Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Why AI Can't Replace Your Marketing Brain (Yet) - CustomGPT VS Niche LLM
192 Plays2 months ago
In this solo episode of The AI-Driven Marketer, Dan Sanchez dives into the intriguing question: "Can AI replace your marketing brain?" Spoiler alert—AI isn't quite there yet. Inspired by a LinkedIn discussion, Dan explores two key approaches to replicating your marketing expertise with AI: CustomGPTs and niche-trained LLMs (Large Language Models). He explains why these tools, while powerful, still fall short of fully capturing the depth and nuance of human expertise. Dan shares his personal experiments and insights on using AI to enhance, but not replace, your unique marketing abilities.
Episode Timestamps:
- [00:00:00] Introduction to the Episode: Can AI Replace Your Marketing Brain?
- [00:01:45] Andrea Bosoni's LinkedIn Post: The Inspiration Behind the Episode
- [00:02:58] The Concept of CustomGPTs: Pros and Cons
- [00:05:45] Dan’s Experiment with a CustomGPT Life Coach
- [00:07:55] Limitations of CustomGPTs: Where They Fall Short
- [00:09:40] Introducing Niche-Trained LLMs: A Deeper Dive
- [00:12:35] Real-World Examples: Chris Walker's and Dave Ramsey's AI Experiments
- [00:15:45] Why AI Can't Yet Replicate the Full Scope of Human Expertise
- [00:18:20] Conclusion: AI as an Enhancement, Not a Replacement
- [00:19:20] Call to Action: Subscribe and Explore the AI Fundamentals Course
Recommended
Transcript
Introduction and Mission to Master AI
00:00:04
Speaker
Welcome back to the AI Driven Marketer. I'm Dan Sanchez. Friends, call me Dan Sanchez. And we're on a journey in 2024 to master AI so we can do more with less and better with marketing, right? Because we always have so much to freaking do as marketers. It's a black hole, I've been told. You can always spend more, do more, and have more talent doing marketing stuff. um And I'm starting to wonder if AI will ever fill that black hole, or if it's possible, right? If we have robots that can do our job for us.
Can AI Replace Human Brains in Marketing?
00:00:36
Speaker
And that's kind of the topic that I want to talk about ah for this solo episode is if AI can replace your brain. The short answer is, I don't think it can yet. But I want to talk about two possibilities where you can use AI to kind of replicate your own brain. Like if you were able to consult a protege, ah can AI do it for you on your behalf?
00:00:57
Speaker
And the the topic of this got started with a post on LinkedIn, and I'm gonna pull it up here visually, but I'll be reading it for all the listeners out there. um From Andrea Bassani, sorry if I mispronounced that, but let me go ahead and read the post, and then I wanna talk about my response to the post, and how to learn more specifically around custom GPTs and niche LLMs, and how both can be used to kinda reproduce yourself.
00:01:25
Speaker
for others or or for yourself or maybe your your direct reports. So let's read the post and then we'll get into it. He says, a random weekend idea I was just discussing with a friend. Let's say I write down most of my marketing knowledge into a few documents or at least the basic key stuff new founders need to do ah to set up the foundations right.
00:01:46
Speaker
Then I train an AI model on the specific data. I should get something out of it that's different from a generic chat GPT answers because it's trained on my exclusive proprietary data. So in theory, the product might be able to give founders basic guidance almost as if they were asking me directly.
Replicating Human Expertise with AI
00:02:04
Speaker
I wonder if it's easily doable and what the final results would look like in terms of quality.
00:02:10
Speaker
So it's a really good question, and I love that people are thinking about the possibilities that are now possible with AI. Like I said in the very beginning, I don't think this is quite capable yet, um but there are two routes to getting this done. Both fall short in a lot of different ways, um but have some have some pros. So let's talk about it.
00:02:31
Speaker
um A few weeks ago, I talked about the difference in AI memory, specifically long-term memory, which I have represented here visually as a big circle, and short-term memory, which I have here visually as a is a very small circle that kind of overlaps with the long-term memory.
00:02:47
Speaker
And in that episode, I talked about how long-term memory is like all the stuff ah chat GPT was trained on or any ah large language model was trained on. It's the whole internet. It's all the tweets. It's all the books and everything else they threw at it in order to get it to its current level of intelligence. I mean, it knows a lot.
