Don’t let an algorithm decide what you read »

Leveraging AI in Expert Businesses

October 18, 2023
The Recognized Authority Podcast Cover

The podcast that helps experts & consultants on the journey to becoming a recognized authority in your field, so you can increase your impact, command premium fees, work less hours, and never have to suffer a bad-fit client again!.

I’m excited to share a re-post of an interview I recently did on the unbillable hours podcast with Flo & Ash. We had a fascinating discussion about the practical uses of AI tools like ChatGPT in content creation and marketing.

In this wide-ranging conversation, we cover when and how to effectively utilize large language models like ChatGPT and Claude. I share my perspective as someone who has been experimenting daily with these tools over the past year.

Key highlights include:

  • Why it’s crucial not to blindly copy/paste AI-generated content without careful review
  • Examples of using AI for brainstorming ideas and restructuring draft content
  • Tips on crafting effective prompts to get quality results from the AI
  • How subject matter experts can refine AI output by adding their unique perspective
  • The importance of iterating with the AI to improve results over multiple rounds
  • Potential applications like summarizing meetings and generating proposals after client calls
  • Exciting possibilities like simulating expert panels to spur creative thinking

Florian, Ash and I also discuss concerns around data privacy, misinformation and hype versus practical use cases. If you’re curious to learn more about how AI can boost your productivity and content quality when applied judiciously, have a listen. I think you’ll find this dialogue enlightening, as we cover both promise and pitfalls. Let me know what you think!

Show Notes

unbillable hours is podcast about professional services marketing. Stick around and listen to our insights, tips, and best practices to improve your firm’s marketing and even – your career!

Guest Bio



ai, prompt, output, gpt, otter, claude, write, podcast, pasting, tools, talking, expert, transcript, people, conversation, blurb, playing


Ash, Florian, Alastair McDermott

Florian 00:00

Alright listeners, welcome to this episode of the unbillable hours podcast. Ash, myself, and Alastair McDermott who’s with us, whose idea was a friend of the show and host of the 40. podcast, I have to say that again, because it’s The Recognized Authority podcast. The Recognized Authority by guys, there we go. And and just by large, also in a space we like, which is you work with smaller consulting firms and single solid consultants, expert businesses, in the marketing business development space, correct. Understand? That’s right, yeah. But we invited you here, because you are also you, I’m not gonna say you have made a name for itself. But we’ve seen a lot of great content coming from you about the actual practical hands on use of artificial intelligence in that space, right in what we would think of as consultative marketing. So we wanted to have you on because we still have no clue how to actually use it in ways which makes sense. And I think it’s timely because we discussed this before I switched on the mic. This stuff like Gen AI, these tools, and their impact on the profession of consulting has been on the news more recently, right, with studies like the BCG, self study coming out and all these things. So So it’s about time, we talked to someone who’s been working with us all along is that’s my lengthy introduction. Do you have anything to add? Did I miss anything? I come to miss anything.

Alastair McDermott 01:27

Thanks for having me on the show. No, I just want to say like that, the way that I think about using AI is not to generate crap, because I think that a lot of people are generating crap. And that’s what we’re trying to do. Okay, I hope you don’t have to believe that.

Ash 01:43

That’s doing work for the moderators, but by generating crap, though.

Alastair McDermott 01:50

Yeah, but it’s to like, I’m all about helping people to build authority and become known as The Recognized Authority in their field. That’s why I called my podcast The Recognized Authority. And like, we don’t become an authority in our fields, by copy pasting rubbish output from chatty PT and just pasting onto LinkedIn. Like, that’s not going to work. So yeah. I think that we can use these tools and like we’ve seen from the BCG study, and just if people aren’t familiar with that, like it found Ethan were like, did this study with with with all the consultants there, and they found that it massively improved, increase their productivity, and the quality of your output? I think they said, increase the quality of output by 40%, by people who were using it. Like that’s a huge increase in quality of work as well. So it can it can increase the quality work. That’s, that’s my message. But you’ve just done it. Yes. And

Florian Heinrichs 02:40

the question we have to you, how exactly dammit, because what you just said, right, that there’s these misconceptions of what do I just take output? Or how to even use it? Well, maybe we can recognize where do I use it? Right? Do I write my emails to a lot? Right? My thought leadership does it right, the strategy? So I think these these are questions people sometimes are not clear about and that’s, that’s what I’ve been seeing here and there is that what AI seems to be doing in some firms is, first and foremost, create confusion in my work, because there’s an implicit or even explicit expectation of the marketing colleagues or the consultants for that matter, to use it. But you know, without proper guidance, it just becomes one more thing on your to do list. And with all that mass amount of tools, so all seem to be doing similar things. equally well, or definitely, well, no, it’s just why don’t we just start right? So that’s, that’s maybe the first question we have, can you can you walk us through because you’ve used this stuff? And you’re working with it? Can you can you share your highlights? Where Where have you found the most value in those in those tools?

