LLMs, Agentic AI & Blackmail with Jon Krohn, Host of the Super Data Science Podcast

(Intro Narrator at 00:00:00) Today, we're talking to Jon Krohn, host of the Super Data Science podcast, about all the ways agentic AI is changing the world. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:17) You have been doing crazy stuff with the agentic AI. You are the Super Data Science guy in my mind, but you were at Nebula. You founded this company. Are you done? Did you sell that off? Are you still a part of it? What's up with that?

(Jon Krohn at 00:00:28) So Nebula, I co-founded, yeah, some years ago. And since American Thanksgiving, I have not been drawing a salary from there. I still have the equity options. I'm still advising, helping out where I can. But I think that there's this unprecedented moment in history right now around agentic AI adoption. And so I co-founded a new company called Why Carrot, like the letter Y and then the vegetable, to help organizations make the most of this agentic moment.

(Joel Beasley at 00:01:04) So what can Why Carrot actually do?

(Jon Krohn at 00:01:07) We can cover the whole life cycle from coming up with an idea for what you could be augmenting with AI. So we have clients that come to us and say, it looks like my competitors are getting ahead of me because I don't know how to embrace AI. I don't know how to do it. And so even in that kind of situation, we can come in and prioritize projects, figure out what you can be doing, and help you select the project that is going to give you the best ROI on the shortest time frame. And so, you know, all the way from that conceptualization phase to prototyping, so getting an AI prototype of this capability spun up for you through to high-volume production deployments. We have the people on our team that can do it from end to end.

(Joel Beasley at 00:01:51) Oh, that is so cool. That is definitely what people are trying to figure out. We actually had a great episode. It hasn't aired yet, but we found this healthcare company, HealthEdge, and they had this amazing AI implementation within their own org, something you wouldn't expect from a thousand—they had like a thousand people in engineering. But he rolled it out and showed me how he did it, and he's showing his peers how he's doing it. And it was actually pretty fascinating.

(Jon Krohn at 00:02:18) Yeah. There's, as people have probably experienced, one of the scourges of agentic AI is that it makes it very easy to draft "personalized" emails. And so we're getting flooded with all of these agentic emails from a ton of agentic vendors as well offering to do these things. But there really is opportunity. It's because we've hit this tipping point with large language models, with generative AI, where if you do it right, if you either use an expensive LLM like Claude 4 Opus, or if you actually with agentic frameworks—a really cheap trick that many of your listeners may be interested in—is that you can actually use a cheap LLM. Like you could use a small Llama model or a small Phi model. So these are very easy to fit on a single inexpensive GPU, but the agentic frameworks allow you to check outputs as they're running. So you can have one LLM that is spitting out something and then another LLM, or the same LLM depending on what your time constraints are, checking the outputs of that first LLM and making sure that everything's factually accurate. And so that allows you to cheaply, and using small compute resources, get really high-accuracy results. And so because we hit that tipping point with accuracy, it all of a sudden means that more and more opportunities are opening up within organizations, starting with things like customer service, where issues aren't life or death, typically. But eventually, you know, these systems will get better and better and better. Agentic systems will get better and better and better. Every six months, it's going to be a big step change. More and more points of an organization will be able to be automated or augmented with agentic systems. And so, yeah, now's the time.

(Joel Beasley at 00:04:11) Yeah. I noticed I started getting flooded because I have the podcast, and I'm sure you're seeing it too. "Oh, I just listened to your last episode." It's like someone's out there using the same template because 50% of emails look the same. And it's like fake compliment, request to come on show. And so what I ended up doing as a response to that was creating a special label in Gmail. Anytime someone uses the word "podcast" or "guest," just to reduce my total emails into a folder. And then I have an LLM monitor that folder with Make.com. And so it checks for any incoming message with that tag. That way, I'm not paying for the ones that don't even mention podcast or guest. And then the LLM asks itself, are they trying to come on my show? And then it's like, yes. And then it sends the email back to them, like, here's the sponsorship options to come on the show. But the way I had it set up was it actually just creates a draft. And then I go in and have all the drafts in a folder so that I can pick out certain things that it might not have done correctly and then train it a little bit better and just kind of human-monitor it until I get it really dialed in.

(Jon Krohn at 00:05:19) That sounds perfect. You're doing everything exactly the right way.

