From 2 Guys in a Garage to 700 Employees with Scott Henderson, CTO & Co-founder at Celigo

(Intro Narrator at 00:00:00) Today, we're talking to Scott Henderson, CTO at Celigo, about why he thinks AI is overhyped, but still needs to be adopted as quickly as possible. Thank you to DigitalOcean for sponsoring this episode. For simple cloud and powerful AI that's built to scale, visit digitalocean.com or just click the link in the show notes. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:29) Alright, so Celigo is the name of the company. You're the co-founder. Why did you start this company?

(Scott Henderson at 00:00:34) You know, I was a young, naive kid back—it was like 2006. I had worked one other job, and, you know, just that adventure aspect of startups. And, you know, I grew up hearing about all the garage companies that started in someone's garage and turned into these big, huge empires. And that just was always really exciting. You know, I'm like the type of person that if there was a gold rush, I would go there and want to go do that too. So just the adventure of it all when I was young. And so my boss, the CEO, he started the company. And I had worked with him at NetSuite before he started Celigo, so I knew him, and I knew he had just started a company.

(Scott Henderson at 00:01:19) So, you know, I just called him one day. I was like, "Hey, dude, let me come join you. I want to do it too." And that's really how it started. And he was actually working out of his garage, so I get to say that I joined when we were in the garage. But yeah, it was just in pursuit of that, of an adventure. And I really had no idea what I was getting myself into because it's a grind. It's hard. But at that time, it was pretty fun and exciting sounding.

(Joel Beasley at 00:01:48) When you made that phone call, you were like—what value did you think that you had to him? You're just going to call this guy and be like, "Hey, I want to join you." What value did you bring?

(Scott Henderson at 00:01:57) Well, at that time, I brought a lot of value. So the company started off as a consulting firm specializing in NetSuite integrations. And I was the primary person responsible for testing NetSuite's integrations. So I knew them better than anyone on the planet. And so I told him, I said, "Hey, you know, I know these things better than anyone in the whole world, and I'm really good at solving automation problems." I still remember telling him that at a Starbucks or something. I said, "I'm really good at taking hard problems and automating them." And that combination, he just got real interested real fast, and then we made it happen right after that.

(Joel Beasley at 00:02:41) And then you just started having those business conversations about what it would look like if you joined, because that's a big risk for you. Like, you have this vision of empires. You've seen it happen. So you have to understand the business side of things, and then you've got your family that you've got to care for. So you had to work something out. How did all that happen?

(Scott Henderson at 00:02:59) Yeah. I mean, to me, the risk—so I didn't have, luckily, I didn't have any family, kids. You know, I had no wife, no kids at that time. The risk for me was I was at NetSuite. They were pre-IPO. They were for sure going to IPO. Very successful company. Tons of career paths. When I told them I was going to leave, they gave me lots of options. They were like a really awesome group of people over there. And so that was the risk. It's like, man, am I just making a stupid mistake throwing away this chance to work at a great company like NetSuite, which is now part of Oracle, for the rest of my career, to go do something that could totally fail? And so that was where the risk factor was. Talked to family and friends, and, you know, they all—I think my dad said it best. He said, "Look, Scott, I can hear the excitement in your voice when you talk about the startup. I don't hear that when you talk about the safe bet." So, you know, that was kind of the straw that made me decide, like, yeah, I've got to take a chance.

(Joel Beasley at 00:03:56) Your dad basically said YOLO.

(Scott Henderson at 00:04:00) Yeah, totally.

(Joel Beasley at 00:04:02) And so it worked out. How is Celigo doing?

(Scott Henderson at 00:04:06) Celigo's great. So we've been—yeah, we're like, what, almost 20 years old. 700, 800 employees. You know, successful business. Always been kind of profitable along the way. We're one of those organically grown companies. Didn't take a lot of funding. Always teetering the lines of profitability. And yeah, it's a good company. It's a lot of good people that work here. So that's probably the most, aside from building awesome products, the people that you work with is the second most important thing to me, and the people here are really nice, hardworking, good people.

(Joel Beasley at 00:04:45) Were you intentional about that at the beginning, or was it just heads down, let's make money, we've got to figure this out?

(Scott Henderson at 00:04:50) No, we were very intentional. So we kind of came at it like the two things that we both wanted to do, you know, was never make someone work on the weekends. We both worked at companies where they would, you know, kind of like that movie Office Space where the guy walks around on Friday and is like trying to get everyone to come in the next day. Both had experienced that. Didn't want to have that kind of culture. Never discourage people from working on the weekends, but never like, "Hey, you've got to work Saturday, Sunday." And then the other big thing we said is we don't want to hire any assholes. So we had some really talented assholes in the early days, and we just decided it's not worth it. And so that was the other thing. And I think from those early decisions, we just have a good group of people that have come and stayed and hired similar good people.

(Joel Beasley at 00:05:44) And so if you were to go back to yourself 20 years ago when you were starting this and give yourself like one key takeaway, what would it be?

(Scott Henderson at 00:05:53) Oh, man. I mean, just that it's going to be a grind. I don't think I understood just how much you have to keep going, and it's never easier. It's just always another hard thing around the corner. Yeah. I don't think I was prepared for that, just that grind.

(Joel Beasley at 00:06:14) It's just a series of getting punched in the face different ways.