Ash 03:43

Maybe we can also structure it. So ideas tool as intended and tool less help people working to help us?

Florian Heinrichs 03:51

Yeah, nice. Yeah. Cool. Well,

Alastair McDermott 03:54

I mean, these I mean, the tools that I’m using on a regular basis are Chachi PT for Claude, which is a GPT competitor. And both of those have billions in funding. The reason why I like Claude is the input token size is much bigger, which means it can take like 75,000 words of text as input and process based on that, which is way more than Chachi, pte, it’s 20 times more or even more than that, I think. So that’s why I like GLAAD. And they used a lot regularly. So the the thing that I find that most useful for are that just a few things, there’s brainstorming ideas, and restructuring text, those are the two most common things that I will do with with GPT. And what I find is I’m using it to make me more efficient, and make make my time more productive. So the the time that I’m spending on something, I’ll spend the same amount of time but they’ll get way more done. And so and it’ll be higher quality. And like I know there won’t be typos and grammatical errors and things like that. So that’s where I’m finding it really super help. For now, one of the things I think that you need to do if you want to get, if you want to get good at using these things, you want to understand how to use them. There’s no manual, even the people who created the system don’t really know how they work and how they can be used, because these are general purpose API’s. So you’ve got to just dive in and just start experimenting with them. You know, just play around with them. Ask them to do things and see what happens when you ask it. Try the same prompts. Try the same question in Bard, and Bing, and Claude and chat GPT and see what the answer is look like and see, which is most useful for you. And, you know, just play around with all of the different things that you can do. Like, for example, I did a survey of 1000 consultants three or four years ago. And I have spreadsheets full of survey results. So I took those survey spreadsheets and they export them as CSV. And I just copy pasted that CSV text into into I think it was Claude because it was bigger. And so I just said, Please pull up pull out the key insights from this, this research data. And so it started processing and pulled out a whole bunch of really interesting things. And I was able to, so I was just playing around with that I didn’t have any particular purpose with that. But I think that you have to do that you have to start playing around and learning how to use it yourself. Like there is no manual out there. Any training course that’s out there, anybody who claims to be an AI expert, I guess there’s kind of two two types of AI expert. There’s AI experts in terms of people who are creating these systems, and to kind of have better understanding of what the system do. Yeah. But even they don’t know much about how we use AI. And so then there are people who are expert AI users, and expert AI users are like blockchain experts, when blockchain came at first, like it’s just people on the edge of the wave. It’s not it’s not that they know much more than anybody else. It’s just that they’re playing around using these things.

Florian Heinrichs 06:47

Very diplomatically. So I want to call out something though, because I think I understand something. Alright. So you started in saying we don’t want to just generate crap output and use it that was like that. Then you went on off saying using Bart because it takes 75,000 characters a word long inputs? Words? Yeah. Yeah. And then you said to use it for brainstorming and for restriction of taxa. What I’m hearing is you’re actually using these machines to help you refine stuff, where stuff is something that had you put in considerable work and thought and so forth. upfront, right? Yeah. Somewhat. I only someone who has written thought taking notes substantially about a matter cares for whether or not the machine takes 75 words, the simple it’s right. So you’re not starting with please write me a LinkedIn post for Thursday. That’s not working. And I think that’s, that’s a compelling insight and very important already. Because it’s the opposite process, right? It’s you’re not using the generative AI to generate something from nothing, or cliche or whatever the bot finds on the web, but you have something like your research data. And now you use it to almost like an editor of sorts, right? To help you pull out interesting things or refine the stuff you’ve written? Would that be a fair description?

