Ideas for How to Use AI with ChatGPT

There are lots of different ways that we can use AI, so I want to start to give you just some ideas. Now, some of these will require paid accounts, so it can be for answering questions, like the things that people think typically about using Google for. Like, hey, I have a question, I want to ask ChatGPT, and it will give you the answer to that.

We can have it, and we'll go through some of these, but I just want to kind of give you an overall thought process, and then we'll go through and see different examples here. But writing, we could have it draft an email, we could have it write an article. Now, these could be starting points, or they could be finished things.

So think of this as, yeah, I really don't know exactly how I should write this. Maybe it can start it, maybe you go back and use it as a first draft and you refine it, or maybe it's good enough that you can say, okay, this is good to send as an email or good enough to use as an article. But when you think about creativity, some days you're just like, I just don't know what to say for something.

And so you can have it start it, but it doesn't mean you have to use it exactly as it is. You can go back and you can refine it. Or you can say, give me some ideas for how to write an email, or give me some ideas of how I might approach writing an article, because maybe you don't know the steps in doing that.

Maybe you can say, give me an outline of things I should talk about, and then you can write your article based on that. If you were talking to a real-life person and you were asking for help, just substitute ChatGPT, and you're asking ChatGPT for that help. You can ask it for feedback.

You can say, here's something I did. Here's an email I wrote. Here's an article that I wrote.

What are ways that I can improve it? Does this sound right? Is this offensive to someone? Like, if you wrote an email and you're like, I don't know, how is this going to come off? You can even say, what do you think of this? What do you think of the tone? Are there ways that I can improve the tone of this? Is this appropriate in a business email? They might say, oh, no, you should change these things for it. If you've written something and you want to make it longer or shorter, and maybe you're like, this is way too long. I'm just not good at shortening this up.

Maybe you don't have the gift of language. You're like, I'm just not a good writer. You can give it the ballpark of kind of like, these are the things I want you to include.

Can you write a good email or could you write a good article on that? You could change some wording, change the perspective, change the tone, because they make this more professional or make it more casual. Right. Maybe something that you're like, oh, this sounds too stuffy.

Right. So make it more casual. You can tell it to summarize something.

You're like, I don't have the time to read this. Can you just summarize this webpage, summarize this article? You generate ideas, you know, those days where you just like, I just don't know what things I could do. Give me some ideas on such and such.

Translations. I was actually just using this the other day when. So I do Adobe stuff.

So I'm on the Adobe forums. And some people were posting in different languages. And so I actually to answer their question, I asked you, I said, translate this question and then give me an answer.

And then they translated it. And then it gave me an answer in English. And then I said, OK, translate.

What language was that? Like, so somebody ask a question in Polish. And so I said, OK, translate, translate this English, write a response in English and translate it to Polish. So it gave me both responses for me to post on the forum because I'm like, I could write this out, but I bet you I could save time and just let you write it out, especially writing in both English and Polish.

Also, I was like, I don't know what this language is. So I just pasted it in. I said, what language is this? Polish.

You could research a subject matter. And if you use the one that is the search, it'll give you citations. If you don't, it won't give you citations.

Writing instructions, you could say, give me instructions step by step of how to do something. And we'll go through it. I'm going to we're going to go through multiple examples.

It's just kind of get us thinking, like, because Transition PD is such a blank slate. When you open this up, like literally like, what can I do with this thing? You could do so many different things. You can have it write formulas.

So let's say you're really not good at Excel formulas. You could say, well, you could upload a file and say, here's my Excel file. And you could tell what you want to do.

Or you could say in column A, I've got this in column C. I have this. I want to generate a formula to do something. Can you write the formula for me and write the formula for it? And you can say Excel or if you're using Google Sheets or if you're writing DAX code in Power BI.

So you can have a write formulas for you and also have it write coding as well. So, for example, one way I used it is there's things that I'm familiar with, but I can't do it myself. So, for example, I'm familiar with the idea of writing bash scripts, which are terminal scripts that can do things for you.

So I knew it's capable that you can write like command line bash scripts to do things. So I wanted all of the computers here at Noble Desktop to be able to download their own class files. And we have a Web Developer who could give me that information of which file to download in a certain room.

But I needed a script that would go find that information, look at the room it's in, and then download and put it into a folder. I know that there are bash scripts, but I don't know how to write bash scripts. They're not very easy to write.

