Prompt Engineering for Better AI with Lior Weinstein

In This Episode

• [03:00] – Lior Weinstein shares his background, growing up in Israel, a tech hub, and starting his first business at 14.
• [07:10] – Lior recounts his worst business venture, buying a spa that failed despite his efforts and the involvement of a friend.
• [15:04] – Lior explains how he initially started conducting AI workshops after ChatGPT’s release to help friends understand AI.
• [18:19] – Lior shares examples from his workshops that help people understand AI, such as explaining how large language models work and the concept of temperature.
• [27:01] – Lior provides practical tips for using AI, such as talking to AI like a person and giving it full context.
• [33:31] – Lior discusses the potential of AI to replace certain SaaS solutions and the shift from buying to building AI capabilities.
• [41:38] – Lior talks about the future of AI, emphasizing the importance of agents with a mission and tools to execute it.
• [56:29] – Lior emphasizes the need for businesses to use AI responsibly and avoid creating artificial intimacy with customers.

Jump to Links and Resources

Lior, it’s so great to have you on the show. 

Great to be here. Thank you. 

We know each other through the masterminds that we’re in, and Genius Network is the one that pops to mind. I think there might be another mastermind that we had in common as well. Did we?

All these things blend at this point. I’m not sure.

Yeah. That’s right. Anyway, I love Genius Network. I’ve had some of my friends from Genius Network on this show, and now you are another one of my friends from Genius Network. So I’m glad to have you here.

Thank you. Thank you, Joe Polish.

Joe is an amazing and amazing marketer who has also been on the show. If our listeners haven’t heard that episode, please check out the Joe Polish episode on the show. So, Lior, why don’t you provide a little backstory about your rise to tech stardom? What were your humble beginnings?

So, I grew up in Israel, and as you know, Israel was a big tech hub. Building tech companies is practically a support for the country. I was an entrepreneur from a very early age. Got into online things, BBSs before?

Did you know that I also started my own BBs when I was 12 or 13 years old?

I have the opportunity to build startups, exit a few, and buy and flip online businesses, improving them through technology or better marketing.

I love those connections. We just had it in one of my as well. When you say that, you clock somebody, and they instantly have these memories.

I wrote the BBS myself, too. I wrote it in basic from scratch.

I got a chance to start work on my first business, like an actual limited liability company, when I was 14, and build actual commercial software by the time I was 17, so pretty early on in the software world. And at the same time, I grew up in marketing. Twenty years ago, everybody called us affiliate marketers, and then over the years, those names changed to, you know, digital marketers, performance marketers and so on. And those were my growth tracks at the same time. So, I kind of grew up in the technology space, and the growth space is pretty much parallel. 

I translated that to businesses that I built, and I took those positions for all of them. So either way, the technology guy, kind of the CTO, or the product guy, integrates what customers want and how to build, and then, of course, the growth and the marketing. So I kind of butterflied between those positions and activities. Now, I have a chance to build a few startups, exit a couple, buy quite a few online businesses and flip them. Kind of make them better and flip them. Sometimes, it’s making them better by changing the technology. Sometimes, it’s done by doing better marketing and growth. I’m utterly unemployable. So this is my life. 

For me, too. How many businesses have you bought and flipped?

Oh, well, or a couple of dozen and as small as like $70,000 forum, to four or $5 million of E-commerce software, that kind of companies. 

So, for the $70,000 online forum, how much did you sell it for?

That one was actually great because I bought it for 70. I made, I think, like 150 on it, and then I think I sold it also for 150, and all of that was in a span of 18 months, and doing a bunch more in parallel. It was one of those opportunistic, nice kind of cash flips.

What was the secret to making it perform much better than it had with the previous owner? 

Updating the tech, adding better features, improving speed, and implementing automatic reminders for users to log in and message—all of this boosts traffic.

So this was a period where I did a lot of this between 2014 and 2017, and you still had this lag. Of all the websites that were pretty good on traffic and SEO rankings, they never updated their stack, right? I think now it’s going to be fairly rare to find. So that was one of those examples. I think it was built on the BB forum or one of those, and updating the tech, adding better features, making it faster, and automatic reminders for people to log in and message stuff like that just really increases the traffic. Similar thing. I bought a network of kids’ websites that the original founder built, like Dreamweaver.

I’m also dating myself and everybody who understands what that means, and with static files and Flash games, he had great content and traffic. And that’s the opportunity to go in, update the designs, add, you know, WordPress at the time, make everything link better, and make new content strategies, HTML5 games instead of Flash, you know. So that was a pretty good playbook for those few years to just get old websites that had a lot of traffic but didn’t change your stack and make it better modern to increase engagements. 

What was the worst kind of mistake, flop, buy and flip, that you made?

So, I would say my worst one is an offline business; I’ve seen this years later. I’ve met a bunch of other entrepreneurs. Everybody has that when you have a lot of successes you think you have, like The Midas Touch, like, “Oh, sure, I can do any business work.” And I ended up buying a spa, like an actual offline business. And it was opportunistic, with the friend that could have managed it, and that kind of a story. And it was terrible, and it was so bad, I think it was like I bought it for 100 or 120,000.

Did you end up ruining a friendship over it? 

No, I pride myself on that. I don’t have any bridges burned in my life now that I’m aware of. But yeah, we ran for nine months or a year, and I just closed it, meaning I didn’t even bother selling it. I literally just shut it down. And it taught me a lot. Also, one of the things that taught me is that I am not an offline kind of guy; that’s not my skill set, intuition, and passion. It’s not that I don’t get some people’s family members. They like the store, right? They like being in a space with people coming in and smiling and kind of being the host. And that’s just not my stick.

That was my worst flop in terms of nothing going right from the acquisition to the operation to the shutdown. And then when I had another interesting one. So when I had the kids’ websites, somebody was selling, and I think it was like coolmath.com or coolmath.net, which is a good domain. It was like 20 grand or something, and it was 10 or 20 again. So we bought it, and I completely just bought the domain, got it over to the registrar, and didn’t think twice. 

AI is not just another tool—it’s changing the way businesses operate at a fundamental level. Share on X

One day, probably six or eight months later, I logged into my registrar for something else and saw a notice around that domain. I’m like, “Hmm, what is this?” So there’s a company that has that trademark. So it’s like, cool math, cool math games, something like that company, I think, out of New York, and they basically sent a letter to ICANN about trademark violation or something like that, basically claiming the domain. Now, the contact details, once we made the domain change, I never updated them. I never updated them to my email or anything, so I never got any of these notices. 