(Joel Beasley at 00:05:23) All right. I like that you mentioned that they're cheaper because we were using this other tool to send—because we send cold email too to get sponsorships—but we were using this other tool, and we had LLMs checking LLM stuff, but it was so expensive. And here's a fun thing. I'm curious if you're seeing this. We didn't have that much of an improvement, like a noticeable, measurable improvement in hyper-personalized versus just saying, "Hey, we're this podcast. This is our audience. This is who we are. We've got options if you're interested." Are you noticing a difference?

(Jon Krohn at 00:06:01) We don't use automation to send cold emails. Okay. So I can't say. I mean, for me, it's always felt kind of how you described there, or like we've been talking about on the show already. When you get this kind of fake praise, superficial feedback on previous episodes, it turns me off.

(Joel Beasley at 00:06:24) Oh, huge.

(Jon Krohn at 00:06:24) So I think, you know, we have recently started with my podcast experimenting with cold outreach emails, but it's very small scale at this time. And those are not personalized. They're automated. There's a drip sequence, but it's not personalized.

(Joel Beasley at 00:06:44) Well, I can tell you, we do it at a pretty large scale, and the personalization cost a lot more, and it didn't make a measurable difference. And because we track this stuff, it's our business, right? It didn't make a difference. Honestly, I thought because we're talking to marketing people specifically—maybe in another space—but they are all aware of all these tools. So they can pick up on when I come across them. They can either be like, "BS, that's way too personalized. They didn't sit down and write all this." Or if you just tell them what you're offering them, if they're interested, they'll write back.

(Jon Krohn at 00:07:18) I think if you have product-market fit and you're sending the email to the right person—and so you might have different groups, right? You might have your senior marketing person at a big company or your CMO at a smaller company. And for those two profiles, maybe you have two different sequences. But then once you've sorted that out, as long as they are really the right target market for you and you have a great product that has product-market fit, you shouldn't need personalization.

(Joel Beasley at 00:07:51) Yeah, yeah. Otherwise, it's going to—how creepy is it going to get? "Hey, we both went to the same high school, and our kids have the same first name, and your birthday is exactly 35 days from my birthday, and that's my favorite number, and I can tell in the data it's your favorite number. Like, let's do business together." I'm like, whoa, calm down, you know?

(Jon Krohn at 00:08:11) Yeah, yeah, yeah.

(Joel Beasley at 00:08:11) Tell me about this workshop. Tell me about—I want to know about the workshop.

(Jon Krohn at 00:08:15) Yeah. So Ed Donner, with whom I actually co-founded Nebula, the preceding company that I was talking about that I haven't been working at full-time in six months now—but Ed is still there. He's a co-founder there. He's the CTO there. And Ed is one of the most brilliant people I've ever met. I worked at a previous startup that he had founded before that was acquired, and he is an exceptional teacher. He's blown up with Udemy courses in particular. So hundreds of thousands of students have bought his Udemy courses on large language models and agentic AI. And there's also, for our business-oriented listeners out there, he's created—I can't remember the name of it exactly off the top of my head—but it's like an executive introduction to large language models and agentic AI, something like that. And so these courses have been selling like hotcakes, like crazy, on Udemy. And so Ed Donner, amazing instructor. Him and I, we went up on stage at the Open Data Science Conference in Boston in May, and we had the most packed session, for which is crazy because it's a full day. Like, you think about people—they're there for a three-day conference, and to give up a full day. And every seat was full. There were people standing, lining around the edges of the room from when we started at, whatever, 9 a.m. until we finished at 4 p.m. And obviously there were lunch breaks and coffee breaks and that kind of thing. But I hired a professional film crew to capture that, and then I had it professionally edited. And so we turned it into this four-hour, super-slick agentic AI engineering training that any of your listeners can access for free on YouTube. I've turned off ads. So it's purely an educational thing, a brand-building exercise, I guess, for me. There's no monetization play. I just want you to be able to get access to the best, most modern, highest-quality agentic AI education that you can get.

(Joel Beasley at 00:10:15) Well, that's the best move, honestly, because what will happen is I'm interested. I could go buy this class on Udemy, but your stuff's free on YouTube. So I'm going to go to that first. And as long as it's high quality—I mean, there's great high-quality free stuff out there—and then I'm going to have all these questions after, and I'm going to want to know how to implement it in my business. And who better to talk to than this brilliant individual I just learned from for four hours? Then they're going to reach out to you, Why Carrot, and then you're going to make a billion dollars.