Alastair McDermott 08:01

Yeah, my friend describes it as a a high functioning, a high functioning intern who is occasionally does acid, because it’s really good. It’s not very experienced, it’s, it’s really good at processing lots of stuff. And then occasionally, it’s going to hallucinate and give us just like something that’s totally weird. So I think that, you know, understanding that the AI is not trustworthy, like you can’t trust what it gives you back, when it gives you back is probably going to be really good. And then occasionally, it’s going to just go off and give you something totally weird. Like, I’ll give you one example. I was processing a podcast transcript, which was conversation where I actually did a coaching call with a client and we recorded it as a as a podcast episode. We were discussing her live streaming on podcast strategy. That’s what we were talking about. And so I was doing some stuff with the transcript saying, you know, he pulled out the key insights from this. And is there a step by step process that we talked about, and just pull pulling out some things? And then he gave me a really weird answer. And I said, Can you describe the conversation that we’re talking about here? And so chatty, be touchy, but he came back and said, Yes, this is a conversation between Richard Nixon and one of his advisors about the Watergate scandal in December, blah, blah, blah. And it was just like, off the wall, like, where did it where did this come from? Like, ended up absolutely nothing. So the conversation that we were talking about, about podcasts and livestream strategy, and that just sometimes happens. Now, I don’t know what happened, you know, like, did a wire get crossed? You know, sometimes, like, maybe there’s this memory buffer and it gets this like buffer overflow or something. I don’t know where this came from. But this can just happen. And if you are blindly using these systems, without doing quality assurance yourself, if you’re not reading the output and going through it yourself, then you run the risk of taking something and using it on your own website or using it in client work or something. So you got to be really careful with these things. The like, you cannot trust them 100% They are very, very good. But you just always have to be careful by using them

Florian Heinrichs 10:06

at the risk of offending real interests on real acid. I liked the analogy, because I think that this was like the first step, maybe sort of using it more for refining things done for or putting your own thinking and effort in first right now, that’s, or at least go back and forth. I think that’s when you said brainstorming, though, you might have the suggestions, but you have to do the work to process the threat and then maybe prompt it again. So doing your own homework there. And then I think the internet is good, because that sort of points to what types of workflows you might build with this. And when you describe it, the I haven’t I haven’t thought about this previously. But I think the way I would describe it as Okay, so Alastair uses it, to delegate tasks. And I don’t know if that translates into English very well. But in German, there’s a difference between delegating something and handing something over where delegating means if I delegate something to you, you just do the task on my behalf. And I do quality control final checkup, and I keep the responsibility for the result. If I hand it over, I mean, okay, I’m no longer take care of this. It’s your responsibility. You take it to the client, for example. Right? So and seems to be very clear that what you say is no, no, we need to do the final check. We need to literally like it’s an intern, we need to delegate it takes off some of the work it takes saves us some of the time. But that doesn’t protect you from at the end really, really carefully looking through the stuff you’ve got back.

Alastair McDermott 11:30

Yeah, and I think that’s a really good distinction. And we don’t have that in English. But but maybe like it the the, the idea that you have to QA QC, quality control, the output is like that is crucial, because that, that if you if you take that into account, then I think that makes a lot of the other things that people are using AI for it makes sense. Because if you use that as a guide rail, then you can start to use it to make some really cool stuff. And I also think that there’s an important difference between how people are using it when they’re at different stages of their professional journey. So somebody who is an expert and has deep expertise in a field is going to be able to use in a totally different way than somebody who’s just starting out and is learning. So some people are learning from what the AI is giving it. And other people are refining it based on their own expertise. And so there’s there’s a garbage in garbage out is a very old phrase to do with computing. And it like there is that you know, you want to be able to like this is where it thought leaders can use it, you can have aI start to draft brainstorm ideas about a certain like, let’s say you’re writing a post or writing a blog post or something, and, and then you will spot as the thought leader, you will spot that it’s missing something important that you would say, and that’s where you get it to add in. And you can get it to add in, and then you can give it some like supporting information around that. And then it can continue to to fill in the gaps around that. And that’s where you can get it to help you do some stuff really quickly. But you are doing that that quality control and you are adding in the bits that it’s missing. Whereas if somebody who is not experienced in the field is doing that, they won’t spot that that thing is missing, they won’t spot and and that’s where you start to lose your voice. Because what it’s really doing is it is averaging all of the human knowledge that it was trained on. And it was taking the average the mediocre average of that. And it’s acting like this kind of very intelligent kind of auto auto responder auto autocorrect where it’s just kind of giving you a completely autocomplete. Yeah, so it’s a really smart autocomplete really just giving you what a thinks sounds like a good answer. And that’s what it does, it gives you what it thinks will sound like a good answer. That’s a way to think about the output.

Florian Heinrichs 13:43

I mean, this is another the maximum. But I think that is that is a way I have started using it as a marketer is to have it for example, if I head into an interview with a subject matter expert on something I haven’t asked, or have a give me questions. And I will share those in advance with the SME and ask the SME to say, Look, these are the questions I would have in mind. You know, tell me which of these don’t make sense or are too banal and add to you like to sort of the idea being that if the chat bot gives me the accepted common standard knowledge than actually dropping that and asking and have done discussion with the SME about the stuff they think was wrongly framed? Or the question that was, that’s going to be the interesting stuff. Because to your point, right, that’s not in this generalized average body of work. You could get out of an AI that is just something I get from the person who looks at this and says that your question five, there is nonsense. We stopped doing it this way. Three years ago, there were good reasons for it. Great. Let’s talk about that. Not forget the rest of the list like that is that is for example, a way I’ve been using it. But maybe you mentioned the dress code, the podcast. So Helen, you said you have to do you have bar does the other guy always look on the cloud? It was called kept. They keep changing and it’s not I don’t know. because I can never

