The following is a rough transcript which has not been revised by The Jim Rutt Show or Carlos Perez. Please check with us before using any quotations from this transcript. Thank you.
Jim: Before we get started with our show today, it’s actually quite relevant. I was going to tell the audience about a Medium post I recently put up called ScriptHelper-001: An Experimental GPT-4 Based Movie Script Writing Program that I wrote. Probably more to the fact that I’d written the program, is I’ve written this quite detailed essay about how it works, the ins and outs of prompting, some pictures of the UI, et cetera.
Those who want some tangible, hands-on examples about the things we’re going to talk about in the show today, that’s a good place to look for it. It’ll be on our show notes as usual at jimruttshow.com. Also, if you want to comment about today’s show, feel free to do so @Jim_Rutt at Twitter. All right. Now, let’s get down to it. Today’s guest is Carlos Perez. He’s the co-founder of Intuition Machines.
He’s got 20 plus years experience in software development, technical consulting, and he’s the author of some interesting books on machine learning and AI. I think I first ran across Carlos on Twitter where he runs a very interesting stream of commentary and highlighting of papers about machine learning. By the way, his Twitter handle is @IntuitMachine. Soon thereafter, I read an early book of his called Artificial Intuition: The Improbable Deep Learning Revolution, and that book has stuck with me.
It’s actually a quite interesting book. He subsequently wrote Deep Learning AI Playbook: Strategy for Disruptive Artificial Intelligence, or your tech business book. Hey, business dudes, here’s how to use that stuff. He’s currently still a work in progress, but it’s available online to download. Artificial Empathy: A Roadmap for Human Aligned Artificial Intelligence. Welcome, Carlos.
Carlos: Well, thank you for having me in the show, Jim.
Jim: Yeah. I’ve been looking for an excuse for some time. We’ve had these interesting conversations over Twitter, and I think we had breakfast one time up in DC several years ago. Putting out this new book, I thought was a good excuse to get you on the show. But before we go on to today’s book, I should also note that Carlos also writes an interesting stream of essays on Medium.
Maybe like anybody on Medium, he’s hot and he’s cold, but he definitely an author on Medium worth following. Today we’re going to talk about a new book that Carlos has published called A Pattern Language for Generative AI: A Self-Generating GPT-4 Blueprint. Well, that’s interesting. When you say self-generating, what did you mean?
Carlos: Yeah. Let me give you some background in this book. When GPT-4 came out, I was doing some experimentation on its ability to introspect its capabilities. I had previously wrote a blog entry on what I would call a roadmap to general intelligence or some capability maturity level. I mapped out about six different levels of AI capabilities with the intention of using it to track progress in the deep learning space, and seeing how far we’ve actually evolved.
What I did with GPT-4, was I had ingested the blog entry itself, and I asked it what level do you think you are at? It cranked out a response, which actually was what I expected the level to be, which was around something that did not have too much counterfactual reasoning but somewhere in the middle of that, so it nailed it. I was quite surprised with that result. It gave me some intuition that this thing, this GPT-4 has some ability to introspect its capabilities.
Then one Saturday morning, I had a tweet that queried whether someone had written a pattern language for prompting. I’m not sure if your audience is familiar with pattern languages, but it was something that was formulated by the architect, Christopher Alexander, to basically come up with what he called a generative language for architectural design. The idea was that as you build this living language, it becomes a language that different architects could use to exchange how basically the motivations for their design.
That was later picked up, I think, at least a decade later by the software community, particularly the object-oriented community, that used it to describe a more complicated, object-oriented designs. That’s known as the Design and Banners book or the Gang of Four book about 20 years ago or probably a little bit more. I’ve always been a fan of that particular methodology of developing languages for tacit knowledge that exists. I was thinking, “Okay. Let’s see if I can apply this to the growing space of prompting that we’re discovering with things like GPT-4.”
But to my surprise, I could actually query GPT-4 for that information and that’s what basically I did. I asked GPT-4 first, I asked it to explain what a pattern language was, which it did correctly. The pattern language template, which consists of a particular structure, the name of the pattern, the context, the competing forces, some examples, and so forth. It has a particular form. Then I asked it generate patterns of prompting, and then it cranked out 10. Then I asked it to generate more, give me another 10.
It kept going until it eventually repeated itself and it got to around, I think, it was around 70 different patterns. I knew of some things that it likely didn’t know so I included those additional ones in the book. The self-generating part is that part where GPT-4 itself was aware of the kinds of prompting patterns that it could actually execute. What’s interesting here is when you see the pattern come out and you see the example, you look at it and you say, “This can’t be right. It can’t possibly do this.”
And gives an example. But when you put the example into GPT-4, it does what it says it does, it can do. It is quite surprising. I did that, but in the method of pattern language design, you tend to place them into higher level categories. For example, in the Gang of Four, they put it under three categories, behavioral, traditional, and I forget the third one. I think it’s structural, something like that. I needed to do something like that. I had GPT-4 take the 70 patterns and say, “Give me a classification for this.”
Unfortunately, when it cranked it out, it was a classification that was I would call pedestrian, something that I didn’t find satisfying. That revealed some limitation in the kinds of concepts that GPT would favor. I think in general, it would favor concepts that would be more commonly known. I didn’t take its advice and instead, basically manually categorized the patterns myself under a framework that was more organic, so to speak.
Now, I also did the same thing with Bard and Claude. It’s in the appendix of the book. What’s interesting is Claude comes up with some interesting ones, a lot smaller set, but Bard at that time was very limited. Surprisingly, had only a few patterns that he could crank out, so it’s an interesting note.
Jim: GPT was intimately involved in both the inspiration and the creating of this book.
Carlos: Yes. Yeah. Where I made a mistake was I should have initially constructed it through some executable language within a programming language, rather than just spit out the text in GPT. Because lot of work that I found out I had to do was take the content from GPT, and basically cut and paste and put in Word document and do all the extra formatting work.
It really wasn’t as self-generating or self-updating as I wanted. My next step is really to just take it to the next level, and then actually put it within more like a generative executable language, rather than a Word document, so to speak.
Jim: Yeah, that’s an interesting point. As I mentioned at the beginning of the podcast, I’m working on this script writing program and it actually started from similar motivation. I was working with a friend of mine, who’s a semi-pro movie script writer. He’s had one produced and he’s had several of them win prizes and stuff. We were playing around with GPT and cutting and pasting, and after a while it just got to be untenable, the complexity of managing all that.
