Airtable CEO: This Is What the Top 1% Do With AI | Howie Liu

The people who are able to do that most effectively will become like almost superhuman, right? You're able to give good judgment and feedback to just like a really effective CEO of humans. You could imagine building a company that you never would have dreamed of without any employees or with minimal employees.

So the tinker mindset is most important first and foremost. And I think the second is really...

I feel like right now we're trying to close the loop with these agents. Like, teach them our taste, teach them how we give feedback, and make them more autonomous.

I do think like there's been this, you know, kind of step function change in the experience that you can create with AI from pre-ChatGPT, but then I think we entered like a mature era of even AI chatbots where the models got smarter. So even thinking about like, you know, the year after GPT-4 came out, you know, we could actually do really interesting things with chatbots. So chatbot maturation, and then I think we started to hype agents a little too early. So there was a year, like last year, people were, you know, they were saying this is the year of the agents, and I think that was when Salesforce Marc Benioff went up and was talking about like, you know, in the future every CEO is going to have thousands of agents, like, you know, agent employees and human employees. And I think it was correct but just early. And so what we finally hit now, I think, and some would say like really it was like around end of last year with like the Anthropic like Opus 4.5 and beyond generation of models, but now, you know, you have GPT-4.5 and Gemini 2.5, like these open source and, you know, alternative models that I think have all reached this level of intelligence that actually enable agents to be like almost humanlike, right, and autonomous.

I don't know, like I've deployed a lot of agents but I can't sleep. It's still in my head like I'm responsible for that agent. I'd rather have more humans.

You know, I think, um, I think that as the models get better and better, we do need to rethink the entire paradigm and like the product form factor to reflect it. So I'll use a parallel here, which is in the development agents world, you know, you had Copilot, GitHub Copilot that was kind of the first, you know, kind of chatbot for developers, right? And it really was that like you couldn't have it go and develop like really complex programs. It was more like you could autocomplete a few lines of code at a time. And so it was a convenience. Then you had Cursor which introduced its own AI agent called Composer, which is different from like the now what they're using that name for is the model they've fine-tuned, but at the time it was a much more autonomous agent. You could tell it go and like build me this file, write me this simple like, you know, kind of script or function and it would do something a lot more autonomously than, you know, the original Copilot experience. I think now yet we've entered into a new era where the agents do enable, you know, kind of you to go and execute in parallel with multiple agents. And so in the development world, the best developers are already, you know, they're not using necessarily Cursor, they're actually now using many different agents, maybe it's Claude Code for instance, running in parallel, and it almost does feel like they have their own company or team running including overnight, right? I actually when I'm developing I use agents that, you know, I try to have them do something substantial before I go to sleep so that, you know, for hours at least, like I'm not wasting bandwidth cycles, um, and they're doing something useful. And so when I wake up they've already completed, right? So I, you know, I think we are moving to a world where the form factor will start to look more like managing a fleet of agents or almost managing like, you know, people. Um, you know, like everybody when they transition from being an individual contributor to becoming a team manager, it's kind of a different role, right? You still need some of the same techniques but also it's quite different. So I think we are moving in that direction. But to your point, I think it's also not yet at a, you know, at a point where you literally just have the agents do everything, right? Like I don't think agents can effectively run an entire company or any substantial company autonomously.

I think that to your point, like on analytics, you know, the fact that agents can help close the loop with, you know, maybe drafting some content, there's still a human review step. You're posting it and then maybe the agents are helping to analyze what's working, what's not working and seeing what's, you know, kind of catching fire out there from other people in the content world. You know, I think it does create this like really high leverage circuit for people who know how to effectively deploy agents. Even if humans are still in the loop at points, I think that the amount of the loop or the percentage of the loop that is more agent-driven is going to increase and increase and increase until at some point it's almost like, yeah, there's a human, but like that human is extremely leveraged and you could imagine building a company that you never would have dreamed of without any employees or with minimal employees. Now you can do that with agents.

And another person can do three other outlets. And maybe even like researching ideas, so like, you know, all the substantive work that goes into like going deep on a topic, like a lot of that can be enabled by agents.

Yeah, so I think maybe going back to our generational metaphor, you know, there was like the chatbot era and there were a lot of products that were built as chatbots. And, you know, now we have agents that are much more fully autonomous. So I first, you know, would separate products that kind of fall into the chatbot era from the agent era. Um, even within the same company like Anthropic, they have Claude like the vanilla version which is much more close to a chatbot. I mean it is actually a weekly agentic product. So it can do a few turns at once but you're not using Claude to go and, you know, work on something for 5 hours without human intervention. Right. And this is Claude like the end-user product. Confusingly they also use that name for their models and as a prefix to their other products like Claude Code and Claude Co-work, but the plain kind of end-user Claude experience I think of as more of a chatbot. Um, likewise with ChatGPT, and that can still be very valuable but, you know, when you think about like frontier agents that are actually able to perform, you know, maybe hours of human equivalent work autonomously, you know, you put into that category OpenClaw, you know, everybody kind of got excited about OpenClaw on Twitter because of what it could do. Um, I would put into that category Claude Co-work, um, and then, you know, products like Perplexity Computer and of course Hyper Agent. And then within that bucket there's kind of, um, there's, you know, I think different options for, you know, depending on what you want. So what Hyper Agent gives you and what we focused on is both the fact that you can do this in a much more team-enabled setting as you kind of refine their skills and memory. We put a lot of emphasis into this like closing of the loop for self-improvement which I think is key because even though agents are out of the box very smart and they're only getting smarter as the models underneath them get better and better.