If I would have Googled and tried to figure out how to write bash scripts, if I had to learn how to write bash scripts, it would have taken me probably 20 or 30 times longer. And all I did was I told ChatGPT, I said, this is what I want to do in a bash script, write the script for me. And it didn't always get everything perfect, but piece by piece, I could take the parts and put it together, and I got it working.

And probably, seriously, like, it would have taken me 20 or 30 times longer to learn how to write those than having it write them for me. And so it works. And now, I wanted to do that.

I wanted to do that for, like, 20 years. I never took the time, because I never had the time to go figure out how to do that. And finally, with ChatGPT, I finally got it done.

Something I had avoided, I'd wanted for years and could never do it myself. And ChatGPT empowered me to do it. So that let me do something beyond my skill level.

Right? Amazing. You can also have it not only write code, but also if you get code from somebody else, you're like, here's some code, explain it to me. And it'll go through and explain what the code's doing.

So you can learn from it. Data analysis. So you want to create some data visualizations.

You want to upload some financial information or something. Have it do some data visualizations, create some graphs, analyze the data, try to pull out insights from it. You can ask it questions instead of you having to do it manually yourself.

You can have it create images, graphics for you. You can give it text files and say, you know, take out the timestamps from this video transcript and just give me just the text. Just extract just the text.

Think of it like your digital personal assistant who you can talk to and have it do things. When you compare this to Google, does Google do any of this stuff? Google is not a doer, right? It can tell you how to do things because you can find articles on the Web to tell you how to do any of this stuff. But Google doesn't do any of this stuff for you.

That's why AI is so disruptive—because now it can help you to do things that would have taken you way longer to do. Or you might not have been able to do those things. So ChatGPT can be more direct.

It's clean results with no ads, at least for now. I do fear that maybe the Web search, it's just a matter of time until ChatGPT wants money and is like, oh, you can advertise to get it into, you know, the top part of our responses. So, I mean, as of right now, there's no ads, but I don't know if that will change.

The paid users, we can search the Web, whereas Google is only a search engine. But just keep in mind, anything can be wrong. Whether it's ChatGPT or Google, anything can be wrong, but especially ChatGPT can be wrong, so you've got to double check yourself.

So let's go through and just see some examples so we start getting used to using ChatGPT. So let's say we want to learn about something. And we can say, tell me how to do something.

For example, like, maybe you're not good at connecting a receiver to a TV. You're like, yeah, I don't know how to do that. So tell me the steps in how to do that because I need to know how to do that.

I'm going to use my paid account here. I'm going to create a new chat. This is a pretty basic thing, so I don't think I would need to waste my 4.0 messages.

I can just go back to the 4.0 Mini, which is going to be faster. It's a pretty simple task. So I could say, tell me how to connect a receiver to a TV.

So it says check your connections, kind of figure out what you have. Ideally use HDMI. If you don't, you could use an optical cable, which is correct.

Definitely that's the better option there. Or like, oh, if you've got a really old thing, you might have to go back to those old RCA component cables, which is true. If I wanted to ask it, tell me what an RCA cable would look like, would it give me an image? So that's an interesting thing.

If I say, what is an RCA cable, for it to explain it, it's going to use words. Now, if I could say, what does an RCA cable look like? Let's see if it generates an image, or no, it's just going to describe it. Now, if I say, create an image of an RCA cable, right? Create an image of an RCA cable.

This is where AI will create an image. Now it's creating it based on what it knows. And I must say, that looks correct.

But just like this, this is not a photograph. AI generated this image. This was created, this image was never taken.

And if you look closely, look at how these ridges are not quite right. But like, that's really quite good enough for me to at least know what an RCA cable looks like. Which is pretty amazing.

But that image was never photographed. It was just created by this computer. Well, not my computer, but by OpenAI servers at this moment.

But like, now if I'm looking at cables, I'm like, oh yeah, those are my RCA cables. It can generate an image. In this case, it's going to be more efficient to just Google RCA cables.

Well, it depends. I mean, if you're in this stream chatting, you know. So let's say, this does use up capacity, yes.

Right? So yes, you could just Google it. Yeah. If you don't want to waste your messages, you know, then maybe just Google that.

If you think it's easy. Yes. That's still using your messages.

Yes. Right. So, you know, that's the only limitation of it.

Google's free, unlimited searches. So, do you need to use ChatGPT for this? I mean, it did work. But do I need to? Do I need to waste my message? Depends on how much you're using it, right? If you're using it a lot, don't waste your messages on stuff that is easily Googleable.

