Ultimate Guide To Auto-GPTs in a Browser

If you use ChatGPT before, but you weren't that  impressed with the results, you're gonna love this   video because this really takes it to the next  level. Automatic GPTs that work by themselves   and get your results without you having to  interact with it. That's the topic of this video,   and what they provide is the following:  usually, if you interact with ChatGPT,   it's you and ChatGPT. You give it commands, it  gives you results. And I find that most people   that are unimpressed with the results didn't give  the best commands to begin with, or they might   have needed internet access to complete the tasks  they were looking for. The apps that we're about   to look at solve both of these, kind of. And I  say that because they're anything but perfect,   but they're fascinating, and for certain use  cases, they're actually really useful already.    So with that being said, it's time to sit back and  relax because I'm about to lead you for the world   of automatic GPTs and what is possible today. Okay, so first things first, there are many

of these, and I would put them into two  categories. There are the ones that run locally,   where you kind of need the basics of coding to  run them. Auto GPT and Baby AGI fall into this   category. And then there's all the web interfaces  that have been popping up over the past few weeks.    That's the ones that we'll be looking at this  week because realistically, for most people,   these are the only ones that you're gonna end  up using. Now, there are different ones. They   come with various limitations and capabilities, so  we're going to talk about that in this video. But   what all of them have in common is that there's  the settings icon, and then you have to link up   your API key for this to work. You can find your  very own API keys here. And before we begin,   I would strongly recommend you set up usage limits  under billing usage limits, and then set these up   to whatever you're comfortable with because it can  happen that you forget about these. They just keep   running and they'll drain your entire account. And  if you want a step-by-step tutorial, I did a video   on Agent GPT a few weeks ago where I showed you  exactly how to hook it up and how to start using   it. But essentially, once you have your API key,  you're good to go. And now we can start exploring.   Okay, so we looked at Agent GPT as soon as it came  out, but now they upgraded it. So you can see they   added different GPT personas that are particularly  good for one task. They also added web search,   but more on that soon. And in the settings, you  can see that we have way more settings now, which   I absolutely love because now you can actually  pick your language. And in the advanced settings,   the most important thing in here is the looping.   So you can pick how often it can prompt itself,   the temperature, something like the creativity of  the model. How much out-of-the-box thinking do you   want? Where one represents super creative and zero  represents super conservative. And then obviously,   you get to pick the model. And I'll just  say, all of these work way better with GPT-4,

and I signed up on the first day and got  lucky, so I have access to it. So if you   don't have access to that API yet, follow along  with this video, and I'll be doing a livestream   where we explore this together with my GPT-4 API. Alright, so that's some nice additions to what we   know already. But beyond that, there's  godmode. space That you might have heard   of already, and this is a different version  of it where it asks for more of your input,   which I kind of like. And at this point, we  should start talking about how these work   and how to use them. So overall, you give them a  goal that you want to achieve. This is not about   coming up with instructions and context like  with usual prompts. This is all about giving   it the final destination to where you want to get.   So instead of saying, "Write me a sales email for   my company," you would say, "Send out sales emails  for my company. " Again, many limitations to this,   and it's not going to work perfectly, but it's  fascinating, and this is where the space is   going. So it's important that we start learning  about this as soon as possible, which is now.   Okay, and then there's a third one that popped up  a few days ago, and this is a web interface for   Baby AGI. And I find this one to be the best,  except it doesn't have access to the internet   yet. So let's talk about that. It's obviously  a key feature in some of these, especially   when you're trying to do research. It's really  good to have access to the internet. Right now,   the deal here is if you were to set up Auto GPT  on your very own machine manually, which honestly   might be too much to ask of most people that never  used code before, you could access the internet,   and you would get all the custom options that  you get from these different apps. But honestly,   even for myself, having that and having this,  I end up using this interface most of the time   just because it's so convenient. And often, tasks  I want to explore come up spontaneously, and over   the past few days, just pulling this up, switching  this to GPT-4, and inputting your objective down   here is what I find myself doing the most.   But if you do need internet access, well,   Agent GPT added it, but right now, this button is  disabled. On the roadmap, it says they'll have it   back up and running permanently within the week.   So by the time you see the video, it should work   again. God mode has internet access, but it really  depends on your input if the results are useful or   not. So this one, I would just experiment with.   And Baby AGI, as of today, doesn't have it yet,   but apparently, we'll be adding it soon. Alright, so we're getting deeper and deeper on   this topic. But before we go into some use cases  and show you what is actually possible, I want to   show you two more apps because you should totally  take note of these. So one of them is iBabyAGI by   Nate Chen on Twitter, and he took the Baby AGI  code and he put it into an iOS app. So if you   go on here, you can totally download the brand  new application on the iOS store and have this   thing with you on the phone at all times, which  is pretty amazing. And a lot of things we look   at is built on top of the shoulders of Yohe here,  which created the original Baby AGI. And by him   open-sourcing it, allowed other people to build on  top. And this is the last app I want to show you