For context, if anybody’s listening to this and you have the same thing, it’s actually a very easy thing to fight, right? You’re like, well, you can explain the transaction, right? Plus, there is nothing on the content itself of infringing or anything like that. But because they sent multiple letters and multiple notices, nobody answered. They already had a judgment, and it was like this judgment in Switzerland, or something like an arbitrator, and basically, I was at the end of it, meaning, though, like this domain is moving on this date.

So, I lost the entire thing. So we paid 10/20 grand, forgot to update the “who is” contact records, they submitted, and they won, and I got it taken from the domain. So that was like, “Whoa. That was an expensive lesson.” So, anybody here, if you buy domains, and I bought plenty of cents, update the contact details on the domain. So that was one of those painful moments for sure.

Wow. Do you use privacy protection on your domains? Yeah, yeah. So you just bought a three-letter domain.com, and I don’t know if you want to share that or any details about it on the episode. 

Yeah, absolutely. So I bought CTO.com’s big three letters. So now I’m helping CTOs and chief technology officers if you don’t know the acronym and the original business was started. So, I’ve been a fractional CTO for quite a few years for different companies. And what is a fractional CTO? I’m de facto the CTO, but it’s not a full-time thing, right? So, I kind of say I can give them 100% of the outcome for a fraction of the cost. So, I describe fractional as being about the cost, not about the impact. 

In reality, most businesses that are not tech businesses don’t make their living by selling technology. They’re just enabled by tech. They don’t need somebody like me full-time. It doesn’t make financial sense, but they can use the experience. So, I have known this for a few years, and now I teach other CTOs how to do it. And when we started that business, I realized, “Man, this would be really fun if we could just be like CTO.com,” and that was another interesting domain story. 

In reality, most businesses that are not tech businesses don’t make their living by selling technology. They’re just enabled by technology.

So, I looked up the owner at the time, and the website had nothing to do with tech. It was like a music staffing placement business in Philly. So he owned it since ’97 or ’98, and that was like the acronym of the business, or like CTO artists or something like that. I looked up the owner, sent him an email, didn’t answer, found him on LinkedIn, sent a search engine answer, and a few months later, a good friend of mine who’s a domainer, and domainer, if you don’t know, is people who broker domains for a living. I’m on his email list, and I get an email, like, ten thirty at night, his newsletter, and suddenly I see CTO.com for sale and make offers. I’m like, well, I’m instantly texted because we know each other. 

What are the odds? 

Or incredible. About two or three months later, I sent that LinkedIn message. 

We’ve had many conversations about those offline, this reality thing, it’s a rigged game. That was a setup for you, my friend.

I think about it, really. What are the odds? Right? The guy’s been owning it since ’98/’97 and never sold it. I never had it for sale until two or three months later. I tried to reach out to him but suddenly got this. I texted my friend and asked him, “Are you the buyer side or the sell side?” He’s like, “seller side.” I’m like, “Okay, we need to speak.” And then, over a few months, we negotiate. And it was a great acquisition on both sides. 

Was it six figures? 

Oh, yes. I think the broker asked me if I was okay with revealing the amount. I’m like, “Yeah, sure, but the seller preferred not to.” But I think for 2024, when we looked at it, it was like early December when we spoke; it would have been the highest-paid three-letter domain for the year. The highest paid acquisition three letter domain for them, so I don’t know what other transactions happened in that last part of the year. We did our mail of the year. But if not, the highest paid certainly topped two or three. 

That was also one of those. He didn’t sell in ’99 with the .com, but think about when he bought ’97/’98. They weren’t CTOs, yeah, like IT managers, maybe he didn’t sell it ’99, 2000. He didn’t sell it in 2008. He didn’t sell it in the COVID Rush. So that was like a big deal, holding that domain for 26 years and then selling it right when I wanted to buy it. It was serendipity, synchronicity. Call whatever you want, but it worked. 

AI agents don’t need to replace employees—they enhance productivity by removing low-value tasks. Share on X

It’s a divine setup, really, 

It is completely divine, and we plan to build the largest community of techies in that domain. So we hope it will be a very big thing in the next few years.

Amazing Congratulations.

Thank you.

Now, I know that you’ve done a number of AI workshops. You did, I think, 40 of them, right?

Yeah. 

And you’re kind of burned out from doing all those road shows, you know, AI workshops, I think, 

Yeah, it was a lot.

But what inspired you to do it in the first place, and then what was the impetus to stop?

It was interesting. So when ChatGPT came out. So this is late November 2022, obviously amazing. Some of us played with the previous model, GPT-2. GPT is one huge upgrade. Just two weeks prior, I was in one of our coaching groups, and I was talking about other AI tools that I use for content creation, right for different things, and then suddenly, it came out. 

Everybody’s kind of buzzing on WhatsApp and our groups, but they don’t know what to do with it. They don’t know how to think about it or how it works. I’m like, “Oh, you know what? I’ll just start doing that. I’ll teach my friends how to think about it.” So the very first talk I did was, like, just a month, first week of January after it came out, and I called it, and that’s the same talk I gave 40 times at this point, which is AI, models and methods, how to think about AI, how to do AI. 

And the talk was very principled, helping people understand the best way to think about it so it’s more intuitive, and not this kind of flash-bang here in the headlights, here are 20 prompts, but really kind of give them the skill. So it started with me trying to make it more accessible for my friends, and then it just picked up. I think the smallest I’ve done it for was like four people on Zoom, maybe three, and the most was 1000s, many 1000s on a stage. And even the last one was actually a couple of weeks ago. It was a 1000-live webinar with this master class that somebody asked me to run for an hour. And it’s great. 

Every time I do it. I like it; I get energy from people seeing the dots connect. I’ll give you some examples of some of the stuff I say that makes people think differently about it. But it’s not my business, meaning I don’t have an AI training company. I don’t have an AI Service Studio. I’m not an AI consultant and could have done all these things. Whenever I give one of those talks, I get a lineup, but I don’t want to be in the professional services industry. I’ve been there in the past, so that’s why it was a lot of fun. 

You’re not going to be replaced by AI, but you will be replaced by people using AI.

