AI Prompts Are the New Source Code with Christoph Cemper

In This Episode

• [03:03] – Christoph Cemper recounts his teenage experience writing codes and how his journey began.
• [07:11] – Christoph discusses the challenges of being a perfectionist and the importance of letting go of unrealistic standards to complete projects.
• [11:51] – Stephan and Christoph discuss pivoting to Chat GPT, using generative AI in software development, and the limitations of early models.
• [16:47] – Christoph discusses AI-powered engineering and its future in software development.
• [20:45] – Christoph highlights the importance of providing users with easy-to-use interfaces and the potential for AI to assist in tasks such as writing tweets or emails.
• [25:32] – Christoph emphasizes how AIPRM can provide independent assistance in selecting the right model, prompts, and community for successful experiences.
• [33:39] – Christoph explains AI tools’ moderation, creation, testing, quality assurance, and prompt drift detection are crucial tasks that cannot be automated.
• [43:54] – Christoph discusses the importance of understanding the user’s perspective when leveraging AI in digital marketing.

Jump to Links and Resources

Christoph, it’s such a pleasure to have you on the show. Thanks for coming. 

Thanks for having me. It’s been a while, and I look forward to our chat because we have known each other for quite some time. I look forward to talking about AI, which has been a strong force not just in the market but also for myself to learn and educate.

Your roots are in software development when you were a kid. Can you quickly share the story of how you started to write code and all that? 

It all started in 1985 with a Commodore 64 for computer games. I got it for Christmas, and that night, I already wrote my first basic program because I learned and practiced that on other computers back then in computer stores after school. That is really something that I enjoyed my whole teenage years. Not just as a hobby, but that became a profession with a partner a little older than me and kicked my ass to get things done. I mean, finished. 

Getting to 80%, 90%, 95%, or even 99% is possible, but the hard part is finishing up and doing the sales. Back then, we finished games and sold them to publishers, which meant quality assurance, testing, contracts, and other stuff that were not as fun as creating code. 

Tailor AI-generated content to your specific needs, from product names to trademark spellings. Don't fear personalization; embrace precision and efficiency in your content creation. Click To Tweet

It’s funny that you mentioned getting 95%, even 99%. I just learned this concept, and it’s a game-changer. It will change my life. I’m quite certain of it. It’s called the spiritual flywheel; you can apply it to business. The person I learned this from is David Ghiyam, who’s a very successful entrepreneur. 

He and his wife own MaryRuth Organics, a big brand with 25 million customers. It’s very popular on Amazon. It’s in all the different kinds of whole foods stores, etc., here in the US, and they’re in other countries as well. He’s very successful. 

When he talks about concepts relating to Kabbalah and business because he is a Kabbalah teacher and a Kabbalist, I definitely pay attention. This spiritual flywheel is where you get momentum because a flywheel, when it really gets going, keeps going. The momentum keeps it going. It’s hard to get it started.

He talked about how getting from, let’s say, 0-95% of a project complete, it just seems to work, and you don’t get really waylaid or sidetracked in the same way as you do in the last 5%. The dark forces all work to sidetrack you in the very tail end of the project because that’s what all the juice is. 

He is like, “If you gain 95 units of energy, you have light from the heavens for getting to 95% of the project done.” So, 95 units of energy for 95%. It’s not an extra 5 units of energy to get to that remaining 100%; the last 5% is done. It’s like 10,000 units of energy. 

It’s measurably powerful just to complete stuff, even if you’re not passionate about it if you’re most of the way done. Another way to think about it is you can’t drive traffic over a bridge that is 99% complete. It has to be fully complete to safely put cars on that bridge. Just finish the dang thing. 

If you don’t want to finish it, don’t keep it hanging around because that sucks energy. All these uncompleted projects suck energy, so you just kill it. You never think about it ever again. You have to make a decision, and that’s hard. 