before we get into the use cases, and that is Baby  AGI for us. So you might have heard of this GPT-4   model that is completely free but quite limited.   Well, that love here on Twitter went ahead and   hooked it up to Baby AGI, and this one only has  the basic functionality so far and requires basic   coding knowledge for you to run this. But the  result would be that you have a simplified version   of GPT-4 running on your local machine that can  prompt itself, where you deploy various agents,   and it just works by itself. So really, that's  kind of the alpha version of the ultimate thing   that you would want to have, right? An automatic  version of GPT-4 that can work by itself and   complete tasks on your computer by itself. I  personally wouldn't recommend this yet; it's too   early. I find this free version of GPT-4 called  GPT for All to be a little too basic for most use   cases. That's why I didn't create a separate video  about it. And again, it's also early days for all   these auto GPTs, agent GPTs, however you want to  call them. So they're not extremely useful yet,   although there are some use cases. So now that you have an overview,   let's finally get into the practical part of  this video. Okay, so I would say, let's just   dive right in with a prompt that we used on the  channel here before: "Develop a marketing plan   for my AI YouTube channel that spreads awareness  of my AI tutorials beyond YouTube. " So up here,   I'll pick GPT-4 as it is the best. You can totally  experiment with 3. 5, but let's not kid ourselves,   GPT-4 is so much better inside of these  automated apps. And then, to begin with,   we'll say three iterations, and that's what  I like to start with. And for the first task,   developing a task list is usually the best. So this is it. I hooked up my API. There's not   much more to this. I'll just say "run," and it's  gonna start doing it. And the amazing thing about   Baby AGI in the web interface, it actually works  and it's reliable. It doesn't get lost in loops,   and the results are very useful compared  to some of these other web apps that are

more experimental, and you get useful results  sometimes. But from Baby AGI, I was able to   retrieve useful results regularly. So, as you'll  see, first, it's creating a task list. Here, we'll   just have to wait a little bit until it's done. Okay, so there you go. The first result is we   have a task list. Here are the tasks: "To find  the target audience, research the competition,   create a content strategy. " And here, you  can really see the difference between just   prompting ChatGPT and doing this. Because really,  getting the best results from ChatGPT and GPT-4   is all about being sophisticated about your  prompts and knowing what exactly to ask. Here,   you just lay out a goal, and it finds the  different prompts that it's supposed to run to get   to that goal. Is it gonna find the optimal path  and the optimal prompts? No, it's not perfect,   and it doesn't have all the context it needs. But  is it going to get you better results than just   running this inside of GPT-4? Absolutely, yes. And also, let me say this: one of my favorite   ways to use this is I just start with something  generic or a short prompt like this, and then   it's going to lay out all the different parts.   And I'm just gonna take these one by one and   craft my own prompts from these, right? Like,  I could take "defining the target audience"   and turn this into a 10-line prompt that is  super specific to my organization, my needs,   and I would also provide the information that I  already know about my customers. And then I could   run this prompt sequence by myself inside of GPT-4  and get similar results. But this is really going   to get you started, and it's just fascinating to  look at. And as you can see below, it's already at   work. It's developing a channel branding strategy:  create a consistent visual identity and tone for   your AI YouTube channel to increase recognition  and credibility. And it's just doing that all one   by one. And now I gave it three tasks to complete,  right? But if I say 50, it's gonna go through and   make this entire to-do list happen for you. Okay, so that's the basic approach. You just   take the prompts you have for ChatGPT and you move  away from providing instructions and some context   to giving a final goal and some context. And just  that alone is gonna get you very far. But you can   go further and actually use the internet search  capabilities of these things. So, for example,   if I go to godmode. space and I paste in the  prompt that we shared in the Weekly Newsletter   that went out yesterday, and it looks something  like this: "Use Google Trends (then the URL to   Google Trends) to analyze the search volume and  interest for keywords related to (and who are   going to replace) the product or service.   I'm going to say generative AI education,   and we're going to say in the AI industry. And  then we round it up by saying identify emerging   trends and potential niche opportunities. " Now, we did pick GPT-4, and we're just gonna   launch this. And as you'll be able to see, God  mode has a little bit of a different approach.    It's going to give us the tasks to pick from. So,  this is more interactive and less automated. Okay,   so here are three suggestions. So, I then fire  11 keywords related to generative AI education   in the AI industry. Sure, use Google Trends.   Yes. And then emerging trends and potential   niche opportunities. Sure, let's launch this. And now it's going to start working. It has   these three tasks that we defined. And by  clicking this button, you can just lean back,   have a drink, because this is gonna auto-approve  all the requests for the next 10 minutes. And a   good thing about this interface is that you're  going to be able to give it some feedback while   it performs. You can add context while it goes.   But as you can see, it proposes the first action,   right? Let's use Google and input keywords  related to generative AI education in the   AI industry. And it's going to Google that for  you. And now that we have auto-approve enabled,   it's just gonna do all these things one by one. And again, don't expect any miracles. These things   are developing on a daily basis, and they came  out a few weeks ago. But this is the direction   everything is going in. Okay, you're gonna  be setting goals, and ChatGPT is going to be   prompting itself towards that goal. So, the skill  of writing clear prompts is not going to go away.    But it's not going to be necessary for every step  of the way anymore because it's going to be able   to come up with these tasks and execute on top  of them. But as you can see here, it's always   going to be able to take additional input from  your side. Usually, giving additional context   in some form is the way to go here. And again,  that brings us back to the basic prompt framework   of instructions plus context. Because if the  instructions that it's doing are fine, you don't   need to touch that. All you need to do is add  additional context if it's necessary. But if it's   going in the wrong direction, Yes, go ahead and  change the instructions. So that would be my pro   tip on how to get optimal results. Okay, and the  next step we're going to look at is Agent GPT. And   what I want to do here is actually take one of my  newer prompts: "Create a budget. " And as you can   see, this is going to help us create a budget for  my $3,500 monthly income by allocating funds to   rent, groceries, transportation, health insurance.   Then we assign categories to the various expenses   we have. And then in the end, we say, while taking  into account saving for a wedding in three years.   Okay, we're just gonna paste this into the goal.   And as you can see, we have the instruction here,   right? "Help me create a budget. " And then all  of this is context, and this is a classic prompt.    And as I said before, you could also reformat this  to be focused on the goal. But in this particular   case, where it's "create me a budget," and then  you have all the constraints, that is kind of   the goal. You want the budget, that is your goal.   So just always post the question, "Does my prompt   result in some goal? " If the answer is yes,  it's probably going to work super well in here.   Okay, I'm gonna switch to GPT-4. We're gonna  have it on automatic mode, and I'm gonna set   the loop to three iterations to start with. And  in this case, I'll take the temperature down to   about 0. 4. And you can experiment with this, but  generally speaking, with something like a budget,   you don't want that much creativity, right?   It's just like if you're hiring an accountant,   you want a super conservative and  orderly accountant. That's why I took the   temperature down here. I'll just call this Budget  Assistant, and we're ready to deploy our agent.   And it's going to start working on this. As you  can see, it got right to it. It added the task of