I started writing this program, and it’s amazing the combination of programming and GPT, how much more you can do because you can cache all these intermediate results and then feed them back and combine them. For instance, when I have my program write the description of a scene, I feed it all the previous descriptions of scenes that it has already created in compressed form, so that it knows the synchronization. I give it the overall synopsis of the movie before it writes its scene.
I’m basically sending it a whole bunch of stuff, which in itself, had previously created. It’s the kind of thing that you couldn’t possibly do by cutting and pasting. I do encourage people who are really interested in doing complicated things with prompts, to move beyond using ChatGPT and sign up for the OpenAI API. It really opens up the scope of what’s doable and I highly recommend it. Of course, I think we both have found that GPT-4 is qualitatively way beyond GPT-3.5.
I sometimes describe GPT-3.5 as a smart but unreliable 13 year old, while GPT-4 reminds me of a very disciplined, well-behaved recent graduate of an Ivy League college, who’s gone to work for McKinsey or Goldman Sachs. It pretty much does what you tell it. It does a pretty good job and it doesn’t just arbitrarily do something else, which 3.5 sometimes will do.
Carlos: Yeah. I think that because of the limited context window, even for GPT-4, there’s a lot of managing of that context that you have to do. If I have 100 patterns, I can’t have that in the prompt itself, but it has to know off of that pattern if it’s going to relate them together.
There’s some management involved where you’re taking pieces of information and inserting it in the prompt so that at least it can basically refer to that information in its generation.
Jim: Yeah. That’s, of course, a continual problem. One of the things I find so fascinating is the idea of the context window as which everything has to go through and come out. At least I have found in the work I’m doing mostly writing narrative and dialogue, is that even though you have technically 8,000 tokens, maybe 6,000 word window, a little bit more.
In reality, it tends to lose its, I find at least, lose its coherence beyond maybe 1,500, 2,000 tokens. That you actually get a better product if you can chop the work up into smaller pieces even than the context window. Have you had that thought as well, that the context window is maybe big enough relative to the quality of the models? In other words, that the models don’t keep coherence for really long prompts.
Carlos: Yeah, yeah. What’s interesting about the design patterns is that it reveals techniques to manage that context window.
It just comes out for some reason, the GPT-4 is aware of that, a particular way of doing things. Things like chain of thought just comes out for free.
Jim: Can you explain what chains of thought is? By the way, our audience, they probably have done some ChatGPT, but probably most of them haven’t done any API work yet.
Carlos: Yeah. The idea of chain of thought is that when you ask GPT or any large language model, actually not all of them can actually perform chains of thought. But if you ask them a complicated problem, it’s more likely to solve it by actually explaining the intermediate steps. By telling to explain the intermediate steps of the solution, so that would be the chain of thought. You would actually have multiple prompts, but each prompt would be some intermediate explanation until you get to the final solution, rather than just asking it to just say crank out the answer.
A lot of times, it won’t do as well. I found that whole notion to be very valuable in the kinds of prompting that I do for understanding concepts. I guess if you’ve seen my Twitter’s feed where I do a lot of these table operations. So one of the things that you can really do very well with GPT-4, is have it render multiple answers to the same question across multiple dimensions. You actually build these tables that say, “Take two concepts or a couple contents and basically, break it up into different features or dimensions.”
You can compare a complicated concept across these dimensions, but you have to be careful in what you do and how you do this. In other words, you have to also do it step by step. You want to add, for example, like one column at a time, and basically build a column and then use that as the base camp for the next query and so forth.
Jim: Yeah. That’s certainly a key part of prompt engineering is doing a step at a time and learning as you go. On this chain of thought, I’m going to again come back and say I suspect the reason that is important is that the coherence range of these models is actually shorter than their context windows.
Even if in theory it could do it within this context window, the nature of the way the correlations work inside the model, the range from first token to 7,000th token is so far, that the statistical coherence in the model doesn’t really work.
If you ask it to tell a story, for instance, the story rolls off into incoherence. Or you have it explain a business strategy for 6,000 tokens. By the end, it more or less forgets what the hell it was talking about at the beginning. That’s again a reason why.
Carlos: That’s also why it’s useful in prompting to sometimes always give them a generated from some outline and then it would generate.
I think you do that in your script writer where you have it build a high-level plot and then you work out the details.
Jim: I’ve used that trick not just in my script writer, but elsewise. It’s actually quite good at taking a chunk of text that either it wrote, it came from elsewhere, and breaking it down into chunks, in the chapters into just to outline this argument. It’ll do it. Then you can give it the argument and feed them the items one at a time and say, “Write this out in some detail.” But it helps to have it give you the stuff before as well, so it has some context.
I call that the sky hook effect. Get GPT to do most of the work for you. In fact, that’s one of the core notions in my program, is first you gradually build up a long-form narrative of your movie and then you’d say, “GPT, turn this into,” let’s say you want 24 scenes, “Turn this into 24 scenes.” Amazingly it does it, and I’ve seen you do that as well. Now to this issue of tables, I must admit, I learned that from you on the Twitter feed. That was the first time I’d seen somebody doing that.
I use it all the time. Another thing it’s particularly good at for let’s say you’re exploring an idea. You’re not an expert in it at all, you’re trying to conceptualize, is let GPT do the conceptualizing for you. For instance, I will often say compare and contrast anarchism, 19th century Marxism and game B, and you choose the aspects in which to compare and contrast them. And choose those which will help with the most contrast distinguish between these three.
I want about seven, and it’ll pick seven categories, seven or eight categories that actually do a pretty good job of distinguishing between them. Then I say, “Fill in all the cells in the table,” and it does. That kind of thing as you warrant, sometimes it produces fairly banal results. But play around the prop a little bit and say, “I want a lot of detail on this.” Then once you have seven you say, “Add another one. What are the monetary theories of these three systems?”
It’ll just add it. But letting it do the subdivision of the intellectual domain is something that works surprisingly good again, as long as the list isn’t too long, 10 maybe at the most.
Carlos: But I think the value of using tables is that it does also give some coherence between all the entries in it, rather than if you were just doing it individually. When it expresses something, it’s in the same ballpark or it’s within some constraint.
I think it’s a better approach to prompting than if you did it individually. The nice thing about it’s that you can actually individually address different cells and columns and that sort of thing like a Excel spreadsheet.
Jim: Yep. Actually, I haven’t tried that. Can you say, “All right. In the second column, third row, please expand on that”? Will that work?
Carlos: To make it more concise, you tell it to label the columns like ABC and then the rows 1, 2, 3, et cetera. Then you can more precisely describe it, but you can also get away if you’re lazy, just say the second column or the third column.
