Augment Stay Human
AI for Sales, Augmentation, and Learning | Account Executive, Ryan Burwell
do my work for me and then I'm just gonna copy and paste. How can you help me become better at what I'm trying to become good at? - Yeah, you might get some results quicker, but you're not really like playing a long-term game. You're not growing, you're not thinking, you're letting your skills at your feet. - Welcome to The Bits of Chris Show. I'm Chris Littieri, a staff AI data engineer and former trader, learning in public. Join me as we explore augmented learning, invest in irreplaceable soft skills and balance professional ambition with personal fulfillment. Whether you're an engineer, a curious learner or someone seeking your sense of enough, this show is about adapting, growing and amplifying your unique strengths. Augment, stay human. Hey folks, just before we dive into this episode, I want to apologize for the audio quality. We recorded this on a rooftop, which in hindsight was a bad idea. Out of respect for Ryan's time, I still wanted to do my best to edit this and get something out there. Won't happen again. Thanks, enjoy. - Ryan, welcome. - So happy to be here. - Who's Ryan very well? What do you do? How did we need to edit that? - Quick 30 seconds. - Yeah, so originally from Williamsburg, Virginia, former brick maker, Colonial Williamsburg, I said I met at the James Madison University. And we're-- - Which is lovely. - Back state champions, rugby, but. So Chris is actually probably the reason that I got moving up to New York. We moved up in October of 2020. So like the first couple of months, I remember you would go into work and I was kind of looking for a job, like kind of a barking gig or whatever, like looking for whatever. - But it was quick. Like it was within a month you had a sales job. - No, I don't think I started. I started at, yeah, at the beginning of January. I think that really only is two months. When you don't have any income, like-- - Right, it just goes like a long time. - Minimum wage, busboy wages from over the summer that nest egg dwindles quick. But no, end of finding jobs through like Jamie's like job board. I was like meeting with a like army recruiter, like day that I got the job offer, like I had gone in like interviewed. I don't have much else lined up. - Nothing's working. Do you want us to turn the army into a house? - I literally took the aptitude test. I'm like, okay, envisioning myself in the army. I got the offer letter for yeah. And it turns out it was a straight cold call in the offering. Grind out and I'm still in sales now. But that was kind of where you put your teeth talking about being rejected. Do whatever, go get the sale. Literally it was month to month. Get a certain number of sales. - Right, what was that like? So I don't know much about sales. So teach me about stuff. What was that cold calling experience like? The hustle hit quote is the constant rejection on there. What was that like? Like that first sort of tenure at yak. - The wild, it was a lot of fun because it was like you're surrounded by other people that are in the same boat as you. And it was like, well, gentlemen like, like you weren't, weren't like a, like boiler around, like sell like. - It was other people like entry level people who were to work at college, made a couple of months. - Not yet shaded by the cold calling industry. - Everyone's just excited to be up in New York. We're in Chelsea Market where the Oreo factory originally. It was a huge warehouse space. Kind of like, is there a nerve falls around at the same time it was a metric for a while where it was like, you have to make 150 phone calls a day. Plumbers, chiropractors, just before Uber calling car service people setting up like, the pitch basically was, hey, we're going to do all the advertising, we have really good partner network. And like, you know, the industries we're calling into, if you can get someone on the phone that's looking for a chiropractor, therapist or veterinarian, that's worth half the amount of dollars depending on the industry. So the pitch basically was, hey, it doesn't cost anything to set up, but you'll pay us like five maker phone rang. Kind of early days of voice recognition software like talking about LLMs now, like it was back for that where it would transcribe these calls and use some kind of, I'm not a. - Back when AI was just called machine learning. - Yeah, like it would basically say, okay, this person said, I'm looking for a veterinarian and pitch basically was, hey, look, if we get you a legitimate opportunity at the customer, you'll pay us a certain amount for call. And then because they use call forwarding numbers, we could prove that it came from the advertising that our company did. And then we would give the customers the recovery. Nice thing is, it's really good service. If you actually know how to, okay, I'm getting a call, I need to be nice to customers, but give you a really interesting perspective because it was, you know, some people use, the sale people could listen to the call. So I think kind of become the account manager, like after you've gone board somebody, and you're the people, like Bob's plumbing business, shout, Frankie, Frankie would be like, hi, how are you? This Frankie's plumbing service happened out there. And then we can send a flumber out for these guys, which, hello, what? And just be, you play back these calls and they're like, oh, this service isn't working for us. And it's like, guys, like, I'm not telling you how to do your job. - Right, maybe you should answer the phone like, hi, how can I help you today? - Then there'd be other people that get it really quick. So like, they would do the math. They'd be like, are you telling me I pay you $30 and $2,000, like, water heater job? Like, you're