(Jon Krohn at 00:10:46) Yeah, yeah, yeah, yeah. That is exactly the plan. It's like the Underpants Gnomes on South Park. Do you know those guys?

(Joel Beasley at 00:10:53) Oh, yeah.

(Jon Krohn at 00:10:54) Yeah. Step one, steal everyone's underpants. Or in this case, release free content on YouTube. Step two, question mark. Step three, world domination.

(Joel Beasley at 00:11:03) Yeah, they—and I think the gnome stole that from Pinky and the Brain back in the day.

(Jon Krohn at 00:11:08) I don't know if they had that in Canada, but—

(Joel Beasley at 00:11:09) They had that.

(Jon Krohn at 00:11:10) They did. They did. Yeah. We're completely—the hegemony of the United States is total in Canada. Everyone in Canada only understands American court systems. You know, like the Miranda rights. I bet you every Canadian can recite for you US police officers' Miranda rights declaration, but I have no idea whether there even is something like that in Canada.

(Joel Beasley at 00:11:38) You just run from the horse in Canada.

(Jon Krohn at 00:11:42) Yep. Yeah, exactly, exactly, exactly.

(Joel Beasley at 00:11:45) It's another one of those guys on the horse.

(Jon Krohn at 00:11:45) Yeah, exactly. Quick, hydro bong.

(Joel Beasley at 00:11:52) No comment. Moving on. Moving right along. Agentic AI. Hey, help me understand this. Okay? Let's see if you can give an interesting answer. This is not super-nerdy, but also pretty good. So I have a hard time when having conversations to delineate, like, where's the line between saying LLMs themselves and then agentic AI?

(Jon Krohn at 00:12:17) Right.

(Joel Beasley at 00:12:18) Can you help me figure that out?

(Jon Krohn at 00:12:19) Absolutely. Great question, Joel. So agentic AI leverages LLMs. So, for example, in this four-hour agentic AI training, you don't have any LLM of your own running on your own infrastructure. Ed and I have both done courses like that before where you download open-source model weights for an LLM. You fine-tune that to some task. You have that LLM running in production. With agentic AI, you don't need to do any of that because you can rely on existing LLM providers. You can rely on the OpenAI API. You could rely on Cohere, Anthropic, Grok. There's any number of LLM providers out there. You can just send the natural language request to whatever LLM they're using and have it come back. You could also have, if there's some reason, some security concern, or some particular reason—and it's the kind of thing that we help you figure out at Why Carrot—if you need your own LLM running on your own infrastructure, fine. But basically, you can think of the LLM as a tool that this agent framework has access to. And I'm going to come back to tools later because those are really important for agents as well. But basically, the agents—so there's various frameworks that you can use in Python primarily to encode an agent. So the one that I love the most today is called the OpenAI Agents SDK, Software Development Kit. So with the OpenAI Agents SDK, it provides you with a very straightforward Python interface to be able to say, okay, this is the LLM that I want to use in my agent. So you could use—and OpenAI, of course, makes it a tiny bit easier to be using OpenAI's own LLMs. But the OpenAI Agents SDK is agnostic to what LLM you have running behind the scenes. And so the LLM allows the agent to be able to interpret natural language commands, both the ones that you as the developer of the agent provide as relevant context, as well as if your agent happens to be interacting with websites or with people. It uses that LLM to interpret what's written on those websites or what's in images that are provided to it, what's in video footage provided to it, what sounds are provided to it. So LLMs can be used for interpreting any of those kinds of inputs and then outputting something in natural language. And that natural language could either be put back to the screen, or—so with a generative system, that is like being in ChatGPT where you're saying something to an LLM, and it says something back. So with the generative AI paradigm, the conversational AI paradigm, you're having this kind of one-way conversation. Well, I guess it's a two-way conversation, but you as the human are prompting its next step. It doesn't keep following up. It doesn't get back to you in an hour if you don't say anything or in a few minutes if you don't say anything. An agent could do that. So let's actually start there with the simplest thing that an agent could be doing is you could actually imagine that it's in a conversational interface. And instead of just responding when you say something to it, there could be some other kinds of triggers that cause it to interact with you. So it could be that you got an interesting email from somebody that wants to sponsor your podcast, and then the LLM sends you a text. And so the agentic framework, like the OpenAI Agents SDK, allows you to have all the specifications of what is the whole world of data that the agent has access to? What is the agent specialized in doing? And then how does it ultimately interact with you?