It gets it right most of the time. But if you want to label it, it’s probably better to just label it to be more precise.
Jim: That’s an excellent idea. Now let’s go back again. We talked generally about prompting and different ways of thinking about it. Let’s take your formalism of a pattern language and let’s dig down into that a little bit.
As you said, you took the 70 and then you added 30, so you had 100 patterns. Then you decided you didn’t like GPT’s categorization, and so you built your own. Why don’t we lay out the pattern language essentially that these things have fall into, and let’s talk about each one a little bit?
Carlos: Yeah. I came up with eight patterns. Creational, transformational, coherence, explainability, procedural, composite, corrective and recombinational and variational. They’re somewhat organically inspired in the sense that you have variation and recombination and some selective pressure, which I used the word corrective in this case. Creation of patterns are really the most basic patterns that you have, what most people would use.
Basically, just you are kicking off a generation for the first time, so you start from scratch. There’s several of them probably, but the most important, probably the most fundamental one, and we knew this back in GPT-3, was what I call the input-output pair pattern. Somewhere in the middle of that document.
Jim: Let’s keep on our outline. Let’s be a good language model here. This is under creation patterns, which is your first bucket, and about halfway through that list there’s one called input-output pairs. Let’s talk about that a little bit.
Carlos: Okay, yeah. This one is the most basic kind, which we’ve known since GPT-3, that if you gave pairs of examples as the prompt, then it can use that to complete a new pair, for example. The evolution of GPT has been that they just added new features like instruct and be able to follow programming languages.
But at a fundamental level, if you want the maximum flexibility, you would go with an input-output pair. What I’ve seen is that there’s certain things that you can’t have, say for example, ChatGPT do, but you can have it through an instruction, but it’s possible through an input-output pair.
Jim: Give an example.
Carlos: Yeah. I had written a Google add-on a couple years ago with using GPT-3, and I noticed that it was able to do more odd kinds of analogies that GPT-4 will just basically punt on, because the analogy that I’m trying to have GPT generate is just too divergent.
It just says punt, these two are not related, something like that. But I forget the actual instance, I’ll have to run it again. But there are lots of cases where a lot of ambiguity, when you give it to GPT-4, it punts, but it doesn’t when you’re giving these input-output pairs.
Jim: Let’s give an example of what an input-output pair is.
Carlos: That would be like, for example, most basic you might do paraphrasing, for example. Of course, today you can always ask it to paraphrase, for example, with a well-known author and it has that information. But what if you needed to paraphrase with someone who’s completely unknown?
You would actually give examples of a sentence, just a bare bone sentence and how that author would say it, and you would give several examples of that. Then it learns the style of that author, even though that author’s unknown. Now, you cannot do that without examples.
Jim: It does like examples and it learns from examples. An example, one of the things I’ve struggled with at first, which was in my program, especially when I have it kick out the scenes, a title and a brief description of each scene, and soon it’ll also list the characters. Because it’s such a pain in the ass to parse that out, I wanted it to kick it out in JSON format. At first, I just told it to do it in JSON format. It was, “Okay. GPT-4, probably good enough, GPT-3.5, a fair shit show frankly.”
Then I gave it an example and then suddenly it got way better. I said, “Put it out in JSON format. By JSON, here’s an example of what I mean, da, da, da, da, da.” In the same way, I also have it put all character names with angle brackets. Again, at first I just said use angle brackets around all character names and it was about 80% correct. But when I then started putting examples, in example, , , thereafter, it got about to be 95% accurate, so it likes examples.
Carlos: Right. That’s an interesting thing too, that it’s also in the same creational patterns that the idea of punctuation. Punctuation is actually very important and you can do some programming, like you were just saying, using angle brackets to say, “This part is something that I want replaced, so to speak, or use it like a template.”
Jim: Could you give an example?
Carlos: Yeah. You could use it for example, as a template, as in you would set up like a form letter and basically say in square brackets you would have a variable name. Then you can assign the variable name before the final generation and it’ll fill it in.
Of course, the most interesting thing, and Claude actually gave an example of this as a capability, I’m not sure if it’s in the GPT-4, but GPT-4 is also capable of doing it. It’s called the, how do you pronounce that, C-L-O-Z-E?
Jim: C-L-O-Z-E, Cloze or what the hell is that? I don’t even know what that is.
Carlos: It’s called a cloze prompt. Basically the thing is, for example, in the cloze prompt, you could say the ship sailed into blank ocean, seagull circled in the blank sky. Sailor grays out in the blank horizon.
Then basically you can tell it to just fill in the blanks and then it will fill in the blanks in a way that makes sense.
Jim: Yep. Yeah, you can do that in computer languages with wild cards, things like that.
Carlos: Yes. A wild card in that sense. A wild card, but it keeps it somewhat consistent because the entries are different.
Jim: Yep. Well, that’s interesting. Yeah, that’s a good point. Yeah. That’s not really a classic wild card because you want it to be context dependent, which I expect it would do. That is cool. Let’s move on to transformational patterns. I found these to be very interesting and I’ve played with them some myself. For instance, summarization and simplification prompts, very handy.
I think one of the things that we’re going to find that’s GPT-4 in the real world is going to be most useful. I would love for GPT-4 to read all my email, summarize it into one paragraph and just give me the paragraphs. Do the same for much of the stuff that floats by in websites and on Twitter and Facebook to summarize whole threads in one paragraph. Then for me to then decide do I want to see more? That’s really a superpower and it does it pretty damn well.
Carlos: One of the interesting examples of these transformational patterns is this idea of compressing text.
Jim: I played with that one too. You could get it to do non-human readable compressions. You can get it to do human-readable compressions. I’ve got the feature in my script writer program because some of these prompts get pretty long. For instance, I can compress the movie textual narrative and it cuts it by about two-thirds amazingly, and seems to lose very little in expressive power, maybe a little around the edges. I only turn it on when I need it.
I’m sure you’ve seen the prompts. Use arcane characters, use emoticons, use whatever you want. It’s not for human consumption, and it really compresses the hell out of things if you give it complete freedom to do so, at least GPT-4 does. 3.5 doesn’t seem to handle compression nearly as well as four does. It would be interesting if in the future, they built a standard compression language into the processor as a post and pre-processor.
Though of course, it would cost them money because they charge you by the tokens so putting in less tokens and getting out less tokens actually cost them some money.