Yeah. Right. Just let it give you the stuff, you know.

So let's say if you ask this, how to connect a receiver to a TV. Say I Googled that. Now, here, this is Gemini, Google's AI.

You know, so they're, of course, that's the equivalent of ChatGPT. They're trying to learn from websites where they link out to how to do it. But let's not compare their AI.

Let's compare it to what people classically think of as Google, which is linking out to websites. So, like, you know, already, oh, these are very clear steps. Like, so, you know, you're bouncing back and forth.

Yeah. You know, going here. Right? Now, here we're starting to get steps.

Yeah. You know? But seriously, like, compared to this, which is clearer, right, there's a lot more you have to look through with this. So as you start to use ChatGPT, you'll start to see where its strengths are and where its weaknesses are.

Like, I didn't need to use ChatGPT to generate an image of that, because a Google search would be easy for that. And it doesn't waste my paid messages, because that does take some time. But if I rarely use ChatGPT, it wouldn't hurt to do that, because I don't have to leave and go somewhere else.

So, yeah. How to write good business emails. Right? So I could say let's create a new chat.

I would still stay on the Mini for something like this, because it's kind of a basic thing. Tell me how to write good business emails. And what's nice is I can say keep it short and to the point in a bullet point format.

And that way, you know, I get the response that I want specifically. And it gives me some ideas. Right.

And if I want to write an actual business email, I could say, you know, I have a client. I have a client proposal for $3,000 for a training on Figma. Please write.

You don't have to say please, technically. It doesn't have feelings. Please write.

Brief email with the details. Ask me anything you need to know before writing it. How do you like that? Ask me anything you need to know before writing it.

Because, of course, it's going to need to know some stuff, right, before it writes that email. So it wants to know, like, okay, what's the client name? What's the date and time of the training? What's the scope of the training? What's the format going to be? Okay, so could you have forgotten to include some of this stuff when writing that email? Maybe you would have forgotten to say, oh, that's right, it's going to be in-person. But, like, it's even knowing that training can be in-person, virtual, or a mix.

It could be hybrid. Because it read that I'm doing a training, right? Is there anything additional? I will say that. The date.

Time is January 12 at 10 a.m. And will be online only. Payment is due up front.

I just answered the questions. And then so it goes through. Now it didn’t put in the recipient name, but I guess I didn’t tell who I was talking to.

They knew it's for Big Corp, right? So it's for Big Corp, but they didn't know who at Big Corp I'm writing to. So they're like, Oh yeah, you have to fill that in because you didn't tell me that. Figma training proposal for Big Corp and the dates.

Like that's a nice subject line, simple, right? And then they put it in there. It will provide this. Here's the thing, right? And I notice I didn't tell it what duration was, right? So I had to tell it that, but they kind of put in all the places where I need to fill something in.

Notice that they understood what we were doing. When I told them that they just knew to continue on that same thing, but notice there was a memory. So it now remembers that I did that training or that I'm doing that training.

So in the future, maybe if I was saying, give me a schedule for something, it might say, Oh, well on January 12, you're busy because it learned that I was doing that training. Now, if I don't want it to remember that I could go into my memories and I could get rid of that. Memories are a paid feature only.

If you're in the free plan, you're not going to have that memory be created. Sometimes it creates memories. And I'm like, no, I don't want you to remember that.

That's not important because you have a limited number of memories and that's something I don't need it to remember. If you see it, you can just clear those memories. If you don't want it to remember that, if it is important about you and you want it to remember that across your chats, you can just let it be.

And then it will remember those important things across your chats. So, right. So that's kind of similar to what I just did there, right? Like, and explain yourself, you know, write a succinct professional business email.

You can take with subject line, telling some such, you know, my company, this will provide this and, you know, give it that stuff. And you can always have that follow-up of, ask me any questions that you need to, to be able to do this. It'll still do something like this, even if you don't have that ask part.

So for example, if I just use this, if I tell it to do that, it'll still do it. It just might not know as much. So sometimes if you say, ask me any questions you need to, you might get a better response because it'll think, oh, I need to know this.

It won't always ask you those questions unless you ask it to, you know, because you just told it to do it. So it does what you said. If you told it to do it, it's going to do it.