only gonna charge me $30. And like, back then, like, like, F&B, like advertising thing with like, Angie's List, it was the other one. Like, home advisor was the big one. The way they would do it is someone would go online, fill out like a little like questionnaire. Like, I'm looking for, you know, this kind of project, like in this kind of time frame. And then that company would blast out the lead to like five different companies. And then like, you as the customer would suddenly get like five phone calls from like unknown numbers. Hey, I hear you're looking for a project bombarding these and everyone would have to pay for that lead. So you'd call these guys up and it's like, no, they'll call you to scandal like you normally would. So that sounds easy enough to sell, but like, when you're like grinding out, like, you might pay 40 phone calls or you leave the week now and try to go through a receptionist. So they had a script that started where it's like, the hardest thing is sounding normal where you're talking through and being like, hey, asking them questions about their business. Like, and you would see these new kids, like, sales, like, what the... We went to JMU, so we kind of like, we're not the premier university in Virginia, but like, everyone's like friendly and like, no, that'll have a conversation with each other. - Right, this is what I love to have to play. - Yeah, once you're there for a while, you're kind of looking good at what you're stomped in and asking the right question. I still use some of that stuff today. I work for John's Controls now, and we've got a different type of sale we're selling building systems to hospitals and mechanical equipment. But it's the same kind of concept where getting to the, how can I help the customer solve their problem? The way that makes a smooth path for everyone. Like, same concepts, but like, it's very different, like, setting. That's kind of how God damn knows what the sale's at. And then, fun thing is, not sales. You see a direct line of like, "Okay, I'm gonna work hard or find new opportunity." And you see the dividend prompted. Yeah, okay. I made that happen. I'm getting rewarded like, economical stuff. That's kind of the excitement of it. And you used to get the, like, we were talking about core. Like, you get the little in the garage. - Like, it's like winning the game or like chasing the stuff. - No, you get like a purchase order emailed over to it. Like, "Oh, that's like fun." Like, you know, it kind of keeps you going. - Yeah. - Well, thanks for that 30 second. - Yeah. - How do you think sales skills, like, would you've learned on that job to be effective? Can you sort of like, describe what worked for you? Like, what do you think your sales process is? And then, or maybe like, that, but then how maybe you apply it to life or how it's kind of helped you as just like a human? - I mean, for what I do now, it's almost more specialized, where I'm not knocking on someone's hospital doors. Hey, let me replace this million dollar chiller in your basement. Like, that's not really the process. Like, I'm not cold calling, like, back in the day, a cold call, a former on the side of the road, give him to read his credit card number off for 12 minutes of like, say, high or high. - So, the Johnson control is like, you're selling big specialized equipment. It's a much more, you're showing the technical specs or like technical understanding, explaining it to the customer and how it's solved. - English major knowledge, not near, but you know enough to, like, get her stuff. You know how to code, but I know how to talk to the programmers. - Yeah, you're good at dealing with people. You take the specs from the customers and bring it. - But no, I think the most important thing is sounding confident, like being a synthetic with the customer. - Don't tell people what you're asking for. - It's not the best way to go about that. There's a cheaper, there's a better way, there's a way that it'll be, have a longer impact. - You know, understand the customer's problem and put yourself in their shoes a bit, feel what they're feeling and try to just say, "Hey, knowing what I know, what my company can provide you, is how we can solve your problem." 'Cause I've become you in talking, you can listen. - Right. - And like, I think it's really interesting. You know, I love startup industry. I see the founder is supposed to be like, "You have a good idea, do you come up with a solution?" And they also have to be the person that's able to sell. - Right. - I think there is a gap sometimes where you have a really good idea, but how it'll articulate it to somebody. It might not really think that they have a problem with, you know, maybe I'm selling a piece of software that we know is gonna save them on, they're onboarding process. But if you're not talking to the right person where that's a real pain plan, or if you're talking to someone, like trying to kind of jam a square bag into a round hole, like, it kind of doesn't really flow the right way. But I think another thing I see in, you know, people that are newer to sales, like knowing when to stop talking, like, you know, you try to, you know, like, you have in your head, like, what do you wanna, it's why my product's really, really cool. But you don't, like, pick the time to be like, "Is the person paying attention to me?" - So while you're speaking, you're listening to your conversation partner's body language and non-verbal, and then while they're speaking, you're listening to what they're saying and, like, trying to grok or put yourself in their shoes and empathize with you. - I still catch myself. The best three words in sales, like, I'll use the, like, every day, is, does that make sense? - Yeah. - Then stop talking. And that's because I'll go up, I'm