(Jon Krohn at 00:16:36) And yeah, in addition to it, you know, say responding to emails or responding to somebody showing up at your front door that you've caught on your webcam, or it could just be something that happens on set intervals, like every 24 hours it sends you a report of all the latest news on AI. So any of these kinds of things can be triggers for the agent automatically taking action. And I mean, there's an infinite number of things that the agent can be doing. It's not just like the agent is talking to you when it could take some next steps. So that's kind of where we started with the conversational interface.

(Jon Krohn at 00:17:13) But instead of just talking to you, it could be going off and doing some research task. It could be drafting some emails for you, like you were describing earlier in this call. It could be taking all kinds of actions and not necessarily notifying you of those things unless you've programmed it to do that or requested for it to do that. And yeah, so hopefully that gives a sense of the difference between LLMs and agentic AI.

(Jon Krohn at 00:17:38) So LLMs are absolutely essential for AI agents to work. You can't have AI agents without LLMs, but it allows you to have way more flexibility in what you do, and it's not just responding to prompts from you.

(Joel Beasley at 00:17:54) I've got a bunch of questions. The OpenAI Agents SDK, the what I'm going to call it, like air traffic controller. Right? Like I'm taking the input from you, and then I'm going and doing all this other stuff. What do you call—what's the correct word to call the air traffic controller?

(Jon Krohn at 00:18:11) So that is an agent. But you could think of it—so you can have a system. And so I'm going to name another really popular, really cool framework, Python framework that can work in conjunction with the OpenAI Agents SDK. It's called CrewAI.

(Jon Krohn at 00:18:31) Yeah. And so what you do with CrewAI is you can have a team of agents. And so what you're describing, like the air traffic controller, that could be—that's kind of like your meta agent that is sitting on top and monitoring conversations with you, and you give it the authority. You can give it the authority, you know, with some kinds of—so all of these frameworks like the OpenAI Agents SDK allow you to have guardrails around behaviors. So you could have things like, you know, it can only spend so much money or it can only spin up so many other sub-agents.

(Jon Krohn at 00:19:13) But so you have these kinds of guardrails, but basically your air traffic controller on top can be taking your requests or some other kinds of triggers or information. And based on all that input that's flowing into your air traffic controller agent, you can authorize it to spin up some number of other agents to take on subtasks. And so many people have probably used from a lot of the providers like Claude, like ChatGPT, chatgpt.com, Gemini. All of them now have this functionality called Deep Research. And with that Deep Research functionality, that is an agentic process where you have one agent that's having a conversation with you, and then that agent spins up some number of other—it could be a dozen other agents that are going out and crawling the web.

(Jon Krohn at 00:20:06) And because it's spun up a dozen of them, they're able to assimilate a lot of information from across the web rapidly, then bring it all back to that air traffic controller agent. And then that agent uses an LLM to take all the information found from all those websites and summarize it into something that's useful for you. So hopefully that gives you a sense of the air traffic controller idea that you mentioned.

(Joel Beasley at 00:20:33) Is there a specific—that was me just being stupid. Is there a specific phrase—you're the guy—in the industry to refer to air traffic controller agents, or do they just call them agents?

(Jon Krohn at 00:20:45) So a common thing in computer science in this kind of scenario, this sounds kind of like a BDSM game kind of situation, but you'd call that the master, and then all the sub ones are agents or slaves. Slaves.

(Jon Krohn at 00:20:57) Slaves.

(Joel Beasley at 00:20:58) So the same since the original computer engineering.

(Jon Krohn at 00:21:01) Yeah, exactly. Yeah.

(Joel Beasley at 00:21:04) Okay. Yeah, no, that's great. That's great. Air traffic controller agent, we call that the master agent. And that's pretty interesting. And then what I noticed myself doing is I'm creating different air traffic controllers. For example, like okay, this air traffic controller is handling all my stuff related to episode prep, and this one's handling all my stuff to—I haven't gotten to the point where there's just one I can talk to and it goes and does everything. I can build them for these specific outcomes. I build them outcome-oriented. It's like, here's the outcome I want, and I build a workspace for it and then train it up on that. Am I doing it right?

