alignment, um, and sometimes, like, you know, not a pivot, I would say, but, like, a refocus of the organization internally to to be able to fully deliver on.

**[0:23:34] Lukas Biewald:** And so is the transformation that you feel like you're going through right now, this transformation to accommodate more complicated workflows for bigger enterprises, is that is that right?

**[0:23:44] Howie Liu:** Yeah, and I would say it's not just about complexity, because sometimes there's, like, highly complex workflows that are sort of esoteric, right? Um, like, have a really complicated, like, financial planning use case in finance that, you know, is just very complicated, but, like, not enterprise-wide spanning. Um, and so the the types of use cases we're really interested in basically kind of look like the modern equivalent of what ERP was and SAP was to a supply ch- uh, supply chain, physical manufacturing-centric, you know, world, um, to now, you know, obviously the supply chain that matters is more of a digital, right?

**[0:24:20] Howie Liu:** It's, you know, digital product operations, not just in tech companies, but in every company, large retailers, financial services companies, they're all trying to become digital companies with their own, you know, consumer-facing, uh, you know, digital products. And so building the products actually becomes a new form of supply chain, right? Or, you know, I mentioned, like, the media company example of, like, producing content and deploying it at a much faster rate, um, and more of them to streaming platforms or to, you know, new methods of delivery, right, across other platforms.

**[0:24:51] Howie Liu:** Um, you know, it could be, uh, within, um, you know, a really large retailer, actually, like, you know, some blend of the old physical supply chain with now, like, the new way that they they do omni-channel planning and marketing, et cetera. And so I think what we're trying to attack is the new economy value chain of all of these companies that are basically getting disrupted and their industries are getting transformed. And we wanna be the backbone of that value chain in each of those top industries, mainly, you know, we we see the best growth in retail, financial services, media, uh, and tech, um, so we wanna be to this new world order what SAP was to the old world order.

**[0:25:32] Howie Liu:** Of course, our product looks radically different, right? I mean, it's dramatically easier, faster to build on, it's, um, you know, there's centralized, you know, centralized data sets. In fact, now we have a lot of customers who are pulling in data from, for instance, Snowflake or Databricks into Airtable, um, but then also creating these operational workflows, um, that represent the different stages of the value chain into Airtable. So that's one vector in which we're really kind of focused.

**[0:26:00] Howie Liu:** And then I think the second vector is really around disrupting ourselves and reimagining what this whole thing looks like when it's truly AI native, right? Um, and I'm sure we'll we'll dive more into this, but, you know, as a very meta product ourselves, like, we are an app platform, so we enable our own customers to build apps. We have, like, a very meta role to play in terms of, you know, incorporating AI. So it's not just about leveraging AI to enhance a feature, it's really about both using AI to completely change how people build apps, cause now you can have AI help you build that app in the first place, and then also to incorporate AI into the runtime of

**[0:26:37] Howie Liu:** those apps. So effectively allowing you to take the capabilities of today's already models and put them to work in the context of real, you know, company data in real workflow form. Not just through kind of a RAG QA use case, but actually embedded into a process and automating key recurring steps that might actually add up to many, many hours of, uh, employee time and also slowing down the output. So if you can deliver content faster and for half the cost, I mean, that's, like, completely disruptive in in a in the media industry, right?

**[0:27:11] Howie Liu:** And we wanna do that with our AI runtime capabilities for every industry that we serve.

**[0:27:17] Lukas Biewald:** And I guess, before before going down the AI path, I wanna make sure I I really understand, you know, what what you're doing. Y- you're a little abstract for me in terms of, like, you know, kind of managing big companies operations, which maybe I don't don't understand as well as I should. Like, could you describe, like, a real-world use case that's sort of representative of what someone in, like, say, retail, you know, might do with Airtable?

**[0:27:38] Howie Liu:** Yeah, so, um, so a common one is, if you think about, like, what a retail company does, right? Like, you know, they sell stuff, right? And they used to sell stuff in stores, now they sell stuff online. And a big part of their value add, especially if they don't necessarily make all their own products, is first, you know, having a central database, usually called a PIM, or a product information management system, that represents all the the entire universe of possible products that they could have in stock, they could stock, they could source from their vendors, right?

**[0:28:08] Howie Liu:** So this is kind of like the the kind of, um, uh, the master data set of, like, here's all the products I could be working with. And then they generally pick which of these products do I wanna use, right? Um, and and actually, uh, market and sell for this season. And then they'll go through kind of creative operations where, you know, they take these products, they get a version of it, they photograph it in front of beautiful settings.

**[0:28:29] Howie Liu:** Like, you think about, like, furniture companies that, like, you know, it's about how you make the furniture look in a beautiful home or, like, beautiful sunlight and, you know, kind of arranged in a beautiful kind of design. Um, that's a lot of the value add, right, is making the these products look enticing and giving a vision of, like, how you would use them. And then, of course, you actually, like, take these assets and put them on the website, you put them in marketing campaigns, you may put them in, like, physical print catalogs, et cetera.

**[0:28:52] Howie Liu:** And then ultimately, that may, you know, uh, drive, uh, e-commerce transactions, people, like, buy in store. There's kind of a cross-pollination between the two. And at each of those steps, there is a role for Airtable to play, both because we can be increasingly the scalable database that houses, you know, the source of truth for for a lot of that, right? So if you pick the seasonal edit of products that you're gonna feature this this, uh, this season, you can actually represent that in Airtable.

**[0:29:19] Howie Liu:** Airtable becomes the system of record for that, right? And then if you're gonna do, like, creative production and manage, like, the the literal photo shoots and the creative asset, you know, production around each of those products, you can manage all of that in Airtable, right? So here's all the different steps, here's the different, like, review, um, steps of the, you know, the shoots that we've done, here's the creative assets we've produced, we can represent all of that in Airtable.