Carlos: But I think there’s an extremely valuable technique because if you need to, you have a limited context window and you really want to put as much semantics in there and how would you do that? Some compression technique would be extremely useful. I’ve also found it useful in terms of emoticons. It’s like working with some raw, semantic vector and then manipulating. You can take a set of emoticons, say five emoticons to represent a concept.
You can have it also say, “Give me the opposite of this,” and it’ll generate another bunch of emoticons. Then you can ask it, “What’s the explanation of this emoticon?” Because you can’t read it. When it conjures it up you’ll notice, it’s almost like semantic vector arithmetic of old days when they used to do the word back. It’s an interesting capability that’s just there.
Jim: That’s a very cool one. I’ll have to try that. I did see somebody, this was the GPT-3 days. They just asked it to produce five emoticon versions of movies. Say give me five emoticons that explain Godfather. Give me five emoticons that explain Apocalypse Now, et cetera.
Then some scientists did research and then turned that around and said, “All right. Now take these five emoticons and okay, what movie is that?” The hit rate was like 80%. It was pretty good.
Carlos: You would think that these models have some internal language that allows it at the minimum to translate between different languages. How does it actually translate between different languages?
It’s probably going through some Interlingua that it knows of, because you can have it translate into a different language that you don’t actually have a dictionary of between those two languages.
Jim: I think we got to be careful. We don’t really know what’s going on in these black boxes and they don’t have any language, they don’t have any logic, they don’t have any memory, they can’t change their state. They’re entirely static. I must say, I for one at least, am totally blown away that these bigger models seem to have these abilities.
Even though we know they’re static, there’s no moving parts, there’s no memory, nothing changes. How does it do that? There’s something about these statistical correlations that’s truly emergent. There’s more seemingly happened than seems possible. I’ve talked to some real experts on this and nobody seems to know. We don’t know how this is working.
Carlos: Well, I have my own explanation as how does it actually track state, for example? Because we know that it can actually track state up to a certain level with regards to if you ask it what’s the result of this computer program, for example? It will do it at a certain level. My explanation is that it works very similar to how functional programming languages work.
In other words, functional programming language are supposed to be stateless, but it has execution state by basically transforming the actual symbols itself. That’s what I think it’s doing to actually, basically emulate state. You can actually see this in GPT-4 and you can run a functional program, and you can predict how many layers does GPT-4 have? I think GPT-4 has something under than 100 layers because it cannot compute more than 100 steps.
Jim: That’s interesting.
Carlos: That’s the same depth as GPT-3, so what’s the difference? I think GPT-4 is just wider than GPT-3, but it’s depth. You’re right, it’s just the same.
Jim: That’s interesting. That’s very interesting. Well, let’s go on to your next set, coherence patterns.
Carlos: Yeah. This is the one I really struggled with, but it’s really this idea that across the prompts you want to have some coherence.
In other words, I guess the best way to think about it is if you’re trying to build a chatbot, you would actually have some prompt that would consistently be there so that it keeps a current identity, so to speak.
You want something that is coherent across different prompts. You are essentially copying and pasting or moving state across multiple invocations of GPT.
Jim: Give us an example.
Carlos: Yeah. For example, in the case of if you needed to, for example, have a character within a chatbot, for example, you would basically explain what the personality of that chatbot would be.
But you would maintain that across different invocations. You get that on the API itself. I think it’s called the, what do they call it, the system?
Jim: Yes. The system prompt. Yeah, system prompt. I use it very heavily. For instance, in all my script writing work, the system prompt does say you are an assistant helping a screenwriter write a screenplay. Now something I’m going to add tomorrow is a style prompt, which will also go in every single system prompt.
Because people have asked me for this, which is, and you are going to write in the style of Quentin Tarantino. I’ve tested Tarantino, Alfred Hitchcock and one others. It’s amazing how much it changes the writing style. You can also say be verbose, be precise, be concise. It does take those style hints quite seriously.
Carlos: Right. It’s definitely much better with GPT-4 or even ChatGPT as compared to the previous version.
Jim: Yeah. I have found that GPT-3.5, you’re better off putting those prompts in the user prompt than the system prompt, because it only semi pays attention to the system prompt.
But four does a good job. In four, they really, really can save a lot of prompting by having good system prompts.
Carlos: But it’s interesting that the API itself is structured that way, because that’s not really how the transformer models are actually structured.
Why does it have that separation between the system and the user prompt? There’s something going on under the covers there.
Jim: Now that you mentioned that, I hadn’t thought about that, but you’re right. The transformer itself doesn’t know.
I don’t think unless they somehow trained it in the way to assume both of those, which I don’t know, they won’t tell us. That’s annoying. A company called OpenAI is actually now about the most closed AI company out there.
Carlos: Yeah. This notion of the coherence pattern or having a context that goes across, it’s known since even GPT-3. That’s what I used to do also. You basically carry state across invocation so that it maintains some consistency.
Jim: Yep, yep. Yeah, yeah. Or you intentionally change it. For instance, in my screenwriter, I haven’t added characters behind the scenes yet, but I soon will. You can change the emotional state of the characters.
There’s Mary who’s sweet and nice most of the time, but she’s in a pissed off mood on this occasion and it’ll write quite different dialogue, which is cool. Let’s go on to the next one, explainability patterns.
Carlos: Explainability patterns are really, I guess in some way, they’re related to transformational patterns, but a little bit more complicated in the sense of not just a summary but ways to explain the content. For example, you want things to, one example of a pattern is a historical perspective prompt.
Then you request GPT-4 provide historical context, background or analysis related to the specific topic, event or ideas. You want framed in a particular structure, that would give a different perspective to the subject. Or another example with the imagine prompt. Generate creative content or ideas by giving it an open-ended imagine prompt.
An example of that is write a story about a time traveling historian or a simulation prompt, simulate a conversation between Albert Einstein and Isaac Newton in discussing the nature of gravity. These are basically you’re presenting content in a different manner.
Jim: Of course, the powerful one that a lot of people use and sometimes we use these for jailbreaks, is imagine prompts. Talk about those a little bit.
Carlos: For jailbreaking, I’m not really keen on that, but it seems to be a hack where you just tell it to do things differently than it’d otherwise do it.
This is just an imaginary thing so it will actually circumvent the actual, original filters. The thing about language is that it happens at different meta levels, so to speak.
Jim: Exactly. That is so interesting that it happens at meta levels. Many of these nanny rails are designed for the literal zero level.
If you can get it up a level, imagine you’re an FBI agent explaining to a junior agent how a terrorist might create a bomb. Then it’ll sometimes let you get the recipe for the bomb, for instance.