like, okay, why, like, I think we're gonna replace all this piping and we'll inflate it with this level of, like, insulate and then I think it'll save us this much money. But like, if you go on, like, diadrot and don't let the other person get a word in edgewat, they're thinking about, "What am I gonna have for lunch tomorrow?" Or, "What would that meeting look like?" And they're not engaged in the conversation. Like, you just waste all your breath. - Which I think is a natural life skill is like, people aren't good listeners, usually. Well, most people think is listening, is waiting for their turn to speak. And, like, as soon as thought comes in my head with what I'm about to say, kind of start tuning you out and just keep waiting to say what I'm about to say. And what I've read, a few books on communication, like, recently, I'll have to look them up. One with Super Communicators by Charles Doig, most recently, and then there was another book about listening. And there was one point that really clicked, and it was just like, you have to trust, you'll have something to say when it's your turn to speak. But to just really focus and try to quiet your mind on what the other person's saying. When they finish, you'll have contact and understood and received so much more because you're just engaged in listening. By trusting that you'll have something to say, it'll come. But you have richer conversations by really listening. But that listening and that pausing were really magical. And one thing I liked even these interviews was, I've never heard you talk this much about, and how else, what you learned from it. - We're normally talking about cool. - But yeah, I think listening in that empathy is a really ancient skill, which kind of comes up, you know, let's talk about AI and Chitachi, and how your company or how you're using it entails. If sales is a very empathetic sort of human human, especially big-ticket items, like what role does an LLM, like Chitachi, but T-plane, how have you been using it? And what do you see as like the kind of most effective use that helps you better? - You like, I'm almost like, I think I could utilize more than I do. I'm almost like overly conscious of exposing like proprietary company information. It's, you know, it's like literally like, okay, I would love just to say, hey, chat to be deep. Here's a spreadsheet with, you know, my 50th account. We got this product that I want to do an introductory, like, hey, we already got X, Y and Z, let's add, you know, A, B and C, like, and just upload it, boom, boom, spit it out, have a nice like email blast that goes out to everybody. But I get nervous because, like, A, that can be kind of unpersonalized or like just kind of like annoying to the customer and doesn't feel natural. That's not really like how we normally go about things. And then the other thing is like, yeah, I know we have, like my co-pilot that, yeah, I can plug in, hey, can you help me find a way to soften the tone on that? You know, like, I don't use it as much like, hey, write me a first draft or things like that. And then it kind of gives them the idea of like using agents and like, I feel like the adoption for like my workflow is kind of tricky because I'm bouncing between a bunch of different, you know, estimating tools and then our sales force, like CRM, and then we have a separate or billing and then there's things that I don't have access to that separate team candles like collection. It's kind of interesting because I can use it for, you know, writing an email or I can use it for assessing like, okay, what's the better way to get this point across? But in the macro of like automating things, like I know that we'll get there, like, and I hope it does because I think that's where the, you know, vast like, there's so many inefficiencies with like just how like my company operates and I'm sure hundreds and hundreds of others like do the same, but it's like getting a lot of people are like people in power can be dinosaurs where you're like, oh, you know, more risk adverse than, you know, like, oh, the potential of like being able to, you know, use several tools to create personalized messages to like all of your accounts and like, you know, have the individual, individual salesperson, like, critique them and mold them and get them through and prove like I've been, you know, like more like software to service companies, like, like, use it like a account wrap or BDR is like, you know, using outreach, like cold outreach emails, but they can leverage AI to send like personalized video methods to like, you know, someone where it's targeted and they have like a use case, like, locked and loaded. I think that would be really like interesting to do in my industry, which is more like, it's more building and like we do a lot with, you know, reducing energy. And like, I would love to find a way, maybe it can help me with this, like find a way where I can say, hey, look, you're all the big building hospital schools, like, whatever and my territory or my mark in New York and help me come up with a strategy where we have a solution that can save thousands and thousands of dollars like the average customer. Like one of the things we do, we have a digital solution where we basically optimize the machines and the basement of like all of this building, like they're called chillers. They put just make cold water and they normally run in an array where you have like three or four machines. Some of them are extra capacity, some are redundancy. We have a software that lets those machines run autonomously and based on outside condition, like humidity, the weather to fine tune, hey, we don't need to be running these things at max after we save energy. And then we have a software platform that you can see. Here's how, here's how much