Ash 15:00

remember is different from Claude. But yeah,

Alastair McDermott 15:03

Claude is from a company called anthropic, and they have a think so. So a lot of them have been doing deals with multiple parties, I think and tropic have, I think they have funding from Amazon, and possibly also Microsoft. So I think that Microsoft are maybe hedging their bets with open AI, I don’t I don’t really know much about all of that kind of the business dealings behind the scenes, it’s probably changing very quickly as well. What I do know is that is that even AI experts have been caught off guard by how quickly things progressed in like in in the last 12 months, since that, that time in November 22, when Chad GPT three came out, I think that was the that was like the trigger point for all of this suddenly, coming into the kind of the mainstream, where people are saying, Wow, this is actually pretty good. You know, all of these things like this is still brand new, and like we are on the crest of that wave by even talking about it now, because a lot of people are not using this at all. So that’s why I say go like go start using it go start playing around and experimenting with it. learning some basic prompts like, like one of the things that’s most important is to understand, like the prompt is what you say to that to the boss, or the Chatbot. It’s what you’ve seen instruction that you give it and what you the instructions that you give it are crucial. And if you give it really bad instructions, it’s going to give you the garbage in garbage out. So you have to actually learn how to talk to it properly. And, you know, there’s some basic things that you can do. Like you can tell it, for example, how it should act like you want to get it like for example, here’s here’s one example where it was it helped me to be more productive. I gave the text of a book that I written. So that gave Claude because obviously, I couldn’t put that into jeopardy. It was it was only around like 15,000 words or something. So as a short, it was a short book I wanted to put up on Kindle. And I wanted to make it free on Kindle. That’s a topic for another day. But one of the things that was holding me up was I didn’t have the blurb, the marketing blurb for the book that I wanted to put up on. And that was holding me up because I had to put that on the Amazon into KDP when I upload the book, because I have the book tax, but I don’t have the blurb for it. And so I just said, Look, I want you to act like a Kindle marketing book expert in the nonfiction business field, and write me marketing text for the Amazon page for the following book. And then I give it the book texts, and it came back with a really well formatted, well written blurb that I could just take and put directly onto Amazon. Now I edited it very lightly, but not much at all. And the time that it saved me because that was the thing that I was procrastinating, I wasn’t procrastinating. But writing the book texts, that was the easy part. For me, it was writing the blurb for it and kind of summarizing of distilling it down. And it did it in like three seconds. You know, so that’s the, that’s where I think it’s really powerful in in, in, it’s taking the friction away from a lot of those, those tasks where we would probably go like as the expert as a thought leader as a as somebody who wants to become known as 30 field, you are creating a lot of content. And you are a true subject matter expert. I mean, that’s what that’s what you should be. But there’s all these little supporting roles that you need. And those supporting roles. Sometimes you go to to third parties, you go to services, you go to other people to get chatty, PT and Claude and these general purpose AIS can fill in the gaps for a lot of those and make it much quicker. So that instead of having a back and forth, like I could, I could hire somebody to write the back cover blurb for a book, like there are people out there who specialize in just doing that. And then might charge $500. And it might take three weeks. And like what what they come back with is amazing. You know, maybe if I go to somebody who’s really heavily in demand, it might cost more than my costs $10,000 or something, and then they come back with something that’s amazing. What I got from Traci beauty, or from Claude is 80% of that. So it’s the 8020. And that’s where I think it’s really cool. And like, I know it’s not perfect, and I can I can improve it a little bit myself, what it’s giving me is much better than I would have done on my own. And it’s allowing me to get that and just put it out there. And then that gives us the accelerating effect, which allows people and this is where I think it makes people who are good at what they do. It makes them much more productive, that once they learn how to use it.