I think now, two years in, everybody feels way more confident and takes it more for granted, but they are still probably underestimating it in some areas. But now the world, I think, is more caught up than it would be now. It’s also different because I have five patents in AI and machine learning. One of them is literally structuring unstructured content, like back in 2017, so I have a lot of brain cycles in this area. Back in Israel, I used my first simulated neural network when I was about 14, the Technion, which is kind of the Israeli MIT. They did a project at my school, and we were in this special project. So we got to work on it for a week or two with neural networks, like building simulation and construction. So my brain has been thinking about this problem and this area for many years, but because I have a strong marketing background and product background, I think it was just easier for me to explain it to people so they understand it, and it’s not just like scary, or it’s not just too much gobbledygook. So it’s been a lot of fun. 

Yeah, what are some examples that you said you see people light up in these workshops when you teach them some new skill, insight, or way of thinking differently? 

Sort of thing. That January, I had this slide, and a month later, I also said it on stage. Many people say it, but at the time, nobody said that you’re not going to be replaced by AI, but you will be replaced by people using AI. We talked about Joe Polish; Joe Polish actually made me say that back in February 2023 as proof, as I said it on stage, and that was like one mental model another one is explaining. So, I do this kind of live audience thing to illustrate how large language models work.

And besides me explaining, like other AI, that’s not just large language models, but I explained to them. Put aside the concept of a problem; just say what the LLM is trying to do is figure out the next best word. And I was joking. You know, it’s kind of an odd and frozen kind of phrase, but what’s the next best word? What are we trying to do? And anything you put in is the before, and anything the LLM outputs is the after. Then I do this exercise, and I tell the audience I’m going to give you a word. Can do this lot. So the first word is old, and I’m going to ask you, “What’s the next word after that?” So, what do you think is the next word after old? If that’s the only word I gave you? 

Man.

SaaS is becoming obsolete. The future belongs to businesses building their AI-driven solutions. Share on X

Great. And then I have an audience, right? So I’m like, “Man, person, news, car, street, dog.” And then, usually, there are some repeats, right? I’m like, ‘Okay, is here like three, news, four, man’s, five persons, etc.” I’m like, “Great. And then I’m gonna give you the next word after old. The next word after old is MacDonald.” So now we have Old MacDonald. What’s the next word after that? And then everybody had, I’m like crazy in statistics, we call this like you have 99% confidence, right? And that’s what it does. And that’s what it does at scale. Old MacDonald had a farm. Now, a phrase could have been that MacDonald was a bastard. Yes, it could have been. Is there a high probability of that phrase? No, and then explain the concept of temperature, which is really what that is. 

So LLMs scrape the world right whatever corpus of text they have and count the repetitiveness, like the pairing of words. If it only saw the word Weinstein after Lior, then every time you put Lior, it’s going to put the word Weinstein. But the one time we would have seen Cohen, then now it’s not going to be 100% Weinstein. It’s going to start with Lior Cohen or Lior Levi, and so on. 

This concept of temperature is how creative or how much lower the probabilistic scale is going to be. It’s going to choose different words, and then you get basically creativity. And it’s quite fascinating that statistics and mathematics mimic humans. It’s kind of weird that it works. It’s weird that you put in a word, and it projects the next set of words, and you think it makes sense as a human. It’s wild. It’s incredible. 

So that alone, I usually do that exercise, and that’s a big coin drop because now people have an intuitive sense of what’s happening. They also understand how there’s, okay, no intelligence. I hate when they call it artificial intelligence. I don’t like calling it that, but it’s a useful label. There’s no intelligence, right? When you’re running statistics. And now, when you understand that mechanism, you can understand, “Okay, well, if it didn’t scan for the words, it can’t have an answer.” That’s why you can’t ask it for stuff about your personal life unless you give it information. It just doesn’t know and can’t reason with information it doesn’t have. 

That’s why when you hear Sam Altman, Jensen Huang, and those guys talk, the big breakthrough is going to be when it can create new information, right? When it can discover new things, right? It’s just this incredible world salad, and we like it, but that’s what’s happening. So it’s not equivalent to a human, it’s not equivalent to intelligence, but it’s extremely useful because the reality is, a lot of human activities are just basic like that.

Now, you mentioned earlier about this temperature, and if a user has more tolerance for error and wants more creativity, can they say you have a wider range for the temperature here? ChatGPT or Claude or whatever AI they’re talking to? Then, it will get more creative and not just keep using Weinstein every time it sees Lior. 

Yeah. There are the products themselves, ChatGPT, Claude is a product. And then there’s the underlying technology. With the underlying technology, you can change the temperature up and down as much as you want. If you go to the developer backbone of open AI, not the ChatGPT product, but if you log in through the developer area, you have basically the same kind of chat window, but you have the dials. You actually have a temperature dial. ChatGPT, for example, is set at 80 right, like 80% temperature, or point eight. So you can just move that dial-up and down, and the lowest temperature is the most probable.

So if you had temperature one, an old person would be the most common pairing after the word old, and then every time, it would have been old. So there are some reasons why you would want to have a low temperature or a high temperature. So, for highly creative stuff, you don’t want 100% temperature for most things because then you also lower usability, right? Because at the end of the day, you’re using the tool to use something like you wanted to write something, 

Then you’re much more likely to infringe on people’s copyrights, too, right?

Large language models love words. The more words you put in, the better the output.

It is at lower temperatures, not at higher temperatures. 

Yes, that’s what I meant. So when I ask for a new song to be written with the lyrics, and I set the temperature low, then it’s probably going to rip off other people’s lyrics.

If it’s going to be super low, the reality is that the chances for a neural network to actually output something that is unique and specific is very low, very low, to the point of improbability. Now, where do we see it? Sometimes, we see it with proper copyrighted terms. So let’s say you make up a word like Dan Sullivan; for example, he makes up a lot of words right as part of his IP in his business. Then suddenly, when you see that word, that’s, well, that’s a Dan Sullivan word; it certainly took it from Dan Sullivan’s content, right? 

The lyrics are a bit different, right? The lyrics are two words, and that’s hard to claim. You took it from a song, but if you have four words in a row. Eight words in a row, that’s way easier now for a neural network to output the same eight words in a row, even at low temperatures; that’s not very problem. Now, in some cases, low temperature makes a lot of sense. So when you’re doing like enterprise applications, when you’re trying to scrape or not scrape, you’re trying to get information from official compliant content, like IRS, tax codes, legal documents, core documents, your enterprise data. You don’t want high temperatures, right? Because you want a direct reflection of something. 