calculating the average cost of a wedding right  away because we need to keep that in mind. And   if you're paying attention, you would realize, "Oh  wait, maybe I could have provided the cost of the   wedding that I wanted to calculate for, right?   Then it doesn't have to do the step, and it's   going to be more precise. " So if you wanted to do  that, you would stop the agent, add that to your   prompt, and rerun it. But let's just go with this. Next up, it researches and suggests a   percentage-based budget allocation for essential,  discretionary, and saving investment categories.    So it respects our categories. That's fantastic.   And only then, it creates a monthly budget plan.    So now it's working on the first thing,  and it came up that in the United States,   a wedding costs around $28,000. And  then it divides it into the three years,   coming up with the number of $777 that you  need to save every single month. Fantastic!   Alright, the next step, respecting that situation,  that context that it just came up with itself,   it starts splitting up your budget into  different categories and gives you all the   reasoning behind it. So look, this is fascinating  and actually really useful if you're creating a   budget and don't have much experience with that.   A prompt like this is really gonna get you going.   And to round everything out, I just want to show  you some of the coolest use cases that some of the   people across the internet came up with. So for  example, this guy hooked up all the GPT-2 11 Labs,   which allows you to produce realistic voices, And  he used it to order a pizza successfully. Why is   it always pizza? I don't know. If you're familiar,  but the first real-world Bitcoin transaction was a   guy buying two pizzas for 10,000 Bitcoin back  in 2010. So I guess we're at that moment with   AI right now. Anyway, here's another one that  I'm experimenting with, and once I perfect it,   I'll show it off on the channel. But you can  essentially use these agents to research your   podcasts and your videos. So here, you had five  different searches in 15 different web browsers,   and GPT just prepared all the topics for  the All-In podcast, which, by the way,   I couldn't recommend enough. It's  my favorite podcast. And then here,   Nicholas shows off how to use Agent GPT, that  we just looked at, for competitor analysis   for specific products. So all these things,  you can explore. I'll include all the links   below. There's so much fun to be had with this. So now you have all the links, you have all the   basic resources, you know how it works. And  even if you don't have GPT-4, don't let that   hold you back. Take the 3. 5, start playing with  it, and see what you get and how it differs from

ChatGPT. Because this is the direction the space  is evolving in. And by you becoming proficient at   prompt design and these primitive versions of  these Auto GPTs, once the better versions come   around, you're going to be way ahead of everybody  else because you're going to understand how   this stuff works at a fundamental level. But if you don't want to do it yourself,   here's the live stream where I'll be  exploring some of these with the GPT-4

API. And if that's not out yet, turn on  notifications because it will be soon.