(Scott Henderson at 00:06:17) Yeah. I mean, you're always dealing—yeah, you're always dealing with the next big problem. Then you eventually think you have it all figured out, and then it's like, "Hey, I came around about four years ago," and just like, then everything had to be rethought out. Like, I kind of felt like I was in my golden years. I had all the tech figured out, and it was just very clear what was going to happen. And then all of a sudden comes AI, and that's just like through everything for a whirl.

(Joel Beasley at 00:06:45) But you guys have adapted pretty well. I've been reading and looking at you—as far as your current product goes, you've been able to leverage the AI more than it crushed you, right?

(Scott Henderson at 00:06:55) Yeah, I think we definitely have stayed ahead of the curve and are on top of it. But talk about a whirlwind of changes and tons of info to decipher and keep up to date with and, you know, what's real, what's not. And, you know, there's just a lot. It was definitely a lot of change really fast all of a sudden, and yeah, it was tough.

(Joel Beasley at 00:07:22) Well, luckily we've got beautiful, infallible, perfect AI agents that can autonomously help us with this stuff, right?

(Scott Henderson at 00:07:28) Yeah, exactly. Yeah, just tell my agent to do my job.

(Joel Beasley at 00:07:31) That's right. Yeah, you say wake up, do job. That's it. That's the function call, do job.

(Scott Henderson at 00:07:36) Right.

(Joel Beasley at 00:07:37) So you said that they started as this consultancy for NetSuite. Now it's a full-blown automation platform. How do you explain to people, like if you're out in public and you run into somebody and like, "What do you do?" How do you explain to them the tool in like 10 seconds?

(Scott Henderson at 00:07:53) You mean like, what does your company do?

(Joel Beasley at 00:07:56) Yeah, like what does it do now?

(Scott Henderson at 00:07:58) Yeah, yeah. I mean, the example I always give people is, you know, let's say you sell hats and t-shirts on Amazon. You have a website where you have a store where you can buy online or you sell on Amazon or eBay. Well, those orders get placed, and then, well, you need those orders to go into your accounting system, into your inventory system, to your warehouse to actually go and ship them. And so I just tell people we're that middle layer. We're that plumbing. Like when the order gets placed and it lives in Amazon, we get that order over to all your other systems so that that hat that you bought makes it over to your house and so on. So we're just—that's how I describe it to my friends and family who are not at all technical. I just say we're kind of the plumbing between all the different websites and applications that are out there that people run their business on. It used to be hard to explain that. Now it's much easier because I think pretty much everywhere you go, people are using lots of different apps. And so they get it, like, "Oh yeah, I don't want to enter data in the same place multiple times." So I understand why you need a product like Celigo.

(Joel Beasley at 00:09:07) Well, absolutely. And it sounds like you're specialized. So the first product that comes to mind is Zapier because that's one of the ones that we've used previously, trying to connect CRM data systems and things like that. But it doesn't seem like you're a direct competitor to them. It seems like you have your own lane that you crush it in. Is that—am I getting it right or no?

(Scott Henderson at 00:09:27) Yeah, that's true. Like Zapier is a really good product for just individuals that want to try to do automations for themselves, for their own work, their own day-to-day work. Like, "Hey, every time I get an email, look at this email and go create a task." You know, that's where Zapier really shines. It's for the individual at a company. Anyone can kind of pick it up and start doing their own little personal automations. We're more like integrating systems, like integrating Salesforce and NetSuite and bringing those two apps together in their entirety where, you know, let's say in Salesforce, you updated 20,000 contact records and you need that to just be synced automatically to NetSuite or put it in Snowflake or, you know, wherever you have the data replicated. That's where our product shines. It's keeping systems in sync in mass plus automation. So I mean, you can do the same type of Zapier automations in our product, but their product is easier to do those individual things.

(Joel Beasley at 00:10:39) When I was preparing for this interview, I learned about something new, atomic agents. Do you know about this? Can you explain this to me?

(Scott Henderson at 00:10:48) I mean, it's just a term I started using.

(Joel Beasley at 00:10:51) Okay.

(Scott Henderson at 00:10:51) Yeah, because we were really thinking about agents from a first-principle standpoint and workflows and how they fit into workflows. And our premise over here is that you want to give an agent the least amount of autonomy possible. Anything that you can do—anything that you can figure out that's deterministic about what an agent needs to do, you should actually code that up. Or if you have a workflow automation tool, map that out so that all the deterministic stuff is laid out and running. And then you want to find the few—the places where autonomy is just absolutely required, and it's minimal. It's just the part that cannot be done through any type of deterministic set of steps. Like, you just need a brain at this step to decide, "I need to do step—I need to do like A, B, C, then A, and then B again," and there's no way to map out what it would be always. And so that's, like, to us, the atomic agent. It's that minimal amount of autonomy that you need that there's no way to break it down any further. It's like impossible to break that down into a deterministic set of steps. And then that's what we build. Like, we call them agentic workflows, AI workflows. There's a lot of terminology out there. But we believe just when you're automating workflows and processes, you want to map out as much as you can deterministically just with all the old traditional, good old-fashioned tools and APIs. And then the parts where you need autonomy, you put in what I'm calling atomic agents that are just like the minimum autonomy possible, because it's the least likely to make mistakes. It's easier for the AI to do its task. And yeah, and then you couldn't do it before. So it unlocked something that was impossible to do before unless you had human-in-the-loop workflows where you send a task to a human to do that part. Now you can actually, you know, just have the AI do some of those things that are—you know, there's some things that are easy for the AI to do.

