Alright team, let’s dive into today's tutorial on creating consistent characters in AI-generated images. We'll explore using a character reference URL for consistency and in-painting to vary scenes with the same character. We’ll also look at a blending hack, though it might not be available forever. Let’s get started and elevate our creations!

I'm not sure how creating consistent characters will really benefit our fizzy drink company. \n
I mean, isn't it more important to focus on our product and marketing strategies? \n
But I guess if it can enhance our brand image or help us in storytelling, it might be worth exploring. \n
I'm curious to see how this tutorial will change my perspective. \n
Let's give it a shot and see what we can learn!

Consistency is key when it comes to creating memorable and believable characters in storytelling. Whether you're writing a book, a screenplay, or a video game, ensuring that your characters remain consistent throughout different situations and scenarios is essential for maintaining their authenticity. In this blog post, we'll explore some techniques and tools that can help you achieve consistent characters in your creative projects.

One effective method for creating consistent characters is demonstrated in the transcript above, using OpenAI's DALL-E model. The transcript highlights how you can generate different variations of a character by using different prompts. However, to maintain consistency, a specific parameter can be added to the prompt, known as the "C dot E F character reference." By adding this parameter along with the character's URL and specifying a desired consistency level, the model can generate variations of the same character that remain consistent with each other.

Additionally, the transcript introduces another technique that utilizes in-painting to create consistent characters. By taking an initial image of the character and manipulating it using in-painting tools, you can create multiple variations of the same character in different scenarios. This technique allows you to cut out specific parts of the image and replace them with new backgrounds or outfits, resulting in a consistent character that can be used in various situations.

One of the advantages of using the in-painting technique is that you can generate multiple images of the same character, each with different expressions, poses, or settings. This can be particularly useful when mapping out different scenarios that your character needs to be in. For example, if you're writing a detective novel, you can create images of the character in various crime scenes or interrogation rooms. By having a range of consistent character images, you can effectively illustrate the character's development and experiences throughout the story.

Furthermore, the transcript suggests that you can use the generated character images as training data for your own stable diffusion model. By refining the generated images and iteratively improving upon them, you can create a more customized and consistent character. This can be especially valuable if you want to have more control over the generated variations and ensure that they align with your creative vision.

It's worth noting that while the techniques mentioned in the transcript primarily rely on AI models and tools, there are other traditional methods for developing consistent characters. These include carefully crafting character profiles, defining their personalities, motivations, and backstories, and ensuring that their actions and dialogue remain consistent with their established traits. These traditional methods can work in conjunction with AI-generated images to create well

Consistent Characters

 Okay, I'm going to show you how to do consistent characters in mid Jenny. So this is an example of a character I wanted to replicate. Like I've got a picture of her and the party just generated this, using this prompt. But say for example, I was writing a story about. She was one of the main characters I'd want to show her in different situations, maybe in the dentist office or at a birthday party. 

Whatever it is. And if we use the same prompt again, we're just going to get different results. And see all of these like very clearly different people, even though they're quite similar. So how do you keep them consistent? So mid Jenny actually added one way to do this. 

So if you click a copy of the link and then let's put in the same prompt again. Actually, yeah, that'd be grab the prompt.  

Just say 

At a dentist office.  

All right. A.  

Okay. So that is a new prompt. And obviously we just run that. We're going to get like a very different person. But no, to keep it similar, just going to click copy link.  And then the parameter you need to add here. Is , 



C dot E F character reference and then paste in the URL. Then you want to add dash. 

You want to say C w. And let's just say. It's, this is from zero to a hundred and so we want to keep it at that's how consistent you want it to be? Let's keep it like 90 on it to be very consistent. All right. Now let's pace that and see. How it works.  And what a man Jenny should do here is keep that character relatively consistent. Over time. 

So that would be really helpful.  

But obviously, while that's running, I'm going to tell you about a different technique that I use. That doesn't rely on this proprietary feature. It uses in painting instead. So I'm going to. And to show you that in a second. Now let me see. Is this look like the same person is. Pretty similar. 

So this is her.  

And yeah, you can see a condo looks the same. So a little bit wrong though. I think to be honest and look, she's wearing the same sequent dress. I didn't want it to be wearing that. I want it to be. In a dentist office. So we'll see if we can address that.  All right. How do I do it? 

What I do, my trick is I want to pan out. I'm going to click pan.  Submit.  And that's going to add some extra bits to the right-hand side.  And we're going to do that twice in order to get a two to one aspect ratio.  

And you can see here that we've added some extra background side.  

And then we need to upscale one of them. Does really matter which one.  And the moon to pan that, again.  

When you get here is basically. Enough space for two of the same character.  

Okay, let me just upscale one of these.  I'm just doing the one with the least amount of noise and the. Background. Okay. So now we have our character. We have, a consistent background on this side. All right. What we can do now is we can do a very region.  



And this is where the magic happens. 

So you can do this actually with any AI tool that offers in painting. So we're just going to select.  This sort of. Basically cutting out this part of the image. So the right-hand side of the image. And instead of. Doing one image with a background on the right hand side, are we going to change it? 

So we're going to say. Sorry. Two images.  Is side by side. The same.  



And you say At an 1975. Priority.  

And then we're going to say.  

This.  Hey. Your sweater.  

Alright.  

And then we're going to get rid of this stuff. We don't need that.  All right. And we've got the aspect ratio two to one.  So let's see if that works and this is not always consistent by the way. Sometimes it does go a little bit crazy, but that should give us the best chance of success. So how does this work? When you're in painting you're basically erasing part of the image and then prompting it for what you want. They image new part of the image to be well. The. Here we go. 

So what we're doing is we're taking instead of this, a wide angle, A shot. We're cutting it in half, or we're saying it's just two pictures of the same. Person right.  

And you can see here that, okay. It didn't get the dentist office saying properly, but it is the same person, right? It's very consistent, which is good. And she is wearing a blue sweater. So that is pretty good. And now all of these have our eyes closed. So one of the things I'm going to try now is I can go back to a very region. Did I say one it's still cut out. You're saying let's just try. The. She's Chinese five for potty left. At wearing the blue sweat 

Staring straight at the camera.  

Alright.  

All right, the, see if that fixes the problem, you're back into any normal kind of prompting techniques that you're going to use to do this. But the key is that. We're in painting. So there's two. Pictures of the same character now. The reason why this is useful is now we can basically cut out. If we drop this into Figma or Photoshop or whatever, We can cut out the different images that we want. 

So we can do this a bunch of times and then get Maybe 10 20 images that the same woman. And there's a few different things. We can do that, but one is we can. Map out all the different scenarios that we need her to be in. But the more interesting thing is that you could then use that as training data to train your own stable diffusion model on this specific character. 

And that's something I've done. In the past as well. You can see that if some of these are better than others you might take a bit of an iterative process here. Like maybe you need to get her with her eyes opened first, keep the sequined dress and then try and work on. Maybe a different background and they'll try and work on a different outfit. And it is a bit of an issue that you're active process, but the really cool thing here is that now you have a method for getting to different. Perspectives or two different behaviors or situations from the same character. 

And that character is consistent. And this works really well. Now the other thing actually I've found.  In general is that you can't do this with uploaded images. There is a little bit of a hack. They have this blend functionality. And if you blend two images together, they're similar. 

You do get a generated version in the same style. If you did need to hack it, I don't know how much longer that. But if he didn't need to hack it and wanted to get a character that you didn't generate that as like an upload character, then you could do that. But for now they've disabled. The ability to in paint, uploaded images. But yeah, hopefully that's helpful for you guys. 

And it gives you a lot of different opportunities to create more interesting character development in your stories.

What is the purpose of using the "C dot E F character reference" parameter in the tutorial?

What technique does the tutorial suggest using instead of the proprietary feature to achieve consistent characters?

How does in painting work in maintaining consistency in characters?

How can the generated images be further utilized in character development?

What is a limitation of the in painting technique mentioned in the tutorial?

There is a random element to diffusion models, so you get a different result every time you run the same prompt, making it hard to keep characters in a story consistent.

Every time I prompt it generates a different person

GoolyBib Headquarters - Early Morning Rush Hour

You're at the GoolyBib headquarters early in the morning and the rush hour is just beginning. You're here to learn how to automatically create blog articles using AI and add translation into French, Spanish and German. Our CEO, Gustav Gieger, has asked us to figure out how to set up a sequential chain step with input and output operations, use custom chains, and output keys to reference the output of previous chains, and finally, how to introspect each step of the chain.

We're in a tight spot here. We need to get these blog articles up and running quickly, but it's going to be a challenge.
We need to figure out a way to make sure that we don't miss any important steps. We need to be able to quickly set up the input and output operations, use custom chains, and reference the output of previous chains.
We need to be able to do all of this quickly, without missing any crucial steps. 
Can you figure out a way to do this?

Creating blog posts with AI can be an incredibly powerful tool for businesses, as it allows you to quickly create content at scale. While many people assume that AI-generated content is of poor quality, that isn't necessarily the case. In this tutorial, you'll explore how to use AI to automatically create blog articles and even translate them into multiple languages.

Using LangChain you can create a SequentialChain which will outline, write, edit and even translate your blog posts. 

The output_key within a SequentialChain is useful because it allows you to reference the outline in future chains. 

Once we have these three separate LLMChain objects, you can combine them into a single sequential chain step. 

After running the chain, you'll get the article outline, the generated article, and the edited article.

Generating Articles Using LangChain

 In this video, you're gonna explore how to automatically create blog articles using AI. At the end of this video, we'll also do a stretch an add-on translation into French, Spanish, and German too. first things that you are obviously gonna have to do, if you don't have it already installed is install Lang. 

Chain Using pip, install, Lang chain quiet at the top. 

After that we're gonna import a range of different lang chain packages, such as the chat. l l m module. We're gonna import l l m chain, which allows us to make lang chain l m chains. And prompt templates. also import the sequential chain. That's gonna allow us to combine multiple Lang chains into a sequential step of input and output operations. 

Let's start by setting up the chat model, which we're gonna use for our L L M, We'll also have to create a template for each of our chains. So to start with thinking about it from a high level perspective, the first chain that we're gonna have is to generate a very descriptive and detailed article outline. So we set up the template variable saying you're responsible for creating a very detailed article outline for blog, post idea. 