AI is a tool to enhance your writing and productivity, not a replacement for your skills and expertise. Use it as a collaborative writing assistant. Click To Tweet

I’m a kind of a hoarder when it comes to projects and ideas. I have a whole 6000 to-dos on my to-do list. It’s pretty insane. It’s draining my energy instead of giving me energy because I don’t finish these things, and that’ll be a game changer for me.

This spiritual flywheel. Getting in there and finishing stuff that is mostly done, but then the things that I’m no longer passionate about, I completely kill them and delete them from my to-do list. That’s going to be a game-changer. What are your thoughts on that? 

I remember the whole GTD thing and to-do lists using approximately 30 different to-do apps, project management apps, or project task trackers like IT software development. Jira is very popular. Of course, in bigger teams, you have to have a way to take all these notes and manage them. 

The truth is, after some years, it just becomes a huge digital symmetry. It’s complete trash and takes so much energy to find the good stuff there that I killed off even the last software development project management tool, uTrack. I mean, not kill off. We still have it as an archive to search for why we decided that way. 

For two years or so, we skipped all that and went back to simple Slack, which is not a project management tool, but naturally, it makes things that are not as important disappear. Things come back to you if they’re important.

It’s like a to-do list bankruptcy, right? 

Exactly. I don’t have a to-do list except when, let’s say, I have to finish my books in accounting, and I have five items where you have to find the receipts or stuff like short-term notes, but not the to-do list for my life, not the to-do list for my business. That’s all a fad, and I wasted so much time. 

Inbox Zero is just the next fad, just like that. Of course, you can try to get there, but what’s the reward? You can try to answer everyone quickly. Why? It wasn’t like that in earlier times. 

In a way, all of that takes energy and does not reward sufficiently. As you said, those last 5% take maybe 10,000 energy units. But that is 100% true, and what you need in this phase needs so much energy because it means saying no all the time, not just to your employees or customers. “We want this feature and that feature.” You have to say no to yourself. 

You have to cut back on your own requirements and standards. You cannot get finished if you want to have everything perfect.

You have to cut back on your own requirements and standards. You cannot get finished if you want to have everything perfect. The perfectionist is the opposite of getting done or whatever that saying is. You have to make those sacrifices.

Perfect is the enemy of done. 

Exactly. That is painful, especially if you enjoy delivering good work. Some of that stuff is not good if you get to 100%, not good by your standards, but the solution is discovering that only you see the difference. 

You show this to random people, you show this to customers, you show this to your boss, and they don’t even notice the difference, maybe don’t even look at what kept you up at night. What is the difference? Unfortunately, many people do not switch to change it to “this is good enough.” It’s the good enough mentality, which is not delivering crap or subpar result. It’s about letting go of your thoughts about what must be done. That’s why a to-do list is so tough for getting anything out the door. 

I remember Tony Robbins explaining that the perfectionist has no standards. Having impossibly high standards is the same as having no standards at all. The reason why is that you get nothing done either way.  

The spiritual flywheel thing is that it takes discipline and energy to get to the finish line, to fully complete the project so that it ships. It needs to ship. If it doesn’t ship, it doesn’t count. You have to cross that finish line, and if you don’t cross that finish line, you’ve expended all this effort, and it’s still draining energy because you haven’t completed the task or the job you set out to do to actually ship the product. 

Once you do, all of that energy releases like it’s a boulder at the top of the mountain, and you just gave it that push, and now that potential energy stored in the boulder has turned into kinetic energy, and you are on fire. You are just unstoppable. 

Let’s talk about AI. I don’t want to get too far into philosophy. Let’s talk about AIPRM and how that connects to your roots as a software developer. You’re a successful entrepreneur and have a great team, a great reputation, and a great brand. I’ve been referring folks to Link Research Tools for over a decade, and wow, what a significant shift or pivot to go after this whole ChatGPT thing with AIPRM. How did that all come about? 