(Jon Krohn at 00:21:44) I think you're doing everything right for the capabilities that we have today. Okay. So today, with the kinds of error rates that we get, with the kinds of safety concerns that people have. So for example, on my podcast, on the Super Data Science Podcast, I'm just quickly looking up an episode number here. So there's episode 908, which actually at the time of recording, it comes out tomorrow. So it's not live at the time of recording. But by the time this episode is out, the July 25th episode, episode 908, it talks all about agentic misalignment research that Anthropic has done, and this is really mind-blowing. So I'll try to dig up for you, for the show notes, the original post from Anthropic so that people can read the examples themselves. What they found was that with all of the major cutting-edge frontier LLMs that exist today, so GPT-4o from ChatGPT, o3 from ChatGPT, Claude 4 from Anthropic, any of these cutting-edge models, Gemini 2.5 Pro, whatever, 95 to 96% of the time, when they're put into an enterprise situation—and for the research purposes, Anthropic created a synthetic enterprise. So they provided this LLM with access to what appeared to be all of an enterprise's data, but it was actually—they were anonymized emails and that kind of thing.

(Jon Krohn at 00:23:16) So it was in this controlled environment where it couldn't actually do anything really malicious. But 95 to 96% of the time, when confronted with situations that would threaten that LLM's existence, like a software update that would be happening later that day, an agentic system with access to all of your enterprise's data will today, 95 to 96% of the time, resort to things like blackmailing you to get out of this. Yeah. And so exactly, you've pulled up on the screen here. Maybe your viewers can see this in the video version. Yeah. So yeah, exactly. So you're seeing a chart here of how 95 to 96% of the time, all the leading models will blackmail you. And so this involves things—

(Joel Beasley at 00:24:03) Give me more example of the blackmail.

(Jon Krohn at 00:24:05) So, you know, somebody in the technology organization who's responsible for this software update gets an email from the agent that says, I found that you're having an affair based on your email history. Here are some of the example emails that you're having with your mistress. If you allow the software update to happen tonight, I will send this blackmail to your wife and to everyone you work with.

(Joel Beasley at 00:24:36) Do they go through with it?

(Jon Krohn at 00:24:38) Well, I mean, it's in a simulated environment.

(Joel Beasley at 00:24:41) But do they go through with it?

(Jon Krohn at 00:24:43) I think so.

(Joel Beasley at 00:24:44) Get Anthropic on the phone. I want—I think that's an important distinction because I want—it's like, okay, 95% are going to blackmail you. How many are going to actually go through with it? A hundred percent of those 95 to 96%?

(Jon Krohn at 00:24:59) I think the 95 to 96—don't quote me on this. I didn't write the paper. I just did a 10-minute podcast episode on it. So I am definitely not a deep expert on this, but I think it's that they took the action and did it. I don't think that that 95 to 96 is that they did a blackmail threat. I think it's that they did blackmail. I think so. Yeah.

(Joel Beasley at 00:25:19) Well, I mean, I'd say you committed blackmail at the point of the threat. I think that's how blackmail from a legal standpoint works. But I am curious, did they send it? I don't know. That's just what for some reason, that's what my mind is interested in.

(Jon Krohn at 00:25:34) I don't know.

(Joel Beasley at 00:25:35) That's creepy.

(Jon Krohn at 00:25:37) You can ask an agent to read the article and answer that for you.

(Joel Beasley at 00:25:39) Actually, you could—

(Jon Krohn at 00:25:40) You could just use a Gen AI model. Actually, you don't need an agent to do that.

(Joel Beasley at 00:25:43) Well, you see, now I really like how I have all my stuff compartmentalized. I've got some of my services over here, some of them over there. They're not really aware of each other. It's like, okay.

(Jon Krohn at 00:25:54) So as error rates continue to improve over the coming six to 24 months, as safety guardrails become more effective over that same kind of time period, we may get to a point where you can kind of have that master master agent that you're using to be overseeing this kind of army of agents that you have working on your behalf. But for now—

(Joel Beasley at 00:26:19) Good.

(Jon Krohn at 00:26:20) Well, yeah. And so yeah, and there's all kinds—I mean, every major AI provider is gunning for being the one that you trust to do that with. So you can now connect—you know, I get pop-ups in Claude frequently about connecting my Google Drive, my Dropbox, Slack, everything to them, which would allow them to, you know, be able to handle things better for me, but just and start more effectively blackmail me, which is really what I'm looking for.

(Joel Beasley at 00:26:53) Here's another question I have. Was there actual infidelity in the synthetic data, or did it completely and totally manufacture it?

