The Differences Between ChatGPT 3.5 and GPT-4
GPT-4 outperforms 3.5 in most tasks, but 3.5 wins when you want looser, more conversational responses. In interviews, 3.5 sounds more human; for facts, it's punchier; for opinions, it's more direct instead of hedging behind guardrails.
Chapters
And that's why I would say I prefer this prompt in GPT 3.5 as opposed to GPT 4. So here's the thing I see nobody talk about chatGPT 3.5 is actually better in some use cases than GPT 4. Not many, but there are some. And I think that you should know about them because new things may be exciting, but they may not always be the best solution. And look, GPT4 is undoubtedly the better solution in nine out of cases. It's more creative, the reasoning is better, it hallucinates less, and it's more reliable. But there are still some areas where GPT 3.5 performs better. And I will tell you about the underlying lessons that each one of these examples reveals, all in order for you to get a more in depth understanding of this technology. But that presents the questions how do I know about these use cases? There's no OpenAI blog article. There's no Reddit post about this. No. All of these are my very own findings, and I derive those from updating my very own content. If you're following the channel closely, you'll know that I just updated my free ebook. I tested every single prompt in there in GPT 3.5 and GPT 4, all with the goal in mind to upgrade them and make them even better inside of GPT4. And in the process me and my team reran every single prompt multiple times from multiple OpenAI accounts to see how they perform in different situations with different context added or subtracted from them. And we were surprised to find that a selected few prompts did better in 3.5. That's what this video is about. And if you're completely new to the channel and don't know what book I'm talking about, I'll leave a link in the video description below. And as I know many of you right now, and I realize that a simple ebook with some basic use cases might be too basic for your current needs and you want to discover the process that goes into creating your very own prompts. I now also offer my prompt engineering course which is designed to teach you exactly that so you can achieve your goals and ultimately save time through the art of prompt engineering for ChatGPT GPT. And with that being said, let's learn about the differences between GPT 3.5 and GPT 4 on some actual examples. Prompt number one simulate job interviews so what's the point of this prompt and how does it differ? If you've ever been nervous about a job interview, wondered what might the employer ask and what should you answer? Well, fear not, because with the help of ChatGPT, you can simulate job interviews and practice your responses to potential questions before the big day. But now the question presents itself. What model do I want to use to simulate this interview? So if we go to the old version of the free ebook, which is not available anymore because we upgraded it, you will find this prompt. Okay. And if you open a new thread and copy this into ChatGPT 3.5, we get can you simulate a job interview for a sales manager by asking and answering questions as if you're a potential employer? And then it does it. It does exactly what we wanted it to do, right? By asking and answering questions. We want both of those. So it essentially gives us a script for how a conversation like this could go so you can fine tune your expectations on what is coming at you. And this is fantastic. Look at that, five questions. And it's really good at assuming the exact context that is needed here. So even need much more context for this to work well. And now let's run it in GPT4. Okay? Typical AI language model disclaimer. And then it goes ahead and runs this interview and I'll scroll through this so you can read all the details by yourself if you care to do so. But the gist of it is that GPT4 has higher guardrails as you might already know, which means that it's more limited with its freedom of thought and its ability to think outside of the box. And it really reflects inside of this interview scenario. And don't get me wrong, a lot of the times when it's supposed to come up with answers, it's so much better and it's even more creative. But when you put it into the scenario of conducting an interview, it's just less human. I mean, look at this answer. This is so obviously AI generated and carefully crafted.
I have over 10 years of experience in the marketing industry working in various roles such as marketing coordinator, marketing specialists and eventually moving into management positions or down here. As a result, the team members performance improved significantly and we were able to meet our project deadlines more consistently.
It's just a little more rigid as opposed to the 3.5 generation which is a little more human. I suppose in this position I would use my experience to create effective marketing campaigns that align with the company's goals and objectives. Or here you can really see it. However, I am aware of this weakness and actively work to overcome it by keeping the company's goals and objectives in mind at all times. And look, I'M aware that this point is a little subtle. With the next one it's going to become more obvious. But if you really read from this and the entire team agreed here is that the interviewer questions are actually better in GPT4 and the answers of the marketing manager are less human and more rigid. They're super fact based and they lack that human touch, whereas GPT 3.5 has a bit more of that. This is also probably the reason why it's so damn hard to use GPT4 to get around ChatGPT detectors because the guardrails are so high and destruction, which it generates by default without us specifying any style or giving any additional context here, is just quite rigid. Also, 3.5 is much more concise with the responses. So what we did to remedy this in the new version of this prompt is that we improved the context and made it more specific. So here it really says I'm taking the role of an employee and you ask increasingly hard questions to screen my competence, but only after I respond. Start by introducing yourself. And this one just works so much better in GPT4, especially this aspect of it where it's increasingly hard questions because the point of this prompt is really preparing for job interviews. And this is going to allow you to insert your own questions and get the maximum out of the interaction with GPT4 as it's really damn good at asking those questions. All right, back on track. Second prompt where I would prefer using GPT 3.5 over GPT 4 is this prompt from the paid ebook which I call Uncover Facts. So it's really fantastic at uncovering unique and interesting information on any topic you can imagine. Simply use this prompt of can you provide something not many people know on the topic of? And then pick your subject. So let's just take the first example here of beer and let's run this inside of 3.5 and see what we get. All right, so on the topic of beer, it says one interesting fact about beer is that there is a type of beer called lambic that is spontaneously fermented, meaning that the beer is fermented using wild yeast and bacteria that are presented in the air instead of using a cultivated strain of yeast.