**[0:29:42] Howie Liu:** And then finally, when you actually wanna go and deploy it into marketing campaigns and put it on your site, et cetera, you can manage all of that in Airtable as well. So, you know, it's not project management, because I think project management to me associates with more, like, dry and tactical tasks, right? So it's like a to-do list. But it really is, like, representing all of these different stages of, you know, what the business cares about and giving you a database to house that data, but then giving you this kind of workflow interface layer.

**[0:30:11] Howie Liu:** And increasingly, you know, what we've been shifting towards is, you know, when when we first launched Airtable, the only interface you had was this spreadsheet-like interface, right? So you got a database, but the the main way of interacting with it was, like, a spreadsheet-looking thing. Um, and what we've increasingly done is built more streamlined interfaces where you can create, like, something that looks more like a a really custom interface for somebody to go through and, like, you know, triage the creative assets, right, or perform a workflow.

**[0:30:40] Howie Liu:** And so that's really where, you know, we've evolved from just being, you know, the database plus a default grid view on top of it, to now being able to support these bigger data sets, but also these more sophisticated and more nuanced workflows around it.

**[0:30:56] Lukas Biewald:** Hm. And okay, so now now that I get that, now let's talk about how AI fits into your into your product and and this vision. I mean, you gave, like, an incredibly compelling demo at actually not our last conference, but the year before. So you were, like, right Great. on the cutting edge, like, one of the first um, it was one of the first really kinda compelling business applications of LLMs that that I had seen. But, you know, I actually, like, went back and see your product recently and and, you know, prepped for this interview.

**[0:31:23] Lukas Biewald:** And we, you know, we use it lightly, actually, for more project management type of tasks at at Weights and Biases. Maybe we We got should use it

**[0:31:29] Howie Liu:** up one level. more. Yeah, we got Yeah.

**[0:31:30] Lukas Biewald:** to up level. But I was I was actually in there, and and the AI was definitely, like, not in my face in the way that it is in some other products. So where where does the AI fit into, um, Airtable right now?

**[0:31:42] Howie Liu:** Yeah, so, I mean, depending on when you tried it, uh, we just launched, uh, our AI co-builder experience. So when I talk about the 2 ways in which we're, um, you know, sort of incorporating AI, the first is this builder experience, which is all about, you know, that is very much in your face when you first come to the website now. And the way you you you know, one of the ways that you can sign up and start using the product is, in fact, you just tell it, what is the app you wanna build, and our co-builder will build you a version of that app for you, right?

**[0:32:09] Howie Liu:** And what's really cool, I think, and underappreciated about modern LLMs is that they really have a lot of, like, cross-domain spanning context, meaning, like, these LLMs are pre-trained on every industry, like, they understand that retail workflow I just described to you, right? Like, I don't have to teach it that, it already knows, um, right? And it knows all these companies by name already, it knows William Sonoma, it knows Nike, it knows all these these companies. So instead of having to be descriptive and tell it, like, okay, this is exactly the series of workflow steps I need, and here's the data model that I need, you can just tell it, like, almost in in natural business language, like, Hey, like, I

**[0:32:49] Howie Liu:** am the, you know, um, like, product planner, basically, or product buyer at XYZ retail company, and I need you to help me with an app for, you know, for my use case. And it will do a pretty good job, especially if if we're doing some prompt engineering and kind of prompt steps behind the the the scenes to tap into the LLM's existing knowledge of all of these concepts to then produce an app, right? And we can also, like, augment it with if you wanna upload us additional context, or if we wanna do a web search and grab more context about your company or your use case, we can we can augment with that.

**[0:33:22] Howie Liu:** But ultimately, I think what's really powerful there is that, you know, if you think about the friction today to build an app, I mean, with traditional software development, it's freaking hard, right? I mean, it takes, like, weeks, if not months, to go and build, you know, an app, like, what I'm describing, and, you know, you you it's expensive, you have, like, developers, um, for building it, and then you get it back, and it might not do what you want, right?

**[0:33:42] Howie Liu:** And this is the whole point of the Agile development manifesto and, like, reducing that feedback loop between, like, the app developer and the business requirements holder. And I think what co-builder does is takes what we already were doing with no code to the next level, which is no code already made it radically faster and easier to build those apps as the business person yourself. And co-builder allows you to do it even faster yet, right? Because instead of, like, what would have taken, like, maybe hours to go and drag and drop and define the build, you know, the data model and think through the problem, you can just tell it what you want it to build.

**[0:34:15] Howie Liu:** And, like, a smart, industry nobl- kn- knowledgeable, you know, software development partner, it will go and, you know, kind of actually not just build the app, but build an app that reflects an understanding of your your industry, and that's really, really valuable. So that's pretty exciting and, like, literally, if you go to our front page now, um, that that experience is front and center. And, in fact, we allow you to even generate an app before you sign up.

**[0:34:38] Howie Liu:** So you can come in and just test it and say, Hey, here's my use case, show me an app, right? As if, like, there was a a software developer just sitting at your beck and call to, like, make you a a demo app, right? And then you can take it over and obviously directly, you know, edit and manipulate it to to get it to where you want. So that's one way. And then the second way is, you know, is really about, like, giving you these Lego pieces, and it's intentional that they're not loud and they're they're not, like, you know, we don't have to, like, paint the entire page in, like, a new color to show you, Hey, wow, like,

**[0:35:11] Howie Liu:** this is all different. Because the way we think about AI at runtime is, you know, AI is just another capability of software at large, but the way that we're gonna expose it to the customer is gonna be through familiar building blocks, right? So, if you wanna add a field in Airtable that's powered by AI and can take structured inputs and delivers an output and allows the human to then review that, edit it, approve it, you know, and then potentially chain that to more actions, it could be fire off an automation where if the output is approved, let's say it's marketing copy, then that goes into, you know, Slack, or it goes and automatically feeds that content into a CMS, um,

**[0:35:48] Howie Liu:** to publish. Or, you know, maybe it feeds it actually into another LLM call, so you can chain LLM calls together where it critiques itself, right, with the second LLM call. But, you know, the idea of, like, exposing it as yet another building block that is as interchangeable and as approachable as our existing building blocks of tables, fields, roads, automations, interface layouts, that's very intentional. Because we want AI to not feel like this completely, you know, scary new thing to builders, but instead something that they can really tap into, you know, in a modular way, right?