Florian Heinrichs 19:21

So I mean, maybe if I can can can close that piece a little bit because those are the large language models is texts, veneration refinement, these types of stuff. I want to hear if there’s any other tools in the McDermott tech stack. But before I do that, I think one one last word. And actually we can really speak to this also a little bit. These MLMs and privacy just mentioned that they used to exist in the these free market models. Yes, you can pay premium for a subscription here and there. But I think the enterprise perspective or if you work in a consultancy where there’s very, very tight data and cybersecurity restrictions because you’re in the financial sector, whatever. Those are often for boton Right, because it’s not entirely clear enough fully transparent, what happens, we have the stuff to put in, right? Obviously, it’s used to train the model, who owns the algorithm, all these things. But I think there’s more and more providers, open AI is one of them. There are others that are now offering sort of closed off walled space Enterprise editions. So if so even if you up to this point, have worked in a place that said, No, overnight whatsoever, you can’t use this chant, chances are that these things will be coming your way or to your firm or become available, right, even within your policy. So

Ash 20:34

yeah, I can come and tell that flow, because what I would say is in the tech space and consulting space, most of them create their own versions. So they don’t need instances in enterprise. I mean, when you’re looking at the space, it’s clearly three different there’s, you know, solopreneurs, as you were saying, Alistair, and there’s the middle space people, and then there’s large scale people, large scale, people have the resources and, you know, people and tech to build data centers of their, yeah, to build their own instances of AI. So they don’t have to, like, worry about the legal and ethical quandary of who owns what, later, yeah, they can work. You know, it’s just like, you know, if you’re working for a large company, if you’re working within their house, whatever you create belongs to them, it’s the same way. So they’re fine with that the mid space kind of blurs the line, sort of like hybrid cloud, I would say some of them have, some of them use existing solutions, others don’t. But I did want to touch back just a little bit earlier to announce to Assange in terms of like, giving, like really good prompts and stuff, because essentially, a lot of what we have in AI is the most efficient search engine in a way because not only it gives you the results that you’re looking for, but it gives it to you in a way that you don’t need to like, note down the results, convert into another format, do all of that, and then take it to the next level, which is essentially where a lot of time savings, especially in a mitten. So it’ll firms that really, you know, they’ve really that time, and this is some a place where Yeah,

Florian Heinrichs 22:13

that’s a good way of thinking of it. It doesn’t just give me the information, it does it in the in output format that I can immediately Yeah,

Ash 22:20

I don’t need to write a report from stuff that I just, you know, looked online, that now I have that attachment format, which, but you also need to know what you’re looking for, which is where your intellectual capital comes in.

Florian Heinrichs 22:32

That’s why we get all these New York Times headlines of all day paying 300 grand to prompt engineers, right? Yeah, sure. Because to be a decent engineering technology field, you got to be as software technician, in order to put in a two page long prompt, that actually makes sense. And then you’ve got to spot the two bugs that are hidden in the code, which comes back to ya can see why they made the top dollars to that they’re not giving you that to come up with clever prompts for LinkedIn posts. That’s not a 300k. Your job? I’m sorry?

Alastair McDermott 23:00

No, no, not at all. I think that one of the interesting things is, it depends on the approach that you take to it. But what you were talking about there is you’re talking about the output that it’s giving you back is in the context that you need it. And that’s, that’s where it gets really interesting. I’ll give you an example. And it’s from something totally different. Imagine that you’re trying to figure out what you want to create for lunch, okay. And you can go and Google just in regular old Google, you can put in the list of ingredients of what’s in your cupboard, or what’s in your fridge. And you’ll get back a whole bunch of recipe ideas, have recipes that have those things in them. Now you can, you know, you’ll read those. And first you need to read the backstory of the chef and the little village. All of that kind of stuff that gives a restraining

Florian Heinrichs 23:48

scroll before you go for the recipe. Yeah, that’s true. Yeah. Thank you, SEO.

Alastair McDermott 23:54

SEO. Exactly. Well done Google. So then you can go to chat GPT, or one of these others. And you can put in your recipes. But you can say, by the way, I’m lactose intolerant. Or by the way, I only have a frying pan, or I only have a microwave available. And it will put it in the context. Now. I’m like, this isn’t the perfect example for the people who listen to this. I understand. What I’m saying is you can actually

Ash 24:17

get it. I have something to add to that. Yeah. Well,

Alastair McDermott 24:22

I mean, I think that that’s really cool. Because now you don’t have to do some sort of convoluted Google search, where you try and look up microwave recipes and the ingredients that you have, and, you know, you just tell the AI what the constraints are. And the context is, and it will, it will do its best to establish to do it. Now. Sometimes it won’t get it perfect, but it will do a damn good job. That’s the that’s the way I see that and I think that that’s more useful to us now. And I think that’s why Google is running scared of this stuff because it’s it’s it’s going to be stopping people doing searches when they realize how good this is. Actually, I don’t know

Ash 24:59

absolutely like A year ago, a year ago, I actually said flow something like, Alright, give me a meal plan for five days, or X amount of money per day reusing ingredients, making sure that it’s x, calorific value per day, etc. And I got a very nice table. I’ve set of tables for like, seven days of what I should cook, what’s healthy, what’s the ingredients, how much it would cost me everything, which is exactly what you’re saying. So that’s what I said, Your example is not bad at all. It’s already in play. And yeah, here’s