(Joel Beasley at 00:13:09) I did my first one about three months ago, three to six months ago. So I was having this problem where I get a lot of emails. I get it because of the show and everything. I get so much outreach. And there were, you know, people reaching out that were trying to get on the show, or I was just missing the emails, people interested in doing, you know, promotions, whatever it was. And so what I ended up doing was I ran the calculator on cost to have an AI look at every one of my emails. And so I did an integration where every time there's a new Gmail email, it will—I've explained to an AI agent essentially what it looks like, the type of emails I want it to flag, and then write a draft reply and then have it pending in a specific folder. So that's the most basic one because, you know, you can't really do it with keywords. Like I tried to do it before AI with keywords and stuff. But the thing is everyone uses the keyword "podcast" like in their footer. So it just creates all of this junk and it's not actually faster because it's got too many things in there that are noise instead of signal. But it ended up working pretty well.

(Scott Henderson at 00:14:16) Yeah, that's like a beautiful example of using AI. I call those human-assisted. Like, you have it do something and tee up the thing to you to look at as a final—like, you're verifying it at the end. But it did all the annoying, heavy-lifting stuff beforehand.

(Joel Beasley at 00:14:36) Yeah. And there were two legs. The first leg was like a standard functional API for new message.

(Scott Henderson at 00:14:41) Yep.

(Joel Beasley at 00:14:41) And the second one was just the agentic review of it. And then the third one was, you know, creating a draft reply, which is kind of almost like a mix of both of them, right? Because they created the text and then injected it. So yeah, it's—yeah.

(Scott Henderson at 00:14:56) We would call that an AI workflow because it's like, it's predetermined what you're doing, and then you just have steps that have AI doing various things. Whereas like, there's the autonomous agent, you know, in its pure form. You would just say, "Hey, your job is to make sure I only get emails that are—that meet this criteria," and you give it a tool to go look at your email. You just give it a bunch of tools, and then you let it decide what it wants to do. And that's, you know, not going to—

(Joel Beasley at 00:15:28) There's a lot of tools out there that are trying to do that right now. Like Superhuman, I think, is one of them. Gmail. Apple Intelligence. Have you gotten that update on your—do you use the standard Mail app on your iPhone?

(Scott Henderson at 00:15:39) No. Yeah, I do. I use it for work. I still have my work emails on my iPhone.

(Scott Henderson at 00:15:43) I use Apple Mail for work and then Gmail for personal, so I don't have to share them. I keep them separate that way.

(Joel Beasley at 00:15:52) So have you seen in the standard iOS mail app the new Apple Intelligence that lays over where it'll filter your mail out?

(Scott Henderson at 00:16:00) No. And so I gave Apple Intelligence a chance and I quickly turned it off. What do you think about it? And this was maybe a year ago. I mean, Apple just seems so behind the curve on this. Like, I feel like they have to almost really impress me to get me to trust them again. Like, Siri still, I'm just like, dude, it's Siri. Can you change songs or—it just seems like it can't do anything. So therefore, I have a bad taste of what anything new they're doing. And so I did try the whole ChatGPT integration and all that stuff, and it still just did weird stuff that I didn't like. So I turned it off. And now I don't know. I have to hear something really cool to get me to want to go turn it back on.

(Joel Beasley at 00:16:44) I don't have that for you today. Yeah. My entire response to all that could be, it's okay. It's something. I mean, it's a step. I just I don't notice it at all in the operating system minus the mail filtering stuff it does for me, which Gmail is doing pretty well on its own anyways.

(Scott Henderson at 00:17:07) Right.

(Joel Beasley at 00:17:08) But yeah, I fully reflect your experience with being depressed about Siri not being what it should be. For me, it's just easier to open up Grok or Gemini or GPT or Claude and just do what I want to do directly into the app.

(Scott Henderson at 00:17:26) Yeah, absolutely. I mean, I think it was Apple's CEO on the news right now saying that OpenAI is really their first competitor in a long time. And I remember when ChatGPT came out, I remember thinking, God, this is what Apple needed so badly if they had had this. It's the perfect usable companion. It's the next evolution of user experience, and they missed it.

(Joel Beasley at 00:17:51) Google missed it too. They were all playing catch-up.

(Scott Henderson at 00:17:55) Right. And I think Google—I don't use Google's AI that much. I mean, their search one I do, but they're at least well-positioned to catch up. I don't know that Apple's—they don't, I don't know what type of research is going on at Apple.

(Joel Beasley at 00:18:08) Yeah. Yeah. But you're right, Google has the bench to catch up. You know, to be honest with you, I've been pretty pessimistic about the Google AI stuff for the past couple years, like Gemini and stuff, especially with the examples that had come out in the news and all of that. But my wife last week said that she had been using Gemini a little bit. She was usually a fan of Grok, and she's been using Gemini a little bit, and she started to like the responses. So last three or four days ago, I installed Gemini. And I got to say, it's getting significantly better than it was.

(Scott Henderson at 00:18:44) Yeah. My go-to when I need a web search, when I need the latest, best info on the web—it seems the best. So I find myself using lots of different AIs all throughout the day, and each one has its own little place. And Google says, if I need the latest info on the web, go to them. The one in ChatGPT is not that great. Perplexity, you know, so yeah.

(Joel Beasley at 00:19:08) I don't go to Perplexity. I cycle between GPT, or ChatGPT, Claude, usually for show notes and writings and things like that, and then Grok. And then now I've been going between Grok and Gemini too. But I haven't gotten to Perplexity. Is that something I should look at?