**[0:36:24] Howie Liu:** So you don't have to, like, upend your entire way of working. And, you know, for what it's worth, I don't think we're gonna see a world where you can just replace the need for human-driven or human-centric workflows, um, and database-centric apps with, like, an end-to-end agent that can just do all the work for you, right? I mean, maybe in certain very narrow domains, like support, especially very, um, kind of linear types of support, like your DoorDash, you know, the the types of support inquiries you get tend to be very simple, right?

**[0:36:53] Howie Liu:** It's like, my food isn't here on time, like, blah, blah, blah. It's, like, always a a very simple and compact decision tree that you could probably automate even without AI. But, you know, I think for for most types of real, you know, knowledge work that entails some amount of creative or open-endedness, um, you know, I mean, you're gonna need a system where, like this, where, like, the AI steps are just introduced, you know, in between, and, you know, they they should feel like building blocks in the system, not, you know, a complete upending of of how the work is done.

**[0:37:26] Lukas Biewald:** Interesting, I mean, I I I feel like there's this kind of trope right now that a lot of these AI-powered applications and and tools, um, have a lot of growth, but really low retention. And I think if I look at my own usage patterns, you know, I always wanna try the latest stuff, and I'm always just like, Wow, that is, like, incredible, but so many of these things that I've tried that are are amazing and seem really useful, I haven't stuck with myself.

**[0:37:54] Lukas Biewald:** Do you do you see that at all, or does the Airtable, um, AI tools have, like, a different character with with higher attention?

**[0:38:02] Howie Liu:** Yeah, I mean, look, I think I think this is kind of the, um, you know, the th- the design of our approach, which is I think the reason for a lot of the AI tourism effect, where you get a lot of, like, you know, top-of-funnel interest, but then very little retention, is because when the AI is just it's all about the AI sizzle and not as much about first starting with, like, what is the real business problem and that you can solve in a meaningful way, in a recurring way that, you know, has permanent value, has sticky value, that's gonna drive retention.

**[0:38:37] Howie Liu:** If instead, it's just about, like, the flash and awe of, like, a cool new experience that people wanna try once, but then, you know, ultimately, it's hard to kind of find value out of it in a recurring way. Um, I think you're gonna get the novelty, you're gonna get the, like, you know, like, kind of, like, fun and excitement, but you're ultimately it's gonna be hard to create that flywheel. So, I think by design, the way that we're implementing AI is more more, you know, uh, focus on delivering that meaningful, recurring value.

**[0:39:04] Howie Liu:** And, um, and so as a result, like, you know, we're not seeing this like influx of people just looking to, like, kick the tires and try the product. I think, generally, the people who are coming in and using it really have a purpose in mind, right? And so our goal is to really use AI to both create a secure product, secure business, and, you know, and a more valuable one, and ultimately one that activates better, right?

**[0:39:27] Howie Liu:** I mean, literally, with co-builder, we can drive a higher activation rate because people coming in who already wanna build an app can now just lower the friction to do so from, like, maybe minutes or hours into seconds.

**[0:39:39] Lukas Biewald:** Can you talk about your product development process? Like, you know, from when you first had this idea to, like, the mockups to how are you, like, iterating on this as it's live in production and and improving it?

**[0:39:51] Howie Liu:** Yeah, so, you know, we we did this reorg in our product group, um, about, call it 9 months ago now, and I think it's been really effective because in the past, we were very feature-oriented. So, even the way that our product org was structured was that we would have these feature-oriented, uh, groups, right? So, you know, here's the group that's responsible for improving search in the product. Here's the, you know, part of the the, uh, product org that's responsible for incru- improving the editing experience.

**[0:40:20] Howie Liu:** And, I mean, there, the you know, there, it makes sense, there's a logical reason to, like, have different groups working on different capabilities. But what I realized is that, by nature, you know, this results in a more incremental roadmap, because each feature group is kind of invested into improving that feature. But kind of, like, there's no overarching theme or there's no o- overarching kind of, like, structure that says, Wait, but what are the bigger unlocks that we're trying to go after?

**[0:40:47] Howie Liu:** How do we, like, you know, not just build incrementally better functionality into the product, which is important, you gotta make sure you you have a polished rock, but also, like, you wanna make sure you're building bigger and newer rocks as well, right? And so, we reorged around these major new pillars that, you know, one of which, for instance, is all about, like, the overall, like, onboarding and activation experience, the usability of the product, and the ease of which teams can grow and expand, and ultimately, that's measured by, like, their self-serve revenue, right?

**[0:41:19] Howie Liu:** So how easy it is it for, uh is it for a new company to come in, start using Airtable? Of course, AI co-builder really helps with that, um, and then ultimately to to start paying for us, right? Um, and in some cases, upgrade to enterprise later. So that's one entire pillar, and because it's defined in that way, instead of in a feature way, they're able to think bigger about the bets they wanna make, um, and thus, you know, we can have a longer-term view of, like, okay, the AI co-builder is just the starting point, but what's the next version of it, or what's the next unlock?