We started with the generative AI more than a year ago. On April 22, GitHub Copilot came out, and we tested it. It was free, and it was magic. I couldn’t wait to pay those $10 because of whatever limitation. I think it was too slow back then in the beginning. We used it in software development from day one. 

While it was based on all the Codex models from AI, the results were exciting already with the suggestions and the feeling that you suddenly have a pair programmer with you, that you have someone suggesting things, was irritating me. You type some code, and then you get this little blurry suggestion that, in the beginning, was annoying as hell to me. I was close to turning it off, but I went through that, got more productive, and used the same APIs for some SEO-specific tasks, some libraries already built, some tools, and some functionality in Link Research Tools that were closed to shipping but not done yet. Parts of the user interface were missing. If a normal user can’t use it and needs to call me to run it on the server, it’s not done. It’s not 99%, it’s 80%. 

That’s when ChatGPT came out, and I used it for a month or so until the end of December, along with Midjourney. At the beginning of January, I had this collection of prompts in text files. I had dozens of source files with my prompts in there. I saw all these AI influencers, so to say that, were pitching and selling their prompt lists, Excel lists with each prompt being one or two sentences, something like “Write me a book that I can sell on Amazon,” not a specification of what needs to be done, so not a real prompt by my standards. 

I was wondering what the fuss about all that is that they went through the roof, and I started thinking to myself, “Hey, where are the people that work like me that use prompts of 20, 50, or 100 lines of code because prompts are the new source code in my opinion. Where are these people? They’re not on Twitter. They’re not on LinkedIn. When I post something on Facebook, I get two likes. Where are all these people?”

AIPRM isn't just about prompts; it's a thriving professionals' marketplace and community where experts come together to shape the future of AI. Click To Tweet

That’s where the idea of having an AI prompting community, an AI prompt marketplace, or a professional marketplace, something where people like me would share their prompts. Nothing like that was around. And so AIPRM had started on January 8th, with 10 hardcoded prompts, my favorite SEO prompts in the SEO space.

Three days later, when it had 20,000 users, I realized, “This was serious stuff, and this could work, so we built that database functionality. The whole marketplace is running on a large server cluster now, essentially a Reddit for AI prompts with voting, downvoting, blacklisting, and duplicate content detection like in Google. All these things that we know from search engines. It’s a search engine for prompts as well. 

It was around mid-January already, and this is it. People came, and people writing good prompts published them just like I did. Of course, some copycats of people also jumped on that idea, but nobody really got enough. 

You were a fast mover, for sure. Where do you think this is heading? With prompts being the new source code, as you said? What about tools like GPT Engineer, which will write whole software packages? I don’t know if software developers will have a job in two years time.

When you run GPT Engineer, you can give it a better or worse prompt.

I thought about that a lot. I thought about what I should tell my kids to study or what would be the education, what would work for them or anyone really in our industry or all the smart people, the white-collar workers, so to say. The prompt development here is really just a bridge. 

When you run GPT Engineer, for example, you can give it some better or worse prompt. You can give it some specific details, like write 10 or 20 lines of specification, or you can actually give it something like, “I want the computer game with asteroids that I want to control with my mouse.”

When you do that, what it does is it generates a questionnaire. It prompts you and asks you questions that it thinks could be relevant to the problem. It’s like a prompt generator or a prompt development tool. I have to say I really like it because it reminds me a little bit of these old computer games where you had to type in instructions on the Commodore 64. “Should we have this? Should we have that? What do you think about these 17 options I came up with?”

It’s really fun to work with it, but it’s still so much work to specify what you actually want. There’s a bunch of people who spread the idea that prompts are a temporary thing and that they won’t be necessary in the future. My take on that is there will always be a requirement for someone who wants something to specify that in a clear, workable, executable way. 