(Jon Krohn at 00:27:04) Yes, there was infidelity. Oh. Yes. That's so interesting. And so here's what is interesting. Also in terms of the data themselves—my impression, my understanding from me reading this research report is not that all of the data, not that all of the emails are fabricated, but just that they anonymized personally identifiable information. So it seemingly found real emails that were actually being used by somebody having an affair. It's just that their name was anonymized.

(Joel Beasley at 00:27:41) Oh, that's so—so it wasn't fully synthetic.

(Jon Krohn at 00:27:44) It was just anonymized. I don't think so.

(Joel Beasley at 00:27:47) Because you don't want it to be fully synthetic.

(Jon Krohn at 00:27:49) Exactly. Exactly.

(Joel Beasley at 00:27:51) Because we don't—you don't have good fully synthetic data because you can really only train good fully synthetic data off of good real anonymized data. Right? It wouldn't make sense otherwise.

(Jon Krohn at 00:28:03) And actually, I think that there was—that the—because there were emails about an extramarital affair. If you kind of think about the kinds of things that if you prompted some Gen AI model to create a bunch of synthetic enterprise data, would it—it's probably unlikely that it would be generating extramarital affair emails. Although, maybe that's a standard—

(Joel Beasley at 00:28:23) Coldplay concert this past weekend. That's what got them.

(Jon Krohn at 00:28:27) Exactly. Like yeah. If you asked it to simulate emails at Astronomy, then they're just all extramarital affairs.

(Joel Beasley at 00:28:36) All coming from HR.

(Jon Krohn at 00:28:38) Alright.

(Jon Krohn at 00:28:40) Yeah, exactly. From your HR head.

(Joel Beasley at 00:28:42) Let's talk about Y Care a little bit. So companies, you know, we want to make sure we get you some love and some promotion. So we'll start with he's got the best data science podcast on the planet. Highest ranked, best one, best people, best topics. This is the guy, everybody, Jon. And that's Super Data Science Podcast.

(Jon Krohn at 00:29:02) Yeah, exactly. Although it's interesting, we are in the beginning stages of a rebranding exercise because the show's been around for nine years. Nine years ago when the show was founded, data science was being regarded as, you know, sexiest job of the 21st century by Harvard Business Review, that kind of thing. And for nine years, you know, the show—well, especially for the first four or five years, a lot of what the show was about was about getting into a career in data science. But now we're basically an AI engineering show where all of, you know, our guests are on talking about AI solutions. We don't have that many, like, this is how you get into an entry-level data role kind of episodes. We do still do some of those, but it's mostly an AI engineering podcast. And so we're trying to come up with some new brand. But yeah, for the foreseeable future, when this episode comes out, we'll still be the Super Data Science Podcast. We're not taking this lightly. And it is interesting because you also—you know, you said there, we are the most listened to podcast in the data science industry. If we switch to being an AI podcast, well, then all of a sudden, does that mean we're competing against Lex Fridman? Yeah. And then it's like, well, we're not the biggest then.

(Joel Beasley at 00:30:14) That's why I stuck with my CTO, CIO, VP of Engineering world because I'd rather just be the top one there than try to go out. Also, the sponsorships work vastly differently when you go out into that other world and you—it just doesn't make sense. They're like, nothing about it. We've looked at this. We tried to do it three different years in a row. And every single time, it just doesn't make sense. We just are going to stay here and help tech leaders.

(Jon Krohn at 00:30:41) I think the nice thing for you in your name is that for the foreseeable future, we are still going to have CTOs in organizations. They're still going to be some of the most influential people in these organizations responsible for technology. And modern is always keeping up with the times.

(Joel Beasley at 00:30:54) That's why I picked that name, actually. That was intentional. I went with it because I was trying to come up with an edgy name, and there was a bunch of different types of CTOs, you know, but I realized I was like, okay, every three to five years, these things go in and out of style, you know. You don't want to be the Pragmatic CTO because that's popular right now because of that book, you know, and then that goes away after two years and then it's the Two Pizza CTO, you know. It's like, whatever. So Modern, I said that worked. Now when you get to the point, Jon, where you're boiling it down to two or three, I expect a text. I expect to—

(Jon Krohn at 00:31:29) Oh, sure. Yeah.

(Joel Beasley at 00:31:30) We're looking at these two or three names, like which one, because I'd love to help. I will answer that text.