And it just might need to leave some blanks. And you can fill those in. If you wanted to do it in an interactive way to be able to go back and say, can you make this longer? Can you change this? That would be a good use case for the canvas because the canvas option lets you go in and say like, can you fix this part? Can you fix this part? You could go in and actually edit the stuff, right? So like in my paid account, if I did this with the canvas one, it'll open up the canvas.

So I could actually add in, I could actually add in my name, Dear Sue. Make this more optimistic. Okay.

It seems I wasn't able to make the edit as intended because the match let me try another approach. Okay. So here's a revised.

I hope the message finds you well. We're excited. Okay.

Yes. Okay. So I'm going to say yes.

Make that change. Fail to edit. Okay.

So changing to be perfect. It never fails. No, it obviously failed this time, but we can try it again.

Right. So I can say, let me select this, make this more upbeat. See if it works.

Still having issues. Okay. So check if you guys have an error.

Well, they do call this beta. So they're still testing this. So it did have a little bit of an issue there, but at least, you know, it's not you.

It's just, you know, they are testing this out. So it's still a brand-new feature that they're working through some issues for. Yeah.

We already went through some of these follow-up messages, summaries of files. So you can upload a file or copy and paste the text in, and you could say, summarize this. So in the class files, I've got some text files that we can use.

So here I've got this Hawaii file and I could either upload the whole file or I can copy and paste text. If I copy that, I could say in a new chat, summarize this and then paste in the text. And they would summarize that.

Now, maybe that's too long. I could say change it to be bullet points. This is where the beauty of AI comes in: you can be specific about how you want to see this or how long you want it to be.

Rewrite that for a fifth grader. Remember the TV show “Are You Smarter Than a 5th Grader?” You might want to lower the education level so it's simpler. So like here, one of the highest costs of living in the U.S., but it's also down here, living in Hawaii is expensive, but it's also one of the wealthiest states.

So pretty similar, but just a little difference in the way that it was written. So it could change the tone of things. The other thing you could also do is you could upload the file too.

Do you have a question? Yeah. If you take me working from the same document and I put in the same job, logically, wouldn't it give me the same response? No. No.

When you do this yourself, you're going to get a different response. Just think of it as if you asked me to summarize a document today and a month from now, am I going to, I'm not going to do it the same way twice. This is not storing the results.

This is on the fly, taking all of this stuff and generating a response from it. But it's easy to think of it as a machine that would repeat the same thing over and over again. But the way that these models work is they on the fly generate that response.

Yeah. Custom or factory settings, and it will not create the same response exactly the same way twice. Which is important because if you ever delete this, and you're like, oh, I can always just ask that question again, you will not get that exact same response.

You might get a similar response, but it might not be the exact same response. If you want this response, you know, you're going to have to either save it over here. Just keep it.

Don't delete it from your account or copy and paste it. You know, there's a little copy button. You can copy it out here.

Because once this has been deleted out of your account, that exact response will probably never be recreated exactly as such ever again. And sorry, I've just got another question on this because we're really struggling with this work. We're trying to get some summaries of reports.

And how can you tell us that? Actually, that's been sitting down before we sent this and just sort of merged down the answers a little bit to be a little more generic. Right. So I like to repeat the question for everybody online just so they can hear it for the recording too. The question is, how do you ensure accuracy? Basically, you can't ensure accuracy other than proofreading and testing things yourself.

The nature of AI is you're asking a virtual person essentially to do this for you. How do you verify that a human did a good job other than double-checking their work? Could they mess it up? Yes.

Could AI mess it up? Yes. It's impossible to ensure that this is right and perfect, that they got the right idea other than this is where the human element comes in. We have to be the verifier of this stuff.

It should be pretty good, but did they always get the points that you wanted to feature? Not necessarily. Could they hallucinate? Could they make mistakes? Possibly. It can definitely happen.

They say right down here it can make mistakes. Check important information. It's said on every screen what you're supposed to do.

You're supposed to check them basically. I would suggest trying to keep it closer to what you're living in and to not change. To change the language or? Basically not start from scratch.

Keep to the text that you're just giving. Are there prompts? Actually, I think the way you said it, where you say keep to the source text as much as possible, you just told ChatGPT. Keep to the source text as much as possible.

Be as accurate to the source text as possible. I think those are good ways to say what you wanted to do. Just say that.

Keeping as close as possible to the source text. Also, if you want to see something specific in terms of length, in terms of format, do you want headings? Do you prefer bulleted lists? How long should your paragraphs be? What tone do you want? Any of those things, when you want to change things, you need to be specific about that or else it will just guess what it thinks is enough. But it might be too long or too short.