(Scott Henderson at 00:19:28) Not yet. No. I would wait until—there's nothing novel there. We liked it in the early days because you could make API calls. So you could put an API call to Perplexity to go get some data dynamically from the web. That was really interesting for us, building workflows and automations where you want to start to throw in web searching and then AI analysis of the info you get back. So they were early there, and that was cool. So we still use them probably for that in some places. But now everyone else has that capability. So I don't know if they have anything stand out anymore.

(Joel Beasley at 00:20:05) You ever come across Venice AI?

(Scott Henderson at 00:20:07) No, I haven't seen that one yet.

(Joel Beasley at 00:20:09) Yeah. That's, it's like the security nerd's AI—unrestricted, unlocked versions of the models so you can interact with them. And it's created by this guy named Erik Voorhees, who did a lot of stuff in the crypto world. But anyways, whenever I have something like a question that I need a good answer on that I don't want there being an audit trail of, I go to Venice. Because they all profile you. They all save your stuff to try to serve you better in the future. And to be honest, I want them to do that. But sometimes there's medical questions that I don't want them to have, or there's legit questions that I have that I don't want it to know that I'm asking or have a history of it.

(Scott Henderson at 00:20:54) Right. Yeah. Yeah. Makes sense.

(Joel Beasley at 00:20:57) Real quick, before we continue, I wanted to give a special shout-out to DigitalOcean for sponsoring this episode. DigitalOcean is cloud infrastructure that's simple to spin up, complete with integrated AI dev tools, plus 99.99% uptime SLAs and industry-leading pricing on bandwidth. We all know DigitalOcean. It's super reliable. It's been around forever. And now with DigitalOcean's Gradient platform, developers can train, fine-tune, deploy, and scale AI workloads all in one place. DigitalOcean is reliable and affordable at any budget. Companies can save up to 30% on their cloud bill when they come to DigitalOcean. Use code DO25 and get $200 in free credits to get started. Simple cloud, powerful AI, built to scale. That's DigitalOcean.

(Intro Narrator at 00:21:46) Now back to the episode.

(Joel Beasley at 00:21:48) So let's talk more about what you're doing over there. What's the project that you're working on that you're allowed to talk about that you're most excited about right now?

(Scott Henderson at 00:21:58) Yeah. I mean, the project that I'm working on right now is—I mean, you've seen Cursor, I'm sure?

(Joel Beasley at 00:22:07) It's on my computer right now. It's up.

(Scott Henderson at 00:22:10) Yeah. So we're building the Cursor for our product. I think every SaaS company is going to do this. It's really the UX of the future—chatting and telling a product what you want it to do. But the key thing is you need to see the product doing what you told it you wanted it to do. So you still have our web app, just like any other, just like Zapier or—you're still going to see that and be able to use that. But you're also going to have a chat on the side. And when you tell the chat to do something, you're going to see that happening on the left. So if you know exactly what you're doing with the product, you're going to watch it and verify that it's doing what you want. If not, if you're brand new, you're going to see it doing it, so you're going to kind of learn. But the user experience—all that work all these years that we put into a great user experience that every company has done—is now the validator. It's how you, the human, see that the AI is doing what you want. So it's still just as important, but it shifts from being the thing that you work inside to how you validate that what you want is happening. And then you're going to just type and say, here's what I want to happen. Just like with Cursor. And the beautiful thing about AI is that it can bridge the gap between how you have something in your head and how you describe it to what that is in the product. It's very easy for AI to map those two things together. Like, I want to build an automation. I want to build a flow. It knows that those things are the same thing in a product. And so it can just bridge that gap. So it's a really game-changing capability for user experience. And so that's the project I'm working on. I'm very excited about it, and it's a lot of fun doing that.

(Joel Beasley at 00:24:04) That concept you just brought up is the first time I've heard it. And it really touched deep. It was like, yes. Is that something you've read from a book or a UI person, or is that one of your thoughts? Where did that come from?

(Scott Henderson at 00:24:18) That's my thought. It's just—

(Joel Beasley at 00:24:20) It's a good one. Yeah. Have you written a blog post about it?

(Scott Henderson at 00:24:24) No. I don't do enough of that stuff. I'm one of those people that's always heads down in a dark room. I take that back. My room's too light, but I'm alone working, thinking all the time. And yeah, I don't get out much.

(Joel Beasley at 00:24:39) Well, luckily we have a clip of it on this podcast, so we'll push it out there to the world. But you're exactly right. I mean, that is exactly what I am experiencing. The interface I was clicking and using, and now I'm talking to the LLM, and I'm using it to watch what it's doing and validate that it's doing what I want it to do.

(Scott Henderson at 00:25:01) Yep, exactly. Because you can't—you know, AI is amazing, but you have to watch everything it does. It is not trustworthy enough to where it's not going to come back with a, oh, here's a better way to do it, Scott. You know? I wish I could say, here's what I'm trying to do. And it's like, oh, well, here's a better way. It goes and does something, and you're like, okay. That was not what I wanted. So you got to be able to watch it.

(Joel Beasley at 00:25:24) And you're like, there's a mistake. You're like, oh, you're right. There is a mistake. Yeah. I'll go back and fix that again. Now I started playing around with Lovable. Have you played with this at all?

(Scott Henderson at 00:25:34) I haven't. No. I've been on Cursor. That's been my main little buddy. And I haven't had a chance to do the other ones yet.