Now, there are a lot of other tools besides the large language models. So, if the large language model is this amazing way to reason, right? That’s what it is, and that’s another, so another thing I say in those talks, I say that LLMS, there are reasoning machines. It is not a fact machine, right? I think a lot of people come in the beginning, and they look at it like Google. It’s not Google. Google is really good at facts, really good at right, way better than anyone, and they suck at reasoning and ChatGPT and the like. They’re amazing at reasoning. They’re not very good at facts, right? Exactly for those two reasons. 

So now ChatGPT as a product is getting better, but a lot of people, like a year and a half ago, were still using it like Google, and they asked ChatGPT questions, and then they quoted ChatGPT. I was just in a workshop a month ago, and somebody was showing a quote that was like quoting a Chat quote. That’s like a fact. It was a definition of something, like a definition of a word, and it used ChatGPT. I like, “Whoa, can’t do that.” And I told him I could tell the organizer. You cannot use ChatGPT for reference. Bad idea, right? Don’t do that. 

That’s almost as bad as using a magic eight-ball.

Just provide LLMs with more context to filter out irrelevant words and generate responses that match the context as closely as possible.

Yeah? Because, in a way, it is because you’re literally, like, randomly choosing the words, not completely random. But it’s not Webster’s, right? It’s not an official, curated definition that humans agree on.

So, what prompting tricks and tips would you like to give in these workshops?

So I teach people that the muscles we’ve used over the last 25 years that Google has been around, 25 or 26, are the exact wrong muscles that you need for large language models. So if we’re trained to think in the snippets, and then, of course, Twitter made it worse, and social media made it worse. Large language models love words, so the more words, the better. 

So I recommend, and I compliment, like I tell them, if you’re like a blabber mouth if you don’t know how to stop speaking, this is perfect. You’ll have a great time because the more words you put in, the better the output. And I recommend that people just talk to it like a person. Just think, if you had a consultant in the room or somebody you want to consult with, you’re not going to come to them like Google and tell them about great restaurants in Atlanta. You’re not going to do that to a human right to a human. You’re going to go and say, “Hey, I want to go on a date with my wife. She doesn’t like Chinese, and we have a babysitter on Thursday. Can you recommend a romantic place that’s not more than 20 minutes drive?” You would give the whole context, and that’s the best way to use it, so just think conversational, not instruction.

And are you talking about it? Are you typing on it? 

So, one of my favorite tips is telling people that people have writer’s block but don’t have a talker’s walk. 

So I highly recommend talking because it’s way more natural, especially now with the products, the voice modes, and everything that is just phenomenal on my computer. When I’m on my computer, yeah, I hope, you know, I type fast and way faster than I speak. And I’m used to typing conversationally. If I’m on the go or with my phone, I talk way easier. So those are by far the two best modalities. Just talk to it like a person. Give it full context. What is the full context? Just everything about it, just like you would describe another person and then talk to it, not type, because it’ll be easier to more natural. 

Then. I would also recommend thinking about the final output, meaning, with Google, again, you’re thinking about the start of a journey. So say you’re doing research on something, and let’s say it’s a travel or business; you’re going to Google for the snippets, and then in your mind, “Well, I’m going to copy this eventually to Excel, or I’m going to write a report, or I’m doing this for presentation.” You’re not telling Google any of this, but you should for ChatGPT.

So if you’re starting a session with it and you’re like, “Hey, I’m working on a presentation for work, or I’m preparing my travel schedule with the kids,” whatever, just say it and then go into the tasks, because it will make sure that the formatted outputs it’s in because that’s the language. Wizardry, right? It can already decide that it knows that if you’re sending it as an email, as a presentation, or as a document, it knows how to manipulate the content so it fits into that format. 

So just say what you’re doing, give it the project. Especially now, these context windows are so large, we’re already getting into this kind of crazy seven-figure context windows, so you can have a single-pronged can be a few books. Well, it’s great, but it keeps the content. So, as you’re having conversations with it, tell it what you are doing or what the project is. Then, keep feeding it with information, keep just iterating with it, and let it do the work for you. That’s some of the new muscles people need to learn. 

Where they’re used to using tools like Google and any kind of a resource hub, and they do snippets of information, they take clips of it, right? They copy it somewhere, and then they organize it with these new LLMs. Just give it the whole thing and let it organize from the beginning when you talk to it like a person, and then you talk instead of type, and you kind of seed it with your intent. Like, what are you trying to do? Those three principles are going to get you everywhere, absolutely everywhere.

Now, like anything. You and I have been in the Google world for many years, and we use Google very differently than most people. And we also, I’m sure, you know, some people that are like Google wizards, right? Those are just phenomenal information finders, right? They know all the modifiers and crazy ways; that’s the same way. 

I wrote a book about that. I’m one of those people. Google Power Search.

Yeah, so there’s the ninja stuff in  LLMs as well, but the ninja stuff, I don’t think, is worth the extra brain cycles that it takes, especially because of what the product teams are trying to do, like the ChatGPTs, the Claude, they’re trying to make it easy for you, just like Google is trying to make it easy for You. So they’re trying to understand what you are searching for when you’re putting words into Google? That’s the same way with the LLMs; they’re really trying to make sure that you speak naturally, and they’re going to give you the best result, as opposed to you figuring out all the wizardries on how to best use it. So I wouldn’t obsess about it. 

I think when I see these prompt packs and 400 prongs to ta, da. It reminds me of the ’90s when people downloaded a list of websites. We had it in a text file on our computer, right? So I think people buy the packs and never use them. I guess that’s the last kind of big principle thing, like trying to make it functional and doing it on purpose. So, as opposed to just vacuum information and prompts that you’re never going to use, if you know you’re working on marketing projects, go get marketing prompts. But specifically, if you’re doing video, go get video prompts. Don’t just hoard endless prompts because the reality is you’re just not going to use them.

You’re not skating to where the puck is going. You know, you should use the Wayne Gretzky quote. 

That’s right.

This is not future-focused prompt packs and trying to get the languaging right. Act as if blah, blah, blah that doesn’t skate to where the puck is going. We’re just going to be able to interface our personal agent, who knows everything about us and that we trust, with the LLM, and they’re gonna have the conversation. We’re not gonna have the conversation.