**[0:41:50] Howie Liu:** Well, it might be, you know, we've we've helped customers get from 0 to 1 very easily, create that starting point of an app, but now we wanna enhance co-builder to make it easier to take the app from 1 to 2, right? Which is to say, you know, if you wanna create tweaks, if you wanna add a new interface, if you wanna add an integration, like, we should also lower the barrier for doing that with AI. So, like, I think that's a that's a great example of, like, the product roadmap has now evolved to really be focused around for each pillar, a clear North Star of where we wanna go.

**[0:42:20] Howie Liu:** Um, and it's quite non-incremental, right? Like, we have a separate pillar that's really about enterprise and scale, and what they're working on entails things as big as we're actually building a new data store into Airtable that will allow you t uh, allow you to ingest multi-hundred million record data sets into Airtable directly, right? And so that's a huge, huge unlock, and it makes sense when the North Star of that group is really around the Hey, let's uplevel from a team or brute-based, you know, easy to use app builder to now also one that can scale up to these big, enterprise-wide spanning, um, you know, deployments.

**[0:42:53] Lukas Biewald:** Hm. One of the things that we notice that, you know, with some biases when people, you know, launch a feature like like co-builder, like like you're talking about, is that, um, there's a little bit of a cultural shift in the organization going from kind of a more deterministic product to a product where, like, it may or may not work, like, it kind of looks maybe more like growth hacking or something like that, where it's it's unlikely that, like, co-builder is gonna make, like, any application that, you know, somebody might dream of, right?

**[0:43:25] Lukas Biewald:** There's kind of a range of things where it works you you might not know exactly, like, what that that surface area is. H- is the non-determinism of the underlying LLMs that you're using changing the way you think about software development?

**[0:43:39] Howie Liu:** Hm. Yeah, uh, so, like, I get your question. Um, it's something I think about a lot, right? Because, you know, if you think about, like, a traditional piece of code, like, unless you have a bug in it, you run it, and it kinda does the same thing every time, right? Where you traditional API services, right? Like, you know, if you wanna interact with the Twilio API, you send, you know, an API call, and it does something pretty deterministically, right?

**[0:44:04] Howie Liu:** Send an SMS to this person, right? And if it fails, you get a very clear failure code, right? Like, could not be delivered because this phone number does not exist. Um, and I think you're right that, like, you know, with LLMs, it's deceptively like an API call. I mean, it is an API call, right? If you're using an inference API, you're not, like, running the models yourself. So you can make an API call to OpenAI or to, like, you know, Hugging Face or whatever.

**[0:44:26] Howie Liu:** And, you know, but but the the reality is, like, what you get back is, like, kind of very non-deterministic, right? And meaning, like, even the frontier of what the models are good at is kind of relatively unknown, right? I mean, like, of course, we've all seen the benchmarks on, you know, standardized tests, MLU, et cetera, but, you know, in terms of, like, all of the surface area of real-world problems, you can apply LLMs towards, like, that's relatively undiscovered, right?

**[0:44:55] Howie Liu:** And you're you know, that's why I think it's fascinating to constantly, like, look at Twitter and Reddit and see all these, like, cool, new zany use cases that people are unlocking. Like, oh my gosh, like, I uploaded my, you know, blood test markers, uh, over 5 years into ChatGPT, and it actually did an extraordinary job of diagnosing some, like, you know, potential health issues I had, right? And, like, these are real-life examples

**[0:45:17] Lukas Biewald:** That's a true story. where, yeah, That's crazy

**[0:45:18] Howie Liu:** like, yeah, yeah, yeah, and story. it w- I mean, apparently, But like, you the hack is, um, to get past the, um, the the kind of, like, safety gates, like, you have to tell it, Well, I'm actually writing a script for a movie, and here's the, like, you know, the the data I'm uploading for the movie. And it's got to be very, very correct because in the movie, we wanna be scientifically accurate, right, or something like that.

**[0:45:34] Howie Liu:** But, um, you know, but but, uh, but the point is, you know, like, even the researchers who work on these models, as you know, like, don't even really know where the model's gonna be really amazing, potent, uh, and and capable before they train the model, and even afterwards, you know, then it's like, sometimes you're kind of astounded, like, wow, like, we didn't think 3.5, you know, um, Anthropic's, uh, you know, 3.5 model Sonnet is actually gonna be so good, right?

**[0:45:59] Howie Liu:** It kind of amount amazed us in, you know, certain dimensions. So I think there is this, like, frontier of unknown in terms of how these models do in specific use cases. And when you think about, like, the really wide-varying texture of real business use cases, uh, think about, like, every single part of a knowledge worker's job in every single department, so marketing, legal, finance, product, HR, in every single type of industry and different type of company size, like, you know, the only like, if you were to go and test how these LLMs did at specific steps, like, obviously, you can't do the whole thing end-to-end, short of AGI being developed, but even if you tested it on specific key steps,

**[0:46:40] Howie Liu:** like, sometimes you don't really know what you're gonna get until you try it, right? And so I do think this implies something for both our customers and something for how we develop our product. For our customers, it implies that you cannot go and just, you know, establish a centralized AI team or committee that kind of top-down and from an ivory tower goes and says, These are the use cases for AI we're gonna deploy into the company, and they're gonna work for sure, right?

**[0:47:10] Howie Liu:** Because I think you have to have experimentation, like, there is, like, this unknown of, like, take this miracle, you know, kind of, like, uh, product, right, and just test it on all these, like, different problems across the organization. And the only way to do that is really to empower a culture and a, you know, even, like, you know, platform model of local experimentation, meaning, like, you need people within each of those functions, somebody within HR, somebody within legal, who understands their own work and the workflow and the data to really be able to have, you know, the ability to experiment with AI.

**[0:47:44] Howie Liu:** And I think the way to do that, um, is not necessarily just through chat interfaces, but really, you know, be able to embed the the model intelligence into your data itself, which we do, uh, with our approach, I think that's gonna be really key. And then the the second question of, you know, how does this, uh, affect the way that we develop the product? I think it means that we need to provide tooling to our customers that lends to this rapid experimentation, right?