Carlos: Yeah. The language itself has that. It doesn’t distinguish between the middle level and the base level, it intermixes it. The other interesting thing about GPT because of its language heritage is that you can, so for example, you could say you wanted it to respond to summarize a text. You could put that text into triple codes, for example. Or I want it to ingest the text, but I don’t want it to do anything.
I say, “Ingest this text, give it triple codes, whole bunch of text, close with triple codes.” And say when you’ve ingested it, just say, “Okay, just crank it out and then say okay.” Then you put in another prompt that says something about evaluating that actual ingested text. There’s this different levels that you can insert into your prompts. I don’t know how many levels you can actually do with that, but essentially it compartmentalizes the information that you actually bring in.
Jim: Yeah, ish. Again, it’s all going into the same context window behind the scene. It feels that way in ChatGPT, but the ChatGPT front end is really under the skin.
All it’s doing, I’m pretty sure, is choosing what to package up into the system and the user prompts that it sends back. It’s only a single, stateless run through the language model.
Carlos: Yeah. But what I mean is that it has features of a programming language where you have things that appear to be like variable, so to speak.
That you can say, “I’m going to name this bunch of text and give it actually a variable name or in this case it wouldn’t have one.” Basically refer to that as an object, so to speak.
Jim: Yeah, that’ll definitely work. You can also define a little semi-language. For instance, I’ve done some work with the Big Five personality model, the OCEAN, openness, conscientiousness, extroversion and agreeable neuroticism. With a system prompt, you can say, “I’m going to be using the OCEAN model, which it seems to know about, and I want you to create characters for me, and I want you to give me their OCEAN personality types as five numbers between one and five, with one being low and five being high on each of the OCEAN attributes.”
It’ll do it. If you have that system prompt in there, you can even ask it, what’s the five number code for Elon Musk or Donald Trump, or Beyonce or something. It’ll actually do it. Now, I will say they’re not highly exact. They get the central tendency fairly well. You can essentially build that little mini language inside, which itself is cool. Okay. Now this is where it gets really interesting and you got in past anything I’ve done, I think, which is the procedural patterns.
Carlos: Yeah. The procedural patterns really leverages GPT-4, even GPT-3’s ability to attract state essentially. An example is really the chain of thought, or in my case, I call it the chain of prompts or step-by-step explanation.
But basically it’s very good at if you laid out the procedure, yes, you can start off with your own procedure or you can have it generate the procedure.
It will follow that procedure to actually generate the subsequent generation. Follow this as your template and it does follow that. These patterns all are around that line, that category.
Jim: Yeah. Why don’t you give an example? Examples are always very helpful.
Carlos: Okay. Well, the last one is very interesting, humor monologue analysis. You’re trying to understand the humor of a joke in a funny situation. You’re trying to have it actually, “Okay, what is the meaning of this joke?” You ask it basically the solution says create a humor monologue and notice as it guides the reader through the mental reactions and thought processes while experiencing the joke. Then you say the pattern should include the following stages.
You list down the stages, set content, establish anticipation, reveal the twist, identify the layers of humor and conclude. You would use that template to analyze a joke. Why don’t scientists trust atoms? Because they make up everything. That’s a joke. Then you basically crank it through that instruction, that prompt, and it works out the details, set the context, establish anticipation, reveal the twist and so forth, until it finally finds the conclusion.
The joke is funny because the unexpected twist and wordplay taking a seemingly scientific setup in delivering a punchline that plays on the multiple meanings of a phrase.
Jim: The other one you have in this group that again, that really got me thinking when I was reading it, is your design thinking prompts.
Carlos: It’s not something I came up with. In other words, this is what GPT-4 said, “Okay. There’s something, it’s called design thinking prompt.” It put together these descriptions in the context. What does this thing do? This thing is to use the example.
Use design thinking such as ways to improve the user experience of a public transportation system, considering the needs of the various stakeholders such as passenger drivers and city planners. Basically, what it does is basically you’re constraining how it’s going to answer it based on certain considerations that are related to your design.
Jim: Yeah. It seems to imply that it has some knowledge about what design thinking means. The example at least is quite terse.
Use design thinking to suggest ways, which I’m at least thinking I haven’t tried it, that if you tell it to use design thinking, it has some sense of what design thinking is.
Carlos: Yeah. What’s interesting is this example, I didn’t come up with this example. GPT-4 came up with that example. If you plug it in into GPT-4, maybe try it for a few seconds, it just cranks out. Right, it cranks out the answer.
Jim: This would be quite an almost like a GPTpedia. What does this sucker know? For instance, in my script writer, it knows approximately what the standard screenplay formatting is, which is very precise. It’s got like 25 rules on what a screenplay, how it should be physically formatted on a page. If you tell it output the actual script in WGA script format, it’ll do a pretty damn good job with only just that. If you give it the 21 rules, it’ll do a better job.
But one of the things it “knows” is that there is something that this WGA screenplay format, it knows it. I don’t know how you would probe to see what it knows like that, but it would be a really interesting project. Maybe it should be a community project. Hey, somebody out there, take this idea, take the idea of the GPTpedia that is essentially a mapping of the procedural things that GPT knows already. That would be very useful.
Carlos: That’s a very interesting thing about GPT-4, in that somehow the OpenAI guys are able to prioritize certain knowledge, such that when you ask it about that knowledge, it doesn’t hallucinate about it. It will give you, I always, for example, I use Christopher Alexander’s 15 properties of living things.
It always gets it right for 15, which is not the case if you use say, some lesser known author, for example. Why is it that Christopher Alexander is prioritized over some random guy on the internet? For example, I’ve asked about my capability maturity model and gave it the link. It does it correctly maybe for the first two levels and then it completely hallucinates about it.
Jim: Yeah. That’s a good question because I’ve noticed that too. If you ask it, “Give me the biography of George Washington,” it does a great job. If you ask it to give a biography of Jim Rutt, it knows who I am but the biography is about 70% wrong, at least GPT-3 is. GPT-3.5 is maybe 50% wrong, GPT-4 is about 80% right, but it still hallucinates. I know it wasn’t trained on me.
They didn’t put any human feedback reinforcement learning into it, but I suspect it has something to do with how much signal it got out of the corpus. There’s probably just a little bit about me, not a lot. Because I only appear in the internet a bit, while George Washington appears all over the internet. The signal is grooved in deeper into the net, something like that.
Carlos: Yes. But the thing is, things like where you actually have to enumerate, for example, the specific properties. In this case, there are only 15, it’s the same 15 all the time. It’s not some mixture of facts. It’s very curious.