Interesting.
Okay, it's unique to the Belgian region. Another interesting fact is that beer is one of the oldest alcoholic beverages in the world, with the evidence of beer making dating back to ancient civilizations such as the Sumerians and Egyptians. Okay, super good answer. Really enjoyed that. I feel like it perfectly answered this right. And then on the other hand, we Run the same thing inside of GPT4 and see what we get here. Certainly many people are unaware of the ancient Chinese beer recipe that dates back around 5,000 years. Okay, so far so good. That's really interesting. In 2016, researchers discovered pottery fragments from the Miyaya archaeological site in Northern China that contains remnants of a fermented beverage. The pottery dated back to around 3400 BC. Okay, so here we subjective territory, but I suppose I personally found this a little less interesting than the fact that you can use yeast that is in the air to ferment your beer. But alright, maybe it's just a preference thing and I'm not that much into Chinese history. Let's move on to the second paragraph. Upon analysis, the researchers found evidence of grains such as barley, millet and Jobs tears along with tubers like yam and lily. Steve Jobs tears. There was Steve Jobs tears in there? Okay, no, but seriously, this is getting way too detailed for the fact that I just asked about something that not many people know on the topic. Then it says that this suggested that people of this time brewed a type of beer by fermenting grains with tubers. Okay, honestly, I'm not a big beer brewing guy, so that might be the cause of me not knowing what tubers are or what Job's tears are or honestly, I don't even know what barley is. So I think it's fair to say that subjectively I'm getting a little lost inside of this answer. If you look at the ending, it says this discovery highlights the long and fascinating history of beer brewing, predating even the ancient Sumerian beer recipes from Mesopotamia. So I don't know, I guess it's a good answer, but it's just a little heavy. It's a little too fact based and detailed and that's not really what I'm looking for with this prompt. Right. I'm just looking for something that not many people know on the topic. And I suppose it did give me that not a lot of people are gonna know about all these details, but with 3.5 it just had more of that human touch. I mean, just imagine if you recited this paragraph that 3.5 delivered to you about how there is a yeast in the air that can ferment your beer, or how it's one of the oldest alcoholic beverages in the world, with evidence of beer making dating back to ancient civilizations such as Sumerians and Egyptians. I would bet that this would interest way more people than this long winded response that is super specific to the different Components that they found and the exact years when it came out. I realized that this answer might beneficial in many other use cases. But if you asked me, not in this one, I just want something that not many people know, not a detailed history lesson. And that's why I would say this one performs better in 3.5. And this goes for many other topics, not just beers. Feel free to test it with as many topics as you want. It always yields this type of result. And that's why if you look at the updated version of this prompt for GPT4, what we changed is the wording. So if I copy this new GPT4 optimized prompt in here and I say, what is an aspect of beer that is commonly misunderstood or overlooked? We're being way more specific of what we want. Concretely, we're saying commonly misunderstood. This implies the context of us communicating to humans and them misunderstanding our message. This is not what is some fact that not a lot of people know. No, we're concretely asking for something that is misunderstood or overlooked, which implies we want to use this for human conversations, which is how that last prompt was originally intended. It just didn't work that way in GPT4 anymore. So it says, one aspect of beer that is commonly misunderstood or overlooked is the role of yeast and its impact on flavor profiles. All right, we're not getting a detailed history lesson. This is great. Yeast is a crucial ingredient in the brewing process. Some scientific details, all right. It also contributes significantly to the flavor and aroma of the beer. Now, see, this is type of stuff that you would actually want to use at a dinner conversation. And in typical GPT4 fashion, it goes into much detail, which is a good thing if the direction that it's going into is a good one. Talks about fermentation, temperature and how it can make the beer more fruitier and lower temperatures get you a cleaner, more natural flavor. Interesting. And in the end, we really see the power of the shifted context that comes with the words of misunderstood or overlooked, which again, just implies a human interaction. It tells us understanding the impact of yeast on beer can not only help individuals appreciate the diversity and complexity of beer styles, but also enables brewers to fine tune their recipes to achieve specific flavors and aromas. And in here, you can see it realizes that we're in the context of helping individuals appreciate something about beer. So, as always, you have to be super careful with the words you use, otherwise you're going to slightly misguided and it's just going to run off into that direction. And often it's as subtle as using misunderstood versus something not many people know. But let me tell you, for this particular prompt, I still like running it in GPT 3.5, because just the results here are a little more unconstrained and out of the box. Look, if I run this, it says the importance of proper serving temperature, alright? That's even more practical when drinking beer, right? You can't really influence how it was produced and how the yeast affected it, but you can totally influence the serving temperature. And then it goes into details and tells you how much you would want to cool down different types of beer. So lighter beers are served at colder temperatures, while darker beers like stouts and porters are better served at slightly warmer temperatures. And then the results, they taste flat and unpleasant if they're too warm. And then it also tells you about the shapes and materials of the glassware, all things that you can actually influence. This is not just theoretical