Notice how we then create our LL M chain using the chat model for our L L M. And the prompt comes from this prompt template. We're also associating a output key. You're gonna see why output keys are really, really useful. Uh, and the reason why is that when we use the sequential chain in the future chains that come after. 

This article, outline chain. We might wanna reference the outline and and generate an article from that. So you'll see when we actually go to the second chain that we're gonna put in our sequential chain step Notice how we have here. This article outline. And this is a big gotcha. From Lang chain. Is that when we're using a sequential chain that we'll show you later. 

We can actually have an output key associated with a chain, and then we can use that inside of a downstream prompt template. As you can see in our second chain here. So then the second chain, the template is given an article, outline, Write a very high quality and detailed article. The article must be at least a certain number of word counts in length. And here's the article outline. 

We also set up our chain in exactly the same way with the prompt coming from the prompt template. We also have an output key of generated article. Uh, after generating the article, it's also important that potentially we edit the article and further refine the content. So we've got a third step here where we say, given an article, edit it to be very high quality and detailed. The article must be at least word count words in length. And we also give it some principles that needs to follow such as, you know, fixing grammatical errors, spelling areas, factual error, making sure it's the right word length, and that it's very detailed. 

We, again, set up a article editing chain using the exact. Same pattern that you've seen above. And also notice the output key is edited article.  And here comes the great bit. So now that you've set up all these individual chains, we can then put these inside of a sequential chain inside of a list. Notice that the only inputs that we actually need to the model, is the blog post idea and the word count. 

, we're also outputting the article, outline the generated article and the edited article.  so once you run this, you can see that we pass in a, we create a variable called chain and we then pass in a dictionary.  Of our input variables. In our case, we had a blog post idea, which was how to become a better data engineer We have a word count of around a thousand. And then this took around a couple of minutes. And depending upon how many chains you put inside of a sequential chain, it will add on extra time to running these. 

So don't put a hundred of them in and think you're doing lots of great work. They generally work quite well with about two to five chains. that are sort of running sequentially. And so you can see here, we can have a look at the edited article and it's. Sort of given us, you know, a title and, some text about data engineering. 

You'll see that when you run this, all of these output keys are available. So we could originally have a look at, you know, like The article outline Or the generated article before it was, uh, went into the editing chain. And so the more output variables you have that are related to your output keys from above. 

Uh, can potentially give you the ability to introspect what happens at different points in the chain. as well as that, let's take it a step further and say,  maybe we want to generate an article and that's great. We've sort of, you know, edited it and made it really good. but what happens if we also wanted to translate that into multiple languages for our marketing team? 

 And so what we can do is we can create a custom chain. , and what you normally do in L chain is you inherit from base classes. So you can see where inheriting from the chain and also a base prompt template. We've also , modified the,  the actual custom chain class and included a range of languages here that you can see that we're interested. 

 In translating to so French, Spanish, and Yemen. Now, if you scroll down here, this is the part where we've really modified the code. So this underscore call function is where you can write your own custom chain logic. And what we are really concerned is we wanna make a translation dictionary. We wanna loop over the self dot languages, which you can see we've defined here. And then once we've got those, we want to then, , get the prompt value from the language. And also the edited articles from our inputs. And that's coming from the edited article. And if we remember last in our last sequential chain, if we scroll up and have a look at the output key, notice how we've got that edited article as the output key. 

All right. Perfect. So scrolling back down, we then generate the prompt value and then we update the translation dictionary with this generate prompt. We then return the output key. And then essentially we run a custom chain. Okay, So we make our custom chain. We put in the output keys, translations. We put our template. As you are receiving an article, must translate into language. and then the edited article. And then essentially what we can do is uh, we could essentially do this. Uh, and you'll see, actually, this language is kind of ignored. Uh, and the reason why is we're not really using this language, bit here the formatting's actually coming for the language. From self dot languages. Right. So you kind of have to decide, okay, I'm gonna add the languages here. 

and that allows me to get around the fact that when we are making multiple requests on our call,  maybe we want the language to be a bit more dynamic.  So you can see here, actually we generate French, banish and German. Right. , so we've got French, Spanish and German. And then essentially all we can do is add that to your sequential chain object. Right. So just right at the end, we put the chain. 

 We put in the language  and remember that's just a dummy key. And so that could be anything here. It could be empty. , and then essentially what we then end up with is a translations key and notice how we've got an LLM result. And we can then do what we've done here, where we've got this dot generations. bracket zero bracket, zero dot text. , and so if we scroll down here, you'll see, these are all the different output keys We've got the blog post idea. We've got the word count. We've got the  language, which is just a dummy key. We've got the article outline the generated article the edited article. And we've also got translations for both. French Spanish and German. So just wanna give you an extra bit of a stretch it's possible for you to make your own custom chains do multiple things in those chains and then return those back out. So this is a kind of good walkthrough from,  just doing a couple of sequential chains and then in the end of the lesson, getting a bit more hands on with, how can you create that custom chain to. really do whatever you like inside of the sequential chain.

What is the main purpose of the sequential chain step?

What does the code snippet `PromptTemplate.from_template(template)` do?

What is the output_key for the article editing chain?

What is an output_key within an LLM chain and why is it important?

Did you know it's possible to easily write, edit and even translate your blog content using LangChain?

AI Blog Post Creation

We're at GoolyBib headquarters, Monday morning. Our boss, Gustav Gieger, wants us to test out all the parameters of the smart bibs. We need to figure out how to use the X/Y/Z Prompt Grid to understand the effects of parameters like denoising when creating images.

We don’t have enough time to test out all the parameters manually, so we need to find a way to do it quickly and see what effect different combinations have on our images.
Has anyone heard of X/Y/Z Prompt Grids? 
Apparently it's a way to test and understand different variables in AUTOMATIC1111. 
We need to find a way to use it for our creative process.
Let’s get to work and figure out how to make it work for us.

The ability to understand the parameters of image manipulation and editing can be a difficult task to undertake. Thankfully, there is a tool that can help make this process easier: the X/Y/Z prompt grid. This tool provides a way of testing different parameters to understand how they work and what effects they have on an image. 

Let's start by taking a look at the de-noising parameter with image storage. De-noising is a process that reduces changes to an image when it is generated. With the X/Y/Z prompt grid, you can select the variables you want to test, and then generate a script to run the prompt multiple times. The script will run the prompt with different parameters, such as 0.1, 0.3, and 0.5. 

When the script is finished, you can see the effects of de-noising on the image. For example, you might see that with a higher de-noising value, the image looks more like a Pixar film. You can also see how the different parameters affect the image in a grid, so you can find the sweet spot for the image. 

In addition, the X/Y/Z prompt grid allows you to understand the combination of two different parameters. For example, you might want to understand the effects of the CFG scale and de-noising together. In this case, you can generate a grid that shows the effects of these two parameters together. This helps you to understand the sweet spot for both parameters, and to see how they interact with each other. 

Overall, the X/Y/Z prompt grid is a powerful tool that can help you to understand the effects of different parameters on an image. It can help you to find the sweet spot for an image, and to understand the combination of two different parameters. It is a great way to experiment with your prompt and to get the results you want.

X/Y/Z Prompt Grids

00:00:00] Okay, I'm going to show you something. I find super interesting and really helpful when you're trying to understand what different parameters do. Just going to upload an image. Let's just take, yeah. This image here. And I want to understand the de noising parameter with image storage. 

If I don't understand that very well, I can just go, denoising is my understanding of it is the less denoising you have. Then the less change to the actual image. When you're generating the new image. I want to understand that. And there's a pretty cool thing you can do which is in here, which is in XYZ plot. 

And it gives you X, Y, and Z variables. You don't have to put anything in them, but in particular you can go in here and choose the different. Different variables that you want to test. And here we find a denoising. And we're just going to put. I want to is your point. threes. [00:01:00] 

I's your seven. Zero. And I just want to understand how they work. Okay. And we're going to take care, include sub images. Cause I want to actually get the images at the end as well. And I'm just going to put this as Pixar 

or Pixar. A. 

All right. And you can hit, generate. And what that does is it runs a script to then basically run this prompt multiple times. So it's going to use the same prompt. And it's going to go through and it's going to run it. for 0.1, D noisy things are when 3d nosing 0.5 and so on? So it's just like a really good programmatic way to mess around with different parameters and see how they work. 

And you can see it starting to run here. I'm just gonna. Come back to this in a second. When it's finished running, to the pause now. All right. We're back and it's [00:02:00] finished and we can see here is the effects of denoising. This one here we can see. It ran with Dino's 0.1, and this is a little bit more denoising 0.3 and now the sun to get a bit more Pixar at 0.5. And now it's a lot more Pixar's or put seven. And then this is like getting pretty close to Pixar. And you can actually see this. It generates a grid as well. She'd see how it changes over time. So this you might know what D nosing does, but now that you can see it visually on your image, or you start to understand, like where's the sweet spot. And I think maybe you could even go stronger, but for me, like this one looks the most visually appealing. These ones still look too similar to the original and, be different for every image. So that's how you do. The X, Y, Z pot. Th the other thing. You could do here, which I find quite useful is understand the combination of two different things. So if we put a CFG scale, [00:03:00] And then we just want to do. Let's say. CFJ G scale of. Five. And 12. And then I'm gonna reduce the denoising Prentice, cause I know. That works pretty fine. So just going to do these three, certain, I should have a grid of three by three pitches. If I hit generate what is going to do is going to create a grid. Between those two parameters. So I'm going to pause again and then I'll come back. Okay. We've got it. His art grid. Here on the left hand side, we have the CFT scale. Which is the classifier free guidance, how much it follows the prompt. And and then you have to noising here. We can see that with a higher de noising, it gets definitely more Pixar. Which makes a big difference. But then the CFD scale also makes a big difference here as well. So it looks like the denoising has to be pretty high in order to really change the image towards Pixar. But [00:04:00] once. Is high. I think going from a CFC scale. For five to an eight, really brings out the Pixar and then to a 12. And it makes it even more Pixar. It does lose a lot of the original image. Like it doesn't have the same. Stance anymore as well. It's all about how you adapt this. Like you can see. If you wanted. Eh, cat that looked just like this, but a little bit more pixel then. Yeah. Maybe this would be a good path to go down. And you get all the resulting images. Underneath as well. And then if you want to replicate them, you can and you have all the details here of which ones specifically this one is, and you can take the seed as well. And just play around with it. That's how you do XYZ plots. Really powerful tool. And there's a lot of different parameters you can add in here. You can even do a Z. Axis different styles, like different like face restoration. You can check whether that works or not. You can do different mask weights prompt orders. Even different models as well, check out different models and you can also [00:05:00] use this to move along the seed. So like sexual and blatant space You could also see the value of do different steps. It really, the world is your oyster. And this is a good programmatic way to really test out your prompts. All right. Hopefully that was useful.