energy you're saving. If you follow these recommendations, that's a hard thing to articulate somebody, even an existing customer. You're talking to someone most of the time in like the facilities department where that could be seen as like, oh, AI is going to take away my job and to say, because we use machine learning for those algorithms, like talking to the crew and kind of elevator pitch is, yeah, like what you do on a day to day basis than how you run your machine, we're going to outsource that the computer system that knows better based on like feedback was said, okay, we're going to put a layer and they're going to be recommendation to accept it would save money. And then at the end of the year, we show you like, how much money you save by following these recommendations versus like what you would have saved with the idea and that's going to go up to someone like in the finance department and say, well, why don't we do X, Y, and Z, like we should pull this back and that's kind of the play that they want to make. So go back to your original question. How do I use AI or how do I use like LLMs versus how I want to use them? I think it's kind of in there, like how can I reach out like cold reach out to these people where instead of spending all day finding the right person like crafting a specific message to them, how can I do that a thousand times over and have in the background next part of sales process the follow up and saying, oh, I need to get my email or hey, we gave you a proposal. Has there been any progress? I can see a short term. I'm sure there's companies that are already working on that. I'd basically say a sales person you oversee these different automated systems that are running in the background to make sure that things run smoothly and you intervene when you need. - So are you saying like a cold email blast to potential leads? Like that's what the LLM would be? - Yeah, that's a use case. I'm not doing now, but if I could envision like. - So it sounds a little bit like spam. - Exactly, yeah. That's why I don't do it now. - Right. - But if you can use an AI tool where it doesn't feel like where it's like. - Right, not necessarily like tricking the customer, but feeding LLM more intelligent information about like, hey, the school, they're probably using this type of equipment and we have this type of equipment. - Or saying, hey. - Let's get our pit on it automatically. - You know, go to JP Morgan's building up that big building that's going up right there. Like read all their financial statements for the last, like listen to all their earnings calls but that's like number of years. Help me find the right person that what I have to sell would be most effective for. Find me a way in to these like. - Okay, so it sounds like a different use case. So like the first use case seems like making more, like on a massive scale, like instead of you crafting an individualized email, using the LLM to craft individualized mouse potential leads. - Videos or you know. - Right. Just having the, taking some of the leg work off of you to find the specific, have trained to do that. And then the second use case seems like help me mine a bunch of publicly available data about a potential lead to figure out who to target. And that sounds like kind of, say use the LLMs as well. Like that's sort of instead of the generative side. It's more like the distillation or the summarization side of things like, hey, search through this corpus of public information and help me pinpoint something relevant which when you have a specific prompt, I'll help you a lot more than most people give it credit for right now. But the hard part is the model next to be asking the right question with the right amount of context. Don't you get what you want exactly? And I think that's part of a thing that most of society is starting to realize like, hey, I'm not good at asking for what I want clearly. And with LLMs, they're dumb robot. They need very specific instructions. Just like computer programmers now program things. Do things very specifically like you have to know exactly what problem solving. So either like as a writer or engineer, that's sort of the art of like, what's the problem solving? You know, what's my input? What do I want and output? And that's how you have to talk to the chat sheet when you can get more specific and have a richer context with what you're asking for. And especially if you can feed it more relevant data to your problem domain to sort of minimize what it's looking at, you can get really youthful things. - That's another interesting idea. Like as we look out, you know, over all these buildings, right? Like there's obviously some data set somewhere. Department of buildings, like I would imagine there's some either publicly available or privately available like, you know, instead of that information that you could run through the system and say, hey, look, based on this neighborhood or this area of Manhattan, which buildings would be the best target for a specific service. If I'm telling windows, like, show me like, you know, old buildings like in this data set that have windows that haven't been replaced that would save them tons and tons of money. What, you know, the installation of like a building that have better inflated windows, race less energy. And one of, there's a bid that we were thinking about working on for our company through NYSERDA, New York State Energy and Development Association. And basically was-- - You're allowed to share this kind of detail? - It's probably available that we're not pursuing it. Basically the public bid, NYSERDA was seeking a company to help them with that of visualization across all the buildings in New York State, which makes a lot of sense, but you're talking like huge quantities, like not just in New