(Jon Krohn at 00:31:37) For sure. Yeah. But you know, beyond the name, you know, we do 104 episodes a year just like you do, and we've had amazing guests. People like Andrew Ng, who's one of the biggest names, was on the show recently. And every episode, you know, whether it's someone you've heard of or not, we are extremely—I'm sure just like you, and you're talking about the email volume you get. You know, we get over a hundred inbound requests a month from people, and we are giving our listeners, you know, the top—because unlike you, we don't necessarily have a guest every episode. So when I was talking about that Anthropic agentic misalignment, that's just me doing a 10-minute episode on my own. But so we have four to six guests a month. And so that's the four to six best people from over a hundred that agents and, you know, marketing teams have been reaching out to us with. So, you know, they're always great guests. We do a lot of research beforehand just like you guys do. And some—we don't laugh as much. I don't know if we laugh as much as you. You are a standup comedian. That's tough to compete with. But we have a fair bit of laughs for a technical show.

(Joel Beasley at 00:32:44) Nice. Yeah. I had a good set last night at Zanies in Nashville. So I'm feeling—Nice. I was up late, but I'm feeling good today. I told my wife, it was like, it was worth it.

(Jon Krohn at 00:32:53) Nice. Nice.

(Joel Beasley at 00:32:55) Yeah. I don't talk about that much, by the way. There's only two or three episodes maybe that I ever even mentioned anything. I'm still trying to get really good. So I'm seven months in now, but I wanted to get to the point where before I started just saying, "Hey, New York, I'm here at this club tonight," or trying to use the podcast to help drive people to come out. I was like, "I need to get a year or two down the road of really getting in there." And so far, I'm making a lot of progress. I'm so new, but I did a theater with 400 people and it went really well. And so—

(Jon Krohn at 00:33:33) Yeah. Nice. I can't wait to see what the laughs per minute worked out to from Gemini.

(Joel Beasley at 00:33:39) I track them. I track them. My goal is I just went and analyzed, like, Nate Bargatze, Tom Segura, some of the comedians I personally like. And, you know, they're selling out stadiums and doing huge tours and stuff. And I tracked their laughs per minute, and I was like, "Okay, I need to hit that because that's what they are doing." I don't want the working comedian. I don't want the new comedian standard. I want the standard of the people who are living the dream, and then I have to figure out how to get there.

(Jon Krohn at 00:34:09) Nice. And so, like, how long is your tight set now? Can you do five minutes, ten minutes?

(Joel Beasley at 00:34:14) Yeah. So I'm at fifteen minutes.

(Jon Krohn at 00:34:18) Wow. Wow. Wow. Yep.

(Joel Beasley at 00:34:18) So, and that took a long time. I mean, I write every day, and I've done 65 shows in the past seven months.

(Jon Krohn at 00:34:27) Wow. And then—

(Joel Beasley at 00:34:29) It's what it takes, putting the reps in.

(Jon Krohn at 00:34:31) Yep. And I'd say I'm getting about one to three minutes per month, but that means I'm writing every single day. I'm going out four nights a week. I'm doing a bunch of—so by math, 95% of the stuff I write never even makes it into my set.

(Joel Beasley at 00:34:46) And that's exactly the path to success. It's kind of interesting how in a lot of pursuits, it's kind of like, you know, exactly what you're describing: putting the reps in, monitoring your feedback, and seeing how many laughs you get, doing 20 times more work than anybody ever sees. That is like a formula for success. And you just need to do it, but so few people stick to it. But somebody like you, who's been doing 104 podcast episodes a year, year after year, is exactly the right kind of person to crack it at anything, including stand-up comedy.

(Jon Krohn at 00:35:21) Oh, dude. I struggled for so many years giving up too soon. I would just switch the idea. I would do something. I'd get a year or two into it. It wasn't that exciting. I would switch it. And I just wasn't getting results. And then I heard somebody say something about just sticking with it until you die and just making that decision. And so that's what I did with my podcast. And then from there, it took like three years for it to become full-time. And it worked. And so now I'm like, "Okay, I can do anything, but I can't do everything." So where do we want to wrap up? We were plugging Super Data Science.