Exactly. Yeah. And I think, um, I mean, one, by the way, as an aside, I think Telegram has a really interesting opportunity to become maybe the dominant platform in messaging for agents because they just have the best ergonomics of all these messaging platforms to be able to deploy bots, right? You know, Hyper Agent has a really first-class Telegram integration, right? Um, and even if you're not a current Telegram user, I might recommend for people to go and like set up a Telegram account because it's such an easy and free way to create both individual and group chats with your agents. Um, and it's very cool actually to see in a group chat setting, you can add even multiple agents that you've created through Hyper Agent and have them jump into the conversation as relevant and even start to talk to each other, which is very cool.

Absolutely. Yeah. I think that is one of the interesting emergent phenomena here which is, you know, as people have gotten really into their agent building like OpenClaw, you know, you see all these OpenClaw fans who have created like a virtual twin of themselves in their OpenClaw instance, right? This is my like virtual Howie agent that has really learned so much about me and actually does have real-time access to the same context that I do so it knows what my schedule is.

Yeah.

I personally like, um, you know, kind of these multimedia related use cases. So what I mean by that is, you know, sometimes for fun and sometimes for like actual kind of, you know, functional marketing purposes. I like to have Hyper Agent go and create like marketing ideas. So, at one point, we actually did do a billboard, a small billboard campaign for Hyper Agent, announcing the launch of it. And I had it go end to end sourcing the actual billboard location. So, we worked with a vendor that had like basically a list of all the billboards that were available.

It was in New York, LA, and SF. And so we grabbed all the inventory and then it had all the specific locations like what street it was on and then Hyper Agent was able to process all of those locations, cross-reference them with Google Street View and Google Maps. So it can literally put them on a map and then even show like the street view point of view of like here's what that billboard looks like on the street. And then best yet because it's very versatile in terms of being able to chain together different tools, it was able to take those street view images and then pass that in to leading image models like Nano Banana and then use that to generate a really high-fidelity mockup of what our billboards would actually look like. Um, you know, so it took real world location shots, mashed that up with like our actual kind of campaign imagery which it created based on the concept and then could help us visualize like here's literally what that billboard on Sunset will look like. I thought that was very cool. And then, you know, I think like we ended up not, you know, like we weren't around long enough to do a Super Bowl ad. We just started Hyper Agent basically around the holidays. But, you know, had we done a Super Bowl ad or maybe for next year, you know, I use Hyper Agent to go and generate some like actual video concepts, right? So, again, it's able to like come up with concepts combining, you know, real product marketing understanding of like what does this product do? We fed it or gave it access to our documentation for Hyper Agent. So very meta, but like Hyper Agent was then able to learn about all of the capabilities and differentiation of Hyper Agent and then to be able to go and come up with really good marketing concepts that it could cut up into different scenes like a screenwriter or like a director of a Super Bowl ad script and then actually generate really high quality production-grade videos using VO in this case from Google, so that it actually looked like a real Super Bowl ad. So I think, um, you know, I like the creative use cases because they're just so immediate and visceral. Um, but of course I also have like many functional use cases, chief of staff use case where it's reading my emails and Slack all the time and pushing me anything that needs my attention. So saying, 'Hey, you just got this really important email from somebody, a customer or a partner or an investor. You should probably respond to this and here's even a drafted reply.'

Okay. So, this is actually a demo account of mine that's realistically recreated based on my actual usage. My actual usage has everything of mine like all of my calendar, my emails, everything. So, it's a little sensitive. Either you can have it go on a recurring schedule or like similar to OpenClaw can have a heartbeat mode where it's always, you know, waking up and checking new stuff and then pushing you messages either via Telegram, it could be via Slack, it could be via email, but you can kind of set it up how you want. Um, I like Telegram because then it just feels like it's almost like a personal assistant like messaging you all the time. You can respond to it and especially Telegram if you have the bot living in your group channels already with your team, then you can just mention the bot and say, 'Hey, you know, go research this thing on behalf of the team.' Go create a caption for this post.

Totally, you go down.

Well, you know, it's like agents consuming the content. But then, you know, also probably agents helping to create the content. And so, you know, it's this funny loop of like what happens when it's really just like our agents posting and consuming all the content on our behalf. Right.