Building a Culture of Experimentation // BrXnd Dispatch vol. 40
A chat with Howie Liu, CEO of Airtable on interfaces, AI, and getting your hands dirty
You’re getting this email as a subscriber to the BrXnd Dispatch, a (roughly) weekly email at the intersection of brands and AI. BrXnd NYC 2024 is coming on May 8. Tickets are now sold out, but we will be posting everything on YouTube after the event.
As part of the lead-up to the conference, I will be running a few conversations with speakers and sponsors. Today’s conversation is with Howie Liu, CEO of Airtable. I’ve been a huge fan of Airtable for a long time, and have written about how to use it as a scratchpad for experimenting with more complicated AI workflows. Over the last year, I’ve gotten to know Howie and asked him to come talk about how they’re thinking about building AI interfaces that push beyond the chat box. The scale at which they operate and the fact it’s a platform that houses data and lets you run code makes it a particularly fascinating space for builders in my view. After agreeing to the talk, Airtable also decided to come on as this year’s title sponsor (which I’m incredibly grateful for).
Today’s interview is a fun little conversation about Howie about some of how he/Airtable are thinking about AI and a bit of a preview of some of what he’s going to discuss at the event in a few weeks. The conversation was lightly edited for clarity.
Noah Brier: To get us started, can you give me a five-sentence primer on Airtable?
Howie Liu: Airtable was founded with a very clear vision of being the fastest and easiest way to build useful apps. What is an app? An app is basically a set of data, some business logic, and then an interface of your design that allows you to interact with that data and perform a workflow, right? And it turns out there are millions of use cases for custom apps within the enterprise that just happen to have been built on spreadsheets, or people have had to rely on niche vertical solutions for a really long time. And we have been and continue to be focused on giving you really the best of both worlds - the ability for the people closest to the work to actually build those apps themselves, but then also to make an app that actually has structure and can scale up and be your production system for things like marketing content, campaigns, product operations, whatever you want.
NB: So, obviously, you and I have talked a lot about AI, and we're both pretty excited about it. What's got you so excited about it? Why have you “lost yourself” in this world?
HL: Two things. One is just looking at AI—even outside of Airtable—it's pretty clear that we've reached a breakthrough in terms of AI capabilities. There's been a lot of AI improvement over the past two decades. In fact, I did a lot of AI work in college, studied neural nets, did some work on that, but it was very primitive at the time. The things you could do back then were very simple: classify this image, detect the type of object in this image, classify this image as appropriate or inappropriate. And some predictive analytics type stuff, like help me predict what movies I'm going to like based on ones that I've already watched. Those were fine but somewhat narrow in application. There were these high-value but very narrow use cases.
I think what we've seen with the latest generation of transformer models, specifically LLMs starting with GPT-3.5, which broke into the mainstream with ChatGPT, is that these models exhibit a level of reasoning behavior that really appears human-like. They can reason like humans on a very broad variety of tasks. And these models are not only capable of this reasoning but they've also been trained on the corpus of basically the entire internet's knowledge. So they're very widely aware of stuff.
We've kind of reached this tipping point where the AI models have gotten actually useful and usefully good at a really broad range of human intuition requiring problems versus very narrow and deep applications only. And the fact that these models have gotten that good and they're very cheaply available—when you actually think about the true cost of running one of these models, it's costing .1 cents to ask this AI to do something instantaneously that frankly would have taken a human hours to do in some cases. Maybe the human, if it's the right human, is going to do a better job. But the speed and the cheapness have just made this an incredible breakthrough.
The second reason why I'm personally excited about it as the founder of Airtable is because of that breadth of capability and the speed at which the models are improving. It turns out this is very well aligned to everything that we've been founding and executing Airtable around, which is this very broadly capable “Lego kit.” We are a “Lego kit” of building apps. And what better pieces to introduce into that "Lego kit" than the most magical AI building blocks ever invented that are constantly getting better?
And best of all, because we're so flexible and we really put the power in the hands of the people closest to the work to experiment with building apps, we're also able to package up these AI capabilities that are constantly improving and put those pieces into the hands of builders who can then experiment with them. Which is really the only way that you can keep up. These models are exhibiting emergent behaviors with every generation that even the people who developed the models didn't expect.
With every new wave of models like GPT-4, researchers will write papers like, “I couldn’t believe that we actually got this model to do this thing.” And so the only way to fully utilize that capability, especially since there is not just one use case for these models but an infinite set, is to have a really flexible “Lego kit” like Airtable to experiment and rapidly build your own implementation of how to use these new AI capabilities in every single part of how you run your marketing operations. Obviously, this applies to more than just marketing; we’re talking to a marketing audience here.