(Joel Beasley at 00:25:43) If you—so I would say, if you're playing with Cursor, you're getting a developer-side view of what this is like. I would recommend if you ever wanted the nontechnical view of what a nontechnical human trying to use this stuff looks like, I would check out lovable.dev, because it's marketed and designed at nontechnical people to build technical applications and websites and things like that. And then, you know, my background is building software for 20 years. So I went to use it to make a basic website and it did pretty good. And then I tried to do some database-related, some more deeper stuff, and it quickly turned into a rat's nest, so bad, so badly, because it takes away all the details from you seeing it. It's not like Cursor where you can actually see it write the test, and you can be really granular with it if you're smart enough, you know? And so I think a lot of everyday users right now are getting this experience you're talking about—the validating—but they're getting it through Lovable where the validation isn't that great.

(Scott Henderson at 00:26:52) Yep. Yeah. I mean, even with Cursor, it'll run away. And when I first started, I got too trusting. Well, I was just hoping that I could do more. And, you know, I just started to let it run and just accept. And then I almost lost my whole code base because I lost the mental ability to comprehend what was going on across all the files and places. And I almost was like, crap, I have to throw this away. But I was able to, through a couple days of a lot of painful refactoring—I had to read every single line of code myself and then bring it all back together, and then I learned from that moment on. And it was a new project, so it was doable. And I now I know, don't ever let it do its thing without watching what it's doing.

(Joel Beasley at 00:27:41) Yeah. And you're exactly right. You had this webinar. I haven't watched it. But the webinar, we'll put a link to it. But it breaks down building of AI agents. Are you in that webinar?

(Scott Henderson at 00:27:52) I am not. No. But yeah, I'm pretty sure I know who will be. It's people that I—so I've done a lot of projects internally where we use our product to build these agentic workflows and stuff. So I'm sure they're going to be showing some of those, and they're going to be definitely preaching our philosophy that, you know, you want to minimize autonomy, make sure you do as much deterministic stuff as possible. You know, again, so it makes the least amount of mistakes, costs are under control. You know, AI can be very expensive. I was doing a calculation the other day of, you know, if you're syncing a thousand records across five different steps and you're using just AI for everything, you know, that's like $30. Maybe more, maybe $50. Whereas a typical CPU at AWS running normal code is nothing. It costs nothing to do a thousand records. And so that's kind of the philosophy they're going to be talking about. And just showing people practical use cases on things they can do—it's not that hard. I mean, all those API platforms have wonderful APIs to go interact and plug those into your stuff. So yeah, that's what I think they're going to be doing there.

(Joel Beasley at 00:29:11) What's the most expensive mistake that you've made with AI?

(Scott Henderson at 00:29:16) Maybe the most expensive mistake—stops at what you're thinking—it's that we didn't get into it fast enough. We lost the cursor and thinking, being hesitant to just, you know, let it go and—meaning let's roll it out to the whole team and tell everybody on the team they can use the most expensive model as much as they want, and there's no limits on spending, as long as it's helping them do work faster. We took too long, I think, to get that in our heads and get that throughout everyone, throughout all the teams, and let everyone go with that. So I think that's the most expensive mistake I made with AI—is not being fast enough to realize the productivity boost it could have for engineering and getting everyone going there. There's just a mindset where people have, oh my god, I just spent $30 in the last hour on a bunch of AI calls. Like, oh, that's terrible. Company finance is going to come after me and, because that's going to add up. It's going to be $500 by the end of the month. But, you know, you just did the work of two of you because you didn't have to code thousands of lines of code. So the CFO is not going to come after you. Or if he does, we'll come with you and help you.

(Joel Beasley at 00:30:34) Leave that mystery. They might, but we're going to back you up. So you're saying that you are in a position today where that's been communicated. That's the culture now.

(Scott Henderson at 00:30:44) Yeah. Just go all out. And even still, there are people that are hesitant. They're just—I think so. And then there's one person that, you know—so we can see who spends the most, each person. And everyone was really critical, the top guy spending the most. And I get it, and I get it because you're just nervous about spending money. But then when you see what he's doing, he's launching this really cool new service. It's all brand new frameworks and programming languages. And, you know, he did something in a very short amount of time that would have taken months. So then it quickly is like, oh, yes. Yeah. No-brainer to do that.

(Joel Beasley at 00:31:21) So I wanna talk a little bit about when you had that realization. What was the moment where you were like, this is unacceptable, the speed at which we're implementing this. Something's gotta change.

(Scott Henderson at 00:31:35) Yeah, it was just in my own experience coding. So as a founder, you find yourself bouncing around lots of different places. You know, you go wherever the business needs you the most. And maybe for a year, I was working with our account management team, like, on helping customers identify more use cases, just helping us get the word out on all the cool stuff you could do with the product. Getting a lot of tools built internally. So I built a lot of these little AI workflows to help account managers understand the customer. Like, where you go and you get what they're doing in the product, you get their subscription, you get analytics from Snowflake, you go run web searches to find out who they are, you look up the people you're meeting with, and you build this little story for the AM to read right before they go into their meeting. So they really know the customer and they have use cases lined up. So I built a lot of tools there. And so I was doing that for a year, not coding. And that's when Cursor and all that stuff came out, somewhere in that year. And so then when I got to a point where I couldn't really help the account management team anymore, I went back into product land. It's like, okay, what could I work on in product? And I started just using Cursor at that point because it was like, oh, yeah, you gotta use Cursor. And that was the eye opening thing. I was like, oh my god, this is crazy how much more productive you can be. So that was the moment. It was just a break from coding and then back in and then using just the tool that's out there at that time and seeing what it can do and realizing, wow, we gotta get this going ASAP with the teams.