What is the purpose of the denoising parameter when generating a new image?

What is the purpose of using the CFG scale when testing denoising parameters?

What is the maximum number of variables that can be tested using the XYZ plot?

What is the optimal denoising value to achieve a more Pixar-like image?

Stable Diffusion is a powerful tool for generating images in any style, but it can be hard to understand what combination of parameters work best.

How can we test the effect of all these parameters?

You’re in your content meeting and the company is discussing adding stock photos to each blog post. Unfortunately there’s no budget for a license, so your boss is asking if you know where to get good free stock photos. You suggest AI model Midjourney as a potential solution.

I’m wondering if AI is good enough?
Have they really come along by enough to be a replacement for stock photos?
Why don’t you check out midjourney and see what we can get to.

Searching for the perfect stock photo can be a pain. Free stock photo sites like Unsplash are great, but they may not always have the exact photo you need, or the photos might be too generic and used by everyone. But what if you could create your own stock photos with AI?

Using AI to create your own stock photos is a great way to make sure you have the perfect photo for your blog or website. With AI, you can create photos that are unique and that nobody else has. It’s also much quicker than searching for stock photos and licensing them.

To get started, you’ll need to find a photo that’s similar to what you’re looking for on a stock photo site. For example, if you’re looking for a photo of people in a business meeting, you can search for that on Unsplash. Once you’ve found the photo you like, you’ll need to write a prompt for the AI. This should include the basics of the photo you’re looking for, like the number of people, what they’re doing, and the type of camera you want to use.

Once you’ve written your prompt, you can then use an AI-driven photo editing tool to create your own stock photo. These tools allow you to change the photo’s elements, such as color, saturation, and brightness, until you get the perfect photo. You can also add effects and filters to make the photo more unique.

Another way to create your own stock photos with AI is to “scrape” a photo from a stock photo site. This means you’ll copy the photo’s URL and use it as the base image for your AI-generated photo. This can be a bit of a grey area as it’s not clear if this is allowed or not. It’s best to use your good judgment when scraping photos and make sure you modify the photo enough to make it your own.

Once you’ve created your own stock photo with AI, you can use it on your blog or website without worrying about attribution or licensing. You also don’t need to worry about using a stock photo that’s too generic or seen everywhere.

Creating your own stock photos with AI is a great way to make sure you have unique photos for your blog or website. It’s also much faster than searching for stock photos and licensing them. Just remember to use your good judgment when scraping photos and be sure to make the photos your own so you don’t infringe on someone else’s copyright. With AI, you can make stock photos that are unique and perfect for your needs.

Free Stock Photos

0:01 Okay, this is how to make your own AI stock photos. So stock photos super important because you know, we don't have time to go out and take photos of everything.
0:12 You know, some of the things that we want you know, we won't be able to go and, and capture, right?
0:16 So that's why these stock photo sites exist. And I think for most use cases, they're really good. Especially have a license, right?
0:24 So this is Unsplash. The really cool thing about Unsplash is that it's free, right? And they have a really permissive license.
0:33 But they have been bought by iStock photos and you know, now they have Unsplash Plus and they have these banner ads.
0:40 And you don't always get like, even what you want, right? Like have such for people in a business meeting. She's not in a meeting.
0:47 This isn't the business meeting I was suggesting, right? Like, this one looks good, but it's un splash. Plus she's not with other people.
0:54 Like, maybe this kind of like, that looks like really creepy and weird. This doesn't look realistic. So, I mean, it's a bit of a pain to find the right type of image.
1:06 And you know, you, you even when you do find it you find, you end up with like the same picture that everyone gets, right?
1:14 So if I search for chart I, I'm, I've seen so many blog posts with this image, right? <laugh> and you know, this image, right?
1:24 This is like really popular. And, and you just kind of recognize these from some of your favorite blogs, I bet.
1:30 Cool. So now what if you wanna make your own stock photos with ai? Well the way you would do that let's find one of the images here.
1:41 So like, this one's kind of the closest to what we wanted. And I'm gonna click in here. You can see the artist.
1:48 So it has like a lot of views, right? It's free to use under an unpled license. So this is where I would say I'm not a lawyer, right?
1:57 Like you need to kind of figure out what your own risk profile is. But from my understanding all photos can be downloaded and used for free for commercial and non-commercial purposes, no permission needed, although they do prefer attribution.
2:14 They cannot be sold without significant modification. So I would say, you know, kind of using like retraining and making a new photo that looks similar.
2:24 But using AI is a significant modification. Although I wouldn't do this, right? Like, I wouldn't create like your own competitor to Unsplash.
2:33 That probably wouldn't be cool. And I think in general, you could probably give attribution to the artist, right? Like they deserve it.
2:41 So I, I'm gonna show you how to kind of replicate the same look and, and feel but with, but you know, kind of make it look your own and and then you own the license, right?
2:52 Like, you don't have to worry about you know like attribution necessarily. If you, if, if you do, do make it considerably different and modify it with your prompt but also, you know, it looks unique.
3:06 Your blog doesn't have that same problem of like everyone using the same stock photo. Cool. So few things that I want to point out.
3:14 So one is you want to really see what the meta description of the images. So I just wrote, clicked, and go, went to inspect.
3:21 And then, so this is the image. And then just gonna click in here. Just gonna drag this across. Oh, back up.
3:35 Okay, let's drag this across. So that there's the image. Let's write, click inspect again. Okay, this is the image. And here we go.
3:45 So here's the alt tag. So we click in here, it's four people watching something on a MacBook, right? So I'm gonna put in here, imagine, and then the prompt four people watching a white MacBook on top of a glass top table.
4:03 So that's gonna generate it. You can actually just say like, what you want, right? I want a stock photo of people in a business meeting.
4:10 And it does like a pretty decent job of that anyway, although that looks a bit gloomy. So that all that is generating.
4:17 The other thing I'm gonna pull out here is the camera. So I think like, that's kind of useful. That's quite interesting.
4:25 There's other metadata sometimes. So you know, you can just press up and then you say oh in the prompt the camera.
4:37 And then you could say like yeah, you actually wanted it to be a business meeting of four people. And then you want it to be like a stock photo.
4:49 Cool. So you know, we're just kind of using this to, to get into like the baseline and kind of figure this out.
4:56 You know, not, not all these are great, right? But we're getting a little bit closer. This one looks like more interesting, but the faces are a little bit weird and this one's facing away, so that doesn't make any sense.
5:09 Cool. Here we go. This is getting better because we caught it a stock photo. It's looking a lot more interesting.
5:16 So that seems to be helpful. And we also put in the camera here, so I think it'll kind of change the style.
5:21 Like we wouldn't have some of these like, black and white things or like this. Here we go. So look at that.
5:27 That is much better. Okay, so like, this is almost there. Some of the faces are a bit messed up. This is pretty good.
5:34 You know, and they they look very happy. So maybe we'll try one more thing here, which is oh, cancel. So let's, instead, let's say I'm gonna do this prompt but then we're just gonna do a negative prompt, we're you say No unhappy.
6:00 So that's gonna make them happy. Yeah, you could also just say happy, but, but like you know I think, I think there's you know, different ways you can mess around with this.
6:11 So that's, that's well, that's generating there's also another way you can kind of do this. And I think this is like a little bit dicier but you know, I'll show you anyway because you know, it's really up for you to decide.
6:28 So I'm, I'm just gonna paste in. Lemme see. So if you copy image address, see if this works, and pacing the prompt, and then you can say, got a photo of business meeting of four people, blah, blah, blah.
6:52 Okay? And then what this is gonna do is it's going to use that image as the base image. So again, you are transforming it.
7:01 I think this is a little bit like if I was the artist, I'd be a little bit p****d off if I found someone's doing this unless like the image looked really different.
7:09 So I think that's the key is like, start with the alt text, start with the image maybe, but you know, make it look your own, right?
7:16 Like it needs to be something that's your own. You can kind of mimic the look and feel and the style.
7:20 But you can see here, I mean, this didn't work very well. That's kind of weird, <laugh>. So I guess getting rid of unhappy doesn't work as well.
7:30 Okay, here we go. So this is a lot better. This is like much more we wanted so you know, we're getting a little bit closer.
7:39 But but, you know, it doesn't take a long time to do this. Sometimes it's quicker than, you know, than all the searching through stock photos or, you know, it's definitely quicker than searching for, you know, like licensing and, and like, you, you just pay once for a mid journey
7:54 a month rather than paying a huge amount of money for licensing bunch of photos again, you know, this is a gray area.
8:02 Like this isn't, you know, there's no rules around this that anyone's figured out yet. So I would say use your good judgment.
8:10 Don't just blindly copy someone else's photo like this. You know, this isn't actually what, what I wanted anyway. I wanted something more like this, right?
8:19 Cool. Okay. Hopefully you know, use your good judgment and and if you do, then you have really unique, interesting looking stock photos that nobody else has.
8:29 And once you develop a style then you can kind of keep using that prompt again and again. So really, really helpful.

What is the goal of retraining an AI to generate stock photos?

Describe the process of using AI to generate stock photos.

What do we recommend starting with when generating an AI stock photo?

What should you be aware of with regards to copyright when using AI to generate stock photos?

What are the benefits of using AI to generate stock photos?

Getting unique, high quality stock photos without paying tonnes of license fees has always been a pain, but that problem just got solved with AI.

Where can we get free stock photos?

GOOLYBIB HEADQUARTERS - WEDNESDAY MORNING

We’re here at GoolyBib headquarters for Gustav Gieger’s Wednesday morning meeting. We’re here to discuss how to use chatbot technology to automate the process of prompt engineering for the company's smart bibs. It's a complex problem, but the team has decided that this is the best way to go. We all know that if we can get this right, it will save a lot of time and energy in the long run.

We need to find a way to generate prompts that accurately describe what we need and that our chatbot can understand.
I know we've been discussing automatic prompt engineering, but does anyone have any other ideas on how we could get this done quickly and efficiently?
We don't have a lot of time to waste. Let's get creative!