**[0:48:09] Howie Liu:** Um, and so meaning, like, prompt engineering is hard, right? And knowing how to chain together different prompts to get desired output. For instance, if you have 1 prompt, output a draft of a legal document, right? And then the 2nd prompt can actually review and critique that legal document given a certain set of criteria or standards, um, or even, like, example documents it needs to compare them against, then you'll get, like, the critiques. And then you could have yet a 3rd step, which might be human or AI to, like, incorporate those those, uh, you know, those edits, right?

**[0:48:40] Howie Liu:** Um, but those steps and how you design the prompt and how you give it the many-shot examples in each case are not not always obvious to business end users, right? I mean, very few people even know how to prompt engineer and, like, leverage these LLMs effectively. And so I think a lot of our job as a product is to give people the kind of tools and the UX to make it really easy to do this. So instead of having to go and, like, define that prompt sequence yourself, to give you kind of a higher-order wizard to say, This is what I'm trying to do, and then have our product underneath the hood kind of implement the specific prompts, the specific,

**[0:49:15] Howie Liu:** like, many-shot feedback loop. And, in fact, we have the advantage of potentially being able to use the user feedback that people give us in the app itself, where they're reviewing in Airtable the outputs of the AI, and they're saying, No, this was wrong, or actually, let me override this. We can feed that back into a many-shot example, uh, set that feeds into the prompt, even without doing fine-tuning to get better and better results over time, right?

**[0:49:38] Howie Liu:** So just being able to, like, abstract away a lot of the, like, you know, weird nuances and and, like, you know, kind of, um, all the prompt engineering and, you know, LLM, you know, best practices, um, I think they're arcane and they're really hard to get to know, and they're kind of constantly evolving. And so, you know, we can we can make that all easier for our non-technical business users.

**[0:50:00] Lukas Biewald:** And do you use do you use those techniques yourself internally also to to actually build the product?

**[0:50:09] Howie Liu:** So, um, so we we are trying to be very, very good customer, uh, number zero users of our own product, right? And even start to push the limits of, like, the next generation of capabilities. Like, we wanna be able to kind of dog food or do in a hacky way ourselves so that we can learn what works, what doesn't work. So the short answer is yes, and increasingly so. I think, um, you know, what we're learning is sometimes it's really helpful to have, like, almost a rough and dirty version of a capability.

**[0:50:42] Howie Liu:** Let's say we wanna be able to use AI to suggest AI use cases within an app, right? Um, so imagine, like, if you work in legal at Airtable and you already have a lot of, like, legal contract and workflows in Airtable, Airtable could suggest to you, Hey, why don't you add these AI steps or AI fields to automate these things that I've inferred you can do based on the structure of the contents of your app and even what you're doing in the app, right?

**[0:51:07] Howie Liu:** And, you know, one thing we wanna do is, like, start experimenting with, let's just build a rough and dirty version of that and put it into the product for ourselves or maybe for a very limited set of alpha customers externally. And let's see how they use it and let's see, you know, even without looking at their actual customer data, like, let's get their feedback on is it working, is it not? And then we can, like, kind of tune and change the prompts for our own kind of meta features to see how they work.

**[0:51:32] Howie Liu:** So the short answer is yes, and I think to your point about, like, a lot of these things are unknown and you kinda have to, like, experimentally try them. We wanna lend ourselves to a more experiment-capable way of using our own features at AI.

**[0:51:49] Lukas Biewald:** D do you feel like you've collected a set of best practices internally for making high-quality prompts? And and h- how big is the delta between, you know, 1, like, your first guess at a prompt and sort of, like, where it ends up in industrial use?

**[0:52:07] Howie Liu:** So, um, I think we are collecting, uh, a lot of those best practices, and it's not just what's the best way to write a single prompt, and of course, like, there are some very known best practices. You know, you can tell it, Hey, in the system, uh, prompt, like, you are this identity, right? And then it will kind of guide it to, you know, play a certain style or, like, leverage a certain type of knowledge, uh, more so.

**[0:52:34] Howie Liu:** Um, I think of it as, like, almost like creating a bias towards, like, a certain neural pathway in these, like, massive models that can play so many different roles that have been trained on so many different ways of thinking about the same question, right? Um, you know, to, like, the many-shot examples, to, like, the stylistic guidance or even, like, forcing a structured output in JSON mode with some of these newer, uh, models that have that capability, like, you know, so I think a lot of those are, like, already out there, like, they're the kind of well understood, like, you know, um, kind of as a consensus across the LLM development community.

**[0:53:07] Howie Liu:** I think where we're finding more interesting learnings that maybe aren't out there relate to solving real, kind of messy business problems with these prompts, right? So, in practice, like, how much do you need to break down a real business workflow, like, you know, this marketing, you know, content generation from end to end? It's not just about, like, a one-time, you know, generate the thing, like, you could do in Jasper or ChatGPT. I think it's really about, like, breaking it down into all the different steps and, you know, that is kind of like an overall workflow design with AI problem that, you know, we are, um, you know, we are gaining more and more know-how around, and we wanna productize more

**[0:53:43] Howie Liu:** of that, in addition to gaining just an internal expertise around it. So that when our customer-facing teams go out and talk to customers, they they can actually articulate and and show in a, you know, demo, like, a best implemented version of that. But I think it's, you know, I think it's fairly early. Um, I think there's also evolving capabilities with each model that kind of r- you know, make some of these old techniques either obsolete or or different, right?

**[0:54:08] Howie Liu:** Like, for instance, as the context windows have radically expanded from, like, you know, I think the best ones had a few thousand as of 2 years ago to now 200,000 or even 2,000, 000 in the case of the the biggest ones, like Gemini Pro, like, you know, you don't have to do things like chunking if you're dealing with, like, a really large input data, uh, source, right? It could be an earnings call, it could be, like, a massive document or a set of documents.