Yeah, and it’s all about seeding context. At the end of the day, that’s really what the exercise is. Just give more and more context so it filters those words that are trying to figure out what words to generate so it matches as much context as possible as you give it. And yeah, it seems like the future is going to the agent universe, which is this holy grail in AI that we’ve had for many years. In theory, it’s not quite there yet, but it’s amazing to see what we have so far. However, that’s another thing. So, depending on how long I have the workshop, I might get into the edges of agents, and I’ll share with you these mental models that are very useful for thinking about agents. 

So you can think about an AI like an oracle. So it’s an all-knowing thing, meaning you ask it any questions, and it just has an answer, right? Because it has access to all the information in the world and it can generate new insights and new information, it gives you an answer. You can think about it like a genie. So you give it a task, give it something, and it is just magic, gets it done, and gives you the result. Lastly, which is that holy grail, you can think about it like an agent; the key difference is that an agent has a mission and an objective. It doesn’t have the how, it doesn’t have the knowledge, and it has tools to figure out the how to figure out the knowledge, and it comes back with the result. This is really the transformation we’re seeing now. 

This is the big hype right now, and in ’25, the big conversation is around, SaaS is dead, you know, like, that’s, or SaaS is dying, rather, because the future is agent work, which is very interesting there are some areas where that makes a lot of sense, and there are some areas, well, not quite, not. That’s quite, it’s like not. It’s not a relevant, like, mental model for the problem or the solution, but we see it. We see people now with agents. They got good enough. 

You can think of AI as an oracle, a genie, or an agent, with the key difference being that an agent has a mission and an objective.

Just this week, somebody in our group asked us about scraping a website if we know a scraper, right? So now he has a bunch of technologies, if they know, scrape. I’m like, yeah, here’s, you know, a bunch of links he got and a bunch of GitHub libraries right to scrape the page and do the thing. Then I saw all the responses and said, “Well, I opened ChatGPT o1 Pro,” and I put it on the web page he had put up. And I’m like, “Can you give me a CSV file with name, address, phone, and email from the page?” It took about three minutes, and I got the entire CSV, clicked, downloaded, and uploaded it back to Slack. And yeah, so that was, like, a case in point, right? 

The immediate response was all the SaaS that we know, right, all these scraping services that have been around for years, and they’re absolutely great. And it’s technology, so it’s like, “Oh, here’s the code you can deploy as well, download, fork, deploy on your computer.” It would have taken him maybe 30 minutes. But then, even for someone like me, I’m like, “Well, let’s just do the shortcut and see if it can handle it,” and it could, so now it can for that use case, this super scraper SaaS, obviously they do way more sophisticated things, but at some point, it’s just going to start eating up to that. Every company that’s not building agents in their workflow will need to figure out what part of their business will lose. That’s just like their daily reality right now.

Yeah, it’s an interesting time. We’re using PandaDoc as a SaaS solution for a DocuSign alternative for proposals and statements of work. I can’t imagine businesses like that having much of a runway in the next two or three years.

Well, we just had, you know, last year, the biggest news, like, in my eyes, was that Klarna decided not to renew a ten million Salesforce license. And Klarna had a lot of innovation. They are among the first to announce that they didn’t, so they had a natural turnover of employees in customer success and customer support. It’s not a long-lived position in most companies, and they announced they’re not going to hire 700, meaning they’re just going to let them organically leave the company and replace it with AI agents, not all of them, but a lot of their cousins. So that was like big news. A few months later, they announced they would not renew this huge Salesforce license because they figured they could just build Salesforce as a monster. It’s incredible software, right? It has all the bells and whistles. But that’s also the problem. Most companies need 5% of all of it, so Klarna just figures we’ll build the 5% we need. That’s huge. 

So, for years in the Computer Space, right? Like ’80s, ’90s or 2000, it was all built. Get engineers, build engineering teams, build your product, build capabilities. Then the SaaS revolution came so well; let’s lease it instead of building. Let’s let other people build it, and we’ll pay a fraction of the cost to build it in terms of usage, right? So instead of us setting up an engineering team and spending ten million building something, we’ll just pay 15 grand a year, 100 grand a year, right? Then we’ll pay on a per-usage basis, per license per seat. Call it whatever you want, meter. Huge revolution is still the prevalent model to this day. 

Every company that does not build agents in its workflow must figure out what part of its business will lose. That’s just like their daily reality right now.

But what’s happening right now is that the market is shifting back from buy to build because suddenly building makes sense, especially now, by the way, at this point that, you know, we’re a few decades into advanced technology as a society, you have incredible engineers, you know, incredible you have, at this point, like, 1000s of just, let’s say, Amazon alone, probably graduated 1000s of directors of engineering. These people manage dozens, if not hundreds, of engineers under them, right? If they’re at that level at Amazon, we can just lease into the wild, and that’s just one company, right? So if you take all of Fang and Silicon Valley, you have incredible technical talent, and now one incredible engineer has the horsepower of dozens; plus, with good leadership, you just don’t need as big of a team. Sort of you hiring what used to be a classic 10-12 person engineering team, a classic startup size, a few back in a few fronts. LinkedIn, some QA, some project managers. Now, you can do the same thing with one or two people and get the same result faster because people care about the effect, right? 

So the fact that you can get the product working sooner, as opposed to, oh, this huge, you know, enterprise, scalable system, no scale is going to come later, and when scale is going to come, you’re going to rebuild it again. Most likely. I’m not talking about, you know, mission-critical, government, NASA stuff, of course, but most products online are not that. They’re just utilitarian, like they have a specific utility. So, that shift from buy to build is incredible. This is also the first time. So when I first gave those talks in January and February 23, I was a startup guy who spoke about it earlier, right? As I come from a startup nation, it’s in my bones to think MVP, rapid jet skis, and not container ships, right? And because some of those talks were in front of the kind of eight-figure revenue business owners, nine-figure revenue business owners, selling them. And I told them, “It’s the first time in history I think enterprises have more advantage than starters because enterprises are incumbents rather than enterprises. Because what incumbents have, they have the business relationships, they have the data. They have the trust. They just didn’t have the speed.”

They have brand authority, too, which is becoming increasingly important as Google tries to thwart all the AI-generated content; they’re relying on more and more people to trust.