Automatic Prompt Engineer (APE) is a technique for using GPT-4 to automate their prompts. APE uses GPT-43 to infer instructions from example responses and specific instructions, and then ranks the potential responses with probabilities. This technology has enabled the automation of a variety of tasks in the prompt engineering domain.

The concept of APE was first introduced in a paper by the name of Automatic Prompt Engineer. Here, they used ChatGPT to infer what the instructions were. They then returned a lot of probabilities to the potential responses, and then ranked them. It was found that APE was at a human level in terms of prompt engineering. 

Since then, there have been a few other papers that have come out, including one that found a better step-by-step prompt. This was used widely on the internet, and it was also using an automatic prompt engineer. The basics of this technology were found in an open source tool called Prompts Royale, which references another paper. This paper utilizes a different approach to APE. 

Within APE, the prompt template has to be given. This includes best practices and variables. For example, when creating a list of product names, the product description variable is included in the template, along with some example values. It is also important to include criteria evaluation, such as helpful, insightful, and inappropriate. Additionally, prompt engineering best practices should be used. 

Once the template is submitted, it is automatically generated based on the variables given. This works best with GPT4, and the temperature should be set to one. It is also important to label the placeholders in the template, as this ensures that they will be replaced with the correct values. 

APE is quite smart, as it can take context from the user and relevant variables. It can also generate unique and innovative names, such as Google and Nike. Furthermore, it follows the principles of prompt engineering, and can also set the variables in the template. 

Overall, APE is a revolutionary technology that is revolutionizing the way people code and automate their prompts. It enables the automation of a variety of industries, and is now being applied to prompt engineering as well.

Automatic Prompt Engineer

  All right, I'm going to show you how to use chatting. GBTA to automate your prompting. And that's crazy because we're using prompting to automate every other industry. Why not? Why if chatting tea is good at all these other tasks and why wouldn't it be good? At prompt engineering as well. 

So this is a screenshot from a paper here. And they called it automatic prompt engineer. They basically used I think it was churchy petite or duty.   Back then. The, to basically infer what the instructions were. If they give example responses and then if they give a specific. A specific instructions. At what would the instructions be? 

So it comes back with a lot of probabilities to these, the different potential responses. And then it basically then ranks them and found that it was, a human level in terms of prompt engineering. Then there've been a few other papers that have come out since then. Including one that found. A better in a better prompt. Let's think step by step. It does pretty well their attempts to track your C and then there's a few other ones as well. 

If you remember, there was going around the internet. This whole like take a deep breath and then think step-by-step. Step prompt that people were sharing. That was also using like an automatic prompt engineer options. So anyway, this is this is like pretty interesting. I think in some cases it might work better than humans that prompted engineering. What I found is at the very least it's very good at getting you up to above average prompt, straight away without you having to think very much.  lEt me show you, I have a couple of these, but.  I'll start with this one. I put this together. Based on a lot of testing, but what I found was useful in the system prompt is two. Actually tell it you're an expert in prompt engineering previously, that wasn't very helpful. But now it knows or potential hearing is is with the latest updates. 

It's updated. I think to April. Last year. So now it knows what prompt engineering is and previously this didn't work, but now it does. And we just basically tell you, be given a prompt template. One of our test cases, you drop us off to most upon template using prop 10 June. The best practices. 

What I've done here is I've gone a little bit further and given it the best practices. I've told at what prompt engineering is, just get it in the right part of latent space, without making sure it's not, I think you have some other prompt engineering. And then these are the actual best practices and these are ones that I came up with. 

As part of my book, I'm ready for early media. So these are the kind of the first three principles. I found that they're quite useful. Then the other thing you have to put in here is the standard stuff like respond only with your optimized prompt and nothing else do cheat. By including test cases in the prompt you have that sort of thing. 

The the place I got this or at least the basics of this without the principles. Is this a.  Tool prompts, Ray out. And I think that they reference another paper that kind of uses this, but they I looked into that code, which is open source. On get hub and you, and just saw what. What was working for them? For a system analysis. 

Feel free to take a look at that as well. They've got a bit of a different approach to this. Anyway what I found is if I put in the prompt template, And what's different with mine as well as I put like these sexual variables in hand. Which is, I think really helpful for me, at least. 

And the way this would work is, if I want a list of product names for, this product then and then these are like some examples. What I found with the with the test cases. Is that you couldn't put them in Jason very easily. You had to put it in more in a Yammel format or markdown format. 

Here we have the product description variable and the template. And we were just trying to give kind of example, values. So here I've put like a shoe that can fit a foot size and then the response would be, these. And I've just given this as one shot. So I've just given it one, one example. 

But then the other thing that I found worked really well was adding these criteria evaluation. You can put whatever evaluation. Criteria you want to have it actually humbles multiple if you want. But I've just used the default, which is the submission helpful, insightful, inappropriate. That I got from land chain. That's the default evaluation criteria at nine chain. When you using their evils. 

I also reminded that prompt engineering practices, best practices should be used. And then I found it's very helpful to include in here any variables. And then you could build this programmatically. Like I have this for one of my libraries I use for testing locally. Where it just automatically builds this prompt up for me, announce them. Based on. You know anything that's in curly brackets here.  Cool. 

So respond only with your PR. The most prompt template, I'm just going to hit submit. I found. Only really worst with GT four and only works quite well with the temperature of one. Temperature higher than that. It starts to get a bit crazy. Temperature lowered since get a bit boring. That's usually what works to me. Here we go. 

It's pretty smart because it's, See here. Imagine you're head of a creative marketing team task of coming to catch independent names. The product is blank. We're looking for unique and innovative names, akin to how. Brand names like Google and Nikhil test system to the respective industry. 

So it's used some examples here. And it, hasn't given much structure in the prompt, but I think it's done a pretty good job. And we could take this 

 I'm just going to copy this and then we're going to set up another one.  

And then we're just going to put this as a.  

Pair of shoes.  

For the size, that was the product we wanted. And then she's going to submit that.  And it comes up with pretty good names.  So there you go.  It's a credited prompt now. I didn't stop that. We have another template, cause I found that it didn't, don't always have like test cases and sometimes you just want. Be able to just do it. Just from a task specifically. 

This one also follows something from prom per hour. But you can see that it's using very similar. A system message. And then you still got the different best practices there. But the difference here is that we say the content must take context from the user from relevant input variables, surrounded by Kelly brackets, I E this and that, that was like the hardest part to get. 

And once it does that, it does a pretty good job. And and the other thing I had some kind of mentioned is these place holders should be labeled in the template as they will be replaced to values. Because one thing I found that it would do. Is I'm here to say, please take the product category that's done correctly. 

That's labeled it. But sometimes it'll just say please take the, and then product category, and then it wouldn't label it. So when we insert product category in here it wouldn't work right. Because it would be referencing product category down here, but that texts wouldn't be in here anymore. 

If that makes sense. Just to show you. An example there.  Since she's going to get rid of this.  

And then shit. So if we put him product category shoes.  

And then we're going to say target audience.  



 

And then a distinct feature is that. It  should say.  

Extra comfortable.  And then then it knows because it can see that like the target audience is up here, so then they can infer from these examples, if that makes sense. That was squeaky. And then here, we got some more stuff. We have to repeat the product category here in this template.   And put in shoes.  

And then talk to the audience.  

And then.  

This would be.  

Cool. So see how that worked. Okay. Cozy comfort her mate. I got it's called a homemaker. That's pretty terrible. But you get the idea. We don't this was a task to generate product names. We could change whatever task went here and it's this could be  You're right. Some texts. In the.  

The author.  And then see what template comes up with.  

You can see it's using all the different principles or the really cool thing here is that it actually sets these variables in. So it makes the, it makes it a real template. And it does a pretty good job of doing that.  

Cool. Yeah, hopefully that all makes sense. There's a lot of firms engineering principles baked into this as well. And you can see that. It does a good job because we've given it a lot of handholding here. And I think this I used to try and do this. Back before the update and because it didn't really know what prompt engineering was. 

It wasn't particularly good at it. But now it. does because it's in the trainee day to say this opens up, I think a lot of prompts engineers. Didn't really, haven't really tried this again recently. And this opens up quite a lot of of stuff you can do. You can also, you can go even further, like you can get it to generate cases like these things that go in here, you can get to generate good examples and you could also get it to rate the responses as well. 

So that's something I've been working on for my thumb library which, should be out soon.  All right. Hopefully that was useful. 

What are some of the evaluation criteria used in Automatic Prompt Engineering?

What is the first principle the tutor mentions for prompt engineering?

What are the criteria for evaluation mentioned by the tutor?

What is the difference between the approach used by PromptsRayout and the one used by the tutorial?

What did the tutor mention as the hardest part to get?

It's funny because even when AI saves me 10 hours, I still resent the 10 minutes it took to figure out the prompt.

My employees aren't good at prompting

CONFERENCE ROOM - PROMPT TEMPLATE WORKSHOP

Alright, team, gather 'round! We've got a new task on our hands that requires some serious standardization. We need to create prompt templates to ensure a clean prompt test. Now, I know we're up for the challenge and eager to please, so let's dive right in. 

First things first, we'll be using a developer account on platform.openai.com instead of ChatGPT. This will give us more control over system messages and access to different models. Trust me, it's the way to go.

Alright, now let's set up a template for a specific task. Let's say we want to tell a joke. We'll walk through the process step by step, making sure we test and modify the template along the way to get the desired results.

So, grab your thinking caps and let's get those prompt templates up and running. We've got some AI magic to work on!

Alright, team, gather 'round! We've got a new task on our hands that requires some serious standardization. We need to create prompt templates to ensure a clean prompt test. Now, I know we're up for the challenge and eager to please, so let's dive right in. \nFirst things first, we'll be using a developer account on platform.openai.com instead of ChatGPT. This will give us more control over system messages and access to different models. Trust me, it's the way to go.\nAlright, now let's set up a template for a specific task. Let's say we want to tell a joke. We'll walk through the process step by step, making sure we test and modify the template along the way to get the desired results.\nSo, grab your thinking caps and let's get those prompt templates up and running. We've got some AI magic to work on! Let's get started! \nGustav Gieger at CEO

Topic: The Importance of Standardizing Prompt Templates in AI Work

Prompt Templates play a crucial role in standardizing and streamlining work when dealing with AI. In this transcript, the speaker discusses the process of creating prompt templates and highlights the benefits of using a developer account on the OpenAI platform instead of relying solely on ChatGPT. 