(Joel Beasley at 00:33:18) So you saw it, started using it just out of the nature of your work, and then you immediately were of the mindset that, hey, we need to get this spread far and wide throughout engineering.

(Scott Henderson at 00:33:29) Yeah. Yep. And whenever I take on a project, I always like to look at what's the latest stuff out there, you know, get a fresh start every time. Because I mean, that's my job too. But I actually like that too. I'm like a gear person too. Like, any sport I do, I gotta have the coolest, latest gear.

(Joel Beasley at 00:33:47) Of course.

(Scott Henderson at 00:33:48) So if I'm gonna go start a new project, it's like, what's the coolest IDE I could do? What's the new stuff I wanna get all set up?

(Joel Beasley at 00:33:55) Sublime anymore.

(Scott Henderson at 00:33:57) Yeah, right.

(Joel Beasley at 00:34:00) Okay. So let's talk about the executive side of getting this information out, this knowledge out to engineering about Cursor. What was it? Was it a memo? Was it in all hands? Was it you instructing your direct reports to cascade it down? Like, how did this actually roll out?

(Scott Henderson at 00:34:16) Yeah. So I have a really awesome SVP of engineering named Suresh. And so I just met with him. We can talk whenever. Anytime something comes to mind, we just reach out, talk, and so I just showed him. I was like, hey, watch this. Watch me code. Watch what I'm building. And just highlighted how amazing it was. And then it was quickly also his realization, like, we gotta get this rolled out. So the next day, I think we happened to have our weekly engineering leadership meeting. We started talking about this with the team. Strong top-down push to get everyone to do this. So I think we got total buy-in from all the directors the next day. Maybe a couple days after that, they wanted to try some stuff themselves. Then we started telling everyone on the teams to sign up, start using it, give us feedback. Then there was a lot of comparisons. Like, okay, we've used Copilot before. Let's try, how does this compare? Get everyone's feedback. There were lunch and learns. So we did a lot. Suresh is really awesome at, once you know you're gonna do something across the org, rolling it out to everybody and getting all the teams involved in buy-in and doing bottoms-up feedback, top-down feedback, just all that stuff. He's kinda like a force to be reckoned with when he's on a mission to do something like that. So I have the easy part. I sit back and just be like, yeah, that was my idea.

(Joel Beasley at 00:35:47) Well, yeah, you hired well.

(Scott Henderson at 00:35:50) Right.

(Joel Beasley at 00:35:51) Yeah. So that's good. Okay. Yeah. Lunch and learns, connecting it to something they already kinda knew like Copilot. Now you guys had been using Copilot, but what? The adoption just wasn't as wide or wasn't as impressive?

(Scott Henderson at 00:36:02) It was before agent mode came out, and it just, they're like, sure, it helps. Auto-complete is great, but it wasn't the same as, you know, Cursor with agent mode where it goes off and does 10 things in a row that were, you know, kinda hard to piece together.

(Joel Beasley at 00:36:18) Yeah. I didn't really use Copilot because in that time period of Copilot coming out to Cursor, I just wasn't actively programming on a day-to-day basis. And when I went back to it, my buddy Derek's like, check out Cursor. And I did, and I created this Rails project that would have taken me two weeks in three hours, and I was just blown away by it.

(Scott Henderson at 00:36:39) Yeah. And especially if you haven't coded in a while to just think about, like, I was like, oh, man, how do you do a PR again? And now I just say, hey, commit this, do this, do this.

(Joel Beasley at 00:36:48) I know. I know. I wasn't even typing git commit. I just had to type commit, and here's the message, and it would just do it.

(Scott Henderson at 00:36:55) Yeah. So that's amazing. And Copilot was nothing like that. So...

(Joel Beasley at 00:37:01) Yeah. So you're in San Francisco. Is that right?

(Scott Henderson at 00:37:04) Well, I'm just south of San Francisco in a little town called Burlingame, right by the airport, actually.

(Joel Beasley at 00:37:11) Oh, very cool. Yeah. And there are a lot of billboards around there. I've been told about pushing to replace employees with AI. Is this true? Have you seen these?

(Scott Henderson at 00:37:23) Yeah. I mean, there's a lot of AI billboards. There's a lot of AI agent billboards. You know, it's definitely talk in Silicon Valley about agents entering the workforce. So it's out there, and it's definitely, you're definitely inundated with AI agent materials. When you drive through San Francisco, when you drive down the Peninsula, you're gonna see it.

(Joel Beasley at 00:37:48) You're a pretty smart guy. How do you think this is gonna roll out?

(Scott Henderson at 00:37:52) You mean, like, AI...

(Joel Beasley at 00:37:53) Agent interviewing for today versus, like, you interviewing and hiring an AI employee?