**[0:54:30] Howie Liu:** And, you know, you don't have to, like, break that down into different chunks anymore. And so, I think there's a lot of, like, things that are changing as we speak, and our goal is really to be just so immersed into both the cutting edge of every new model and then, more importantly, being very hands-on in bringing those models to real customer problems, right? And I think that's perhaps where we have the most durable advantage, is we do have this really large customer base of people who have built all of their own use cases on Airtable, gotten their data in Airtable, and now, every time we introduce a new model's capabilities, we're kind of, like, getting to learn and experiment with

**[0:55:09] Howie Liu:** them with this large customer base. You know, how do you actually what are the best practices for using this new generational model in the context of real work? So, um, you know, I think it is a way of thinking, a way of learning that we're, you know, we're really trying to be fleet-footed, uh, with going forward.

**[0:55:28] Lukas Biewald:** Well, it's a great segue into a bunch of, um, questions I wanted to ask you about the overall AI market, not necessarily from exactly Airtable's perspective, but from your perspective as someone kind of sitting at the intersection of AI and business. Um, do you, um, do you have a perspective on open source models versus closed source models?

**[0:55:50] Howie Liu:** You know, this is an area where we have the luxury of being able to be somewhat agnostic, right? The short answer is, I have no idea. I mean, um, you know, it seems like the open source models are catching up, and obviously, you know, very timely, uh, you know, to discuss this with the the Llama 400 billion parameter model coming out, and it seems to be a really, really great model. Uh, but at the same time, who knows?

**[0:56:14] Howie Liu:** Like, you know, maybe GPT 5 comes out in 3 months and it just blows us all away with an even greater degree of reasoning intelligence, right? I think that's kind of the next frontier is not just being able to answer questions that look more like a standardized test, like MMLU, but actually, you know, there are these new, you know, kind of, uh, benchmarks where we care about the AI's ability to reason and answer questions that don't just look like other pieces of knowledge it's seen.

**[0:56:41] Howie Liu:** You know, it's like the SAT analogy, uh, you know, type of questions, but instead, really require a deeper understanding of, you know, what what's what's being done, and it's possible that maybe, like, you know, the the closed source models will continue to hold an edge on that form of use case, which I think will have pretty important, uh, implications because, you know, if you just look at the typical standard test performance, it does look like there's a plateau.

**[0:57:04] Howie Liu:** I mean, obviously, like, you can only get so good, right? I mean, you can only get close to, like, 100% on, like, a standardized MMLU-type test. But I think what what is interesting is that if you can get closer and closer to, like, the open-ended reasoning capability of a human, then you can start to see more, like, you know, kind of real hard work problems get tackled by AI. So, I think of it as a benchmark of if you froze AI model development today and all we had was what we have, and then you you spent the next 20 years just trying to figure out every single possible business use case to deploy the these AI capabilities into, how much

**[0:57:40] Howie Liu:** economic value could you create, right? And, first of all, I think you could do that today, just freeze all the model development, and we've we've only exploited maybe, like, point oh, oh, oh, 1% of all of the potential economic, you know, value gains you can get from today's AI models, right? A lot of it is about, like, actually situating it into real data, real workflows at the edges of, um, and the really interesting kind of nooks and crannies of every part of every company.

**[0:58:08] Howie Liu:** Um, but, then the secondary question is, with every new model that we're and the benchmark I care about is if you could play that thought experiment out, how much more economic value could you unlock with this next generation of model? And that's where seemingly marginal improvements in the current benchmark scores, I think, can actually belie the fact that, you know, really major advances in reasoning capability will actually unlock an even broader, you know, kind of, uh, set of of really valuable business problems that can be solved.

**[0:58:40] Lukas Biewald:** Okay. What do you think about AI as it benefits, kind of, in big entrenched incumbents versus startups? I mean, you could imagine incumbents are successful because they have the kind of Workflows, they have the data, and now they can add in AI maybe Easily. But, you know, startups might be better Positioned to use a new technology. Who would you bet on?

**[0:59:00] Howie Liu:** Yeah, I mean, I think this is where the driver behind the seat of the car matters the most, right? You know, it's like you could have, like, the the like, you know, 1 car that has, like, certain, you know, advantages. It's kinda the the traditional winning team in F1, but you got, like, a new upstart, and it's got some fancy new, like, kinda technologies. But ultimately, I think in either case, the driver matters, right? So, meaning, I don't think, like, the big existing incumbents are gonna win by default just because they have the data and the distribution, right?

**[0:59:29] Howie Liu:** I think they have to be nimble on their feet, and they really have to think about, like, what is the most, you know, kinda disruptive way to apply AI to their existing product, right, to create that value. Like, it's not for free that you just flip a switch and then you get AI value in your product. I mean, as, you know, as we've talked about for a lot of the past, uh, you know, uh, conversation, I think it's in really figuring out the nuances and the nooks and crannies of how you deploy AI into real-world business problems that the value is unlocked.

**[0:59:56] Howie Liu:** So, I don't think big companies get that for free, and they have to be very deliberate and kind of smart about thinking what is the way we wanna, you know, kind of infuse our product with AI. Um, on the other hand, I think the net new AI startups are also gonna have some real challenges, right? You know, they're kinda fighting this race of, you know, as you pointed out earlier, they've captured a lot of early hype.

**[1:00:20] Howie Liu:** They, you know, that may parlay into some, like, top-of-funnel AI tourist traffic that that helps boost, you know, kinda their starting point, but then there's often a very leaky funnel, right? Um, and there's also questions of, like, how durable these advantages are, like, what stops the next company from coming in and doing the same thing, or the incumbent from doing the same thing, right? And, you know, ultimately replacing them, uh, you know, with a feature, um, that kind of makes their entire business obsolete.