So, the fact that they have the brand, trust, and data, and now, suddenly, they have a huge speed. So now they can just get people part-time, or a few people full-time, to quickly build on stuff that startups never have. Startups take years to build trust. Startups usually build trust by either extreme innovation, creating something so new and so convenient, like what ChatGPT ends with, or by visual appeal. Oh, this is, like, fresh, fresh visuals, fresh esthetics, fresh copy, like, fresh perspective on something, and then people excuse the lack of functionality for the fact that it’s pretty or convenient or stuff like that, even though it doesn’t have as much functionality as the incumbent products. 

But now, you can use any of these tools to create an extremely visually appealing brand. So, the visual is out. Even an amazing blacksmith still has a full-time job. The top of the top will always have a job. The best brand guy, best designer, best coder, best lawyer, all those people who will have the best doctor and will have a job forever. So, the real difference is in the bottom, in the middle, where the company can now have a great visual product in hours. They can create fully functioning pieces of the cone also in hours and sometimes in minutes, now we see. Startups have a problem, and we see now how many, like every week, ChatGPT and the other tools are releasing features and the cemetery of VCs that put in money for startups that just became a feature in these big products. 

Being really innovative right now is very interesting. I would say the big opportunity, though, is exactly for the promise of SaaS. You know, in the early 2000s, remember how so many people just still so many people made money by creating a guide on how to learn German on this, like, small, little website, and they made money with ads or, like, an ebook, or stuff like that.

I think that’s happening again. You have that resurgence of people able to monetize their insights and their passions because now they’re getting the tools where the most common programming languages just become in English, and they go into all these AI product generators, no longer co-generators. They create full products. It’s like with design and database and functionality and username, passwords and automation, like the whole thing, and it’s Abracadabra. 

You just share what you want to do, and it’s like you have an entire product team working for you and fabricating it. Then, people can add payments and all those modules and create their own businesses. Small businesses can have the same functionality as large businesses for a fraction of the cost, truly a fraction we just had in the news recently, right about also these models that even I saw a funny tweet that ChatGPT got its job taken away by AI, with this huge transformation because people are figuring out these crazy efficiencies, like AI was already super efficient, and now they’re discovering 30x more efficiency on the super efficient stuff. And now we’re. And ’25 and ’26, it will probably be 50 times more efficient. 

That’s actually another thing I used to have in this line. I still have this line in the presentation where I showed this example of an image with 30 random objects organized in the shape of the word love with a Hawaiian sunset background, right? It’s like seashells and flip-flops and that kind of thing, right? And I was saying, Let’s think about what’s happening here. Let’s say you worked with a good designer, not a great designer, and you called him up, and you sent him an email and had a request. It’s like, “Hey, can you please create a graphic for me for my whatever website? And this is what I have in mind, okay, 30 random objects, Hawaiian sunset and so on.”

Let’s say he replied to you on WhatsApp in a minute and said, “Yep, no problem. Got it.”  And let’s say he charged 100 bucks an hour. And if any of you work with designers, that’s not a reasonable price for a great design. And let’s say he was like, “Yep, got it, and extremely efficient was available right now. And he sent you back the results after an hour.” Okay, let’s assume that. Let’s give you the benefit of the doubt. That’s what happens in that scenario. And then I go, and you can take that prompt just said, 30 random objects organized in the shape of the word love with the Hawaiian sense of background, and put in ChatGPT, put it in whatever your favorite image generated tools, you’ll instantly get four options at least. 

The market is shifting from buying to building, as building now makes more sense—one exceptional engineer can achieve the work of dozens.

And if you want a 20, just say 24 options at least that are spot on, 30 objects with the word love within a second. And from a computing perspective, that was like a fraction of a penny people talk about in business growth when they have these paradigm shifts about, like, 10x, right? We hear this all the time: 10x growth, 10x, 10x exponential. And I’m like, people don’t understand; this is a million times cheaper. This is a billion times cheaper, right? And a million times faster. This is not 10x. This is the underestimated portion of what’s happening with AI that people are just thinking about. Oh, I’m gonna move from 20% growth to 40% growth. Oh, we’re gonna save a little bit of money. I’m like, no, no. This is a billion times cheaper and a million times faster. 

You need to look at a log graph to see the straight lineup because we can’t wrap our heads around exponential growth. However, if we look at a log graph, we can see exponential improvements and technologies. We can see the rise in the price of Bitcoin, and we can see it going to the moon because it’s just following a straight line on the log graph,

Yeah. And even on a log graph with AI, it’s also looking to look like a hockey stick, because in for some activity, like the best thing I recommend to people like in their organization, the day life, I recommend they do, like a time study or an activity inventory, where they start just line by line. You know, you wake up in the morning. What do you do? You kind of check emails, you write a letter for something. You go to a meeting. You take not just, just kind of scenes in your mind. Just fight him line by line, and then put a little note on how much time you’re spending on this. Five minutes, 10 minutes, 20 minutes. And try to abstract for a week or a month? Well, I spent two hours on this a week, five hours on it. And for some of those acts, to kind of try to figure out what those useful spotlights are that AI can really make a dent. And then you’ll see that for some of those activities, the two hours will become a minute. It’s not two hours; that is one hour. Two hours becomes a minute. And there’s a lot of activities in our human life and human society where that is the impact. It’s that level of impact, and it’s profound.

So, give me an example of two hours becoming one minute.

Meeting notes, meeting reviews, etc., are available right now for me, for example. So I work with multiple companies, and as their CTO or chief revenue officer, I tell them I don’t do standing meetings. It’s a big thing. In the beginning, I told them I don’t attend standing meetings, which is just not happening, but I’ll send my AI, and my AI will go to all the meetings. At the end of the meeting, it depends on the meeting. There’s a set of prompts, and I get a snapshot of what I need to know, and I get it on, like telegram and email, and if there’s something I need to know, then it highlights it for me. 

So I call it like smoke detectors, and that’s huge because what was the alternative? The alternative was either me going and actually, like listening analog being in the meeting, or somebody else going, taking notes and then sending me the notes like an assistant or someone like that, or somebody in the meeting would have had to do it. I can flag me. I don’t have any if you look at my calendar for a phone; I don’t have anything like that, not alone. Huge time saver.

But then you also have the problem of hallucinations. It picked the wrong things to highlight from the meeting and came to the wrong conclusions. I mean, that’s going to be less and less as that hockey stick goes further along on that log graph; we’ll see way fewer of these mistakes. 

