(187) The OpenAI Religion: 90% of AI Use Will Shock You - YouTube https://www.youtube.com/watch?v=uUtDqtqnQ6E

Transcript: (00:00) Where do we think AI will have the most dislocative effect? And I think personal health is the biggest one. That includes mental health. It is like this basket of companionship more than anything. I've been reading up on some of these situations where people have had Jack GPT level psychosis. That scared me. (00:21) natural insane extension of this would be like let's rewire the human brain to optimize it for AI AI revenue research prompts tech growth record comparisons H-1B visa research track payback in tech I'm a normie and you are not Hi everybody welcome to topline it's Sam Jacobs CEO Pavilion I'm joined by Aid Zaman the CEO of sales talent agency and today's guest star co-host Kyle Norton the CR Oro of owner. (00:50) com and the host of the revenue leadership podcast. Today we're going to dive into three topics. The first chat GPT usage and how it's evolved. How is everybody using AI now that we are over 3 years into the AI revolution? Second, lovable replet and all of the vibe coding vibe coding platforms. Their search traffic has dropped. (01:15) Is it the end of vibe coding as we know it? And finally, what's the problem with Europe? Why can't anybody build a meaningful company in Europe? and what is everybody in Europe's problem. Those are the three topics we're gonna dive into today on Top Line. I like how we're taking shots at Europe, but we also mentioned lovable off the top. So, yeah, exactly. Well, there's some exceptions. It's not a perfect, you know, we're just going for clicks. We're going for views. We're going for long-term engagement. (01:37) Um, let's dive into it. So, Austin, let's kick us off. What's happening with Chad GBT? So the National Bureau of Economic Research put out a really comprehensive and really interesting uh report on how people are using CH GPT and how that use has evolved over the last 12 months. So we're going to talk about this in a second. (01:59) Before we do that, what we're going to do is I'm going to share some things with you guys and see if you can guess correctly or not how this use has evolved. So let's do a couple of those and then we'll jump into the actual conversation. The first one is work versus nonwork usage. (02:19) So for context, this study only looks at the use of people that have a personal chat GPT account, but a lot of people, myself included, use our personal accounts for work rellated tasks. So there's work and nonwork related tasks. In June 2024, Sam nonwork was 53% of the total use of chart GPT and work hence was 47%. In June 2025, do you think nonwork has become 23% has stayed flat at 53% or increased dramatically to 73%. (02:55) I was going to say 75. uh some kind of partoal law much higher usage of personal than work you are right it is 73%. Why do you think that like why why were you so sure that it's so much high I was surprised by this because I think the work use cases are becoming more opaque as the quality issues with AI become more like more visible. (03:26) I think as more people use AI for work, they understand that there's a lot of keying that needs to be done and it's actually not that easy to build AI workflows. But I think on the other hand, we're going through our day and every time we have a question, every time we want to talk about a cultural matter, every time we want to just brainstorm on something, chat GPT is right there. (03:45) So, I find myself talking to it about any number of insane, stupid personal things, but and I do use it for work, but not nearly as much as I use it for personal use. And I uh I I knew where you were going with that question, and I reflected that uh my personal patterns were probably representative of the global human population. You were right in that call. (04:04) I was the right call. Um, it's interesting because I've noticed my personal use has dropped ever since I've been reading up on some of these situations where people have had CH GPT level psychosis. That scared me. Like I'm actually using it a little bit less for personal use. Since I read those, they were seemingly normal people and then weird things started happening. (04:28) But we'll talk about I'm gonna we're I'm gonna, by the way, before we go to the next topic, but keep asking your questions, but I've got a final question. I've got a surprise question for the two of you. Um Kyle, so there are three use cases that make up the top three use cases for chat GPT. One is writing, the other is seeking information search and the third is practical guidance. What Sam was just talking about. (04:52) We know that nonwork is the number one use case and so practical guidance obviously has to be number one. That leaves seeking information and writing. Of those two, which do you think is the higher use case? Seeking information would be my guess? Why? Uh I think for many it's it's replaced Google as your homepage for getting answers. (05:22) Certainly has for me and I think a lot of people who are habitual uh daily users that it's just our instinct now. Even though that is the same for me, I I use it for search more than anything else, search and research. But writing trumps seeking information as a use case. Yes. Very interesting. I I was listening to an A16Z podcast recently and they were talking about within the practical guidance category, companionship is actually like the the top use case of all the things people use chat GPT for. (05:58) It's it is like this basket of companionship more more than anything. Uh as in like personal companion. So as people are becoming lonier, they're turning to chi to kind of fill that void and personal advice using it as their Yeah. Using it as their pseudootherapist or uh executive coach is a is a