“Write me a book that I can sell on Amazon.” It’s not that class of specification, so either it’s guided requirements analysis, or someone who can collect his requirements to abstract problems can describe this as a prompt engineer, as we call it today, or some system needs to collect this information. This system can be machine-driven or a bunch of requirements engineers who still ask customers, “What do you want to do today?” And that can be semi-automated. 

In the end, even without an AI, even if you have humans sitting there, you need to say what you want and say it right. It needs to be a blue car. Otherwise, you get a red or green one, and then people are all frustrated about it not being how they imagined it. The transformation from our requirements to our wishes, from our dreams that we want to achieve, cannot be removed.  

When you look at some systems, components, or technologies, and you know something about the flaws there, you have an idea of what a problem could be that you fix.

Midjourney does it in an interesting way. You type some words minimum, and you get four suggestions. You get four images, and you let people pick. That’s a bit like GPT Engineer, AutoGPT, or Baby AGI; they all have these loops, and now the code interpreter from ChatGPT can do that as well. It can ask you all these questions and, in an iterative fashion, get to a better specification of what needs to be done and what the output should be. 

We would never be happy if we could just get some output. That’s a problem with all the ChatGPT haters. “Oh, I typed in my problem, and the output was terrible.”

There’s an old adage in programming: garbage in, garbage out. 

Yeah, exactly.  

What will AIPRM do to stay ahead of what’s happening with GPT Engineer, AutoGPT, and all these other things? It’s such a moving target.

It absolutely is, but we’re in the best position because we have all the people building the prompts, revising the prompts, and all the lists I mentioned in the beginning. They’re static, they’re a week old, they’re two weeks old. We not only see the usage of what people are working on, I mean not the actual chat content, but we see the prompts they’re using. Do they work on marketing personas? Do they work on programming? 

We now have 1.6 million or 1.65 million users. We have a huge traffic influx of these stats. We see what people are doing and how often they’re doing that. 

While all these automated tools are quite interesting from a technical perspective, very little of that is currently visible to our users. To keep up with that is number one: learn about it all, use it, and fight the flaws. Usually, when you look at some systems, components, or technologies, and you know something about the flaws there, you have an idea of what a problem could be that you fix.

The original AIPRM was the problem of me sitting there and thinking about it. I got this magic machine, this ChatGPT machine on the other end, and here I’m handling a bunch of text files in my text editor. This can’t be real. I’m replacing variables in there by hand. That’s a simple UI feature from the beginning. What we see right now is that AIPRM users have very basic needs for customizing the output to their business, professional, or product. 

For example, we have this feature where you can give AIPRM a profile or multiple profiles where I could work as Christoph personally. I could work as Christoph for LinkResearch Tools. I could work as Christoph for AIPRM, a company not in Austria, Europe, but Delaware. I founded in 2023 with this and this and this. 

AIPRM has other plans and backends, but there are a couple of tools where you could give an API key or connect to some.

I can now switch all this context information with one dropdown and pass this on to anything I do with ChatGPT and any prompt, which means it will never come back with generic answers when filling in a company name or my name.

When you work on emails, you don’t get something you still need to post-process, but you already get detailed, targeted, personalized results. This is what many users struggle with, still struggle with, or don’t know they can do. It sounds scary to give more personalized information, data protection, personal information, etc. But we’re talking about product names, we’re talking about abbreviations, we’re talking about your actual precise spelling of the trademark that needs to be used. You don’t want to edit all these details in the output. 

I’m passing that to ChatGPT, and I’m saying ChatGPT here because it’s the big one right now. AIPRM has other plans and backends, but there are a couple of tools where you could give an API key or connect to some lesser-known engine, and that’s all interesting, but there are no users there. 

I mean not the broad users. We’re targeting small businesses, not AI engineers. We’re talking to lawyers, dentists, and plumbers doing their work, so they don’t need to have the web girl or the social media guide do it? They’re just writing their tweets or stuff like that.