The speaker starts by using a common task, which is to ask for a funny joke, to demonstrate the limitations of ChatGPT in this particular area. They explain that while ChatGPT may not excel at telling jokes, Chat GPT with GPT-3.5 Turbo (ChatGPT+GPT-3.5 Turbo) is capable of doing so. 

To standardize and templatize the prompt, the speaker suggests creating a developer account on OpenAI's platform. They emphasize that no coding knowledge is required, and the cost is usually cheaper than using ChatGPT. The advantage of using the developer account is that it allows for more control over the system message, which can be customized according to specific template requirements. 

The speaker mentions that switching models is also easier with the developer account. It provides access to context models, fine-tuned models, and even the latest models like GPT-4. The parameters can be adjusted, such as temperature, maximum length, and stop sequences, to experiment and optimize the AI's responses. 

To illustrate the process, the speaker demonstrates how to create a prompt template using the "talk like a pirate" example. They show how the system message can be modified to suit the template's purpose. Once the template is saved, it can be shared with others by providing a link or sharing the code snippet with developers. 

The speaker also emphasizes the usefulness of prompt templates for testing purposes. They highlight the ability to easily delete and rerun templates, as well as test multiple templates simultaneously without hitting API rate limits. This flexibility allows for efficient testing and iterative improvements. 

In the transcript, the speaker provides an example of a prompt template for generating funny jokes. They explain the structure of the template, with the system message containing general instructions for the task and the user message specifying the specific joke topic and style. The speaker advises against providing multiple examples in the prompt and suggests relying on the message stream for past examples instead. 

In conclusion, the speaker stresses the power and effectiveness of prompt templates in standardizing AI work. By using a developer account on the OpenAI platform, users can create

Prompt Templates

   Here's how you create prompt templates to standardize your work when you're working with AI. So this is chatty with T and so you have a task. I'm using a standard task that I like to do a lot of testing for which is to get charged for you to tell me a funny joke, because it's particularly bad at this task. Usually when you ask it for a funny joke, it just gives you a dad joke or it gives you lots of disclaimers. 

You can see here that it squirmish when you ask it to emulate a comedian and a, this doesn't sound like Dave Chappelle here. And it doesn't. It's not as funny as you would normally get. But Chad GBT is actually capable of this. 

Now I'm using this as a, just as a placeholder task. Like you might be writing a blog post with AI. You might be. Writing emails, you might be creating reports, whatever it is you're doing for your work. Once you find a template that works. You want to start to standardize and templatize it. I don't recommend doing that in chat GBT because chat GBT is personal to you. There is a teams version of Changi petite, but the problem is that. They have a lot of system instructions in chatty beauty that. Make the results nonstandard. 

If I put something in my chat GBT, it might not give a similar results to yours. So what I recommend you do is you go and create a developer account. You don't have to be a developer. You don't need to know how to code. But if you go to platform to open out.com, you can sign up. And then you will have. I actually have to pay for this. Per API call, but put the corporate credit card on there. 

And it's pretty cheap unless you're doing lots of volume. So it usually comes out of way cheaper than paying for Changi petite. Now the benefit of using the playground is that you get to control the system message. So you can put whatever you want in here in this part of your template. 

And then whenever you create another template, say. You want to. Just.  Create another chat history. Then you can put whatever system message you want, or you can put no system message in there. And that's much better than what you get with Changi petite because whatever you put. In your customer instructions, that's also added to about a thousand words that they put into the template as well. 

This is a lot more stable. The other thing that's useful as you can switch models much real easy, and here. So it's not just G T for juicy 3.5. But you also have the, one of the context models. You have access to your fine tune models. If you're doing anything like that you can use the newest stuff like GBT four. Preview. I'm going to use Before here, you can change some of the parameters as well. 

Temperature. Maximum length is you can set stop sequences. Whatever you want to play around with, you can do that. So this is really powerful and, You prompt it just the same way that you would prompt LGBTs. Just going to copy across a joke here.  

And then submit. And then it just responds in line just the same as it does with chatting petite now. One of the things you can do is when you change the system, as soon as you can say Pirate.  



And when you hit submit, then, straight away is changing the behavior of the AI and that's much more painful to do in customer instructions. 

But the really powerful thing you can do is if you like this template, you can click save. And then you can say, talk like a pirate.  Maybe we put some description here and then you can make it available for anyone to use. So I can save that now. And then I can share it so I can click. On here and copy the text and then anyone else. Who has that link and go use it. 

The other thing is you can pass this to a developer as well, so you can click view code. And right now. Yeah, you can see straight away that it has the actual code that the developer needs to run. Including the system message and so on and so forth. It helps you make that leap from a template to an actual product. Now I showed you a template earlier and he said, I have a ton of these templates saved. 

So I'm just going to go through what a standard template is and what part goes into. The messages versus the system. So this is my funny joke. Structure. And just to show you how well it works and just going to hit submit here. What I'm asking is for Dave Chappelle to tell me a joke about robots telling jokes. So matter.  And the joke has come up with here is that you ever hear a robot tell a joke, man, it's funny in a way that makes you question reality. 

Like why did the robot go on a diet? Because he had too many bites. Ah, man, I wish I could laugh, but I'm too busy thinking about the impending singularity. All right. We can run this again and again, and that's part of the nice thing is you can just delete it and then run again. And so it makes it much better for testing. You could also open up multiple ones of this. Because you are paying for each API call you can just submit multiple ones, right? 

It's not going to block you. You're not going to hit the API rate limit, whereas with a chat GBT, plus sometimes you will. So I use this for testing wherever possible.  

You can see how often it's good and then you can change the template and then test new changes, which is really helpful. So that's one of the main reasons sees this now, what have I done to set this up? So a one is I have the system message and usually I don't start with the system message. 

I try and do it in the user message first. But once I have something working, I'll move it across. Now the system message, what it's supposed to be for is really the template itself. Ignore the specific thing you're trying to ask it to do but the general task itself, that's where these instructions go. 

So the task that we're trying to get it to do is to tell a joke, but if you were writing a blood post, you'd have instructions here for how to write a blog post. Now I'm here. I have a few different things like writing an original joke on the topic. And I've tested a lot of this wording and found that this is what works. The really useful thing they have here is that I'm asking it to deliver a structured data. 

I have. The input that like the topic of the joke. And then I have the output which I've asked it to give a disclaimer, because when in my testing, I found that if you allow it to give you a disclaimer in the data that it responds. Then it's much more comfortable telling you a funny joke and it doesn't have. That disclaimer. In the actual response like this. And so that means that when you're using this template, if you're a developer, you can then just copy, you don't have to, you can drop the rest of the the disclaimer and everything. 

You can literally just grab this and, and as a developer, you know how to pass Jason, so you could. You could grab that text out.  Cool. So that's the system message. And that, that is the kind of general instructions. For the task itself. But then the specific instructions for this task is it goes in the user message here. 

If I wanted it to tell a joke about breakfast and Chris rock. And then I submit that here. If I wanted to tell a joke about bicycles and Robin Williams and estimate that there now, why do I have all these in here? You noticed? The one thing I actually asked it to do was to tell a joke in the style of Dave Chappelle about robots. 

But I had all these other ones in there already. And that's by design actually you can trick it into believing it's already responded successfully to these things. So one of the mistakes I see people do is that they put lots of examples in the prompt in here. Like they'll say like input, output, input, output or they'll put it in the system message. But I would say you just want to talk about the format itself. At, rather than a specific, given an example of a specific joke. The actual use, the actual examples that you wanted to give of pastimes has done a good job is in the message stream itself. 

And I think in general, it listens to this much better. It feels more natural. It's essentially saying it doesn't know that it didn't say these things, basically selling it. Hey, this is what you said before. When I asked you for Chris rocker breakfast, this is what you said before. 

When I asked you for Robin Williams, the bicycles, and it goes, oh, okay. I guess that's how I respond now. And it will give you a pretty good. Example back, I find it follows the examples much better when they're in this messages stream, rather than formatted in here. It's also a little bit cleaner because you have the general instructions and then you have the specific thing you wanted to ask, and then you have the examples or listed in here. Cool. 

This is a really powerful and if you wanted to see one of these templates, You can hit save, and you can also save as new or update an old template. I'll share I'll share this one with you guys. So I just update that and then I'll share that in this video, 

Why is it important to create prompt templates when working with AI?

What are the advantages of using the playground for AI testing?

What is the purpose of the system message in a template?

Where should the specific instructions for a task be placed in a template?

What is the advantage of including examples in the message stream of a template?

Real prompt engineers use OpenAI Playground not ChatGPT.

Can we ensure a clean prompt test?

ULTRA OFFICE – PROMPT ENGINEERING MEETING

I'm familiar with language models, however how can I create completely custom tools and agents? 
I want to understand what is happening in the reAct (chain of thought) pattern?
It would be great if we have the flexibility to create custom agents from scratch, just in case we need to do something that langchain doesn't let us do easily.

Chain of Thought (CoT) is a novel prompting method designed to encourage large language models (LLMs) to explain their reasoning process by providing exemplars that demonstrate reasoning. This approach often leads to more accurate results in tasks like arithmetic, commonsense, and symbolic reasoning. However, CoT's performance gains are only observed when used with models of around 100 billion parameters, as smaller models tend to produce illogical chains of thought that worsen accuracy.

ReAct, introduced by Yao et al., is a framework that allows LLMs to generate both reasoning traces and task-specific actions in an interleaved manner. By generating reasoning traces, the model can induce, track, and update action plans, handle exceptions, and interact with external sources like knowledge bases or environments. This enables LLMs to retrieve additional information, leading to more reliable and factual responses. 

ReAct has proven effective across various tasks such as question answering, fact verification, and interactive decision-making benchmarks, outperforming imitation and reinforcement learning methods while maintaining human interpretability and trustworthiness.

reAct and Chain of Thought

 Hey, welcome back. And in this video, we're gonna have a look at how we can potentially use and create custom tools. Uh, using Lang chain, Uh, we're gonna code up a version of react ourselves. So we really understand what Lang chain and a lot of these kinds of auto G B T and baby AGI are doing underneath the hood. 