massive use case apparently. (06:25) Uh Olivia Moore, who's one of the AI investors there, is a good follow on Twitter, and a bunch of the uh bunch of their AI episodes are great. But uh I I was shocked by how big this this space is because it's not really a use case I I use it for, but it seems like for a bunch of the normie population, it is it is like their biggest their biggest use case. That's me. (06:51) That's me. I think it's really interesting how uh people are getting lonely and lonely and this this obviously helps with that. We haven't found like a great um implementation of it as a product for an exec coach or a therapist yet, but like there is this this massive thing like executive coaches are only really available to people that make a lot of money, but they could be most useful to people that don't make a lot of money or even like a mentorship type of AI. Um, but no one's really released. I haven't come across it (07:23) because I don't think you need I don't think you need much more than what is in the model. I don't think you need to to put a wrapper around it. You can use a mixture of panel or mixture of excerpts approach or um just with like decent prompt engineering. The the the post training they've done for this specific use case is clearly incredible. (07:45) Like out of the box, it functions really really well. Let's look at work related use only Sam. So again writing, practical guidance and seeking information. Which do you think is number one when it comes to work? Writing. You are right. Writing is 41.8% of the way that people use it at work. And of that 41.8% 34% is our team to chip to do the writing. (08:17) And so with writing you can say you can ask it for advice and feedback or you can just say write me this email. The greater majority of people are just abdicating to AI and say can you just write me this email? Can you write me this document? And so we see this word AI work slob. It's it's we've all seen those things right the email that you just look at and instinctively you don't want to read it. The document you get that's 18 pages that says nothing really. That's really what's happening. like the majority of (08:45) people are just like completely abdicating simple prompt copy paste send. It's why uh it's why Fixer is so popular. Even though you know for me as somebody that uh I consider myself a decent writer, I don't actually use the the autodrafting functionality within Fixer. But uh but I but when I imagine the millions of people that aren't naturally good writers, I imagine that's a massive market. As somebody was saying over the weekend, a real estate agent going back and forth on one to show up for, you know, the viewing doesn't, you know, (09:18) that might save hours and hours and hours of time. So, I believe that. Kyle, how does this map with your own use or what you've seen? Maps pretty well. I I Chat GBT/Claude I use as a co-writing a co-pilot writer all the time. the newsletter for the podcast, for example. (09:45) I've spent hours and hours on the prompt that I use to turn the transcript into newsletter content, and I've got a whole workflow that I that I like. anything that I do repeatedly uh performance reviews for example I'll I'll spin up like a quick uh and for that was chat GPT but a quick project I'll uh use prompt cowboy to help me create a halfdecent prompt that I'll do a hand edit on um so why are you using prompt cowboy versus just asking it to write you the prompt and then just copy pasting that prompt I do both. Um, prompt cowboy I find I can be like super lazy with. (10:24) So I just I I use super whisper. I just hit the function key. I like blather away and then it gives me like a much better format. And oftent times I'll I'll take that and then put it into chatt and like have it uh make some edits for me. Some of these things I I don't know they just become habit and you do it this way. (10:48) I was I was doing all of my metaprompting with chat GBT and now it's sort of split. Certain projects will flow from one model to the other depending on like what I what what I think about the outputs and as the models change. So I find this to be one of the challenges that we're dealing with as an organization cuz I think we've been pretty forward thinking in terms of using all the tools and trying to see what they're good at but what they're good at keeps changing. (11:12) So an example of this um is that Perplexity for a while was a better deep research platform for this thing that we do. Um and up till maybe a few months ago it was better by a a margin. Now it's no longer the better place but the comfort everybody has is using Proplexity. They're like I like using Proplexity. (11:32) It's not that bad. Like it's only a marginal difference here. And so I think a lot about in our organizations, it's not just getting people to use AI, but it's having that like mental flexibility to say, I'm also going to keep switching between these things as something gets better at something else. (11:51) Have you noticed this type of the stickiness becoming a problem? And how does one create not just this energy to use AI, but this dexterity to keep switching things? Let me question the core premise of what you just said. I think the advice that you would give somebody to keep switching tools because one might be 1% better is terrible advice. What if it's 10 20% better? It's still terrible advice. (12:14) Really? I think 20% is a good margin of like benefit to like switch something. I I think you're I think this I I think this is the dog the tail wagging the dog. I have to say asset I think that we're lost in the woods and you know we're solving for using the best AI tool that is marginally better than the other versus getting solving the problem for the client that we're working on. It's a good