It is interesting because I am joking about it. Actually, all it does is hallucinate. You know, hallucination is a human judgment. We look at the phrase and say, “No, it’s not Stephan Weinstein; it’s hallucinating, right?” Like we put that judgment from its perspective, it’s pushing out words, statistics, and probabilities, right? So, it is the fact that you agree with something. He said it’s kind of interesting. That’s what I said earlier; it’s kind of interesting. So, all it does is hallucinate. So hallucinate, the way we use it as a label, as we have to describe, we’re getting a result we don’t like, and that’s separate from getting something that’s untrue. 

Also, it doesn’t have a sense of truth, right? Doesn’t have morals, doesn’t have ethics, it doesn’t have standards. It just has corrections, like bumpers, right when you’re bowling or something. That’s the best that it has. We look at something; two humans can look at a statement and say it’s true and not true, right? Put aside, you know, what AI can do or should do. But certainly, I think business meetings are way easier than other stuff because there are some expectations, you know, there’s a topic, there are subjects, there are speakers, if it’s a good meeting, you know what a meeting is about. And that’s where prompting really becomes super relevant. And then where the models make a difference. So this is where you see a difference between something like right now and in modern times, between an o1 or o1 Pro and a Sonic model.

And then when you’re doing these automated prompt systems, just have at it. For years, my assistant was labeling my emails so that when an email comes in, we have one of four categories. Action needed, high priority. Action needed, low-priority content, just FYI, right? Or content that is not important, such as social reading or something. At some point, I think we should train an AI to do this, and that’s what we created. 

So, I created an AI agent that has a very extensive set of instructions. It’s like you are an AI agent for your wine team. He runs these companies. He does these things. These are his family members. These are the things they’re like, plus a lot of examples of this kind of email. VO would have labeled that; why? Because of this kind of email, a lot like it’s very, very, very long, okay, like, multiple, multiple, multiple pages, and then any email that comes in, we have, like, a make and an AI agent automation combo. Emails come in, and I run multiple inboxes, around 12 inboxes. An email comes in into either one of them, takes the email, sends it to the AI agent, the agent returns a label, and then goes back to Google and applies a label. And that’s my inbox, right? And my assistant hasn’t labeled an email for me in over six or eight months. 

Where do we get this prompt? 

That’s the thing I can show you one-on-one, but it’s an elaborate chain of prompts. It’s not a single one, and this is where, going back to, how would you explain it to a person? Right? Just talk to it like a person. That’s the right mental model for agents. And then you train, and then you use these agent tools, which there’s a lot like, I like right now and this modern time, I really like Relevance AI, Crew AI, Lindy AI, these are all really solid agent platforms that you can quickly create that kind of products and services for you. And, of course, there’s the entire open source universe, which is just open source, which is where it’s at. This is where the industry is going to go. There’s no doubt this is something we’ve been speaking about for a very long while, but now you’re actually seeing this. It competes in the big commercial space. 

So you can fight hallucinations right now, you have very strong models that can contain very large sets of context, and because you can give them these huge prompts, you can give them like a 100-page prompt that contains all heuristics, like everything you can think about. And that’s pretty good. I would say my agent note-takers and the prompts are way better than any human I send. No doubt. Are they better than me in terms of catching some esoteric word or somebody on Zoom, just like biting their lip when somebody says something? No, they’re not, by the way; in some cases, they’re also paying more attention than I do, I would say, but I get my time back. There’s nothing perfect. Obviously, if it’s a mission-critical meeting, I’m there, but that’s why I referenced it as a standing meeting. It’s a kind of regular operating meeting rather than deciding who to hire, who to let go, and what to do with the company. You know, those meetings are still coming,

You mentioned earlier about context windows and having a million input token limit. With the LLM, you should take advantage of that. You should utilize that to upload hundreds of pages of policies and sample deliverables and things like that so that it has a real sense of what you’re trying to do and what you’ve produced in the past so that it’s not just inventing from scratch.

That’s right. This is why large companies or established companies have so many advantages. If they’re well run, they have so much SOP training, SOPs on a lot of knowledge that can be extracted out as instructions, so absolutely just shove as much as you can. Even right now, it’s not at the point that any human, like the fact that you can put in the tax code and ask questions always while referencing the tax code, no human can do that, and that’s why medicine is, you know, I have a friend. He just showed me that he specializes in very specific kinds of medicine and fertility, and he showed me how just normal ChatGPT was able to find unique anomalies in these X-rays that he has in the clinics. It’s actually a version of enterprise, a kind of private ChatGPT they used that three radiologists couldn’t, and ChatGPT did, and that’s already now, so I can’t even know if you try to project five years from now, let alone.

Even six months from now, right, the speed of change is; it’s hard to wrap our minds around it.

Yeah, I would say one of my favorite topics, though, to talk about is the frame of my talks, which is usually not the end of work; it’s the end of boring. I framed it like that because I’m trying to get buy-in from the audience because the entrepreneurs and the CEO want AI and want to push the level, but then you have people afraid of losing their jobs, right? 

So framing it as boring, and nobody likes boring words, because boring means something different to everybody, and everybody has parts of their job that are boring, and they would very much prefer not to do. So I frame it like that and tell them I really want you to be aware of artificial intimacy with the other AI. I think that’s where people miss, right? And I think the biggest opportunity of AI is for stuff like this to be more present, and I kind of say fewer bits, more atoms, so it can really express itself. I had one of the questions I had that first time in February. I was on stage, kind of talking about this on February 23, and somebody asked about SEO. Right favorite topic? Aren’t I afraid that there’s going to be this inflation of content or something online? So I told him it is confusing content, authenticity, or stuff like that. I told them one last time that I checked every time you go to Google, and you get about a Google result for everything you search. 

AI, unlike humans, maintains higher quality standards than most people… when it’s running.

So, I would argue there’s already an inflation of content. If anything, you’re going to get higher quality content because AI, unlike humans, has better quality standards than most humans, you know when they run. But we talked about this issue, and then I talked about ghost authorship, and told them, Listen, a lot of books, if not most books you see right now, are not written by the authors, per se and the written but, but the insights in most cases, hopefully, is from the author, meaning the person had the insight, but they didn’t have the words. They didn’t have a rapper. 