You might have to do a follow-up. The more accurate you are in the way you describe what your output should be, the more you'll get what you had in mind yourself. But I can't say there's some magic thing to stop it from hallucinating because it's the nature of AI.

When you fundamentally understand that this is a system that has learned patterns in language and tries to predict next words, does it truly understand what it's doing? It's debatable. Is it really smart at doing things? Yes, it can be very good, but also it can misunderstand what you're saying because it doesn't fundamentally truly understand it. It gets very meta—like, well, who truly understands anything? People can think they're totally right and be completely wrong because they thought they understood it.

I know it gets very meta, very existential, but it's true of AI. So, you can't prevent it from hallucinating very easily. It was interesting, actually, there was one company that did figure out pretty much how to eliminate it.

So, it was always thought that these LLMs, these large language models, that you couldn't eliminate hallucinations. And they were trying to do stuff with legal things. So, in the law, you can't hallucinate because that's not allowed.

There was actually one lawyer who started using ChatGPT. He had it come up with his argument, citing case law, and it just made up a case that sounded correct. It wasn't a real case at all.

He got in trouble with the judge because he didn't double-check it to make sure it was a correct case. The judge was like, this is not a case. He's like, sorry, judge, this was the first time I was using ChatGPT.

And they're like, that's not an excuse, and dismissed the case.

So, yeah, he got in trouble. So, there was a company that was trying to eliminate hallucinations. And what they found was that eventually they could eliminate all hallucinations by going through and really testing the prompts and being super, super, super specific with the prompts to such an extent that they had about a thousand different tests for each prompt.

If you were exact enough with the prompt, they could eventually get rid of all the hallucinations through a very complex series of tests. That they could basically get the hallucinations down; they got rid of them. But it really just goes back to when we talk about prompt engineering.

The more accurate and detailed your prompt is, the more accurate your output will be. But it's very difficult to eliminate hallucinations. So, we should just double-check what we're getting out of it.

So, it can be very good. So, I don't want you to think, oh, this is crap. I can't trust it for anything.

It can be very good. But I also don't want you to think you can blindly trust it either. Like, it gets a lot of stuff right.

But sometimes it's like, no, that's completely wrong too. It does both. Like, one time there was something really hard I wanted to do, or there was something in PowerPoint, I thought, I'm like, I don't think I can do this in PowerPoint.

Maybe I just don't know how to do it. But I was pretty sure you couldn't do it. So, I asked ChatGPT, and it's like, here's how to do it.

I'm like, I bet this doesn't work. But it was very confident. It's like, do this, this, this, this.

And I go to try it. I'm like, yeah. If you just read it, it sounds correct.

But those features don't exist. So, if you didn't test it and you sent those instructions to somebody, when they try to do it, they would come back and be like, this does not work.

So, when I tested it, I'm like, yeah, this doesn't work. And then I said it doesn't work. And it's like, oh, I'm sorry.

Here's how to do it. And basically, it came back with something else that sounded right, but it didn't even know that it's not possible.

It didn't come back and say, no, you can't do that. It kept writing things that sounded right, because it didn't really understand that you can't do that thing that I was trying to ask. But then there are other times when I've asked it to do things, like to write some code for me, and it comes back with like 60 lines of code that work perfectly, and I'm amazed that it can do this with some niche language that most people would never even know existed.

I'm blown away. I'm like, wow, you commented every piece of code, and this is what it does, and it worked perfectly. So, it can be amazing and yet dumb at the same time.

Which is why you always have to check, because any given response, you don't know whether it's the genius or the dummy with that particular response. Through experience, as you start to learn how it creates the output, you will start to find where you can trust it a little bit more and where you shouldn't. But ultimately, I have to test everything that it does.

If it gives me code, I'm going to test it to make sure it does what I hope it does. I'm not just going to run that code without testing it. I'm not going to take an email and send it out without reading that email.

I would never trust this in its current state to just autonomously do things for me. I would not want to write a chatbot that's just writing responses to people and giving car prices. For example, there was a car dealership that put ChatGPT on their website, and some guy got it to sell him a car for a dollar.

He convinced the chatbot to sell him a car for a buck. They didn't have to sell the car for the dollar. They were like, no, that was a chatbot.

But he got the chatbot, ChatGPT, to sell him a car for a buck. He persuaded it. But, yeah.