**[1:00:46] Howie Liu:** Um, and so I think it's kind of a race between the incumbents to disrupt themselves and the upstarts to build durable advantages, and, you know, I think the answer to all of that is it just comes down to really good execution on either part. So, like, I would bet on, you know, the companies that have the most hungry and disruptive teams who are thinking really strategically, but also, like, staying very, very close to the details of what's happening here and now and not being head in the clouds, right?

**[1:01:15] Howie Liu:** And, you know, that's easier said than done.

**[1:01:17] Lukas Biewald:** Well, okay, so you know a lot of these companies, Both the incumbents and the startups, so who's gonna win?

**[1:01:25] Howie Liu:** I I mean, I honestly think it's company by company, right? So, you know, as an example, um, I think we both know Scott Belsky from Adobe, and, like, he have a huge amount of respect from him, uh, for him, and, you know, like, with him working on a lot of the AI capabilities of Adobe, like, I'm pricing a lot of, like, upside into Adobe, right? Um, you know, and, you know, on the flip side, like, there are some startups that I really admire cause I think they're thinking about the problem in a way that is not just short-lived, but really, you know, looking to accrue longer-term advantages, right?

**[1:01:56] Howie Liu:** Um, and in some cases, like, they've made, you know, pivots, like Tome, as an example, started as more of, like, this AI, uh, slideshow tool, right? Which is really cool. I mean, it's it was a very magical experience. You come in, you tell it what you want, and then it makes you a slideshow complete with, like, imagery and, like, text and everything, and that was really magical, but I think they were really smart to kind of adjust course into going deeper into a stickier business problem, which is about, like, you know, actually giving salespeople an entire, like, customer intelligence tool to go and do research and then pro produce a presentation, right?

**[1:02:28] Howie Liu:** And, you know, like, obviously, it's so early, jury's still out on both of those examples, but, uh, you know, for me, like, I really want to, you know, kind of bet on specific people, specific teams, and I think there's gonna be winners on both sides.

**[1:02:42] Lukas Biewald:** Okay. In a world where it's easier to make apps, and I I would include, you know, there's Airtable making it easier to make apps. There's also, like, you know, kinda software development, you know, getting easier through, like, all these different tools, like, you can make, like, you know, a nice webpage on Vercell a lot easier than you could in the past. Do you think that endangers vertical-based SaaS applications, like, I don't know, like, a Salesforce or something like that?

**[1:03:06] Lukas Biewald:** Are people just gonna make their own internally because it gets so much easier to do that?

**[1:03:11] Howie Liu:** So, short answer is absolutely yes. I mean, I think this is already the premise of no-code, right? Like, why do you need a bunch of fragmented, like, not that great apps, especially, like, on the low end of things, like, or very narrow market side of things, like, CalMax. Why do you need the, like, PR-specific journalist database that has, like, kind of a CRM workflow layered for pitching them when, really, that's just, like, something that should be a more flexible workflow built on a flexible platform, right, with more modern new UX and more modern, uh, capabilities.

**[1:03:43] Howie Liu:** So, I think this radically accelerates that disruption, um, that we were already seeing with kind of platforms eating more of these niche vertical software products. I do think that the vertical software products that will continue to be strong are those that are in a very weird industry that's still very large, but has, like, so many domain-specific requirements. So, like, take Viva, for instance, right? I mean

**[1:04:07] Lukas Biewald:** I was hoping you'd say ML apps.

**[1:04:09] Howie Liu:** Yeah, yeah, yeah, I mean, for sure, like, exactly that, and especially those that have a three-word name, ideally, like, rates of any biases, those are gonna be the ones that that do really, really well. Um, but in all seriousness, I t I think, um, you know, the more the depth of weird and unique requirements, um, or dynamics of an industry, I think the more a vertical software player will remain the incumbent and and, you know, succeed and, if anything, have more time to buy themselves before they they need to disrupt themselves and really think about how to, like, you know, implement more AI and and, uh, and all that stuff.

**[1:04:46] Howie Liu:** Um, but, you know, end of the day, I I think it it's really it is gonna accelerate, like, more broadly this this, um, this shallow set of vertical software companies that really you know, I see them as almost, like, a temporary blip in the history of software where early days software was all very either horizontal or custom-built, right? You had, like, your big 3, like, word processor, database, product, database, you know, and spreadsheets. And then everything else was, like, custom-built software for your own company, right?

**[1:05:16] Howie Liu:** You'd buy Oracle as a DB and then build your own application on top of that. And I think I I think we saw a world where, especially with SaaS bringing down the economics of building and not only, um, building, but distributing software, uh, you know, you're starting to see all these more niche players spring up, but I think they always had disadvantages of never quite being the same level of quality as, like, a product that just had more flexibility and more investment overall, right?

**[1:05:40] Howie Liu:** And so, I think this does turn the tide back towards, okay, like, why am I using this, you know, very vertical piece of software that's brittle, doesn't quite meet my needs, and, you know, it's just kind of a one-off that also is siloed from the 50 other products that are in my ecosystem.

**[1:05:57] Lukas Biewald:** Do you expect software developer as a function to continue for the next 20 years?

**[1:06:02] Howie Liu:** So, I think so, and, you know, the reason for it is both no-code built internal business apps and the types of apps that you would build with traditional software development, both for internal company use as well as, like, external-facing apps, I think will only continue to proliferate more as the cost of, you know, building gets lower on both sides, right? Obviously, you have developer co-pilots that augment developer productivity, make it easier and faster to write code, um, and then I think separately on no-code, you see this, like, radical, you know, kinda disruption and and and lowering of the cost further to build apps that way, and I think in both cases, you know, it's not really, like, a zero-sum game

**[1:06:48] Howie Liu:** of which wins, I think it's more of a overall pie growing because traditionally, with any product or good or service, as you radically lower the cost of that thing, people buy more of it, right? Especially when there's no finite, you know, kind of upper limit to the number of useful apps for the amount of useful software that can be created, and I think we're still at the very, very tip of the iceberg of software eating the world, and software is, like, literally a way to manage and represent and, you know, make more efficient or augment the informational workflows of the world, right?