It's really important that we go beyond just creating prompts with language models and we've been tasked with. Creating some custom tools. So let's have a look and dive into, what we are looking at. So we are first importing the regular expression package and this open AI. Thing from Lang chain. And basically that's a way for us to call the GPT three model inside of Lang chain. 

Our objective is the following. We need to hand code a simple framework for understanding what the react pattern is with language models and how we can use this to build custom tools. React focuses on a couple of different things, such as chain of thought reasoning. A good example of chain of thought, for example, is where you are doing. Uh, what is Leonardo DaVinci's wife's age squared. 

We first have to break that down into a series of sub problems or sub thought. First we have to search for what Leonardo's uh, Leonardo's wife is. Then we have to find her age and then we have to square that ages number. And so we can kind of think of this as a chain of fo a chain of reasoning. And so the react pattern follows a similar kind of process where we observe the environment. We interpret the environment we've a thought, We decide on a relevant action to take, and that could be doing a Google search. It could be using a calculator tool. 

We act on the environment and then we repeat steps one to four until we've either found a solution. i e I found the final answer, or we hit some sort of max iterations inside the loop. To start with, let's explore how we can extract kind of the payloads that we might ex. Uh, get back from a language model when we're working with text. So as the language model kind of gives us these relevant actions that it might want to take with an input into the relevant tools, such as search on Google search on Wikipedia or a search on Google. 

Uh, we wanna catch at the last action input. Uh, And the reason why is uh, when chain of thought is happening, we're going through a se sequential. Uh, series of thoughts, but we're only interested in the last action and action input that we then need to put into the language model. And we've got some compiled red rejects expressions to, to deal with this. 

Uh, and this is for both the action and the action input pattern. And so we can use dot find all to find all of the relevant expressions. And then we can say, get me the last one. If there is actions else. none, and do the exact same for action inputs. We can see that when we run this better of code. 

We get the last action being search on Google and the last action input being some other query. Let's have a, just a quick look at what these redx is, is, uh, and and sort of breaking it down step by step. So the co I basically is an inline flag that makes the red rejects case intensive. So it doesn't matter if the language model uses action action or any kind of capitalization, we'll be dealing with that. 

Then we've got the action part of this, which is looking for the keyword action. Um, and as well as that, we then have the backslash slash Uh, back slash S star. Which matches against zero or more white space characters. Uh, we then have the colon. Uh, and then as well as that, we then match against zero more white space characters. again, we then have a capture group set up after that. Um, and this is basically denoted by the parenthesis, So it matches against all of those, which are not a new line character. Um, so we're matching everything apart from new lines. 

Um, and so that's kind of what the carrot is doing here. Uh, it negates the character class. So we're capturing everything after that co. Um, apart from anything on a new line, we then run it on multiline. So basically this pattern should be uh, app, pen, you know, applied line by line. Um, and yeah, you can kind of think of, this is a bit more explanation on how that carrot is kind of working. 

Uh, if we have a look at the action input pattern, you'll see, it's very similar. So avoiding the capitalization with the, uh, you know, the question mark and the eye, and then we've got the action input and then we're, you know, any white space before the colon. And then we're having the colon, then any white space after, and then we have this capture group that's capturing everything, but the new line. 

Uh, and so this allows us to match both the action or the tool name that we want to do, and also the inputs into that tool. And so we can kind of wrap this up in a nice python function called extract Glass action and input. And you'll see, given the previous text, we're able to find what the action is, which is search on Google. And uh, this is the tool. And then what we want to put into the tool is the action input. So some other query, um, so given that we kind of need to find out whether the language model has the final answer. 

We will use the following template. Uh, I will, I found the answer, right? So we want the language model to say, Hey, I found the answer. And so we will also build some rejects for this, uh, which is very similar. Uh, so we found the answer and you can see we've got this final answer is now showing up, given some text. So we've worked out how we can. 

Uh, you know, basically extract the tool name and the tool. Uh, inputs and we've looked at how we can figure out when the language model is truly finished with the chain of thought reasoning. Uh, so what we went want to do is also make a function to extract the final answer and see if it exists. And we can see once we've done that, we've got the final answer. And so now we've got all the, all the components. 

Uh, needed to kind of create this react pattern. Um, so we we define a fake tool. uh, And I would encourage you to, to play around with fake tools until you really understand the pattern. Uh, we then add the function here. And we, and give a, a description of what the tool is. The base prompt is very, very interesting. It's it's important that we never hallucinate the tool result. And we wait with a stop sequence. 

Uh, so that the tool. Uh, you know, the language model doesn't try and hallucinate what the tool might produce Uh, Let's have a read of the base prompt to understand this. You will attempt to solve the problem of finding the answer to a question. Use chain of thought, reasoning to solve through the problem using the following pattern. 

Observe the original question, create an observation with the following pattern. Observation, observation text, create a thought based on the observation with the following pattern, use tools to act on the fort with the following pattern action and action input. We also specify and tell it, do not guess or assume the tool results instead, provide a structured output. That includes the action and action input. You Have access to the following tools. We then say original problem is Jason  Darrio with a partner. We then make a call to uh, a language model. So we get the the keyword arcs and we put our stop sequence of tool result. Uh, and then as well as that, we then can print everything out. So we we call the language model. Once we then extract the action of the tool and the action input. Uh, so what is going to be inputted into the tool? Uh, and then we get the function with the input. So we get, we call the function with the input, get the tool results. And then we've got some print statements here to show what, what tool the agent decided to take given the thoughts, um, and, uh, and given the tools that it's been given. 

Uh, and so then we've got the final one here based on the provided tool results. you either provide the next observation action input, or the final answer if available. If you're providing the final answer, you must return the following pattern. I found the answer fine. So let's have a look at the print statements. 

To start with given the question. of this one here, which is uh, Is Jason Dillo with a partner. The observation is Jason Dillo is a popular singer. It might be possible to search on Google, to find out if he has a partner. Search on Google, Jason to relay partner. The agent, therefore has opted to use uh, The following tool search on Google and the, the input into this tool is Jason Dillo partner. The tool result is Jason Dillo Doesn't have a WI or partner. And remember our tool result came. From whatever the function we called inside of the language model. Right. Uh, the second prompt shows. Uh, you know, 

Based on the provided tool results. Jason Dillo doesn't have a partner. Eva provides the next observation. Right? So that's the part of the prompt and the model outputs. I found the original answer. Right. And so we found the answer here. Uh, and, and it's worth noting that we've done this without a for loop, just to get a good grasp on how the react pattern is working. However, we could go through a series of thoughts, Uh, and you can imagine adding a calculator as an extra tool. Um, into your arsenal. Uh, and that would allow you to then take the age of this partner, for example, and, and square that, or multiply that by a certain number. Uh, we are able to give language models a range of different tools, such as web browsing. And you'll find this is the case in some of the latest plugins with chat G P T where you can do web browsing. 

It's important that you're able to understand this pattern, uh, and this is a really great way for you to get. To grips with what's happening inside of auto G P T and baby AGI, and really understand kind of how you can create a single agent, Uh, from scratch. 

Now that you understand how to create single agents using the react pattern. There's. an in enhanced version of this on site of Lang Chain's documentation. You'll find that, for example, you can go to the agent section and they've got everything on this getting started. So we have the ability to load tools, create an agent type initialize, an agent. 

Uh, and you'll see that the zero shot react description. is very similar to what we were implementing with our custom base prompt. Uh, and it's just good to know that you're able to do this rather than solely relying on land chain Land chain's gonna be great for when you need to do a single agent. Uh, in a very standardized way, but having an understanding of how the react framework works underneath the hood is gonna allow you to choose when to use Lang chain or when to build your own custom single agents.

How did we extract the payloads from a Language Model?

What is the purpose of the extract_last_action_and_input function?

Are you looking for an efficient way to understand the reAct pattern? 

Learn how to quickly create single llm agents using regex, Python and OpenAI.

How can I understand the reAct pattern?

Client meeting in the East Village, at the Dogshed hot dog stand

You’re talking to your client, the owner of Dogshed, a local hot dog stand chain located in the east village. She wants you to build a personalized ad for every keyword!

Ok I know you probably can’t justify building an ad that’s relevant to every keyword
That’s 205 ads, and I know that’ll be lots of work
But my little cousin showed me this AI thing - GPT
It’s amazing
Why can’t you use that to do it? 
Then you can just edit whatever it gets wrong
If we get it working we can expand out to more keywords easy


Including the keyword in the ad can seriously improve conversion rate and lower ad costs on Google. However it can be time consuming to create a personalized ad for every single keyword. You might have hundreds or maybe even thousands of keywords, and no copywriter wants to endlessly write ads in a spreadsheet all day. Often you just use a template and insert the keyword for each row, but this makes the ads sound robotic and uninteresting. So as a tradeoff most marketers will agree to write custom ads for a handful of the top keywords, then use a template (or no personalization) for the less important ones.
This was fine given that writing ads is a highly manual process, but now has become relatively trivial to do this at scale with the advent of AI models like GPT-3 from OpenAI. These large language models (LLMs) excel at lots of tasks, but in particular they can handle arbitrary copywriting tasks with very few examples given. Using GPT-3 in Google Sheets can make the process far simpler to get results across a wide range of inputs, and get back structured results for use later. This will let you write a personalized ad for every single keyword in your account, with the keyword inserted more naturally.