That is where I still see a huge potential. With all the developments, there’s a huge potential of people being scared, like, “I’m a lawyer, so I must be a good writer. I don’t need someone else to write for me.” But something similar was thought when spell checkers were introduced in our text processing systems, correcting the grammar, correcting the Oxford comma, etc.

Back then, in the ’90s or ’80s, that was probably scary, too. The transition hasn’t happened yet and won’t happen very fast. People will slowly get accustomed to the power of AI helping them just like text processes, correcting or expanding that. 

Do you think AIPRM could be essentially like Grammarly tagged onto  Microsoft Word, but you’re attached to whether it’s ChatGPT, Google Bard, Bing Chat, or whatever? It’s that essential or near-essential layer that helps the marketer or small business owner leverage the power of ChatGPT or whatever generative AI platform.

We also launched at the end of June, shortly before the holiday week, AIPRM everywhere, where, at least in the browser, you can use your favorite prompts with just a right click or context menu click. You select some text or activate the live crawl and TL;DR file prompt or function to crawl the page you’re on and give you a five-bullet summary of the highlights in there. That’s there already. 

That reminds me of the Ghostreader feature in Readwise Reader. It will summarize the page that you’ve added to Readwise Reader if you’re familiar with that read-it-later app.

I don’t know that app, but it makes total sense, and it’s something that you might want to automate for everything if you have a read-it-later app. Again, we’re talking about the average show that doesn’t have a read-it-later app with AIPRM next to Google Translate. Just like you mark some text in Polish on a website and translate it to English with one click, you do that with AIPRM already. It could expand it, or yeah, translate it, or everything that chats can do, or other LLMs will be able to do so.

We are running on ChatGPT now because that’s the huge viral platform that everyone is on. But AIPRM is more than that, and we have already shown that even inside ChatGPT, we’ve got GPT-35, GPT-4 with code, and GPT-4 with browsing that they took our offline again. Having a good prompt requires the prompt but also the right model. 

Learning and testing it in the future by automatically testing the quality of all these prompts against all these different large language models is what AIPRM is built for.

Learning and testing it in the future by automatically testing the quality of all these prompts against all these different large language models is what AIPRM is built for and what we are doing already in a way that gives users the chance to have a successful experience. There are a lot of prompts if you run them against the wrong model. The output is just not good. 

This reminds me of a tool I’ve used to get health and travel insurance. It’s called Squaremouth. If you go to, let’s say, a travel insurance website, they’ll give you a quote, and you can buy your insurance. If you go to Squaremouth, it will show you all these reviews and ratings that you can filter and make different decisions about what coverage you want. 

Does travel insurance require, let’s say, baggage insurance and medical insurance? Do you need to be escorted out of the country to a more suitable hospital if there’s a big problem? Is that included? To what limits? You can make all these decisions around what your coverage requirements are, and then it will narrow it down to just those insurance options that meet your criteria, and then you can buy the insurance right through their platform. 

I see AIPRM in a similar way of being able to provide that independent assistance and selecting the right model, the right prompts, and the right community. That’s cool.

The community aspect where the authors have their profiles in AIPRM and then run their own Patreon for users of their prompts is a sub-community, or on Reddit would be a subreddit. We haven’t supported that in any structural fashion yet, but it comes down to the question of: How many prompters in the US are good prompt writers today and are proficient? 

We have some really smart guys in the customer base, but it needs momentum for every vertical or every industry. AIPRM would provide that space for industry leaders or experts to fine-tune that for their community or industry. That’s something that I’m thinking about. 

That’s a great idea to allow a prompt engineer like a solopreneur to set up shop and turn that into a money-making venture without having to go onto a freelancer website and offer their services. They could get supporters like Patreon has. 

There’s a prompt marketplace where you can buy a prompt for a couple of dollars, and others are picking up on the idea that someone would buy a prompt here and there.