Florian Heinrichs 25:33

where and then I really want to close out on the yellow, and we might come back to it because it’s such a big deal. But I saw this example you sent me ash. And I was like, okay, cool. I’m going to repeat this. And I’m going to do a marketing plan. So gave it a lengthy prompt about stuff and said, Okay, I don’t know what the budget was, is it? Okay? Please sort of give me a set of reasonable initiatives for this. And a cost estimate per and make sure we keep it under $800,000. And similar to your table, it spit out something that when I read it, I was like, well, that’s not very detailed, but pretty plausible. Nice. I could wing at least one meeting with it, for someone calls me out. But then I added up the budget figures in the columns and realize it spent 1.2 million instead of less than eight. Okay, so it really was not good at doing the songs. So this is one of these instances as the right way. You really have to. It’s the intern PC.

Alastair McDermott 26:20

Look. Yeah, this. Yeah. And going back to Ashley’s example, as well, the one thing I would say about that is I would check all of those calorific values, for example, yeah.

Florian Heinrichs 26:31


Alastair McDermott 26:32

what it will, it will give you a possible looking output. And that’s the problem is, to an untrained eye. It can be it can look right, even when it’s not. Now, maybe it was right, in this case. Absolutely. And that’s great. But sometimes it’s not. And that’s why it’s, you’ve got to

Florian Heinrichs 26:52

I read a read an article from The Verge, I think it was are from abroad or not, someone is actually very

Ash 26:58

visible in the video image once, right? Like if you’re using metadata, you can basically see if the pupils are not the right spot for the eyes and things like that. I mean, yes, you can make amazing Medtronic prompts, but you have more chances for errors that people who aren’t not at all the untrained eye, you don’t notice.

Florian Heinrichs 27:19

Yeah, that’s a nice read about it. And someone wrote about it. If you need to Berger in motherboards, I’ll have to find it and put it in the show notes. But someone tried to describe what it isn’t that the citizen actually generative AI as much as a fluent bullshit generator, when to addressing the point is just a sigmoid VAT says it’ll create something that sounds very plausible and presented with all confidence, right? It says, oh, no, this is the conversation between Nixon and someone else, then you have to be the person that says, yeah, it’s not so okay. But we’ve we’ve talked about the quality assurance aspect before. I’ll give you

Alastair McDermott 27:54

an example on that. Okay. In the demo that I did last week, where I did, like a shootout between, between claw chat GPT being inspired, I asked it to generate a list of links, and the list of books, and I gave it some input. And none of them were able to actually do it. But Claude gave me, Claude and Bing, I think both gave me output that looks close to what I was asking for. And one of them actually made up a book that I had written that I that I had not had never written. Yeah, but it made up a plausible book title for me, based on very little input, and

Florian Heinrichs 28:29

10 Secrets of podcasting success by 2023.

Alastair McDermott 28:35

Like, if somebody like if somebody searched me, let’s say, let’s say, I’m going on your podcast as a guest, and you ask Chad GPT, to write a bio on me, for example, it could just make up that stuff. And you may not know that that’s not true. You know, you may just use that. Now, I’m not saying that you will. But you know, this could happen where you think it’s giving you something? Oh, yeah. You know, I’ve got Mark Schaefer on he’s written eight books. And here’s the eight books that he’s written. And seven of them are correct. And one is wrong. Like, who knows, you know?

Florian Heinrichs 29:04

Yeah, so buyers beware with the large language models, but maybe we can click into quickly some other tools you’ve used and experimented with, you mentioned, transcriptions that that’s why I wanted to go there. Because we hear I’m using descriptive descript or the script, or I don’t know how you pronounce it for editing podcasts, which is a also a sort of language model takes the video watches it transcribes the audio stream of the audio from it. And what I find is increasingly good English. It even does German reasonably well, at this point, that was a pain point for a long time. And then you can edit the audio the video by editing the text. So it’s me, copying, pasting moving stuff around like I would do in a Word document and it actually does that to the audio file in the backdrop and outcomes something that is probably not audio engineering pro levels, but it’s good enough to be this. Thank you for staying with us. Listen this but that’s that’s it. So that would be an example of another quote unquote AI I’m using. Do you have any other tools besides?