Now, if I'm in a free account here, notice when I hover over this, it says 4.0 Mini doesn't support attachments. Try again. So, I've already hit my 4.0 limit on my free one, starting at 3.03. So, I've already messaged the 4.0 too much.

See how it says I've hit my free plan limit? So, in my free account, I can't upload files anymore. If I hadn't wasted my messages in my free plan, I could upload files. So, I can show you that in my paid plan.

In my paid plan, I can upload files. So, I could also—you can connect it to Google Drive or OneDrive if you want—or I could upload from my computer. And I could go into, and I could attach Word files, plain-text files, rich-text files.

I could upload them. And I could just say, summarize this. So, if I don't want to copy and paste, I could just upload the file and say to summarize it.

So, I could have put in a PDF and said, summarize it. Pull out the key important points. It makes it more convenient because it's hard to copy and paste an Excel file with the columns and things.

Same thing with a PDF. Like, how are you going to get a PDF to copy and paste it? That might be challenging too. One of the benefits of paying is you get more messages.

If you're just doing it for free, you're not going to be able to upload many files. You can also say to summarize a webpage. So, let's say this webpage here.

I could say, summarize this. And notice it's going to say browsing. That was quick. And then it does cite the source.

Would that work in my free account? I try to do this in my free account. So, it did say searching the web. So, it does work here.

I do have a link there. So, this did work even in 4.0 Mini. It still could browse the web and search that website.

Okay. So, that's good. Even falling back down to the Mini still worked.

Again, notice the difference. Just to point out, every time you ask it to do something, it does it in a different way. So, this is where if you have a certain preferred format—say you always like things to be written in a certain style—later we'll see you can write some custom instructions where it will always remember certain things that you like, like always favor bullet points at a fifth-grade level or always be professional or creative or whatever. We'll learn later that we can do custom instructions so that we don't have to keep typing them in every single time.

For improving your writing, you might say, proofread my writing, fix any spelling and grammar mistakes, and make suggestions to improve the quality and clarity. You can upload an example of your writing, which is nice. So, let's say I take that and create a new message.

If I'm using the free version, I could just copy and paste the text in. Or if I have a paid version and I attach files—or if you're still using your free version that still has some leftover 4.0 messages—you might be able to attach files even in your free version. Then here, I've got some writing to improve.

It could be a Word file, text file, PDF file, rich-text file. So, it gives me the revised version of it. So, here's the revised version.

Suggestions for improving clarity and flow. So, this is what it did.

It restructured a few sentences to make them more concise and easier to read such as changing “making processes more efficient and life easier for humans” to “making processes more efficient and life easier for people.” Replaced “artificial intelligence” with “artificial intelligence.” See, there's two l’s versus one.

So, there were some spelling mistakes. Importance. Efficient without an “i” to with an “i.” Okay. How many spelling mistakes were there? Okay. It said there were seven spelling mistakes.

Now, I'm curious. Do I believe it? I don't know. One, two, three, four, five, six, seven, eight, nine, ten.

No, it's complete—no, it's hallucinating. I had more spelling mistakes.

Right? That's not necessarily a spelling mistake. It's grammatical. But there are eleven spelling mistakes.

It sounds so confident. It's like, oh, there are seven. There you go.

But look, it lists nine. Okay. This is where you have to double-check stuff, right? It says there are seven and gives you nine, but there were eleven.

Okay. Now, did it fix them all? I'm pretty sure it did, even though it doesn't really know what it did, per se. So, if I create a new file here and paste it in there, it fixed them all.

It just doesn't really know what it did. It sounds correct, but you should have a healthy skepticism about what you ask it, because sometimes it sounds correct, but it's wrong.

Especially when you ask it intelligent questions like that. Having it rewrite the stuff? It's kind of good at doing that because it's good at writing. Knowing what it did—requiring higher-level consciousness—sometimes that's where it gets tricky.

As you start to get used to doing things and double-check, you start to realize what it's good at and what it's not. It did the writing task fine. The questions about what it did—because it's not conscious—it doesn't really know what it did.

Here's a famous test that a lot of people have given it: How many R's are in strawberry? Let me just ask you: How many R's are in strawberry? Three. Oh, it actually gets it right.

Okay, so for the longest time it said two. It said two for the longest time. So maybe they fixed that, but for the longest time it would just say there are two.

They're always improving the models, right? They can give it more training; they can help it out. People classically say this is the worst you'll ever see this.