00:03:05
Speaker
um But you know and I know if you just ask it to write a blog post on Product marketing for B2B sass. It's gonna give you a pretty vanilla post um And it might hallucinate come up with some weird things in there sometime because the long-term memory like our human long-term memory is It's kind of fuzzy, right? We do better when we load up our short-term memory with ah relevant facts and outline of where we're going um and research that we've done ahead of time so that we can pull them into writing an article. AI actually needs the same thing, which is why in that episode actually broke down like in order to make AI perform better,
00:03:43
Speaker
you need to learn at loading its short-term memory with all the stuff that you would want. and Things like your prompts are actually loading in short-term memory or chain prompts, loading in documents, ah web searches, custom GPT instructions, docs in your custom GPT or using the custom GPT API. In order to feed the short-term memory everything it needs to know and so it can create content or perform the actions that you want more reliably without without relying so much on the long-term memory which causes hallucinations because again the long-term memory is fuzzy just like in our memory and I think this becomes an important illustration when considering the two different routes you can use AI to replace
00:04:24
Speaker
your brain and essentially have AI be your your chief representation of yourself either to customers or your team or even yourself, right? um some There's some pros and cons. um So there's two different routes to creating this. One is creating a custom GPT loaded up with all the documents ah for loading the short-term memory. um And I've tried this before, and it works. It's it's it's better than just a vanilla chat GPT, right? Like just asking chat GPT, oh, what's Dan Sanchez's thoughts on this? You know, it might be like, I don't know how much of my blog posts were loaded into its long-term memory, but probably not very much.
Strengths and Challenges of Custom GPTs
00:05:01
Speaker
um but it'll it'll might take a guess it might do a bing search to try to go figure out who dan sanchez is and what dan knows um and it might scrape a few blog posts and then get back to him but it'll be pretty rough uh it's probably not going to know a lot but if i loaded uh a custom gpt with maybe like a library of my blog posts and all of my linkedin posts and ah maybe some other strategic frameworks I have for answering questions and dealing with problems. Well, it it do it probably sounds a little bit more like me. And to be honest, I've actually tried this. It was one of the first custom GPTs I made where I wanted to build a life coach and I armed it with all the information that I had about myself like life goals, strength finders, blog posts, LinkedIn posts, so I can understand. And I've added a lot of information.
00:05:47
Speaker
And I found it it was pretty hit or miss when I would ask it questions. Like, hey, based on what you know about me, what should I do? What would be the best way to leverage my strengths for X? you know And it would give me an answer. And it would do it based on the knowledge. But honestly, I just found that it was it just wasn't very reliable for problem solving and troubleshooting. So while I tried to make that life coach out of a custom GPT, I found that it just didn't do too well. And honestly, I've had a few different custom GPTs that I've tested where it's like Einstein or whatever. And it's kind of like, eh.
00:06:17
Speaker
Eh, it's okay. It's better than just relying on long-term memory, because it's at least pulling from some short-term memory um stuff, but it still doesn't do quite as well. So that's using a custom GPT to replicate ah your knowledge and expertise. It is certainly a lot better. The problem with custom GPTs in general is that they're just bad at general purpose stuff. Like, where custom GPTs really excel is having highly specialized tasks that they do, and they just do it over and over again.
00:06:45
Speaker
turning a podcast into a blog post, turning a blog post into multiple LinkedIn posts or whatever it is that has a task and it has a set step-by-step instructions on how to execute the task within the custom GPT.
Niche LLMs: Reliable but Limited
00:07:00
Speaker
That's where they excel and they do fantastic and can produce ah blog posts that sound like you, pretty like to they do a great job of sounding like you, I'll put it that way.
00:07:09
Speaker
um That's where custom GPTs excel. Now, on the other side, there's this option out there, and I've not done this personally, but I've played around with other people who have ah other versions of this that other people have created, and it works pretty well. It's called building a niche LLM.
00:07:26
Speaker
And it's taking an existing large language model and then customizing it with a lot of your data and your knowledge so that it's not even loaded into short-term memory. What you're doing is customizing the long-term memory of an LLM, like chat GPT, with your proprietary data and knowledge.
00:07:45
Speaker
And so when you don't even have to load the short-term memory, you can be like, hey, write me if it was customized on all the stuff I ever put out there, and it would know about my my knowledge and framework for how to grow an audience, for example, which I've created a whole course on. Maybe I load all the course knowledge into it, all the LinkedIn posts, all the blog posts, everything I've ever created around audience growth, and I train it, ah this LLM on it, and then it becomes Dan Sanchez's audience growth tool or consultant and you can ask it questions about audience growth. Its ability to take all that knowledge and then give custom information back based on that now long-term memory because you've now trained it on that um is much better of course than a plain vanilla custom GPT. But there's some downsides with it because you've trained it into long-term memory it's still a little fuzzy
00:08:36
Speaker
It still struggles to break down small thing ah large projects into small tasks because it's the the current limitation, it's the current constraint of AI and where it's at. In the future, it won't have that constraint and they're already figuring out ways to help it break down into think slow so that it can actually break down what are fairly complex projects down into small manageable tasks that it can then
Future of AI in Complex Projects
00:09:00
Speaker
do really well with.
00:09:01
Speaker
For example, writing a marketing plan or maybe an audience growth plan is one of those ones where you kind of have to break it down into small steps um to understand the market that you want to get into, the business model behind, why you're creating an audience, the pain points and objectives of the audience, how your expertise can correlate to those things, how you can create content around those things. You see how there's like multiple steps that it can't just think through and solve each one of these. Now, if you created a custom GPT to solve each one of these, it'd do pretty good.