Yeah, exactly. We have things like that. There’s a prompt base, a prompt marketplace where you can buy a prompt for a couple of dollars, and some others are picking up on the idea that someone would buy a prompt here and there. That’s the reason why I built AIPRM. The prompt source code is valuable, but you don’t want to engage in these microtransactions. People need more than that, and we’re working on providing not a service but something of a lot more value than just having a prompt. 

As you can imagine, we got a couple of people copying our stuff and cloning the whole thing. On Apple, there’s a guy who thinks copying our logo one by one is not a copyright or not a trademark violation and things. We got a lot of stuff going on from these free riders, which the lawyers are happy about, and when they get taken down, they’re not so happy about that. 

That’s not all that you see. I envision AIPRM as being a professional’s marketplace or a professional community. It’s not just about the prompts but about the people doing stuff in there, and that’s developing quite nicely. I can see that getting more relevant and important over time as more average shows, as I call them, take that step to try things. 

We’re all geeks. A lot of the users of AIPRM are geeks and technology people, marketing people who are used to trying stuff. But I want to enable this conversion from a problem to a solution via a prompt for normal people who cannot express themselves well enough. If you want a prompt generator as well, but you cannot solve everything with code. My Expert GPT is one example that tries to come up with an approach, but I can guarantee there will be a few plumbers sitting there for half an hour to tell the machine how it should do things. 

You cannot shortcut this collection process. It’s cool that you can do this, but nobody sits in front of a computer for 30 minutes to tell it what the output should be. This whole moderation, creation, testing, quality assurance, and detection of prompt drift is our job. You cannot put this in an Excel sheet or some machine. 

This is the first time I’ve heard the term “prompt drift.” Can you define that for our listeners? 

Absolutely. Prompt drift is a term coined that describes the problem of writing a prompt or the language model you prompt with behaves differently over time. For example, ChatGPT had a release in March, then an official release on June 13. But when you open ChatGPT in July, you still see something about a date in May. 

Generative AI needs clear, executable prompts just as software needs precise code. Click To Tweet

I still don’t know which model they have underneath ChatGPT compared to the models you can run via API, but they all sometimes provide very different results for the same version. Nobody has this under control, and nobody talks about it because it’s rendering the product, a large language model, which is very risky. 

When you develop software and use a large language model, you develop prompts for it. If a new version comes around like that recent OpenAI, they announced they would turn off the old model because it’s very costly to run and keep. It means you need to fix your software. You need to adapt your software. It’s like someone pulls the rug under your software.

If you had based your marketing strategy on Google+, then Google comes around and says, “We don’t want to run a social media thing anymore.” It’s not that harsh, but it’s close to it. I heard some cases of really frustrated programmers or developers in companies.

Prompt drift describes how the output drifts away from your original one, changes, or gets broken. Just today, I was testing a complex thing against the OpenAI API, and it came back with unusable gibberish stuff. It’s not written differently. It broke suddenly. 

This whole moderation, creation, testing, quality assurance, and detection of prompt drift is our job.

That is something nobody has under control right now, not OpenAI, and nobody’s monitoring that except maybe AIPRM. When we do that, we cannot only tell our users that this prompt is not up to date. We could even take them down like we do for anti-spam. We also do this for anti-prompt drifting, so to say very soon, which is not easy, but somebody has to do it. 

That reminds me of something you created a long time ago, a metric in LinkResearch Tools called Link Velocity Trend (LVT). That was such an important metric to add to the SEO community, to the discipline of SEO, so that people could see how their link velocity was dropping off and how the search engines almost certainly were tracking that too and seeing how you were no longer relevant to the global conversation anymore like you used to be. 

Now, with prompt drift, by measuring that and having a metric for that, you could be the leader in this generative AI field in making sure stuff still works and doesn’t break. It’s like Windows 95 suddenly gets deprecated, and your whole software is written on Windows 95, and it doesn’t work on the new version. Your whole business is in trouble. 