Jim: Well, question it. Is that repeated enough in the corpus that it’s grooved in deep, that it doesn’t fall out of that groove?
Or have they programmed it in with the extra prompts, the RLHF prompts where they had humans tune the results? They’ll know.
Carlos: If you ask it like the chapters of the Art of War, it won’t make a mistake on the chapters of the Art of War for some reason. But if it was some other random person who did strategy, it’s going to hallucinate the chapters.
Jim: That was something else that could go into the GPTpedia is what things does it seem to crisply know versus what kinds of things does it tend to hallucinate about? But I would love to know that what would happen if you walked through Wikipedia and took the entry title and said, “Tell me what you know about thus and such”?
Then made some automated assessment on or actually I let GPT do it itself. Say, “Compare this to what Wikipedia says. Is this close or not close?” There’s a way you could build, you could bootstrap a GPTpedia, automate it and let GPT do the work?
Carlos: Have you heard of the story of how Jeopardy, is it called Deep Blue?
Jim: Deep Blue, yeah, yeah.
Carlos: Yeah. That played Jeopardy. Do you know how they actually solved it?
Jim: No, I don’t.
Carlos: They realized that all the responses in Jeopardy are all Wikipedia entries. They only had to say the Wikipedia entry, do a search in the questions, and then spit out the Wikipedia entry. I mean the name, sorry, the name, I mean the title.
Jim: I’m a pretty good Jeopardy player and here’s why. I saw long ago the meta rule, which is the Jeopardy answer is always the obvious one. Don’t overthink. It’s always the obvious one. It’s always 100% of the time, they have to have a disciplined editorial team that never takes the non-obvious answer.
I wonder if someone’s tried putting GPT up on Jeopardy, see how well it would do? You need a front end to get it to format things the right way, what have you, but I bet it does pretty good. All right. Let’s move on to your next category, which is the composite patterns.
Carlos: Yes, right. This is where it gets interesting and the following chapters all work off the composite pattern. This is where we’re actually using collections of things that are all manipulated within the prompt. This is where I have the generate table where you actually have these objects that have features in them.
Then you ask a single question that applies for the entire table. This is basically prompts that include collection of things that are in some way structured through the use of tables or other mechanisms like punctuations and that sort of thing.
Jim: Right. One of the things that you talked about here we alluded to them a little bit before, is the idea of prompt formulas.
Carlos: Yeah, yeah. Basically, the idea is that you would create an enumeration and then from that enumeration the entries, you would create a prompt from it. Then that prompt will generate additional text based on that prompt. They say it’s a level of indirection that’s being applied. In the case of example, you would just say, “Using a table of historical events, create prompts for essay topics.”
It would take historical events like American Revolution, Industrial Revolution and so forth. Then it would just generate the prompt, how did American Revolution War impact the political landscape of the other nations in the 18th century? It generated that based on the historical event. That is a prompt and that prompt, you can have it generating more prompts.
Jim: Gotcha. Yeah, that makes a lot of sense. Now, two topics that probably fit together and I think we all use these intuitively.
But you guys did a pretty good job here of calling out what they are, is the idea of inpainting and outpainting.
Carlos: This one is something that GPT-4 didn’t come up with. I came up with it and basically making the analogy that you also see this in the image generators.
In an image generator, an inpaint means that within a image you would ask it to fill in the details of a section within that painting.
Say it’s an image of a person and their hands are wrong, you could basically tell it to redo the hands, for example.
Jim: It’s also true in text. You can say as I’m finding in my script writer, for instance, just recently yesterday, there’s a scene where the lovers meet in a coffee shop.
I say, “Add in that the woman pulls a flask of whiskey out of her purse and pours it into the guy’s coffee,” and it filled it right in just right where it belongs. Quite remarkable.
Carlos: Right. It fills in texts within an existing body or within existing image. It fills it in. The outpainting is the kind that is the one that would follow it, for example. It would just be a continuation of the text.
In the case of images, it’s known that it’s much more difficult to do inpainting correctly because the constraints are much stricter and it’s probably true also for GPT.
Jim: Yeah. Though I got to say I’ve been finding the equivalent of inpainting in text, surprisingly good in GPT-4 worth trying. Corrective patterns, this is an interesting one.
Carlos: Yeah. This is the most interesting thing about GPT-4. This is the one that truly distinguishes itself from the previous version and it’s ability to correct its mistakes. Or actually when it renders something, you can tell it, “Something’s wrong with this. Can you correct it?” It will decide on what it needs to correct and correct it.
You see that in working with code and you see that also even the explanation that in text that it generates. A lot of this surprisingly, is all generated, these patterns are all coming from GPT-4. I don’t think I invented any of them. Which one do you think might be interesting?
Jim: Let’s think here. Here’s one that I actually use in my program, which is the error correction prompts. For instance, I mentioned that I have an output quite complicated JSON list of lists, when it turns the movie narrative into some arbitrary number of scenes.
I found amazingly, that if I had a system prompt that says, “I’m going to send you a broken JSON, please fix it.” About 90% of the time, it can fix it. That’s amazing to me.
Carlos: How does it actually do that? That’s the question that I was trying to figure out. How does it actually correct anything?
Jim: Especially if something as odd as a broken JSON, which can be broken in lots of strange ways and it somehow has enough patterns. Because again, it isn’t running a program, it isn’t running OC, it isn’t running anything. All it is is correlations between words, tokens at different ranges. That’s all it is, so how can it fix a broken JSON file? How does it debug code?
I can see how it debugs simple things in code like syntax, but it can actually do a better job than that. It’s again, this mystery, what does this emergence that is occurring once models get above 15 billion parameters or something like that, that allows them to have these human seeming powers, despite the fact that there’s no moving parts? It’s amazing.
Carlos: Yeah. I have some explanation for this corrective ability. We’ll get to it in a subsequent chapter, but there’s a very powerful prompt that actually reveals it. We’ll get that in chapter nine on recombinational patterns.
Jim: Yeah. The other one that’s actually I’ve used a little bit just playing around is you called gap analysis prompts. Say for instance, “Here’s an essay, what logical gaps are there in it?” It’ll do a pretty good job.
That’s not perfect, but it’s a pretty good meta editor essentially. Not a text editor so much as a developmental editor, tell you where the gaps are in your argument, for instance.
It’s surprisingly good. All right. Let’s go on to the next one, recombinational patterns. An interesting one I’d never heard of, something called Six Thinking Hats.
Carlos: I haven’t heard of it before. This was generated by GPT-4 itself.