(Scott Henderson at 00:37:59) Yeah. I mean, I think we're a ways away. Yeah. But I think we're a ways away. And there's a lot missing. Like, everyone's so impressed by the intelligence of ChatGPT and all these models, and they are very impressive. But you have to think, there's a whole other set of things needed for AI to be your coworker. And maybe the biggest glaring gap that I'm like, and I sometimes Google it, and I don't find much info on this. And I'm like, maybe I'm just missing something or what, but it's the ability to learn fast. Like, I could tell an employee on my team, hey, you can't do this this way. And maybe they'll make one more mistake. But after that, it's ingrained. You know? It's like, and they take that with them forever. Humans just, you just tell them stuff and they learn. And the AI doesn't have that ability to learn. The memories they are doing right now are nothing like human memories where you're just constantly absorbing new info every day, adapting to changing environments and still being able to use all your intelligence. Like, AI is just smart. It's Einstein in a box, but every time you give it a problem, you gotta give it all the context. And so I feel like that's just a big missing piece is the ability for these things to learn fast, like, meaning immediately. Like, you tell it something and it takes it to heart and that's what it does from here on out. So that's a big gap. And then the other gap is just they would need to be everywhere we are. Think about when you're in a business, how many decisions are made in the conference rooms? So you'd have to be in the meetings with everyone in the room, at the water cooler, at someone's desk, in phone calls. You have to be in the email chains. You gotta be in the Slack threads. And just like humans, if AI is gonna be effective, it's gonna have questions. It's gonna be IMing people. It's gonna send you an email or call you and say, hey, I'm working on this project. Can you help me out with this one problem? I don't know where this info is. It's gotta be able to do all that stuff as well, which means we need new hardware devices, and we need software agents everywhere that things are happening. And then what was the other thing that I was thinking about? Oh, and dealing with conflicting info is another tough one that, you know, when two people are talking and have different opinions on what should be done about a problem, AI's gotta be able to somehow know, like, oh, that's the CTO, they know this. That's the CPO, they know this. Therefore, this is probably the right answer when they're saying the opposite things. And maybe even reading body language. You think about humans. And when I say something, if I'm, the way I say it, my, the way my face changes, that is communicating a lot of info that an AI would miss unless it's actually watching too. So you think maybe there even needs to be video to where it's gonna be watching too. So for it to be a coworker, it's just there's a lot of things missing that are, so I don't worry a lot about the coworker that are coming in to replace the people until I start to see, like, when the first bot IMs me and asks me a legit question, I'll be scared.

(Joel Beasley at 00:41:25) Yeah. No. You make a lot of really good points because a lot of the technology today is just translating the audio waveforms down into text and then decisioning and doing operations based off of the text. Whereas, you know, I've got kids. They say no or yes. It doesn't really matter what word they say. I know if they're telling the truth or not based off of the tone of their voice and the shape of their character and all of this stuff. Right?

(Scott Henderson at 00:41:53) Yep. Yeah. So there's so much there dealing with conflicting people. That's interesting. I haven't thought much about that one. But I would like to point out something you said. You said people learn really fast, but that's got the caveat that you have great people. Right? So some people are listening and they're like, what is he talking about? And the answer is Scott has great people at his company.

(Scott Henderson at 00:42:14) They take to heart what I say, hopefully.

(Joel Beasley at 00:42:19) Now memory retention. Let's, I wanna talk a little bit more about that. Have you gotten into, do you use Grok at all? Or...

(Scott Henderson at 00:42:27) I have not used Grok actually because I'm not on X. And it used to be, but then I just, it was so distracting.

(Joel Beasley at 00:42:35) Thing too.

(Scott Henderson at 00:42:37) Yeah. So I haven't used Grok yet.

(Joel Beasley at 00:42:39) Okay.

(Scott Henderson at 00:42:40) And should I? I mean, is it, do you think it's a worthy one to try? I like Claude a lot, and I like ChatGPT and Gemini. So I've used all three. Is there a place for Grok?

(Joel Beasley at 00:42:52) I like Grok personally. But that's mostly for ideological reasons. I like the base layer of Grok is to be maximally truth-seeking. Because there's a lack of transparency on how these, on ideologies that are put on top of these models after they're trained. So they get trained and then OpenAI and Claude, they have the boot, whatever you call the boot instructions, that's not available for us to see. We don't know what they're telling it on fairness or what they mean about equal, like, we don't know what they're putting into it. Explainability, I think, is gonna be something that's exciting to legislate where they're forced to tell you the truth about how they're programmed, but that's a whole other conversation.

(Scott Henderson at 00:43:43) Yeah.

(Joel Beasley at 00:43:43) But the reason why I brought all this up was because I use the Grok workspaces. So I know, because I've seen Josh, our producer, use it, that Claude has the workspaces as well. But as far as memory retention, so, for example, when we, in this episode and the recording happens and all of that, we'll take the audio files and have them transcribed and then take that transcription, and it runs through a workspace in Claude, which has a series of preprogrammed things that it runs on it. And so I'm not having to teach it every single time. It's got this workspace. And anytime I create a new conversation in that workspace, it's got all that preprogrammed stuff in it.

(Scott Henderson at 00:44:27) Yeah. Yeah. That's awesome that they're doing that. OpenAI also has a similar thing, project level memories. There you go.

(Joel Beasley at 00:44:35) Yeah. Yeah. And it's, so it's...

(Scott Henderson at 00:44:37) But that's not what you were talking about. Is that not satisfying?

(Scott Henderson at 00:44:39) Summarize. The AI will basically take the whole transcript and try to figure out what it thinks is important, and it'll save a couple sentences. You know? That's it. And so it's, and then, again, the more sentences you have, the more rules you put to an AI, the more things it has to think about, you know, the less effective it gets. And here's the interesting thing I read, and I have no idea if this is true. So it's just an intuition that it's probably kinda true, is that, you know, the average human has a hundred million tokens that they use every time they decide something or say something in their context, you know. And then the models are, I think maybe they allow a million tokens max, but I don't know how good they are at a million tokens. A lot of them are much smaller, like a hundred thousand tokens, two hundred thousand tokens. So it's just, again, you need bigger models. They gotta be able to handle more stuff, and then we need better ways to keep track of all the memories other than just, again, a sentence. I think it's probably saving minimal stuff there. I don't know for sure, but I feel like it's not like a human memory at all.