York, like across the whole state. In New York, once the fear had reducing carbon footprint, like, to do that you have to see which buildings are the biggest offender. You can say, okay, if we have this disparate data sets that we've plugged in for one, like, pain in glass or a terrible term, like, it's overused. But like, have someone like worked for NYSERDA say, okay, like I can see that this building, like, you know, would qualify for different redates or like qualify for a loan would save them money. Like, how do we, you know, sort of reaching out to the building owner or the organization that operates? - So you're gonna say, like, Ellen's fine, not the right tool for that. It's a good view for like, for recognizing that. But yeah, that just sounds like an interesting way to apply data to finding for leads. But, well, thank you for the walks through sales and kind of how you use it at work. What about your personal life? Like, how do you use it at home? I go on and on about my many use cases of like, what do you think of it? Like, generating sort of like content? Like, do you think of two-day eye content? Do you use Chachi Beauty for like personal things? Like, what's been interesting or fun ways you've used it? - No, so when my wife and I got married are, you know, George, right, our mutual friend, he told, he was our, our officiant. He was the one that married us. He told us like, kind of had a Chachi Beauty do a lot of the heavy lifting on the bones of my little speech or the speech that he was gonna give. Like, me and my wife wrote our own vow with that without Chachi. Yeah, well, he kind of used that like it's jumping off. - Allegedly. - Allegedly, but yeah, like in my personal life, like, I love using Chachi Beauty. So like, you have an idea. Like, I wanna know something more about that and just kind of using it as a jumping off point. Like, hey, look, like give me, you know, like, let me pull it up like the last class I used. - This is great. - Let me see. It just makes it so convenient. It's a prompt. How hot is "Strader Joe's Hop & You're a Hot?" - "Strader Joe's Hop and You're a Hot Sauce" is quite spicy. Hop and You're a Pepper's typically have a Scoville heat rating between 100,000 and 350,000 Scoville heat units. Mine goes on. - When you asked that question, what were you hoping to discover? - I wouldn't know my follow-up question is, how would that rank on the hot one scale? Like the, you know, the chicken wing eating, where would that rank as my mouth is burning from "Strader Joe's Hop and You're a Hot Sauce?" - You wanted to know how you stood on the, yeah. - Like, where would that rank? That's not something you really Google unless someone's ridden a blog post specifically out there, but, you know, Chatsubitee could say, "Yeah, based on, you know, Top and Yarrow, "I have this mini Go-Ville to interpret." - Well, so that's interesting. And you believe it's like, it's the answer, right? - It could be wrong, but that satisfies like-- - The confidence and the fact that you got a specific answer, it felt good. But the thing is, for Chatsubitee to know, the answer, someone did have to write a blog post on it for that to have been trained on. - Right. - So it's like, for these sort of specific, kind of like needle and haystack type look-up, through you're looking for like a reference piece, it might not be good to go to Chatsubitee for it because they could just make it up. Like the actual answer there, about the better off going to a trusted source. - Two, but I think for me, that's a low risk. Like, you know, I'm sure your kids are like, "Yeah, well, why is grass green?" - Right. - I added some like, chlorophyll, like-- - Yeah, something about photos to get you learned, fourth grade. - Kind of like, you know, not to undercut chlorophyll. Probably more important than the heat rating in the old hot sauce there. It's like, okay, that gives me enough, now that makes sense and it's kind of like, your brain is like-- - Okay. - I mean, don't need no specifically, I know generally-- - You can answer that sound implausible, so you are happy to believe. - Exactly. - Right. - I think the way the LMs work, right, sort of very roughly projecting the next word based on context provided and the prompt you provided. And so for a response, like, you know, Habanero, hot sauce is some number on the scale. Like that number doesn't necessarily need to be the exact number for it to be a plausible response. But at the point me to a scientific study, or where can I look up a fine study? - Well, so that's a very good question. So GT is a user-facing application that has additional layers on it besides just an LLM model. So there's different ways it can fetch information to retrieve your query. And so there are custom GPTs in the GPT store and there are other apps or models that will be called from the UI based on your prompt. So if you were going just in LLM directly, you would just be getting that's most probable text response. - But based on whatever it was trying to do. - Right, but with chat GPT, sometimes it'll go look things up on the internet for you and return late. - Retrieval augmented. - Very good. So that's one example. - I follow a great email in this letter. - But yes, there's that. And then like there's an app from consensus, I believe, is some third party that has like a research app and basically like index on research paper. And so if you were to use that app, that's basically what it does. It's like a more effective search because instead of like Google where your context might be just a search term, but now you're having like many more words that go into your search query. - Right. - And then these words are translated into numbers under the hood, which are then looked for a search against other documents or chunks of text that