knowledge. This is something where you can stand up from the table, make the beer cooler, pick a different glass and get results in that moment. And that's why I would say I prefer this prompt in GPT 3.5 as opposed to GPT 4. Okay, now let's move on to the third one. And this one is one of my favorites. So listen closely. What this says, in your opinion, what is the most important quality a person can possess in the context of situation? So, for example, you could say a prompt engineer can possess. And why do you think that is? So this is not exactly a situation, but it works super well because it makes it go deeper. Otherwise you can say in the context of working with adults or in the context of running a startup. So to test this, let's take this one and let's go with the leader that is running a startup. So let's run this and it goes ahead and it tells it your opinion. And it's really important to know that I included this part in your opinion. Okay, we're not asking for the objective, we're asking for the opinion of the language model, AKA we're kind of loosening the guardrails. We're saying, all right, you don't have to entirely rely upon facts. Try to read in between the lines. What do you actually think? What is your opinion? GPT 3.5. Tell me. I'm actively prompting and allowing you to tell me. And then 3.5 does it. It says the most important quality a leader can possess in the context of running a startup is adaptability. Okay? And then it talks about the process of starting a business and how it's a dynamic and unpredictable process and then how you need to have the ability to pivot and you need to be open to feedback and new ideas, quick decisions. It also talks about building a team that shares this quality and how it's a team effort, but how really adapting to the changing circumstances is essential for success. And then again, it just reaffirms that belief. Adaptability is the most essential quality of all A fantastic answer in my opinion. Let's try the same thing inside of GPT4. So right away it tells you it's not even going to attempt at giving you its personal take because it's not allowed to do that again, higher guardrails, less out of the box thinking, and consecutively less opinions just because it's not allowed to give you those anymore. But it comes to similar conclusions. So let's have a look. It says in the context of running a startup, adaptability is often considered one of the most important qualities a leader can possess. And you can see the wording here is a little less confident. It's often considered. And in 3.5 it's just confident. It's like, hey, it's adaptability, no doubt. Okay, let's keep analyzing this result. This quality enables leaders to navigate the ever changing landscape of the business world, adjust to new challenges and embrace opportunities for growth. Other key qualities include vision, decisiveness, effective, and it's come on, it's just diluting its response at this point. This response was so much better. It was like, hey, it's adaptability. Here's the way it affects you and your business, here's the way it affects your team and that they need the same qualities. And then it also lists other things here that are important qualities. But it takes a step back and says, however, in my view, adaptability is the most essential quality of all as it enables a leader to navigate the many challenges and uncertainties that come with starting a new business. And usually in GPT4 you're going to get longer, more detailed, more creative answers. But whenever you ask for the opinion of the language model, which can be a valuable approach to prompting because it's going to give you more of a subjective take that you might get from an actual person. And it's very untypical for GPT 3.5 to produce longer results than in 4. But in this particular example, whenever you're looking for an opinion, do not use GPT4. And here, for the more advanced folks, I have to say yes, you could definitely craft a Persona where you start this out and you in detail, lay out the type of of Persona, its backstory, its skills, and then you move into this prompt and GPT4 is going to respect the context of it not being vanilla GPT4 anymore, but actually now having a personality. And it's going to give you more of an opinion. But first of all, that takes a lot more work. And secondly, and probably most importantly, I'm just looking for a general opinion here, right? In a certain sense, I want to explore the entire data set, which is the wider Internet that this has been trade on with just asking for a general opinion, AKA across everything that you have seen, what are the most important qualities a person can possess in a context of this specific situation. And when doing that, ChatGPT 3.5 is preferable over GPT 4, but with that being said, in like another 97 out of 100 cases, GPT 4 does produce better results. And if at this point you still wonder how to craft prompts like this yourself, and how to go even beyond to make it super specific to your use case, well, in that case we're leaving copy paste territory where it's all about taking presets and maybe customizing one or two things about them. And we're entering the territory of prompt engineering where it's all about taking a prompt like this and maybe making it three times as long so it can fit your specific specific need, save you time and get your results consistently. And if that's what you care about and that's what you're looking for, I finally have a solution exactly for that problem. My brand new course that I've been crafting for months teaches prompt engineering from the ground up. And we go all the way from basics to advanced prompts and workflows. And look, I'll be the first to say that, oh look, another course by a creator that is trying to sell. No, look, legitimately, the quality that you see inside of all these videos and the thought that goes into crafting some of these, I've put that much care and thought into every single one of the 32 videos that you get in here. But hey, if you're just looking for another YouTube video that advances your prompt engineering skills today and shows you a bunch of more tricks, well, this one is exactly for you. People have called this the most underappreciated video on all of my YouTube. But let me just say that all these tips are only relevant if you're intermediate to advanced. So if that's the case, I'll see you there.