Alastair McDermott 30:07

Yeah, like, I’m a bit skeptical. So I’m a former software engineer. And so I’m a, I’m a bit skeptical. I’m also a big sci fi nerd. When I see these things actually being called AI, like, they’re not really AI. Yeah, I know, like, a reporter from the New York Times has a conversation. And you know, the the chat GPT says it wants to commit suicide or something, like this kind of stupid rubbish. It’s like, yeah, you’re, you’re taking it down a line with the conversation. And it’s trying to give you a plausible answer to the conversation that you’re having. That’s what’s happening. This is not a self aware, intelligent, autonomous AI that’s out there. Yeah. Skynet or something, you know. And so, yeah, I sorry, I that’s a bit of a rant. But like, when I hear people calling these things, AI, it’s not really AI. Like, like nobody’s encryption software. It’s it’s really smart software, you know, and it’s really great what it can do. But like the transcription stuff, I’ve been using otter for years, and I really love otter, we had

Florian Heinrichs 31:05

IBM’s dragonfly or whatever, there wasn’t a 90s. Remember where you could dictate something to the software? Yeah, yeah, it was crap. But it existed.

Alastair McDermott 31:13

That was that was okay as transcription software on the market for decades. And like, I used to use that as well, I actually dictated part of a book using that, but 10 years ago, and it worked really well for what it was, but like it was, so it was so clunky. Whereas now like I’ve got the otter app on my phone, sometimes when I’m driving, I’ll click record on otter, on the on the otter app, and I’ll just drive along narrating some thoughts. And when I get back to the office, I will take it, it’s already in Otter online. Sometimes, and this is getting a little bit nerdy, but sometimes I’ll dictate a chat GPT prompt, at the start of that, and say, these are some thoughts about such and such, please format this in the form of a table, or please give this to me as a blog post outline or whatever. And then when I get back to the office, I can just take that text, I export the text from otter with just export to clipboard, I can just paste it directly into Chechi putty or cloud or whatever. And boom, there’s your there’s something done, you know, like, that’s where it’s just allowing you to be more productive, and more efficient. And that’s where I think, particularly for independence and solo people, it’s, it’s going to allow us to turn stuff around much quicker, we’re going to be able to, you know, I have a meeting with a client, I can put together a proposal based on that meeting, at you know, in in an hour, whereas before, it might have taken me half a day, you know, might have taken a day could have taken a week to get back sometimes. And just being able to turn around based on that conversation. Now sometimes depends on the scenario. If you record the conversation, which I record a lot of my Zoom calls now with with with Fathom dot video, which is a free, and I’m concerned about it being free, have been recommended by some people who are who are pretty good. So I’m going with it for the moment. But I know that the old adage is when it’s free, you’re the customer. But this this service by them dot video, what it does is it does a recording resume call and gives you a transcript and does an AI summary. And you can you can send the recording and transcript in summary to the other person on the Zoom call with one press of a button. And that’s a big deal. Because that just makes that just makes things a lot easier. But you can take a transcript like that. And like you could say to jack up, here’s the sales conversation that I had, please, you know, act like a an expert sales coach in the b2b space and critique my handling of the sales call, for example. Yeah, and it’ll give you a good one. Yeah, that’s cool. Yeah. So that’s just that’s just one one thing that you you could do, because of general purpose systems, you can use them for anything. You know, that’s the cool thing. I

Florian Heinrichs 33:45

never I think we were I think we were getting to a sort of a robust framework where we’re saying, you know, delegates don’t completely let it run its own thing. Put in the work before and after. Right. And I think you said a word early on, we were discussing prompts, which I think was interesting is that the prompting is crucial to get right. And you said, we said two things, I think which are essential, you have to give context that will improve the results. And you have to specify the constraints, that was another word that also improve the performance of Yeah, and talk about you have to return to you in turn analogy, right, I have to return to you. It’s an analogy, if you’re very good at explaining what you want from the intern, obviously, the outcome will be much better than if you just tell them Come on, you know, the presentation is not good enough, fix it without can

Alastair McDermott 34:28

also iterate it like some people take the approach where they write, you know, a 500 word prompt. And you’ll see that like these chat, CPT cheat sheets and guides and stuff that people are pitching since we a lot of people are pitching this on LinkedIn at the moment. But I tend to take an iterative approach where I write a smaller prompt, and then I iterate based on the feedback. And so I’ll ask it to make changes to the output based on what I see in front of me. And so we’ll start off with I’m kind of keeping it more I don’t know we’re agile, maybe Yeah, but I’m Not pasting in huge prompts, and then just copy pasting the the output and using it straightaway.