also have been translated into numbers. And then they're compared based on different algorithms to look at the similarity. And then those relevant chunks of text or relevant research basers are added with your prompt to an LLM for then to reason about. So like the system prompt that takes some research paper and says like using these three things respond to this query. And so that's a way to sort of reduce hallucinations, which is when like DPT just kind of makes stuff up or like kind of wings it. What do you think about the internet? Like, you know, you're not supposed to believe everything you read online anyway because like, you know, how are you trusting? So for people to expect anything less than LLM's, you can kind of the silly, more or less. But the thing with like retrieval augmented generation is nice because you can cite some of your stories. It's not perfect. I think there's still room for hallucination, but it reduces the plasticity of some of it. And then at least you have a source that you can click through due and then go read it and then determine if you trust that source. Rather than it just being sort of like a black box or it just getting you an answer. - Yeah, I'm gonna try to find another set of prompts that I've been playing with chat UBT on the kind of gets to something somewhere for that. So I'm no longer making eye contact with because I'm trying to scroll on my phone to find it. - So we were talking about I'm considered doing an MBA through NYU and I'm trying to weigh right now whether that's especially in this world, like you can use chat UBT or you can use the next generation two, three years from by the time I would graduate with that degree, how far, like how close are we to like, you know, what AGI where is that, is that not even worthwhile or is it still like a credential that like, okay, a sign of app, like, you know, be beyond even what, you know, a future where AGI is like. - You're using chat UBT to help reflect and kind of thing. - I'm trying to find a set of prompts that I use but basically I uploaded, you know, a curriculum that kind of had like overview of like, okay, this is what you're gonna learn in the course. So you're gonna learn finance, you're gonna learn marketing and then as each one of them has a snippet and I'm sure they call it the compute for the model, the author because it's big three pages of dense tech. I'm like, okay, create me a reading list. It would teach me beyond what's in this course let me see if I can find it. - So you were using GPT to sort of generate an open source curriculum that might replace what the 10 VA program was. - Exactly, but then it generated like, okay, here's a list of books that you could read. It would, you know, and then I just kind of like, using it out, I'm like, okay, give me a five book for each one of these courses. And then it's like spitting out tons and tons of stuff. And like, we were kind of talking about this before. Like, can you prompt like a chat UBT or something similar? Like, are you reading through every little thing that it's picked out? Like, or are you just kind of saying like, okay, I have this somewhere that I could go back to and I'd say, okay, here's, here's, you know, my summer reading list now is gonna be easy accounting with, you know, Johnny or whatever the book, like, recommend it. Yeah, so like, and then kind of going off I'm tangent there, but yeah, like, where does that kind of like play in? Can you use that to teach yourself? Like, can you use these models to say, hey, I want to, you know, improve my skill set, like, help me point me in the right direction. Like, we were talking about Duolingo before using these models to say, okay, I want to, the prompt is, pretend you're my French teacher and we want to have a conversation in French and we can use the mic and talk back and forth and correct me if I'm pronouncing something incorrect. It's just an interesting way, conceivably accelerate learning. You know, catch me on where I'm making mistakes, help me improve and kind of using it in that way instead of using it, okay, do my work for me and then I'm just gonna copy and paste. How can you help me become better at what I'm trying to become good at? - So, I couldn't even agree more and I think you're touching on something I am been thinking a lot about, which is LLM's, Genervic, do a lot of cool stuff but if you're sort of using it as a crunk and trying to remove the work that you're supposed to be doing from the equation, you might get some results quicker or you might get things done faster but you're not really playing a long-term game. You're not growing, you're not thinking, you're letting your skills atrophy at the expense of checking a box to get something done and that kind of level of effort shows and people can quickly sort of tell when you copy. But you get that sixth sentence when you see come across like a medium article and you're like, oh, this is clearly just like some overly verbose nonsense that was literally just copied out for what? What was the point? The person publishing this article like didn't get any benefit from that. No one's really gonna like it or engage with that piece of content 'cause it's sort of just the same median regurgitation of junk on me. Might not even be coherent. So, I've been thinking of her, I've been calling it like augmented philosophy and so I've applied it to a few different areas but the one I like now is I'm calling it being an augmented learner which is like use generative AI to supercharge your ability to think, to learn, to consume and process information. And by that I mean use it as the tool and lean into the thinking but use it to go faster. Like if you think about a calculating for doing math, like okay, yeah, you don't have to do math by hand anymore so you kind