point. (12:40) I think I don't think human brains are wired. I mean, maybe listen, it's like the the the natural insane extension of this would be like, let's rewire the human brain to optimize it for AI because at any given time there might be a new tool that's emerging that's slightly better than the one that we've just spent the last six months training ourselves to use. (13:01) But I I think that we're I was just trying to think about this like would you teach if it was if it was 2000 would you be teaching your team like listen guys I want you to run searches on Alta Vista Los Yahoo and Google and then we're going to compare all five results every time. (13:23) It's like what are you let's let's use what we have a winner will emerge we don't necessarily need to be the ones that are arbitrating the winner. we can leave that to the market to a certain extent and when the new tool emerges and everybody consolidates around it then we can all pick the winner as opposed to this is an interesting point so clicking into what you're saying there's a certain amount of improvement on what you're used to that something has to present for switching to make sense what type of improvement do you have to see so if I'm using something it's good enough but something is 100% better then (13:52) I switch obviously but what's the what's the minimum level of improvement at which point switching is the right thing to do. I think if I is there a world in which I would have become a math major and then become an AI researcher and then make $3.5 billion working for Zuck maybe. But uh in my current uh world where I haven't really worked on statistics in a very long time although I did I did okay when I would when I took statistics class my point is the the the algorithm would not just include how much better one tool is (14:21) versus the other but it would include some kind of uh overall like you know alpha or beta or some kind of overall estimate of the overall volatility of the market in which that tool was operating and the rate of change of that market and then the half-life or decay rate of you know the permanence the moat of any of any poss possible tool. (14:39) My point is like one tool could be way way better in just you know in six in two days and you look at it and you say okay everybody shift to this tool and then there's another tool that and now you're causing all of this thrash within the organization. So I would say if you incorporate if I'm going to assume that this tool the new tool is going to maintain its advantage over the existing tool and I would say maybe it has to have had maintain that advantage over like 6 to9 months otherwise I'm just everybody's running around chasing tools and not solving problems for the client and then I would say you know 30% better (15:13) like you know I I think 10 to 20% is really hard to measure and again this is this is probably the nature of like a lot of work that business schools should be doing right now. Like we need the new quarter, we need some new organizational theorist uh to help us understand what is uh the the the way that we should be thinking about using tools versus driving outcomes and productivity because I think right now uh to the point of what we're talking about, we're all so terrified that we're losing productivity by not AI optimizing (15:45) our organizations and it's still not immediately clear that that's what's happening. But what we definitely are doing is causing all of these people within the organization to run around like chickens with their heads cut off trying out every new tool. And and let me answer this a slightly different way. (16:04) I I think if there is a workflow that is so important to your business that it could be improved that you could see material business ROI through a 20% or 10% model improvement, then that has to be something that you are centrally managing. (16:23) And I wouldn't be advocating that it's a process that you're just letting your teams decide on how they do in their own workflow. And so if this is like candidate research and if candidate research changed your close rates or placement rates by 2x between good and good and not good, well then you should be building an internal tool that is model agnostic and in the background you could be changing the model. (16:47) And so, you know, we hired this GTMAI lead and this is one of the things that he's doing using prompt layer is being able to like test and iterate our core prompts in the background with proper evals to see, okay, if we change these prompts in this different way or we add this additional context, we're going to get this type of lift. So, that the end the end output that goes to the rep is just going to get better and better over time. (17:18) But the person that should be making those decisions is your GTMAI lead or rev and the rep there's no difference. They're just using the same interface. The front end for the rep is the same. The back end is what you guys keep switching around. If it's so important that a uh 20% improvement on the model actually matters for the business because I I I agree with Sam's point. The the the cognitive switching cost is significant. (17:43) Yeah, you only get a certain number of like change credits every month or every quarter. You can only do ask people to change so many things. And if you're asking them to switch from perplexity to chat deep research, you know, you're spending some of your change credits and now you can't have them change the way that we do candidate discovery or outreach or whatever it is. (18:02) And so I would be thoughtful about only asking people to make change when you really see the ROI there. And if there's something that you see yourself iterating in this way, that should really be a centralized process versus something