Exactly. This is a problem developers have today for their prompts, and the systems users will have as well. There was some personalized AI product where people could have fake friends or AI friends or something like that that changed overnight where some people essentially lost a friend, where the way it responded to them based on the history, based on the time invested with this AI friend seemed lost. 

This could have been overly dramatic, and I didn’t follow through with all the details on that one, but it comes down to the problem that underneath the database was changed. It’s just switching from one database to the other without being able to test all the millions or trillions of possible outcomes.

I don’t know if OpenAI has this on the radar, but because there’s no solution, they might have it on their radar with a “don’t talk about it” thing, just like Google doesn’t want to talk about links. It’s still the one thing that matters so much and is easy to manipulate. 

Prompt drift describes the problem of writing a prompt, or the language model you prompt with behaves differently over time.

When I say easy, I mean it’s still possible. It’s not like 20 years ago, which would be surprising, but it’s still the one thing that matters on the web: content and links. When you have content and links, you can look at a link graph. You can look at how many other websites or people point to it.

Of course, if there are spam links, you need to discount them. That’s not innovative in 2023, but the large language model operators also crawl the web. Now, what do you think of how they are scoring their documents? Of course, they can calculate many in-document metrics like TFIDF, a document where word frequencies and things like that can extract some entities from there. 

When it comes to the question of how popular this document is, what are you going to do? They may be using the page from the 1998 formula at least, but somehow, this machine needs to get some input on how important it is to be able to give it more weight. All these weights are similar waves needed for a search engine.

If you want the ChatGPT or all these other large language models, they have a different form. It’s not like a search index with an inverse index on the documents you can search by words, but the original input is the same: the web and information by humanity. Maybe Midjourney scanned all the comic books out there and maybe gave it some deviant artwork websites to suck in with a lot higher priority.

Maybe OpenAI or ChatGPT did some manual tweaking to give government sites and edu sites more weight when it comes to the trust that they give these documents. But these concepts of what is probably the right information cannot be distilled out of the information itself. This is something I don’t see anyone publishing because, on the one hand, it’s a trade secret of how you do that. 

It’s a very important trade secret, just like we don’t know the exact filters between the OpenAI and ChatGPT, having all these security measures when it says I cannot respond to you on legal or health topics. It’s the same thing that Google has to do with the data. 

In a way, it’s very similar to all my work on link research tools when it comes to trying to find out how to reverse-engineer it. Prompt engineering is something like that. Change the input a little bit and see what comes back. In a way, this reminds me of SEO in 2003, where whatever you did on the website, whatever you did with links, you could see results, sometimes in hours or the next day. 

Yeah, you could reverse engineer what was working and what didn’t. 

We are there now with AI models if you want. This will not remain as it is because it’s too easy to see and find out.  

To reverse engineer, they need these platforms to keep it a black box to create a moat, a durable competitive advantage. 

Yeah, which is interesting because Facebook, the LLM, leaked. It got stolen, and then the next day, Mark Zuckerberg woke up and said, “Okay, we’re going to open source it all, and it’s all free for everyone.” It’s an interesting reaction that is the only possible because everybody had it on his heart is go radio from the torrent. Everyone has been using it anyway. 

What they did not open source is how they got there. How they trained it? That’s maybe, you know, where they said, “You can have this, and this is from 2020. We got our own better one. Let the academic society and the open source play with this while we learn from their findings and use our modern ones.” Maybe something like that will make sense.

When you develop software and use a large language model, you develop prompts for it.

Interesting times, for sure. We’re out of time, so I want to ask you two quick questions before we wrap. One is, is a ChatGPT API key required to use AIPRM? And the second oneis, there a way to feed your database of all the stuff you’ve created? The listener, the viewer, can they upload that somewhere, still have it somewhat private or hopefully not getting leaked by OpenAI, and then have that be the basis for a whole bunch of output from ChatGPT? 