NB: One of the questions I get often is, given the breadth of possibilities and the power, how do you prioritize use cases?
HL: As a platform and as a “Lego kit,” we've always really benefited from our customers, the people who go and build with the Lego pieces. We sell, in many cases, just the generic red bucket of Lego pieces. And it's really the creativity and the business process ingenuity or acumen of the builders within every company we've worked with. By building primitives or Lego pieces with AI, we actually put the capabilities of these models into people's hands.
For example, we already give you the ability to build an AI field or an AI automation that can use whatever prompt you want. We give you a bunch of templates or suggestions, but ultimately, if you want to make this a contract review prompt you can write one—you can build one into the primitive. If you want to make it a campaign brief generation step, you can do that.
The second thing is that I actually think what we found in marketing is that there's a pretty consistent structure or architecture to how marketing operations work. There is a basic marketing supply chain architecture that applies to almost all B2C companies and almost all B2B companies. And what we have done is started to come up with a point of view, a lot of it customer-inspired, of how in each part of the marketing supply chain, what are the ideas, at least as a starting point, for how you can use AI. And of course, you can continue to experiment and innovate and use AI even more at the edges or in the specifics of it all. But there are some basic obvious points where AI can be really powerful.
NB: Of those, do you have any favorite examples of either internal things you're all doing or external things you're seeing from customers?
HL: I think a lot of people put the focus on using AI to ask questions, so you can ask any question in natural language and get an answer. Those are more of a reporting, querying use case. And I think those are fine and useful, but I actually think what's really powerful is to take something that is part of a recurring process that required a lot of human energy or even creativity, and rather than try to completely replace the human, do a first draft of something with AI.
As an example, if you have a global marketing organization, and in the B2C context, you're running these global campaigns that take a lot of planning but then also require a cascade into regional versions—not just literally translating that campaign, but translating the concept, taking a global concept, global campaign concept, and even sometimes global assets, and then translating them into a region-specific version that plays to that audience, that region's customs and culture and what we know has worked in the past.
It turns out AI is pretty good at this. If you feed in the global campaign brief and assets and also feed in some context about this region and what we know has worked historically for our company in this region, AI can generate a regional campaign brief and even literally show examples of what that could look like, creating concept art, examples of what a local influencer could literally be saying about a new product launch and why they're excited. Those are examples where it's not just automating some redundant, tedious part of the job. It's actually helping to spark better creativity and ultimately produces a better campaign that creates more impact and drives more output value, in addition to accelerating the time and saving the human costs required.
NB: Well, how about one last question—if you were to assign everybody one thing to read about AI, what would you assign them?
HL: I would actually say don't read, go and experiment. For anyone who does not already actively use the leading models, either ChatGPT or the other GPT-4 level options, start experimenting. And obviously, I'm going to propose you experiment in Airtable where you can actually see it in the context of structured data.
So take a list of campaign concepts, take a list of PDFs representing different work documents, customer briefs, and have AI start performing analysis of these things. Really push AI to the limits of what it can do, because I think what you will be surprised at is how wide the frontier of AI capabilities already is. You don't have to wait for tomorrow. Ask AI anything that you would ask your human coworker—it's probably not going to do as well, but it's going to be fast and surprisingly good.
And that's where the profundity of some of these use cases is, especially when you integrate that capability into the context of your existing data and a structured recurring workflow where it's very easy to interact back and forth with the inputs and outputs. That's where the real magic is happening.
A big thank you to all our 2024 sponsors, without whom I couldn’t pull this thing off. Airtable enables any team, regardless of technical skill, to create apps on top of shared data and power their most critical and unique workflows. The Brandtech Group The Brandtech Group is a marketing technology group that helps brands do their marketing better, faster, and cheaper using the latest technology. Brandguard is the world’s #1 AI-powered brand governance platform, Focaldata is combining LLMs with qualitative research in fascinating ways, and Redscout is a strategy and design consultancy that partners with founders, CEOs, and CMOs at moments of inflection for their organizations. Plus, big thanks to McKinney, Inuvo, Dstillery, Canvas Worldwide, and Persistent Productions. If you’re interested in sponsoring the 2024 event, please be in touch.
Thanks for reading, subscribing, and supporting. If you have any questions, please be in touch. If you are interested in sponsoring, reach out, and I’ll send you the details.
Thanks for reading,
Noah