that's just left to the team. I love I just love what Kyle said about the the centrality of the workflow. (18:20) If it's a core strategic workflow that it should be always optimized, but if it's not, we kind of have to live with the fact like, hey, maybe they're using LIOS for 6 months even though Google's been around. But, you know, a year later, Google is clearly the winner. you know, uh, page rank has clearly clearly won as like the dominant algorithm. Okay, now we can all switch to Google. Google won. (18:38) Okay, this makes a lot of sense. Kyle, as you've been at the forefront of implementing AI and your go to market team, what's new and diff what's new or something that has changed over the last few months in terms of how you're using this and some benefits that you're getting out of it? Anything new and interesting that's popped up that wasn't around or wasn't possible before? um or are we a neutralization phase where like you already have the core stuff built out and now everything is improving incrementally. So I don't think this is I don't think (19:13) this is driven by model changes or where the market is going. This is sort of more I and I think this is the same for most companies. you know what you're doing different now than six months ago is really driven by like where are you on your AI adoption journey where are you on that maturity curve and so for us we just hired we just hired YG and so one of the things that he's doing is trying to get the right infrastructure in place so we've all been implementing AI pretty willy-nilly in a bunch of different ways from a personal productivity standpoint there's the the centralized stuff like Momentum and (19:51) Avara and one mind that is uh centrally deployed But what the reps are doing is a good example of where it is like pretty all over the place and we need the right infrastructure to do the evals and the testing and the iteration in a thoughtful way. And so we're building some stuff. (20:17) We're building some new interesting use cases um without a doubt. But we've also taken a bit of a a more holistic view to look at do we have the right underlying infrastructure. You know, we've in-house built a whole bunch of code to do scraping and enrichment and and uh lead scoring and what have you. But now the question is like, well, should we honestly just replplatform to clay? Should we should we look at we're thinking about building our own uh sales tech stack and so like now are we going out and and you know what is the foundation that we're building these things on. So we're we're actually (20:53) looking at the the building blocks today more than say more about building your own sales tech stack. What does that mean? Well, our business is unique in that we sell to SMBs. We sell to mom and pop restaurant owners that um aren't on their email all that much. (21:18) And the sales loft outreach Apollo of the world are all built for selling to people who buy by email. And 90% of our meetings are booked over the phone, but there's not really a great sales tech stack for that. Um, sales loft is even if you just look at the UX of the of where the rep lives. It's all about the email, but but it's not really a UX that is meant for having a live conversation with a customer, where the information lives and um how the screen is built and what information you could populate. (21:52) And so I believe there's a future and if you're a founder and you want to build this, we'll be your first customer and I'll uh give you my whole mental road map. Some of it is actually in PRDS because I've been trying to convince people to build it. But um I want to build something that's native for people who sell via the phone that is like AI native. And I've looked around a bunch. (22:11) There's not really anything great. And so we might try to build this as like a skin on top of Salesforce or through Chrome extensions plus like moving all of the cold emailing into the background. Um, which which we're in the process of doing. So we're thinking about like reinventing the whole like one of the biggest disservice that these email automation companies did was this propaganda that crawling people is just it works. (22:40) So it doesn't work to such an extent that everything all our workflows are now built around email, email, email email. Like the fact that you're saying this is as a salesperson just sad that you don't have a product that prioritizes picking the phone up and calling people and and enabling that at scale, efficiently, effectively, etc. (22:58) It just feels very very sad to hear that. Yeah, I wouldn't blame the I wouldn't blame the tech platforms for it. I think once emailing got a lot easier, you had a whole bunch of reps that were like, I can live behind my keyboard and not have to do the scary hard thing of making a cold call. (23:20) And I I actually think it was more driven as a reaction to the market than it was these companies being like, we're going to just build email products because that's what we want. A little bit of propaganda like here's the response trade over. Wasn't that um stop before we go I want here's here's what I here's my question. Um can I ask you I want both of you to open up your chat GPT instance and I want you to open up the left braille and we're going to read and as long as it's not deeply personal um you know like um terrible test results or something uh we're going to read the last 13 titles of the GPT (23:57) conversations that we've had. I'll go first so that everybody can feel approve postpromotion and you can ask questions about any of these. Approve post promotion NFL kicker accuracy increase Buffalo Bill name origin scaling smarter outline chess mind training techniques stable coin value and use book summary request send book via Bluetooth Kindle charging