AI Copywriting

0:01 Hey, I'm gonna show you how to use GT three to write your ads and write them at scale, right? So we have 200 different keywords here for hotdog vendor, right?
0:12 One of our clients. And obviously we can't go through every single keyword and write them individually so it'll take too much time.
0:18 Also, a lot of the keywords don't have, you know, a huge amount of spend, right? So it's like 37 cents. So, you know, what you would do normally if you didn't have ai is to just, you know, rank them by costs descending, and then just write personalized ads for you know, for these ones that the high spending ones.
0:36 So like all beef hot dogs, who would include the word Al Beef Hot dogs in, you know, in the word.
0:41 You know, and then yeah, he's an Applegate Organic Hot Dogs. This is competitor. You would say. You would say like, Don't eat Applegate organic hot dogs.
0:49 Eat us instead. Yeah, Or if it's a, if that's something you sell, then you could kind of incorporate that. Cool.
0:56 So so you can see a few that we've manually written here. So it just said you know, try dog shared or Beef Hot Dogs today in the East Village location.
1:07 Choose my hundred sources. This is a competitor example. So seven 11 sells hot dogs by local at dog shared based in the East Village.
1:15 No more. Seven-eleven hot dogs pick from my hundred sources. So each time I'm including the keyword. You know, this one was all it was beef hot dogs buy hotdog, you know, And then and then this is like seven 11 the competitor so you can experiment with different prompts but what I found was this one works here, like right short ad descriptions, max 90 characters, which includes the keyword.
1:39 And then and then I kind of break it up here with these dashes. So this is like the template and I'm giving it three different examples.
1:51 But then I'm basically kind of, you know, in attaching on like the next example. So you can see here GT three I'm pushing in this template and then sand says, caning it together.
2:06 And then I basically press control enter to add another space, like a line break. Then I put the keyword so that the keyword is this bacon wrapped hotdogs grill.
2:19 And then the add description and then that's right end the prompt. So that part is the prompt, and then that will send that to G PT three.
2:27 And that will kind of come back with the response.  The other two parameters here, these are again, something to experiment with.
2:34 This is how many tokens you want back max tokens. So I said 20 here. I think for this length of our description, that seems to work.
2:41 And then the temperature. So basically how random are the results? Like the, it's 0.7, which is the default, but you could have it like really straight list, which is like around 0.2.
2:54 Or you could have it like really wild and creative, like 0.9. Cool. So here are some examples of what what it came back with.
3:03 And you know, if you wanna kind of try it with any of these, you know, here you gonna try a bagel dog.
3:10 I've just copied and pasted, and there we go. We have an ad straightaway. <laugh>, get a bagel dog at dog shed local local in these village, I'm, and pick, pick from a hundred a hundred sources.
3:22 So it actually provides a pretty good response. Now the way this works, by the way, if when you open this, your sheet will won't have an API key here and I'm gonna delete this API key afterwards.
3:36 Yours will say you know API key here and you just need to basically click through to open ai and create an account.
3:47 And then you'll be able to get your API key in there. It costs like 2 cents for a thousand tokens, I think.
3:53 So it's you know, each one of these is like, not, not a huge amount of you know, not a huge amount of of data, right?
4:02 So you know, it's not gonna, it's not gonna cost that much. Probably like a few cents, right? So here I'll just show you my usage.
4:12 So this is, today I spent a dollar 31. I've done like, you know, hundreds of these, right? And in testing.
4:17 So it you know, it's, it's, it's a really, really cheap opportunity, right? Or a really, really cheap way of doing this.
4:26 And the really cool thing is you can literally just like, if you want to, you know, you can just copy and paste, you know, and drag this all down, and then this grading <laugh>, you know lots and lots of ads.
4:37 You can have 200 ads like you have here. Not always does it work, right? So in some cases it gets it wrong.
4:47 You know, like if I just delete those like for example, you know, this one here gets it right, get all beef hot dogs, a dog shed.
4:57 So that's really cool. And I picked up on the hundred sources thing. All American hotdog carts for sale, that's pretty good.
5:04 It handled that like it's selling the actual cart rather than the hotdog. Although one problem is it's saying, you know, choose from a hundred sources, but that's not relevant specifically for this keyword.
5:14 So what I recommend is you go through and actually kind of copy the ads once you've done them. So I just select them and press control C and then control shift V.
5:26 If it's control shift V then it's just pasting the actual text itself, right? So so now when you kind of update the function you know, it's not gonna update anything here.
5:41 So here, like if I change this to, you know I change the temperature a little bit you know, here we go.
5:51 Now we have a better version. I think all my hot dogs are back, but this is, you know, competitor say, I'm gonna copy this across.
6:02 And then I would like kind of change this and kind of take, take an edit and say, Can you just edit these?
6:08 So by, by all be of <inaudible>, should I pick from a hundred sources? And then we could say like  Then we could say I'm a,  Or you buy, buy.
6:40 And then there we go. So I just copied the, that function across. So yeah, now, now we know that like this is within the right character.
6:50 We can also use a trim function. It's gonna get removed the spaces from it as well. We could use red Xs if we wanna remove stuff but, you know, kind of let you figure that out.
7:03 So once you've gotten you know, the, the variations that you want what I recommend is like, rather than kinda leaving it in here I would, you know, copy and paste values, and then I would copy it across into another spreadsheet.
7:17 There's two reasons for that. One is every time you, you know, kind of change something here or, you know, change your prompt or whatever then it'll recalculate everything and you'll lose all the variations that were good, right?
7:31 But then you also, the other reason is that your API key is basically like your password to your account. So you don't wanna share it.
7:39 And you can always go through, you know, and you know, you can go into API keys, you can delete the one that you were using, and that's fine.
7:52 So now this wouldn't work you know, if, if you wanted to use that, it's just not there anymore. So so yeah, it's you know, obviously have to use if you want to live dangerous dangerously, but or if you're sharing this internally, but what I would recommend is you don't share the actual generation spreadsheet.
8:08 You keep that safe to yourself. You just share the output and then let someone else, like a human error to clean it up.
8:15 Cool. All right, So hopefully that was useful.

What’s the advantage to using GPT-3 in GSheets?

Is AI actually good enough to write adcopy?

Why is it useful to do this sort of AI copywriting rather than writing it manually?

What does the ‘temperature’ mean when sending a prompt to GPT-3?

Every PPC marketer knows tailoring your adcopy to your keywords is best practice for increasing conversion rates and lowering cost per click. However, nobody wants to sit there and write hundreds of unique ads.

Make me 205 ads

UPSERT HEADQUARTERS – THE DAY BEFORE END OF QUARTER

You’re in a boardroom at UPSERT Headquarters and your boss, William Winters, Brand Strategist at Whipple, is asking you to figure out a way to protect the company’s data from Open AI. You’ve been researching the LM Studio and think it might be the answer.

We need to keep our data safe from Open AI.
I've looked into LM Studio, and it seems like it could be the answer.
But we need to decide how to use it.
What are our options? Do you have any ideas on how we can protect our data?\

LM Studio is an open source model created by Matter that is used to run models on a local computer. Despite being a smaller model than Chatpetite, LM Studio is slower and doesn't usually follow instructions as well as GPT-4 does. However, a new model created by Mystro 7 is getting closer to the results that one would normally get from GPT-4. 

For those looking to use LM Studio, it is important to understand the technical aspects of the model as running into errors is something that is common. Despite this, the text generated by LM Studio is still decent and with some editing, can be turned into something worthwhile. 

What makes LM Studio so unique is that it was created with the intention of creating narrative and character development. This is done by having two paragraphs per section. This allows for a unique fantasy style that is not found when using other models.

Despite the capabilities of LM Studio, it is still not as good as the free version of Chatpetite. This is why many people are hoping that the model will get better, as having something that is as good as GPT-4 would be extremely beneficial to those who want to create interesting stories. 

Overall, LM Studio is a great model that is able to generate interesting stories and narratives. Despite its shortcomings, if used properly, LM Studio can create some unique stories. It is important for those using the model to understand the technical aspects as well as the capabilities of LM Studio in order to have the best experience. With some practice and patience, LM Studio can be used to generate some great stories.

LM Studio

 Okay meet. Lama this is an open source model from matter. They release it which is really great because they spend all the money training it and then just released to open source. That means I can actually run this somewhere on computer and then there are different. Sources online where you can run this, like on hugging face or replicate. But I really wanted to run it. At my own computer, I'm using something called L M studio. Which you can download is just a Mac app. And and then once you've got that, Set up, you can download anything from hugging face. 

Like this is a downloaded the Lama to chat 13 B. Model and you can prompt it just like you would  So just going to paste in my prompt. And see how it does. What I've found in general is that obviously running it on your local computer is a little bit slower. So even though this is a smaller model than chatting petite that the response has actually come out relatively slower. The other thing is that it doesn't follow the instructions usually as well as GPT four, does. Does. These models, these episodes model is getting closer to what you would get from the free version of challenging.  And in fact, there's a new model out by Mistral which the Mystro seven. The model, which is actually pretty good. And pretty close to chat cheap T levels if you're comparing it to the free version, but nobody has managed to get to the actual results that you would normally get. From GPD for. 

So while this is generating, let's familiarize ourselves with what we came up with jazzy beauty. You can see. Wrote some pretty interesting stuff here. And, we have two paragraphs per section. And it's a unique fantasy style. With actual narrative and character development. Let's see if we can get that We run into an error. Sorry. You do run into this sometimes, unfortunately running these models require some technical ability in some cases. 

But I think in general, I would say these models aren't quite there yet. They're not as good as what you get from chat, GBT, even the free version. We're all hoping that these models will get better now because it would be very cool to have. Something that's as good. 

Now you can see here that the text isn't bad, it's come up with some interesting stuff here. It has, pretty decent introduction. But again, I prefer T's version.

What is the name of the open source model used in the tutorial?

What is the name of the software used to run the model?

What is the name of the open source model from matter?

What is the difference between running the model on a local computer versus online?

What is the name of the new model released by Mistral?

We're just now getting to the point where OS models are as good as ChatGPT (the free version), except they have one major advantage: no data sharing with OpenAI!

I don't want to share my data with OpenAI

GoolyBib Headquarters - Strategy Discussion Post-Lunch

At GoolyBib HQ, Gustav Gieger, our CEO, introduces DSPy Prompt Optimization for smarter baby bibs. He explains prompt engineering essentials, system design, and optimization strategies like chain of thought and co-pro, emphasizing the complexity yet vast potential for enhancing our prompts. Inspired, we're ready to apply these techniques.

Prompt engineering is crucial for our success in creating smart bibs for babies. Even with AI, we need to carefully design our prompts using DSP Y. It's a complex process, but it can significantly improve the performance of our prompts. Let's dive in and unleash the potential of DSPy Prompt Optimization!

DSPy Prompt Optimization: Improving AI Prompt Engineering

Prompt engineering plays a crucial role in training AI models to generate accurate and relevant responses. The recent development of DSPy (DeepSpeed Prompt) Optimization offers a framework that automates the process of prompt engineering, making it more efficient and effective. In this blog post, we will delve into the details of DSPy and explore its potential in optimizing prompts for AI models.

The transcript provided introduces us to DSPy and highlights its significance in the AI community. It emphasizes that even when AI is used to generate prompts, it still requires prompt engineering to design the inputs, outputs, and evaluation metrics. DSPy simplifies this process and enables prompt optimization through various strategies such as chain of thought and retrieval.