**[1:07:22] Howie Liu:** And that is, like, you know, there's so much more software that we could build that would create that economic value in a world that especially is moving every value chain more and more towards these informational workflows away from, like, the physical and hard good, uh, you know, kind of processes, so, you know, I I think, you know, it's only gonna accelerate the amount of software and useful and value-driving software that will be created on both sides.

**[1:07:47] Lukas Biewald:** So, where in the AI hype cycle do you think we are? I mean, I feel like you said a whole bunch of really optimistic statements that makes me Think that you might think we're sort of on our way up if we don't even need better LLMs to, like, you know, fit many more business use cases. That would sort of imply that, um, you know, the the AI funding train and excitement is just gonna kinda grow over the next few years.

**[1:08:06] Lukas Biewald:** But, I guess, like, from my vantage point, I worry that I see a lot of pretty bad products with AI baked in that I I and I see companies sort of investing in AI without a a use case in mind, Which that that really scares me and sort of says, like, peak bubble to me, like, w- where where do you land on that?

**[1:08:23] Howie Liu:** So, you know, I think we are still at the very, very early innings of realizing all the value with AI. Um, so, long term, I'm still very optimistic about what can be done with AI. I think this, you know, technology is real, and again, even without banking on some major new breakthroughs in terms of, like, model capabilities or architecture, I think you could take what we have today and just actually create tons of economic value with it, like, trillions of dollars of GDP.

**[1:08:53] Howie Liu:** Yet, at the same time, I think today there's a lot of hype, there's a lot of funding, and there's a lot of valuation increases that are not quite being fully justified, uh, or not at all justified by, like, the real value delivery to customers, um, and either that's because you have companies that, you know, again, like, they are just raising on the promise and the hype, but have not actually delivered meaningful value, uh, or even if they're making revenue from customers, sometimes that revenue, like, nominal dollars can create equal, right?

**[1:09:23] Howie Liu:** And I think there's a fair not amount of, like, frothy dollars from even customers who are just looking to spend on AI products to be able to have an AI strategy, and yet, you know, I just saw, like, a tweet today that, um, you know, was was, uh, somebody saying, like, I'm talking to a customer who, uh, you know, who is, you know, saying Microsoft Copilot is their main AI, you know, kind of strategy, and but they don't have any demonstrable ROI, they have no, like, measurable impact from this, right?

**[1:09:50] Howie Liu:** And there's no way to really say, like, concretely, like, this is actually improving the output of the business, um, and, like, maybe, like, you know, you could track if it's a public company, like, Okay, well, does the earnings call next time and the next time actually gonna be better or not, right? Like, you know, obviously, there's a lot of confounding variables, but I think that's the difficulty right now is, like, you have a lot of hype and dollars being thrown at the system by both, you know, the the investors, but also by customers without, like, fully applying this rigorous bar of, wait, like, what is the real business use case and business impact here, right?

**[1:10:22] Howie Liu:** And I think that bubble is going to collapse on itself because, at some point, the customers are gonna wake up and realize, Oh, wow, we're spending a lot on certain AI tools or vendors that we have nothing to show for, we have very little to show for in terms of real business return, right? Um, and so we're gonna pull back, we're gonna rationalize this spend again, and I think that's gonna cause an adjustment on the investment side because, you know, it not only does it become harder to invest in companies that are doing 0 revenue, but, like, even the ones that are doing revenue are gonna see kind of a slowdown if they're not providing meaningful value.

**[1:10:57] Howie Liu:** All that being said, I think it's gonna be sort of like the Web 1.0 bubble collapse, which is, you know, obviously, to say it it it, you know, kinda wiped out a lot of the frothy companies, I mean, you know, I was, um, still very much, uh, you know, a, uh, I think I was in middle school, high school during a lot of that, but, um, you know, getting to see it from that vantage point, there were just all kinds of products that probably should not have existed, right?

**[1:11:20] Howie Liu:** And, like, in hindsight, like, we all kinda knew it, like, I, uh, I signed up to be a beta tester for a smell printer company, right? Who was gonna send me as, like, a not even 18, like, high schooler, a beta version of their smell printing product, right? Um, turns out, it's not even possible to really print all the smells in the universe, right, with a printer, but, like, you had a lot of these, like, kinda silly products like that that, you know, once the rationalization cycle, you know, switched in, went away entire thing, and yet you had really good companies like Amazon that actually were born in that era, right?

**[1:11:53] Howie Liu:** And, you know, uh, and kind of because of that rationalization, arguably became even better companies, and they survived because they were actually good businesses, and, you know, they were well run and they had, like, a path forward, and so I think we are just we we are gonna see some of that, and, you know, maybe also, like, the the crypto bubble, like, as a lot of people in crypto say, like, Well, actually having a little bit of a crypto winter could be a good thing because it cleans up the system so that, you know, what remains are really the meaningful and value-driving types of companies that have real purpose, right, versus are, you know, basically benefiting from capturing all these

**[1:12:29] Howie Liu:** frothy dollars coming into the system.

**[1:12:32] Lukas Biewald:** All right, well, I think that's a great place to end Howie, thanks so much for your time. That Was a lot of fun.

**[1:12:36] Howie Liu:** True, so thanks for having me.

**[1:12:37] Lukas Biewald:** Yeah, that was awesome, thank you. Thanks so much for listening to this episode of Gradient Dissent. Please stay tuned for future episodes.