You don’t need them because you are just using the standard ChatGPT. There will be other used cases in the future where you need an API key, but those are not released yet. When it comes to personalization, the data that you pass to ChatGPT today is still very limited. But of course, having a whole repository of all your company documents or contracts is the use case that we are prepared to ship and deliver. 

With product development at that scale, it’s really hard to ship things in a way that doesn’t break the user experience. For example, it takes three weeks until everyone gets a back fix that we do in the extension. With that being said, every move we make is a lot riskier. Having to wait for weeks until everyone has a bug fix and answering support questions for three weeks for what you know is already fixed is really annoying, but that’s how these extensions work. 

I have some solutions for that, and we’re working on something like that precisely as you described it because that’s the missing piece of the puzzle. When it comes to data security, having run everything out of Europe with GDPR and running data centers in Germany, I guess I’m best prepared for compared to some competitors sitting somewhere in Alabama or so.

That’s something that I want to see in AIPRM very soon. Many users ask for that, and they will get it because that’s where we’re going. I just don’t have all the features and all the details. Also, because OpenAI is changing at a rapid rate, we’ll get there eventually.

Yeah. That’s awesome. Congratulations on all the success and over a million users of your platform. It’s incredible. It’s an inspiration, so congratulations. For our listener who wants to download and install the Chrome extension for AIPRM, where do we send them?

Prompt drift describes how the output drifts away from your original one, changes, or gets broken.

Go to aiprm.com, and you should have a really big install for free button, which is important to understand that all 1.6 million users can use all the prompts for free. You don’t have to pay. You don’t have to sign up. Only when we talk about more advanced features do you need to sign up so we can store data, and even more advanced features cost a couple of dollars. The majority of the users are still using it for free. 

I don’t plan to change that. If you have more than 3,000 prompts to select from, that’s plenty for most. That’s it. 

Awesome. 

aiprm.com and install. A second later, you just need to confirm one, “Do you want to install this thing or so,” and then just open ChatGPT and click some buttons.

And also linkresearchtools.com. That’s also an important tool for the toolset if you’re doing SEO. I’ve been a big proponent of Link Research Tools for over a decade. Congratulations on that success, too. 

Thank you. 

Alright, thank you, Christoph. Thank you, listener. We’ll catch you in the next episode. I’m your host, Stephan Spencer, signing off.

Important Links

Christoph Cemper

 

YouTube Videos

 

Previous Marketing Speak Episodes

 

People

 

Further Resources

 

Your Checklist of Actions to Take

Understand the power of AI prompts. Recognize that AI prompts can be a game-changer in various aspects of my work, from content creation to software development.

Prioritize clear specifications. When working with AI prompts, I need to be specific in giving instructions. Provide clear and detailed prompts to get my desired output.

Adapt to AI-driven changes. I should remember that AI tools are continually evolving. Stay informed about the latest developments and adapt to changes in AI technology and applications.

Embrace AI as a writing aid. Recognize that AI is a tool to enhance my writing skills and expertise. 

Personalize my AI experience. Recognize the importance of personalization in AI outputs. Tailor prompts to my specific needs and add context to improve the relevance of AI-generated content.

Address AI prompt drift. Stay vigilant about prompt drift — the phenomenon where AI models produce different or unexpected outputs. Understand that prompt drift can impact the reliability of my AI-generated content.

Monitor prompt quality. Monitor prompt quality and test prompts for multiple AI models to ensure successful outcomes.

Beware of database changes. Database changes can lead to unexpected issues in AI systems. It’s essential that I monitor and measure how these changes affect AI outputs.

Leverage open source opportunities. Consider open-sourcing certain aspects of AI technology while retaining crucial proprietary elements. This approach can help me foster innovation and community collaboration.

Remember that AIPRM can be an essential tool. Think of AIPRM as a valuable addition to my toolkit to enhance my productivity when I work with generative AI platforms like ChatGPT, Google Bard, or Bing Chat.

About Christoph Cemper

Founder of AIPRM, LinkResearchTools, Link Detox.