of lose that little bit of skill but now you can go do like much more complicated math problems with it because that part is shacked it away and if you can find ways for AI to fill this sort of monotonous repetitive things like searching for you, like finding the right information or starting, like getting a not a blank page to start with like, hey, I wanna learn how to braid hair. Give me like, where do I start? Well, like give me like a curriculum from easy to hard. Like give me 10 different like braids to do or something. Now I don't have to spend time looking up how to get started. I just go, okay, the first one sounds good enough to me. What is that? And then you start drilling into the learning. I think for things like writing, if you can not just do like the baseline prompt of like write an email to this lead and give me the thing but if you can sort of specify have this lead, this is the angle we're trying to go, this is their need, this is what we can offer. Now, you know, write an email that like does this, this kind of format, talk to this type of person and then once you get it, then you iterate on it. You edit it with it or you do like the last mile as well. So let me ask you this, you know, as someone that has young kids, how do you see this kind of technology when your kids are in elementary school, middle school? Like, how are they, you know, is it gonna be a odd module like sitting like listening to what the teachers get planning to the kids with a vision like, oh, someone's not paying attention. Like I'm gonna reach out be like a sifting teacher or is it gonna be something like, oh, I'm stuck on how to learn this problem. One of the big things, like my mom's a kid and one of the big things that she always pushes on is, hey, look, when kids are learning how to read, it's such a inflection point where like the kids that like learn how to read by themselves, they have the tool set where their knowledge is gonna increase rapidly. They can, they're, like, a lot of times those kids kind of get like, okay, you're gonna go, you read your chapter books over here, like read whatever you wanna read, like we're gonna help the kids that still like need their phonetics sorted out. And then they're the ones they suddenly get such a big head star or like there's that divergence where they can, you know, they're reading about like, they're interested in snakes, they're to read about snakes, like what they're interested in, they can use books as a path to deepen their knowledge where it's like kids that are behind like, they're struggling to tread water and like. - So I have two things to follow up. Like first, that sort of inflection point you talk about, I think is the same thing with using AI to learn. Like you're turbocharging the way you can consume and process and understand information by working with LLM or working with some sort of AI model to take in information, summarize and ask followup questions on the points that are confusing to you and get a very tailored personal sort of approach to learning where if you take a course, you're getting sort of like just in case information, like you're learning everything that you might not need to know, but when you can focus your approach to learning on a project on a problem you have, ask specific questions for like just in time type learning, you are learning things immediately applicable to a problem you have, which means you're also going to apply it right away and use it and like it's going to stick much more. Do you get this sort of personalized like quick go deep on something you have and really move the needle in an area with gender value and kind of cuts to what's relevant quickly and keep moving your edge of knowledge faster. So when it comes to education, you brought up a great point which is like we saw this with my daughter and she started a year in one school and is now ending in a different one. She wasn't the bottom causing disruptions or needing a lot of attention. She wasn't the top able to go off on her own yet. And so there wasn't enough resources to sort of tend to where she wasn't helped move her along. Whereas if you have a gender-to-vei assistant in the classroom, they could give kids one on one time and say, oh, like you're struggling with words that make this sound or like you haven't figured out the difference between B&D and they can find your individual sort of hiccups and give you that just-in-time specific looks to get over your hump, continue on your journey. - Just imagine like a company coming in like, hey, every kid has like an iPad set up on their desk. And instead of like just punching in like schoolwork, it's just observing the student like, oh, I can see you based on like your facial expression that that concept's not landing or like you were paying attention like when we're going over like the letter, like the sound and S makes. - So as dystopian as it might sound to have like cameras on kids, for a little kids in schools that are under a resource, it's gonna be huge because kids with attention type disorders are not even disorders, but just like tend to be more of the wandering type. There's kids that benefit from redirection and more attention on the work they're doing. And so if they can have somebody or something who's watching them to then like sort of summarize to the teacher, like which kids are falling behind. It's not like they're really taking quizzes in kindergarten or first grade. And that's just not a scalable. But if you can have this AI type of test personalized with giving the teacher reports and the teacher can tune where they give the extra attention or get a better map of the land. Like I think similar to any field like AI and LLMs are like a way to scale yourself, give more personalized, just be able to process and consume more