It will get better over time, because as it gets more training, as it learns more stuff—think about a person developing over their life. Hopefully they learn and grow and get better, right? A five-year-old is not as smart and can't do as much stuff as a 50-year-old, assuming they've grown and educated themselves. Same thing with AI.

The models continue to be trained and learn, so they are improving. So, just because something didn't work today doesn't mean it won't work a month from now, even in the same model.

The reasoning models got it right, because they thought about it more, whereas the faster one didn't. If I go back to the Mini… look, it contains two R's. Oops.

That's funny. “R” appears twice.

See, there you go. I even chose the reasoning model—the faster reasoning model—which, in theory, should be smarter. In theory.

If I choose the better one here, rather than the faster Mini one…but see, the reasoning models think; they do stuff. That one's better.

If I go back to the regular, the non-reasoning model of just 4.0 Mini—yeah, see, two R's. Generally, the full-fledged models are better. So, both of the Minis failed.

0.1 Mini failed. 4.0 Mini failed. The Mini models—they're faster, but they are not as intelligent.

But you have more messages with those. For simpler tasks, you can go with the Mini models, where you have more messages. The 4.0 was better.

The 0.1 Preview was better. So, the full-fledged models, which you get fewer messages for, are generally smarter and get things better. If it's a simpler thing, you might be able to go with Mini.

If it's more complex—especially reasoning—you've got to go with the higher-level models, where you get fewer messages. These are the good mistakes to see because you understand that different models will give you different answers. So, rather than accuracy, in terms of writing and getting more complex and varying writing from a more generic response, what model would you suggest? Is it worth it?

I haven't done a lot of writing with the 0.1 models because my understanding is that they're more used for reasoning. If I was trying to get questions answered from information that I was providing—like if I was uploading a file that I wanted it to go through and then ask questions—if they're hard to answer, higher-level questions, I would go with the 0.1 models because they take some time; they think about it. Depending on the complexity—the nuance—of that question.

If it's a simpler question, 4.0. GPT-4 is a general model that's good for most things, and it's perfectly fine for writing, especially writing and so forth. GPT-4 should be fine for most things, even answering most questions about the writing, summarizing, those sorts of things.

It's your general Swiss Army knife—good with lots of different things. I reserve the 0.1 models for the hard stuff, where you need to think through things—hard-to-answer questions. And clearly, the regular full-fledged model, the one that's in preview release right now, is going to be smarter at those responses than the faster Mini one.

But I reserve the really hard stuff for those, especially since I don't get a whole lot of messages—50 a day or 50 per month for those. Does that help? Yeah.

Well, maybe I'll follow that up with the notion that people say you can tell text generated by AI versus text written by a person because of the generic nature of the text. Right. So, if you want to avoid that…some people say they can tell it's written by AI because maybe the word choices aren't the ones you would use.

One way to work around that is in your prompt you could say avoid certain words if you find it using certain words you wouldn't use. The other thing is you can also—later today we'll see—we can create custom GPTs if you have a paid account, and you can give it examples of your writing to train on and then say “use this example of my writing and create a response in my style.”

Even if you're not creating a custom GPT, you could upload a file or paste in some text and say, “Here's an example of my writing. Please write an article on this topic that uses my style of writing.” Now, if you want to save that so you can use it repeatedly, save it as a custom GPT.

Think of a custom GPT as a pre-trained transformer. ChatGPT has been trained on all its knowledge, but you can make a custom GPT that's trained not only with what ChatGPT already knows, but you could add your own material. Maybe you tell it something about your business, your writing style, and now this custom GPT—which is just for you—also knows about your business or your writing style. Then every time you talk to it, it has that knowledge of base ChatGPT plus whatever else you've given it.

But even if you don't make a custom GPT, which requires a paid account, even in a free account you could say, “Here's an example of my writing style, ” paste some text, and then say, “Please write an article on this using my writing style, ” giving an example. That way, hopefully, it doesn't sound as generic and sounds more like you because it's looking at the language that you use.

Another thing as far as getting it to do things for you: you could say, “Hey, I've got some timestamped text here.” There's this video transcript that I have.

One thing that we often have to do is we record a video, put it onto YouTube, download their auto translation, but they put in all these timestamps, which I want to get rid of. I just want the text. This is not something I typed in.