Jim: That is amazing because I had to go look it up. I went and read the Wikipedia article on what the hell is it? Six Thinking Hats or something.
Carlos: Six Thinking Hats.
Jim: Never heard of such a thing, so explain what that is. I haven’t tried it properly yet, but it sounds like it might be interesting.
Carlos: Right. It has an example that it generated. Use the white hat, provide data and information about the impact of plastic waste and the environment. Apparently, white hat is a particular mode of thinking.
Facts and information so you can use that within your query. You could call it white hat, a red hat, black hat, yellow hat, green hat or blue hat. It will explain things differently based on it.
Jim: Yeah. Okay, here we go. Six Hat Thinking from Wikipedia was developed by Edward de Bono. Blue is big picture, white is facts and information, red is feelings and emotion, black is negative, yellow is positive, green is new ideas.
Carlos: Which is interesting, because why would GPT-4 actually, explicitly know about this? It’s vague. An example, it says using the red hat, share your feelings about the plastic pollution. People who aren’t familiar with this wouldn’t know what red hat would mean.
Jim: Exactly. It’s one of these things that it knows. That we don’t know what it knows, but it doesn’t know what it knows, which is actually even more interesting, but it could in theory. There’s capabilities that it has, right?
Carlos: Yeah. That’s an interesting thing about this. It’s possible that sometime in the future they might basically put the nanny rail so that you can’t find out.
Jim: Yeah. Well, I don’t know, could it do that? Because you could do probes to see if it could behave in certain ways.
Carlos: Right, but this is not a probe. This is actually asking GPT-4, “What can you do?” I’m giving you the menu. GPT-4 is giving me the menu of ability. It’s actually explaining to you. You meet a contractor and you’re asking, “What are your skillsets?” It tells you what your skillsets are.
In other words, you can find out more about it, by basically asking it, “What can you do?” I just ran the red hat using the red hat share your feelings about plastic pollution and it knew what red hat meant. It pointed to the Six Thinking Hats.
Jim: I just typed in six hat perspectives on climate remediation and it’s writing out the six different perspectives.
It obviously knows what it is and it knows how to take that and apply it to a question. This is part of this GPTpedia, it knows about six hats.
Carlos: It’s not something that I guess it told me what it was. It told me that I could do this, which is amazing.
Jim: That is really, really interesting, I got to say. Another one that’s in this group, attribute listing prompts.
Carlos: Yes. I think this is what we were talking about previously, where you would basically list the key attributes of an effective leader and would come up with the attributes.
We didn’t tell it what it was. It just conjured them up. Very useful if you’re actually going to do comparisons. And like you said in one of your tweets, where you actually had to compare plan B with other plans.
Jim: Yeah. Choose its own aspects, which was the interesting part. Another one that you have in this list, a problem restatement. I’ve actually used that to get around hallucinations. I found I was going to publish this, but I got too busy on other stuff. Which is you can concentrate answers and get most of the hallucinations away by doing the following. Which is to ask GPT, to paraphrase your question, have it write 50 versions of the question, which it will do, and they will be identical.
Then ask the question 50 times, and then take the statistics on what was the answer that was the most common? As it turns out, at least for most questions, right answers are statistically much more probable than any given wrong answer. Even if in the domain it’s wrong two-thirds of the time, the right answer will be much higher count usually than any of the wrong answers. Using problem restatement by just literally writing a sub-routine that says, “Send off to GPT to paraphrase this 50 times.”
Then capture the 50 paraphrases and then feed them back, capture the answers, pull out the facts, tabulate which one’s the most common. We’ll give you in one case, a factor of five better set of answers to some borderline questions that it tended to hallucinate on. That was quite interesting. Let’s move on to the next one, variational prompts.
Carlos: Yeah. The chapters that follow the composite patterns are all working with sets of facts. The recombinational one was basically you would just basically combine facts in different ways to generate new facts. But the variational ones would be where you are contrasting the properties within that set, within that collection, within that composite object. Let’s see as an example, let’s go for the gameplay. Okay, flipped interaction.
The description of this is flipped interaction, inverted interaction. This is also known as an approach to interaction with GPT-4 that focuses on asking questions and rather than generating output. In situations where the user desires a deeper understanding or exploration of a specific topic, concept or idea. The traditional interaction patterns may result in surface level or one-sided output lacking depth or meaningful engagement.
The solution would be you encourage users to ask questions that provoke thought, reflection and critical analysis, fostering deeper and more engaging interactions. As an example, instead of requesting GPT-4 to write a summary of a concept, ask a series of thought-provoking questions related to that concept. Yeah. Given the collection that you have, you’re trying to find contrast between the different items in that collection.
It’s related to, similar to when you had your example of looking at the plan B. You’re trying to see contrast between the concepts.
Jim: It’s interesting, that reminds me of the very first prompt I sent to GPT-3. I guess it was actually the original ChatGPT, so that would’ve been 3.5. I gave it the 12th grade English essay, which was to compare and contrast Conrad’s Lord Jim and Moby-Dick and it did.
It did a quite classic 12th grade honors English essay comparing and contrasting. It did both in a very formal way, those two novels. That was quite interesting actually. Literally, the very first thing I typed into ChatGPT.
Carlos: I think one of the values of this, is like they say, one of the patterns is multiple discipline prompts. That GPT-3 is very good at translating different domains into your domain. Example, let’s say you’re not familiar with the terminology that’s in the linguistic domain and you’re coming from a psychology perspective.
Sometimes they use terms that are the same terms, but they have different meaning. It is very useful to in basically mapping a domain that you’re interested in, into the vocabulary or language that you’re familiar with. You can make that mapping, which is interesting. It’s a very useful bridge for acquiring new knowledge.
Jim: All right. Now let’s move up to your last pattern. Actually, these are some of the most interesting. These are head stretching. Modularity patterns.
Carlos: Yeah, yeah. Now these are not generated by GPT-4. These are outside GPT-4. This is not even something that you can actually invoke within GPT-4, unless they provide the capabilities. When I wrote this, they didn’t have plugins yet. But it’s something that you would do, for example, programmatically, we had something like LangChain or that sort of thing. These are outside of unless they built it in as a default.
Like plugins are now available, but these are not natively supported by the actual language model, but more of external features that you can add on to it. Things like using plugins, like select the tool. So select the tool, it would be the equivalent where GPT-4 or GPT-3 would select, for example, browser for example. When you run a prompt, then it will actually choose to use a browser to answer that prompt and how does it actually do that?