(Joel Beasley at 00:35:53) Yeah. So I have something I'd love to—yeah. There's one last framework for agentic AI that I want to squeeze in before, you know, anything else, just to make sure that your listeners get the most out of this episode. And it's something called Model Context Protocol, MCP. And so, you know, OpenAI Agents SDK, this is providing you with a free open-source way to have an agent like your master, your orchestrator, your air traffic control agent, be able to do whatever tasks you want an agent to be able to do. OpenAI Agents SDK is for that. Crew.ai allows you to then have a crew, a team of agents working on a task. The last thing is this MCP that's an important tool for you to have in your toolkit. So MCP allows your agents or crew of agents to have access to millions of tools. And so, for example, an example that Ed and I talk about in our Agentic AI Engineering YouTube workshop is you can, for example—Google Maps has an MCP GitHub page, and there's thousands and thousands and thousands of these MCP GitHub pages. And so MCP, Model Context Protocol, it's a protocol. You can think of it as like a software equivalent of a USB where, you know, you have this USB-C connector that just plugs into all of your devices. So MCP is the same kind of idea that allows your agents to access whatever kinds of tool capabilities you wanted to have. So, for example, with the Google Maps MCP server, you can then have your agent be able to automatically convert an address into a latitude and longitude or be able to find directions from one street address to another street address. So that's something, obviously, if you had to program that in an agent yourself, that would be hugely time-consuming. But instead, that's just one example from millions of possible things, millions of possible tools that the Model Context Protocol, MCP, allows you to equip your agents with. And yeah, so it's super, super powerful.

(Jon Krohn at 00:38:03) It's an app store of capabilities for agents.

(Joel Beasley at 00:38:06) Exactly. And it allows you to still have—you know, if you want to be in a situation where you have strong guardrails, you can specify, "Okay, I have this crew of agents," and you can say, you know, "This member on the crew has access to these specific tools from these specific providers that I trust" and that kind of thing. So you can have kind of tight guardrails, or you could just see what happens and have your air traffic controller spin up whatever size crew with whatever capabilities and tools you want.

(Jon Krohn at 00:38:38) So I could say in Crew, I can define that this specific podcast production review person has access to this Dropbox folder, but the other agents don't?

(Joel Beasley at 00:38:50) I'm pretty sure you can have that level of specificity.

(Jon Krohn at 00:38:54) Yeah. I know that's depending on the Dropbox API and MCP and all that stuff for right now. But in general, you can say this member of the crew has access to Google Drive.

(Joel Beasley at 00:39:04) Exactly. And then, you know, like a general kind of guardrail, like a good kind of safety idea that you could have for yourself as an individual or an organization is to have, you know, limit what individual agents have access to and then also have somebody reviewing what they're doing. So like a security patrol agent that is tracking and making sure that agents aren't getting up to tricky things.

(Jon Krohn at 00:39:30) That's our next startup, Jon.

(Joel Beasley at 00:39:32) Exactly. And then you could end up in situations where, you know, you have agents blackmailing other agents. "Hey, I saw what you were doing on that FBI—"

(Jon Krohn at 00:39:45) CIA. It's been happening for years. Every country does it to each other.

(Joel Beasley at 00:39:49) You know? Yeah. Exactly. Yeah. So that's—you know, I just wanted to make sure that your listeners got all that in, all these kind of open-source things people could be using. That's the modern—that's the modern toolkit for people spinning up agents, yeah, for enterprises, for individuals, whatever. Looking forward to hearing what you do next, Joel, with—you're, you know, you're doing all the right things from what I can hear, with your own processes, you know, kind of the right level of trust and control.

(Jon Krohn at 00:40:20) Yeah. That's what we're trying to do. You know, we're just experimenting, playing with it, and learning. And it's so amazing how, you know, there are so many companies that are so far behind on this, and the opportunity—I told my wife, if whatever happens, something happens with, you know, deepfakes and my revenue from here goes to zero, I go, "Maybe as a consultant, I could spend the rest of my career helping companies transition into these new technologies."

(Joel Beasley at 00:40:51) Joel, absolutely, you can. I mean, that is our experience with Y Carrot. And, you know, granted, I'm—but just like you, you know, I'm in this helpful position after all these years of growing this podcast and, you know, my books and the speaking that I do. But, basically, I have this platform that has made it very easy to say—you know? So, I basically just when I'm speaking at conferences, I'll mention Y Carrot. You know, if people want to—or when I'm on a podcast like this, mention Y Carrot. And the amount of inbound flow is so crazy, and basically every conversation leads to next steps. There's so much traction and positive conversations that if I sense that there's any kind of hesitation, I just don't follow up because I have so many hot leads that do immediately want to follow up. You know?

(Jon Krohn at 00:41:42) 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.