(Joel Beasley at 00:45:52) Yeah. It's better in some ways, and it's worse in other ways.

(Scott Henderson at 00:45:55) Yeah.

(Joel Beasley at 00:45:56) So I do agree with you too that the more knowledge you give it, the more things you give it, the less effective it gets. And that's a real problem. And to me, I'm wondering if that's gonna tell us something about humanity and intelligence in general. Because humans work the same way. If you flood an employee with too much stuff or a person or a peer with too much stuff, it just gets wonky. Right? They can't do a good job. But if you break it down and do what you were saying earlier, where you have these atomic agents and they're just one-to-one, I mean, people, that's how we're designed as a human structure. Like how civilization is designed. We're really, really good at one thing.

(Scott Henderson at 00:46:39) Yep. Yeah.

(Joel Beasley at 00:46:40) And I wonder if that's a property of intelligence. Like, I wonder if through technology, we'll find out that this is something that happens in intelligence. Intelligence is meant to be these little Lego blocks stacked up. It's not meant to be a single thing.

(Scott Henderson at 00:46:53) Yeah. That's a great point. And then you can get real philosophical and say, then, how it all comes together. Maybe there is some greater form of intelligence that we're just not even aware of, just like the machines would probably have lots of little agents doing stuff. But then would there be somehow how that comes together into one form of intelligence? I mean, I've read, like, or watched videos, you know, of ants, you know, like how they all communicate and act.

(Joel Beasley at 00:47:19) Well, now you're my best friend.

(Scott Henderson at 00:47:21) Yeah.

(Joel Beasley at 00:47:21) Okay. Yes.

(Scott Henderson at 00:47:22) So is it similar? We don't know. I think, yeah. But you said one thing that I wanted to touch on that was really interesting too with Grok and how it's truth seeking. And I think that's one of the things I also read recently is the way AIs are trained is they're rewarded for sounding confident and penalized for saying "I don't know." And that just sets the stage for lots of fibs. And I don't, so maybe I would love Grok because I hate when AI says, like, "Do this," and I'm like, "Are you sure? That seems wrong." And they're like, "No, do this." And then I do it, and then I go and say, "Well, here's what happened." And then it'll say, "Oh, you were right." You know, like, it's just very confident about telling you what to do when it doesn't know. And it doesn't just say, like, "I don't know. I'm not really sure about this." And that's a big problem.

(Joel Beasley at 00:48:12) Very frustrating. It does it a lot with API endpoints. Like, if I ever need documentation and I'm trying to, you know, get help there, it'll just tell you, "Oh, the Facebook API is this." And it's like, that's not in the docs. That looks like a convention that could be.

(Scott Henderson at 00:48:29) Yeah. Exactly. Like, I have to say, like, "Don't invent," like, you shouldn't have to say this, but, like, "Hey, how do I authenticate this API? Don't invent something. You know, if you don't know, tell me." And even then, it's still, you know, hit or miss.

(Joel Beasley at 00:48:42) Yeah. Well, not to get too nerdy, but I have found that if I give it the link to the exact version of the docs that I want and I constrain it to only use that version, it'll go through the right steps typically. Yeah. Yeah. But that's interesting. It's like, how do we start working with this technology where we kind of game it to get what we need out of it?

(Scott Henderson at 00:49:01) Yeah. And that's where we're at. We're all gaming this thing. Everyone's trying to find these clever little tricks to get it to do what you want. And the top researchers I've read say it's very hard, like, very hard to get an AI to do what you want. You know? It's not like there's no answer. It's just it's hard to do.

(Joel Beasley at 00:49:16) Yeah. It's trained on my wife's data. No, I'm kidding. Okay. Alright. Off topic. I want to touch on the ants thing real quick. Okay? So I've often thought, like, if you walk up to the ant and you're like, "Hey. Your motivation is because of this," the ant would be like, "Shut up. I'm doing it because I want to do it." And it doesn't understand how it fits into the larger structure. It's just kind of going about and following its instincts and things like that. But then there's this queen ant. Now, how much have you studied it? Do you know how they're communicating, or is it still a mystery?

(Scott Henderson at 00:49:50) I thought it was, like, via smell. Like, they really...

(Joel Beasley at 00:49:53) They say that.

(Scott Henderson at 00:49:55) Yeah. So that's all I seem to remember. Smells and...

(Joel Beasley at 00:49:59) But do you see how intricate and complex the architectures are that they build underground? It's like smells?

(Scott Henderson at 00:50:05) To me, the one I was, when there was a threat would pop up, then all of a sudden, the whole, like, swarm does something. They start to all behave a certain way. So I just, that's just crazy how, like, they are collectively reacting dynamically to a situation. And I'm like...

(Joel Beasley at 00:50:24) Like the moon launch. Yeah. Like, when we were building a rocket to the... Have you ever seen the movies when they go back and they show you what America was like when we were going for the rocket race and all of that? It's like everybody, like, everyone, everywhere was talking about. You and I were both alive during nine eleven. Right? Like, yeah. Everywhere, everybody just comes together. That's a unique feeling that happens once in a decade. Right?

(Scott Henderson at 00:50:49) Right. Yeah. It's very rare.

(Joel Beasley at 00:50:50) Yeah. Yeah. Alright. Well, dude, this is great. You are surprisingly amazing to talk to.

(Scott Henderson at 00:50:57) Yes. It's fun.

(Joel Beasley at 00:50:58) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.