There’s an example there where in the prompt itself, it has meta information about the tool. That’s how it actually selects the browser, because every plugin has meta information that that’s inserted into the prompt itself. That’s how it actually determines I’m going to use the browser now. Not only just the browser, but the particular methods within the browser itself.
Jim: Interesting. Well, let’s use this as an opportunity to talk a little bit about the plugins. Godddamn OpenAI still hasn’t given me access to the plugins, at least doesn’t seem to have. What have you found the good, the bad and the ugly?
Carlos: Actually, I have not used it much other than the Wolfram plugin. Yeah, I look at it as more like a convenience thing that you don’t have to jump outside of the tool to actually continue with generation.
Jim: As opposed to wrapping the API. Again, as I mentioned, a lot of the work I’ve done has been orchestrating props in and out outside of API. Your LangChain does a good job of that. What’s the other wild one?
Auto-GPT, that just keeps learning and searching and trying to find answers to problems by iterating on prompts. It may turn out that that is more useful than plugins, we shall see.
Carlos: It’s the inverse. In the LangChain setup, the orchestrator, so to speak, is the LangChain, the programming language, which would orchestrate it. In the plugin, it’s the reverse. It’s GPT-4 that’s orchestrating whether it’s going to call the plugin or not.
Presumably, you might have more intelligence if you’re using the plugin, as compared to a LangChain approach. Anything that you have in a LangChain, anything that I called, you can make into a plugin with GPT-4 and you can insert it in. If you want a more powerful orchestrator, then you might use the plugin method.
Jim: Gotcha, gotcha. I’ve got to get my access to that. Now, what’s the syntax for invoking a browser in GPT?
Carlos: There is no syntax. You just say you’re going to use it.
Jim: What do you, say use browser?
Carlos: When you start to chat, you just say the plugins that you want to use and that’s all. Then it just infers by itself which one it’s going to use. You don’t have control.
Jim: I gotcha. I gotcha. I’ve just typed in use browser and it gave me the usual horseshit. I’m an AI and text-based model, don’t have the capacity to use a browser.
Carlos: You can’t tell it to explicitly use a browser. It just decides on its own.
Jim: That’s weird. I think I like better the ability to use a browser outside of GPT, than feed the results back into GPT.
Carlos: You would think how you hinted, maybe you could put in some hints. I haven’t explored it enough. That says, “Okay. For this case use this, but this case use this other thing.”
Jim: Right. Well, we’ve gotten through your categories. Now, we’ve got only five minutes here. Let’s step back a little bit. Your last chapter, you get a bit philosophical.
You call it katas and meditations. Bob’s going to let you talk for a while. I’m going to shut the hell up. I talk too much sometimes. What have you learned about the art of prompting from this work and other work that you have done?
Carlos: Maybe I can explain by going through this chapter. Katas are basically in martial arts, that basically procedures that you learn over time. You have this also in programming languages and in the sense that if I need to learn a library, the really good ones come with katas. Basically, there are basically exercises that you perform so that you actually learn how to use a library. But in this case, that’s what exactly is in the katas, which are basically it’s like a problem set. Then you just solve it using I’m giving you the kata and figure out how to solve it.
Then the other part are these meditations. In this case, I’m exploring complex questions and I’m using GPT-4 to actually help me explore those questions. One of the first meditation is how do you explain a joke? In this case, this is a joke where the question is, here’s the joke. A pair of cows were talking in the field. One says, “Have you heard about the mad cow disease that’s going around?” The other cow says, “Makes me glad that I’m a penguin.” The question is, does GPT understand what’s funny in this joke? Basically, it works out how you would do it.
What you would do is you would use the composite pattern and the ability to generate the table and ask it what are the possible explanations of why this is funny? After you get the possible explanation, can you rank which is the most likely explanation? It turns out if you rank it by American humor as versus British humor, the rank is different. What is funny for an American would be different from the British one. Then it explains why. Essentially, GPT-4 can understand why this explained a joke. Another meditation is is sensory grounding necessary for general intelligence?
We work at that detail again with GPT-4. I give some background there in terms of what it is. Okay. What are we going to do first? In this exploration, I work with seven properties of agency that Kevin Mitchell proposed. He’s a neuroscientist, and he has a framework for agency that has… Actually, this is incorrect. It should have been eight, I found out after I interviewed him, it’s eight. There’s one that’s missing. But it is this, the properties of thermodynamic autonomy for system endogenous activity, holistic integration. Low level in determinants in historicity, agent normativity.
Those are the properties, but GPT-4 isn’t familiar with those properties actually, because it’s probably a new publication that Kevin Mitchell wrote. You actually have to tell GPT-4, “What are these properties?” You get all seven of them, and then you use the seven of them to create a table. That table will have the explanation and example of that particular property. The amazing thing is that you would actually just give the name of it, that property, and it will conjure up an explanation of it. Apparently, the explanation is actually an accurate explanation of it.
In this case, you have the agency property explanation and the example. I wanted to see, “Okay, if these biological agencies translate up to higher level human cognition.” When it gave the examples, it just gave examples that would actually cross different levels of biology. In the first case, it would talk about a cell. In the third case, it would talk about a person. It’s at a different level, but I wanted it to be at the same level so I can think about how does agency relate to higher level cognition?
You actually have it aligned and move up in the biological scale in that particular column. There’s the pattern align analogy such that it gets to a certain level where at the end, it’s at the level of human experience. You’re taking some basic concepts that are fundamental for biological agency and seeing how they actually correspond to human level cognition. The question is, is agency essential for higher level cognition?
Jim: Gotcha. Gotcha. Well, I think this has been a very interesting roll through the book. I think it’s probably stimulated people’s minds a little bit about the power of prompting and CPT for world. Any final thoughts on how people can become better users of these tools before we sign off?
Carlos: Well, I would plug my book because it gives you the landscape of a prompt that you may or might not have encountered before. It’s like this is the menu and this is what GPT-4 has told you that it can do, so that’s a starting point.
Then I guess the next things is really how would you combine these to actually build more complicated solution? That’s still something that we’re all learning to do. But the idea is that with the pattern language, you basically have a vocabulary to actually mix and match to create new designs.
Jim: Just to remind people, the name of the book is A Pattern Language for Generative AI: A Self-Generating GPT-4 Blueprint.
As always, we’ll have a link to it on the episode page at jimruttshow.com. I think with that, I think we’re going to wrap it up, Carlos. This has been a wonderful conversation.
Carlos: Thanks very much and thanks for having me.
Jim: Yeah, I’m glad to have you on. I’ve been looking for an excuse. I’m glad we were able to do it. We’ll have to have you back sometime in the future.