information quick. And I think that's where this augmented philosophy comes from is like you start giving yourself superpowers. If you wield this like the tool that it is rather than getting lost in the hype of its magic or being sort of a negative doomsday or type of like I can't do anything yet or it's trash or make stuff up. We're not even giving it the time of day. I think that's wrong. I think both of them are kind of wrong. And it's like, no, no, like this is happening. Every company and every person is talking about it. There's a lot of money working on the better and finding applications for it. And the coolest things I found are the most interesting things I found is like how did it work? Like literally you can push the edge of your knowledge so fast now. I've had an iOS app, I've wanted to build. I had a quick conversation with it on the train ride into New York this week. And like I have the outline of how to do it. And I have the steps of how to set up my environment. Like I'm not an iOS developer. So like really foreign to me. But I don't need to know that much of iOS development because it's really just sort of like syntax. Having the programming background, you don't need to learn any more languages. Like learn one really well. Maybe two, but you don't need to learn other ones now. You can just sort of work with co-pilot or work with the GPTs and generate that stuff for you. And then you know enough of like what questions ask, what pieces need to fit in. And if you don't, you can sort of get a quicker picture. Well, and also like I remember this is years ago, but wanting to do something similar and like hitting a wall when you run into like Apple's style guide for what gets accepted into the app store. This was going back probably 10 years ago. That was such a, and I was a design major. So a lot of the stuff made sense to me. But if the pages and pages are acceptable guidelines of like, okay, you have to have this kind of format to make the app accepted into the store. And that was just like, okay, it's overwhelming. Like you have to have to read through it all. Whereas now you could say, give me some script. It would conform to like, you know, this set of documents. This, the over outline of like what I'm looking for. And one of the other things you've had me thinking about, and you probably know a lot more contextually. So I hear like interviews with like the BAM Altman or like these guys that are on the forefront of this stuff and a lot of times they talk about the bottleneck being compute and like compute and energy kind of go hand in hand. And my company does a little bit like we do the cooling for data and our stock is going up because there's huge demand for these data farms, you know, help build out this localized compute. I don't think the leaders of my company really know what's actually going on there. But how do you think about what makes something require a lot of use? What are some things we just talked about? What's a heavier lift? Are there different architectures that make things more efficient? Or like what draws more compute demand? Right, bigger models. Okay, so it's the more data going in. Right, so basically what these models are, are like giant equations, billions of variables that have different weights and are computed every time like an inference needs to be made. Like basically these numbers go in and a whole bunch of math hat. And all those math are like cycles in the GPU somewhere. A computer, basically a calculator and like that's bit. Those are binding like zeros and one. One's zeros. Right, so that's all happening under the hood. Like everything on top of it, all these programming language we have are different like abstractions. All ultimately get down compiled to machine code, which literally tells the hardware what to do. Like certain instructions on the chip, what to do. Like how to move different bits between different pieces on the hardware. And so the bigger you make these models, the more numbers need to be crying. The more things that you need to be manipulated, the more you need to keep in memory while you're processing all of it. And right now what we've seen with GPT general or just these large language models is the bigger you make them, very roughly the better result you get. But that's sort of the path we're on is like, let's just keep making them bigger. Like more data, more parameters. And that's kind of work. No one really knows how much longer, bigger for better will be. It could be further, it could be not like we don't know. So that's sort of the direction going. And one thing that's making it hard right now is these things are very like the lead time to getting a new factory to produce these type of chips online is very long. So there's just a huge bottle, like buying videos, just like become the most, one of the wealthiest companies in the world is because they were the only ones who were positioned to make it. And it all started with gaming. And then it crept up a bit with cryptocurrency mining 'cause that again is very heavy math equations based on what they're solving. And now these GPUs for crypto is kind of also being applied to training and inferencing on these large language models. You could do in these computations in the equation. 'Cause all model is is this equation with parameter set. And then you have to feed in your input, which is always converted to numbers, goes through the equation, numbers come out, and then it's reconverted again back text, video, audio, image, whatever. So that's really where the compute need comes from. - Which, now I'm very good at seeing something very high level that I don't understand like how it actually works. - Ryan, thank you for the rooftop recording. - Right, thank you so much, we'll talk to you. - Thanks buddy, it's fun.