So they pay a writer who’s very good with words, just like you pay a musician to write a melody, right? That’s very good with melodies, right? But you’re the one that came up with the insight, with the direction, and so on. And I told them, I don’t think it would have had as much of a loop in your mind if you just paid an employee to write something like, for some reason, you think that using the AI means it’s something else than you paying an employee. Well, if it’s not you writing, and we’re just arguing on the semantics, if it’s the whole of humanity’s statistical knowledge about writing is writing, or Rhonda is writing to me ethically. It’s the same. I don’t think you should present. I think when something is automatically generated, it makes a lot of sense to say something about it in some cases; in some cases, no, it doesn’t matter because a lot of content gets published. We see white papers and PDFs, but nobody has authorship of them. The company can release a 50-page guide, and it’s not like Mary wrote it, right? 

So, I don’t think it makes sense in all cases, but I think it makes sense in some cases. If you’re generating medical advice content, you should probably tell people, “Hey, this wasn’t written by a doctor,” because they’re going to give it a different sense of trust, maybe more, right? But I think you should let humans know if they’re about to make a big decision based on content, but I think the problem is when people mimic when people use it. That’s why I see it on LinkedIn. It kills me whenever people have an AI-written post, and then they have AI-written comments. So it’s like the machine creating the post and then replying to it. I’m like, That’s ridiculous. That’s, to me, artificial intimacy. Don’t do that. You should absolutely use AI to write your social posts. 

Use, utilize brainstorm polish, and make it better, but it should be your content. And I called him out in a couple of conferences. I’m like, I know I was like, I’d given names. I was unjustly about it. But I know you, and I know you, you don’t talk like that, and you don’t have that idea that the voices that people put out are very artificial, you know, it’s not authentic. And then you meet the person in real life, and you’re like, “Oh, you don’t sound like that at all.” And we’ve seen that, in other stuff, when people write books, and it’s not really them, or they publish content, it’s not really them. I think that mismatch is a problem, and I think AI is actually exacerbating that and creating more artificial intimacy, like a sense of, I feel you, and I connect with you, and we see this cold marketing now as well, cold email, cold reach outs, like people trying to hyper personalize with AI, but we also see that it’s failing. 

One thing that bothers me the most is when people talk about the pitch that you can make an AI version of yourself impossible. And I’ll share a last comment on that. The reason why it’s impossible is that it’s impossible to take it like you’ve written a lot of stuff in your life as an author, right? As a person, it’s impossible to take all of your output and figure out how to recreate it, and the reason is that your decisions are not your choices, just like you can’t build an airplane by looking at it from the outside, the fact that you whatever words you say like I’m using words to say right now, but what I’m not saying is 100x more in terms of the choices that I have to say. And that’s why it’s impossible. It’s impossible to take somebody’s writing and actually make it look like the same person speaking at the best. It mimics and echoes. 

But here’s the thing about humans: we’re very good. We’re like, you talk to your wife, and something is like 1% off, just 1% something a little bit wrong. Immediately, we’re so good at our discernment of authenticity and alignment. And AI, with the current technology, it can never get there. I can never get annoyed when I see that pitch; I get annoyed because it either tells me somebody is bamboozling, or they just don’t understand, and then they’re presenting as if they haven’t something. So a lot of opportunities. It’s really a transformation. I’m glad to be at this age.

I’m glad to be a podcaster in this age of inauthentic AI-generated content because I think podcasters are probably going to be the last to be replaced by AI.

In this kind of podcast. Well, we already saw the NotebookLM, which is auto-generated. Do you see that? 

I know that’s creepy.

I know we’re out of time, so I want to respect the time we are and our listeners as well. I’m curious. Do you have a video of your workshop posted on YouTube somewhere? Or can we somehow get our listeners access to it? How can we learn from you?

I wish all of those were either private workshops or that I had done them as part of webinars for other people, but I now post a lot of content and insights on my LinkedIn. Please feel free to connect. I love meeting new people and educating myself about the topic, so I’ll certainly find opportunities to do more public service workshops over the next year.

Awesome. Well, thank you so much, Lior, and if our listener, perhaps they’re a business owner, perhaps they’re a techie, and they have the chops to be a CTO or fractional CTO. Where should they go to potentially work with your CTOX fractional CTO company?

Yeah, we’d love to help you. Just go to CTOx.com if you’re a company that can use a fractional CTO you want. We say we can give you 100% of the outcome for a fraction of the price. We can help you find the best fit. We have hundreds of them. And if you’re a CTO thinking about a career change, please talk to us.

Awesome. Well, thank you, Lior, and thank you, listener. Now take this information and do something amazing to make the world a better place. We’ll catch you in the next episode. In the meantime, have a fantastic week. I’m your host, Stephan Spencer. Signing off.

Important Links

Connect with Lior Weinstein

Apps/Tools

Book

Businesses/Organizations

People

Previous Marketing Speak Episode

Previous Get Yourself Episode

YouTube Videos

Your Checklist of Actions to Take

Integrate AI agents into my workflow. Identify repeatable business processes where AI agents can handle tasks and regularly assess which manual processes can be automated by AI-driven systems.

Shift from SaaS to AI-powered Custom Solutions. Instead of relying on SaaS subscriptions, explore how AI can build customized tailored to my business.

Use AI to automate meeting notes and reviews. Deploy AI-powered transcription tools to summarize meetings and set up AI prompts that extract only the most relevant information.

Optimize AI-generated content for authenticity. Use AI as a tool to enhance and structure ideas rather than replace original thought. Avoid completely AI-generated posts without human input.

Improve my AI prompting skills—experiment with adjusting “temperature” settings to balance creativity and accuracy. Talk to AI like a human rather than issuing short, robotic prompts.

Maximize context windows for AI training. For businesses, feed AI models with company-specific knowledge, policies, and case studies.

Harness AI for business efficiency, not just cost-cutting. AI should be used to free up time for high-value creative and strategic tasks.

Regularly evaluate how AI trends are reshaping my industry. Companies that fail to integrate AI risk losing competitive advantage.

Avoid artificial intimacy by using AI to enhance, not replace, real human interactions. Ensure AI-generated responses and messaging still reflect personal authenticity.

Connect with Lior Weinstein for AI and business transformation. Visit ctox.com for a fractional CTO or explore AI-driven business solutions.

About Lior Weinstein

Lior Weinstein is a serial entrepreneur with a background in growth and technology. Currently fractional CTO and CRO for 20-100M/revenue companies. Life’s mission is to dissolve people’s self-limiting beliefs so they can be creative and pursue happiness. CEO of CTOx, the Fractional CTO Company.