Florian Heinrichs 35:06

When I have to I do I tend to do the same but for the sake of repeatability which that is a very dangerous game, because it’s putting the same prompts three times to get tend to get different outputs. But I’ve done a experiment of just the approach you just set. And then when I had the input, I was ready. I was happy with the also the output I was reasonably happy with. I asked it to summarize the conversation we had in a new prompt that would reproduce the end result with a certain degree of approximation and then that I put in a file for reuse if I need it. I have yet to retry one of those I’m not got it was a fun experiment, to your point of playing with it right? Say, can you help me write a new prompt that would give me the last result immediately, as opposed to you know, all this back and forth we had, because that’s one of the power was still at the end of them. But anyways, that’s the power of these models is that within a stream of chat in the room, it will be contextually aware, right? So if you start chatting about the topic, return a week later, you can just listen, you can just continue the conversation will remember everything in that line of

Alastair McDermott 36:08

return to a certain degree. And it’s all depend on on what the size what the token size is. And I don’t know exactly how all of this works. But it’s a bit like the concept of of a buffer in software engineering, where there’s a certain amount of space, when you run out of space, anything new, you add knock something at the far end. And so it will start to forget, and I think that’s when they will start to hallucinate. And so, and that’s why I think that it’s, it’s an old, old school buffer overflow. But so what well, I think that like, you have to be careful about how you’re using it, you have to check all of the output, you have to be aware of it’s giving you an answer that is based on the average of all of the things that it’s been trained on, which may be wrong, which may be mediocre bland, and it’s going to answer you in the way that you ask it to answer you. And so if you’re not specific about that, it won’t be specific. And then the other thing that’s frustrating people don’t don’t understand don’t get us is something you mentioned earlier, is you can give the same problem two or three times, and it’ll give you three completely different answers. And that’s, you know, I don’t know exactly the, the logic behind how that works. But it’s something to bear in mind that you know, that you can, that you can play around with it. But I think that like, the number one thing I would say is just are playing around with these tools, like that’s the number one piece of advice I’d give to anybody is just start playing around with these tools. Don’t take what it gives you verbatim and copy paste it onto LinkedIn or somewhere like that. And and until we know more about the legalities, just be careful about not putting in flight information in there, you know, yeah, I put it I put all my own stuff in there. But I don’t put any client stuff in there. That’s gonna be important until we have those enterprise models that you talked about.

Florian Heinrichs 37:50

Yeah. All right. I think that was interesting. I think there’s, there’s a we could this could be an indiscretion. But we’re we’re nearing we

Ash 37:57

could go on about this topic forever. But we don’t have that much time. But I

Florian Heinrichs 38:02

think there were a good set of examples, references metaphors for how to think about how to use it. And you gave us a few very cool examples of doing things like I like the dictating stuff, as I’m moving around, and having that ready to snap,

Alastair McDermott 38:15

I’ve got another one that I think is pretty cool. I’m gonna record I’m actually going to go live after we get off this call. So people listen to your podcasts can find it on my LinkedIn, it’ll be a video, my LinkedIn is gonna be a short one, but five or 10 minutes, but it’s just about it using experts in the fields, and creating a fake panel discussion and using that as a brainstorming tool. So it’s kind of cool, because you can get like, what would this person say about this topic? And you can kind of use that as a brainstorming tool. Let’s say you’re trying to come up with ideas for blog posts, you’re trying to come up with ideas for LinkedIn posts or something like that. And what would these people say about this topic? So that’s another? What would Vidkun Stein

Florian Heinrichs 38:51

say about large language models?

Alastair McDermott 38:57

I said I made all of these where I said Aristotle, Socrates and Plato are having a heated debate about the future of AI. Please, please imagine that they’re that they’re discussing this and please write the transcript of the conversation. And you know, it gave me it gave me that, you know, so it’s wasn’t a Greek

Florian Heinrichs 39:15


Alastair McDermott 39:16

It wasn’t a Greek.

Florian Heinrichs 39:19

Yeah, interesting. Interesting. Well, we will have to bring you back some time. Whenever you say you have fun and new, new, new and really crazy way of using us until then, I think, Ken, do you have any I know you’ve written some very short sort of guides or books and put them out on the air stuff. So where can people find some of your thinking some of your work around this?

Alastair McDermott 39:38

You can find me on the recognized spelled the American way with the Zed and you can also find me on Amazon if you search my name, Alastair McDermott, and you’ll have to figure out how to spell that.

Florian Heinrichs 39:50

links in the show notes. So I just Yeah, I wanted to give people the idea. Well, that was that was great. So thanks for stopping by. I’m looking forward to I’m following your stuff post these things I’m looking for stealing more of your AI recipes. And you’re right, I think the general ideas or the concepts are more useful than the ready made. Here’s the 500 word prompt template, which by definition, will give me different results than whoever posted this. So yeah, thanks for stopping by. That was excellent.

Alastair McDermott 40:17

Awesome. Thank you so much. It’s been a pleasure to be on.

Florian Heinrichs 40:21

Cool, and that’s it right? Time for your line flow.

Ash 40:26

It’s tough. I’ll stop the recording

Florian Heinrichs 40:27

right here then.