00:09:29
Speaker
um But a large the long-term memory has a hard time thinking through all the different steps you would need to do in order to do it well, even if it's trained on the way you do it. It will still struggle to do it well, but it will do a much better job of utilizing your method in just one shot in order to get it done.
00:09:47
Speaker
So I'm scrambling a little bit because it's ah kind of hard to explain how this long-term memory works, but that's what people have done. and ah I've seen two examples of people using niche-trained LLMs. One was Chris Walker, who launched ah a... I haven't tested this one, but I've heard good good things about it.
00:10:04
Speaker
He loaded all the information from his podcasts and blogs and articles and case studies and data that he had in refined labs into his membership area and trained in LLM to essentially act like him. So you can ask the question and it'll give you answers like Chris Walker would give the answers, which is really helpful.
00:10:23
Speaker
um but it's still not going to break down a problem like Chris Walker would answer the problem. It's more it's more useful for a Q and&A type function, where custom GPTs are great for very specific things, ah training a niche LLM to behave more like you based on your data and your frameworks and stuff.
00:10:40
Speaker
um It's a better for long for more generic purposes, but it still doesn't know how to apply and think through what should actually be asked. because you might I might go to Chris Walker's thing and be like, hey, write me a demand gen marketing plan for a company. here's Here's a few details about the company.
00:10:58
Speaker
um and it'll write a plan, and it'll do it based on Chris Walker's framework for demand gen. But if Chris Walker had been sitting there, Chris Walker would have stopped and been like, hold on a second. Before I write you a plan, here's five questions of information I need before I write the plan, right? That's different than how Chris Walker would operate, even though it's much more closely aligned with Chris Walker's principles.
00:11:21
Speaker
um Another example I saw this was Dave Ramsey. I have a lot of friends that work there because I live in Nashville, Tennessee, and that's where his company's located. So I have a lot of friends that work at Ramsey Solutions, and they have a Davebot that's not publicly available. It's just a little internal thing because they're testing and playing with AI and just, I don't know, they're smart and they're they're staying on top of it internally for the day when it becomes more reliable to to do something with it externally.
00:11:47
Speaker
um So you can't test this but I've seen it because my friends have been over and i'm test and I've tested Davebot against just asking chatgptbase to give me Dave Ramsey answers to financial questions since he's a big finance finance guru out there. um i'll ask I've asked questions and I've asked the Davebot questions or at least my friend has on
Limitations in Problem-Solving
00:12:07
Speaker
their phone where they can tap into it and the Davebot is much better at answering in a way that I know Dave Ramsey would answer.
00:12:13
Speaker
Let's put it that way. Uh, it's much more reliable. Is it a hundred percent reliable? No, no. Cause it can still hallucinate. It's still, you're still pulling from long-term memory in order to do it. So because they're a trust based brand, they don't want to release something that they don't have a very, very strong degree of confidence that it's going to be reliable like 99.999% of the time. Right. Um.
00:12:34
Speaker
But that is the advantage of loading a custom niche LLM with all your knowledge and then training it and giving it time to process it all um for people to interact with. It's great for Q and&A, but it's still not the same as replicating yourself. It's still not the same as identifying the question behind the question that needs to be asked or asking someone who who you're consulting um that might be going in a direction. They think they need to go to this direction, you're like, actually, based on what you're saying, you don't want to go in the direction, you need to go in this opposite direction or a slight deviated opposite like road you need to take, right? And that's the power of a consultant. That's the power of someone who has the experience and the knowledge to be able to guide a ah business in the case of the LinkedIn post we were reading like later. um It's still good for a Q and&A, though, I find if you want to know if you want to have direct answers for certain questions, loading the long term memory is helpful because it'll have much more reliable answers.
00:13:31
Speaker
um So, can you replicate your brain? Not yet. But you can build a custom GPT that follows step-by-step processes based on your frameworks and knowledge um to execute very specific results.
00:13:45
Speaker
And you can load long-term memory with all your knowledge so that it can give general Q and&A tile answers. But it still can't replicate your ability to identify the actual problem and solve for it, regardless of what if they ask the right question or not.
Conclusion and Call to Action
00:14:03
Speaker
That's still the very human thing that we all can do, luckily. But AI is still working its way there. And eventually, i I imagine I will do an update episode where I talk about how to replicate your brain on AI. And that'll be fun. But for now, these are the limitations we have and can still do some amazing things with with where AI is currently at.
00:14:23
Speaker
So if you've liked this episode, ah please give it the star rating it deserves in the ah Apple podcast or Spotify or wherever you're listening to podcasts. If you're listening on YouTube, please give me a thumbs up. It makes it all all the work worthwhile or give me a give me a comment or something. um And if you want to dive more into AI, please go find my AI fundamentals course. I promise you will learn something you haven't known before about AI.
00:14:46
Speaker
um go to aidrivenmarketer dot.com slash course and you'll even be surprised by how I've designed the course with hyper personalized course instructions based on the information you put in the form. So if for anything go check it out for that reason and get started today.