light issue VS1 clarity ranking PDF order form summary AI bubble sources and evidence itch relief in Spain you're like the average user Ah, that's hilarious. Which was the most interesting response out of those 13? (24:47) Uh, I I well, you know, the most interesting, the most in-depth was the stable coin discussion cuz I I just really didn't I was like, give me the primer cuz we need to get up to speed on this stuff. And then I was happy to know that somebody while yesterday while we were watching football was saying what is the origin of the name Buffalo Bills and I was like well sounds like Buffalo Bill Cody let me guess that that's what it was and I was correct. (25:13) Ah okay next go then we can see cars mine is hyperscaler CCO search criteria model releases and products AI revenue research prompt tech growth record comparisons H1B visa research cat payback in tech oh my god I'm a normie and you are not I am just thinking of one thing these days uh gates software revelation HubSpot growth and market Palunteer research time conversion UK to EST drug management influence uh startup exit rates in the US. You are a very boring person. (25:51) You are a one-shot my All right, Kyle, your turn. So, you will quickly uh discover that I took a bad spill on my bike last week and I got stitches under stitches under my chin. Uh because I have wound healing advice, insurance plan comparison, shoulder separation tests, check for ticks, cuz we found a tick on my daughter yesterday. Awesome. Oh Yeah. Uh sales training recommendations. (26:23) Bermuda Triangle, myth versus science. Uh, that's that's a that's a slippery one. We could go down that rabbit hole. It just changed the whole docket. Um, PLG and salesled B2B text change request. Uh, learning style analysis, signature identification tips. I don't even know what that one is. (26:50) Um, building corporate athlete mindset. Uh, high yield credit spreads. There we go. I thought that was a very fun exercise and I appreciate both. I think that was fascinating participating. How do you feel that two of your creators are just falling all over the place? He fell off a bike. I fell down the stairs twice. What's going on here? I don't feel like you're my creators, just my friends. (27:14) And I feel like, well, I mean, honestly, that's where where do we think uh AI will have the most uh dislocative effect? And I think personal health is a bit is the biggest one. And and and I think that includes mental health. I think to Kyle's point earlier that uh people are using this for therapy. (27:31) Therapists are using AI to uh model responses when they're actually in sessions with patients. And I think that there's uh lots and lots of people that are using it for medical advice. And this is where, you know, this is where the great danger of of AI comes into play, which is that the the key feature of LLMs is not accuracy, but persuasiveness. (27:49) And so you just have to make sure that you're double checking some of these things because it will definitely make up some stuff. Sometimes it'll be generally correct, but generally correct typically is uh you know, is this thing allergic to uh to my dog? Uh it doesn't you don't need generally correct. You need specifically correct so that you don't kill your dog. (28:10) Sher has um Sher done some tests and she found that there's these these growths that she has. Um and so she gets the report before she sees the doctor. She puts it into Chach GPD. Chach GPD makes her feel like this this lady is about to die. And so she was very concerned, very upset, like pulled me out of the office. (28:27) We were like, "Holy the world's falling apart." I was like, "One second, let's just give it a breather. Like you're going to the doctor tomorrow, have a conversation with the doctor." Goes to the doctor. Doctor's like, "Yeah, there's nothing to worry about. This and that. This is not a problem." And so I've seen it do that. (28:44) And at the same time, I've seen, you know, I've told the story of how my mom would take recordings from her conversations with the doctor, upload it, and be able to go deeper and then understand the thing a lot more. So, I've seen both ends of it. It's so powerful, but it can also cause a lot of anxiety and be very wrong. Yeah. (29:02) I I think one of the most important changes it with respect to health is, you know, certainly in the Canadian health care system, you really have to own your own healthcare outcomes. like the the system just wants to get you out. There's not there's a very low quality of care overall and unless and and so you don't get things explained you that well. (29:24) You don't get the in-depth um understanding of what's going on. And so chatbt is such a valuable tool to be able to take what you hear from doctors, take what you see online, synthesize it, be able to ask a hundred questions. Like I've asked a hundred questions about all the like cuts and scrapes I have all over me, like sending pictures. I know exactly what to be doing at this particular stage of the healing process. And you'd never get that from a doctor. (29:48) And so I think it's in Canada, you go to the doctor, they treat you like your IQ is 20. Like everyone's IQ is just 20 and that's how they're giving you information. Like when I when I get hurt, I don't even go cuz like most of my injuries are in areas I've already have had injuries before and I know what they're going to say and it's going to be so surface level and basic that you're just like you start not wanting to go to the doctor over here. That's why you're doing that H-1B research. (30:12) Yeah, exactly. Exactly. Well, folks, another great episode of Topline. Uh, smash the subscribe button on YouTube. Give us five stars where you listen to your podcast and we'll talk to you next week. Bye everyone.