This is something that was generated from a video. But, you know, even just taking all the text from this, think about how tedious that would be to extract just the text out of that. So I could either copy and paste it, or I could upload, assuming that I have some messages where I can still upload stuff.

And I could say either something like extract the text from this, or remove the timestamps from this. Either way, could be an example of how to do this. And so I could say extract the text from this.

Or I could upload the file. And there it extracted the text. But I don't like how there's some line breaks in there.

So I'll say fix the line breaks. Because see, it was just extracting the text literally. Yeah.

And now it's getting rid of, but see, this is where it's so smart. It's kind of figuring out like, okay, let's keep the sentences together. Right? Because like it didn't remove every single line break, because that would be kind of annoying maybe.

But in a sense, like, oh, that's like a sentence there. That's another sentence. Right? So like it kept the sentences together.

And that's like a tedious thing that like nobody would enjoy deleting stuff and everything. Like I'm glad to give AI stuff like this, because nobody's desire is to sit all day just cleaning out time codes from text. So perfect example of how to use AI as your personal assistant to do this stuff.

Yeah. So you might have to say also remove the numbers. Sometimes the numbers would still be there, because it would extract the dates and times.

Or I actually chose just times. That's where I don't always know exactly what you're going to get, because when you do it, it might be different from what I get. If somebody says, here's an exact prompt, just know that, think of that as a starting point.

And when you ask that, you might get a slightly different response than them, or maybe a majorly different response. And you can always tweak that. Think of all these prompts as just starting points for ideas.

And if you wouldn't write it that way, like here, I gave you two different ways to write it. But if you use slightly different wording, like just maybe you say, give me the text, like it could still use those things. There's nothing special about these exact words.

This is just how I would write these things. But get me the text from this, would also work. It understands natural language.

This is not like command-line software that needs exact commands. It understands natural language. Now, one thing that I found when I was having it do some things, ChatGPT does not want to try to let people do bad things with copyrighted material.

So they do have some safeguards. And there was an article that I just wanted to remove. It was an article that I fully had the rights to do this with.

But there were many little footnote references, like five, six, seven, eight, all spattered throughout the text. And I wanted to get rid of them all. And I was like, oh, can you just remove the footnote references? And they said, my guidelines prohibit me from directly copying or regurgitating unmodified content from web pages or documents.

This is to respect copyright and intellectual property rights and to encourage users to engage with the original sources for complete and accurate information. If you have specific questions or need a summary of the content, I'd be happy to help within these constraints. So there are some constraints as far as what ChatGPT will do.

And I think in this case, they're thinking, if we're transforming it, because it is a transformer, a generative pre-trained transformer, if you're not transforming it into something new, it's kind of thinking, well, why are you involving me in this? Because I'm basically just regurgitating it as it is exactly. It's the unmodified part. So I think the extracting of the text worked because there was a big material change.

Half of the content was times. But maybe if it was a more minor thing, maybe it would fail to do that because it would say, oh, I'm basically just copy-pasting text. I can't really do that.

So there are parameters. You might get refusals to do certain things. So they do have some safeguards for generating images, for example, that we'll see later on.

You can't generate images of famous people, not with ChatGPT because ChatGPT will say, no, I won't produce. If you said, give me a picture of Donald Trump doing something, they're not going to make a picture of Donald Trump doing something. They might make a picture of a man doing something, but they're not going to make it look just like Donald Trump or any well-known celebrity.

Are there other AI image generators that will do that, that have no safeguards? Yes, but not ChatGPT. ChatGPT has some safeguards. If you don't want safeguards, you got to go to some other AI.

ChatGPT is like, yeah, I'm not going to just let you do just anything you want. I do have some integrity of things that I don't want people to do. Now, interestingly enough, it can produce some types of files.

So you might not even think to, like, say, save it as a Word file. So like here, if I want this, I'll say, save it as a Word file. Granted, I could always just copy and paste it, but if I download that file, there's a Word document with that text in it.

So they created the Word file. You just might not ever even think to have it create a Word file for you. Is it limited to the Microsoft suite, or can you say save it into a Pages file, like the Apple version? Let's see.

Save it as an Apple Pages file. I've never tried to do that. Let's see if it can do an Apple Pages file.

No, it does not know how to do that. Behind the scenes, it actually writes Python code to do this. If Python can do it, then ChatGPT can do it.

So yeah, not that you have to understand Python, but this is Python code. So Python needs to be able to do it. What's crazy is they wrote Python code and executed Python code to generate a Word document.