To begin using DSPy, the first step is to define the task. It is recommended to initially experiment with the OpenAI API to familiarize oneself with the complexity of the framework. For example, one can attempt to create a model that generates funny jokes on a specific topic. By setting up the task and configuring DSPy, it becomes possible to generate initial results.

The next crucial aspect is defining the prompt signature. DSPy offers flexibility in defining the inputs and outputs of the prompt. This can be done in a prediction signature format, specifying the topic as an input and the joke as an output. Alternatively, a class-based signature can be created, allowing for more advanced functionalities. This signature serves as the foundation for prompt optimization.

DSPy provides modules for prompting strategies, and the choice depends on the specific requirements. For instance, chain of thought involves asking the AI to provide rationale before generating the final result. Another strategy, retrieval (RAG), incorporates context from a search or external tools to enhance the prompt. These strategies can be combined and customized to create a tailored prompt pipeline.

Once the prompt pipeline is established, it is essential to evaluate the performance of the prompts. DSPy offers multiple evaluation metrics, including programmatic, human annotation, and synthetic evaluation. The choice of metric depends on the nature of the task and the availability of resources. Synthetic evaluation, where AI assesses the quality of the generated content, proves particularly useful when prompt optimization involves creative tasks like joke generation.

To train the model and optimize the prompts, it is necessary to gather data. DSPy provides various options, including manual entry of examples or using AI to extract examples from external sources. Synthetic examples can be generated by instructing AI to mimic the desired prompt format

DSPy Prompt Optimization

What is DSP Y and why is it important in prompt engineering?

How does DSP Y automate prompt rewriting?

What are the different optimization strategies in DSP Y?

What is the purpose of defining a task in the DSP Y framework?

Which optimization method in DSP Y improves the instructions without changing the examples?

Sure AI has saved me several hours by automating my tasks, but now I'm spending a tonne of time trying to optimize those prompts: can't AI do that for me too?

It's actually taking a long time just to write my prompts

We are in the boardroom at Whipple Consulting getting ready to discuss strategy for our client's thought leadership posts. We have the data from our research and a plan for getting the most out of our professional writer. Now we just need to figure out how to make the posts unique and save time in the process.

I know that we don't have a lot of time, but I think we should take a look at a process called repurposing. We can take existing content and build on it to create something new, and it should save us a lot of time.
Plus, it should give us the unique content we need to stand out.
What do you think?\n
Janet Jones at Copywriter:
That sounds like a great idea! I think I can take what we have so far and build on it quickly. 
What kind of timeline are we looking at?\n
William Winters at Brand Strategist: 
I think with the help of the professional writer, we should be able to have our posts ready in a couple of days. Let's get started!

Thought leadership posts are an important way to establish yourself as an expert in your field and gain recognition for your work. They can also be a great way to grow your business and network with like-minded individuals. But how can you create a thought leadership post that stands out and gets noticed?

One trick I’ve found to be useful is to rewrite my post in the style of the blog or website I want to be published on. For example, if you’re writing for the HubSpot blog, you can read some of their recently popular posts and rewrite your post in their style. You can also use the SEO keywords of their popular posts to make sure yours ranks well in search results. This helps you tailor your post to the exact audience you’re trying to reach.

Another great tip is to use analogies to explain complex concepts. This makes your post easier to digest and more engaging for the reader. For instance, when discussing the concept of dynamic pricing, you can use the analogy of painting to illustrate how it works and why it’s important. This helps to make the concept more accessible to readers who may not be familiar with it.

Finally, it’s important to treat your post like a conversation. Share personal stories, anecdotes, and opinions to make your post more interesting and engaging. This also helps to make your post unique and stand out from the crowd.

By following these tips, you can create thought leadership posts that stand out and make an impact. Not only will you be able to demonstrate your expertise, but you’ll also be able to establish yourself as a leader in your field and build relationships with other professionals.

Thought Leadership Posts

Uh, really the world is your oyster. Um, but, uh, I find this as a really powerful way to, um, basically, you know, get it in a good format. Uh, one, one, actually one trick. You can use by the way, um, is if you're running a guest post or, you know, you're kind of wanting it to get it past someone like an editor.

Like say you wanted to publish on the HubSpot blog. Uh, you know, you could actually have it rewritten, like read it yourself, have a rewritten in the style of the HubSpot blog. So it's like much more likely to get published. You know, you could actually, even if it's a smaller book that, that, you know, change, which he doesn't know about, you could actually then just go and, you know, copy.

A bunch of the recent posts are the ones that ranked the highest in SEO. Um, and then paste them in here and then ask it, um, you know, Tell me, you know, Briefly describe and bullet points. Like what the stylist. Cool. So. Here we go. Now it's been rewritten and I think it's much better. It's much more concise actually. So in the dynamic.

AGC pricing. Is it, uh, its appeal. Okay. Yeah. I've had my share of encounters. Now it talks about lotto, which is cool. And yeah, I think this is much better, right? Like it's much cleaner. Um, no, it's it's uh, yeah, here we go. It gives a really good analogy. It's like trying to pace it and price it. Painting.

Yeah, this is brilliant. Okay, cool. Uh, now we're at a much better post. And, um, now the, I think the ethics of this is much better as well. Like, you know, I really don't like the idea of just having chats empty. Right. All my posts. Right. Like I actually have interesting stuff. I think it's a show with the world. And, um, and now using it as a ghostwriter, I think is much more ethical because I'm contributing to the training data, right. I'm not just consuming it. I'm actually giving chatty T a lot of direction here saying like, here's what I think about the topic and, you know, hear my opinions. I think this is much more likely to rank on SEO. But it's, uh, more importantly, uh, going to be more uniquely interesting to other human beings as somebody to be just like the average of the internet. Uh, like you would get, if you, if you didn't follow this process. Cool. So, uh, hopefully that's useful to you and, uh, yeah, I'm using it. Um, yeah, a fair bit now. Uh, and, uh, you know, finding that saves me a ton of time. You know, this, this post. You know, it's taken us 24 minutes. I've had to explain it all to you, but, um, you know, I find that I can actually write like a four hour posts sometimes in, in half an hour, which is pretty amazing. Uh, you know, normally take me half a day, uh, or even a day to write some of these things. And now I can just get it out of my head and onto paper and, you know, use my professional rider to, to make it professional for me. Uh, and that helped me move on quickly to the next post.

What is one trick you can use to increase the chances of a post being accepted?

How much time can be saved by using this ghostwriting technique?

What does the author think about the ethics of rewriting content?

You don't have enough clients coming in, but you're also too busy to write content to attract new clients. You tried AI but it comes out sounding dull and uninteresting.

I just started consulting... now how do I get clients?

We're in the GoolyBib headquarters on a late morning. As the new employee, Gustav Gieger is counting on us to learn this new technology and use it to make their wearable technology smart bibs for babies. He's asked us to learn how to use the Segment Anything extension to quickly and precisely mask out objects in an image with a single click.

It's essential that you learn to use this technology quickly and effectively. We need to be able to make the most out of our resources and time. Knowing how to use the Segment Anything extension will help us get the job done faster and with more precision.\n 
Are you ready to get started?

Are you looking for a way to quickly and easily segment any image without the hassle of manually masking it? If so, then you’re in luck! Segment Anything Inpainting is a powerful tool that can help you get the job done quickly and easily.

Segment Anything Inpainting is an extension for AUTOMATIC1111 webui that allows you to quickly and easily segment any image with the click of a button. Simply upload the image to the software, and it will automatically segment the image into different parts. It can detect people, objects, and backgrounds, making it easy to separate them out and create complex masks.

The software uses advanced artificial intelligence algorithms to detect and segment the image. This means that the results are often accurate and require minimal manual editing. This saves you time and effort, as manually masking an image can be a long and tedious process.

The Segment Anything extension also has a feature called “Grounding Dino” that can be used to detect people or objects in an image and automatically generate masks for them. This is especially useful if you’re working with a large number of images and need to quickly create masks for each person.

Overall, Segment Anything Inpainting can be a powerful tool for quickly and easily segmenting images. It can detect people, objects, and backgrounds, all with the click of a button. It also has the ability to detect people in an image and automatically generate masks for them, for later use in inpainting.

Segment Anything Inpainting

 Okay. Let me show you how segment anything can be worked with in painting. So I'm just gonna click on the upload here and they're going to take an image that I wanted to in paint. Just gonna. Upload. 

That's the image. Now if I wanted to segment out her. It would be a bit of a pain, right? Cause I'd have to. Directly, mask this whole image and get it perfect. That'd be a real pain. And Africa around our whole body to do that. And nobody wants to do that, sir. 

Gonna get rid of that. And what I'm going to do is said is get this segment, anything extension. So the way segment anything works is if I click on it, And then I upload that same image. I could just click even just sometimes just once in the image and hit preview and it should actually mask out that dress or that person. 

And whatever it is we're trying to get. So while that's running, I'll just show you quickly how to install it. The there's an extension cord. S D Segment anything if you search for that. And in here you have instructions for installing it, but you also have the different models. I've downloaded. 

Specifically this one, that 1.25 gig Sam feet. I, and that's what I'm using. You install it like this, you go to settings. Or necessarily go to extensions, available load from, and then that's going to. Load this and then I'm just going to. And tick installed because I already have it. 

So here we go. Segment. And you can see Segment anything. So that, that's where you install it from you. Click install. Here. And then you just restart ex. Automatic 11, 11. So let's just see if it's worked. Yeah, look at this. It's actually segmented out. And it's got given three guesses. One is the dress, one is the person. 

And I'm going to say, like this one has done a really good job. In terms of masking split out specifically her. And then I'm just going to say copy to in paint and switch in paint, upload, and then I'm going to say. Why. 

Around. 

Generate just while it's doing that, I just want to show you where it goes. So when you download the model, And to go to extensions and then you want to find. Segment anything, and then you want to put it into the models. Sam here. 

Oh, here we go. I've got this backwards. I generated a white background. So my prompt was the mask. I actually intended it to be the other way around. So that's fine. That should be an easy. Thing to do I just click I actually sorry, not that one. I just want to. In paint, not masked. 

So it just hit, generate. And that's going to flip it around. Hopefully. Chevelle to just generate the background. But you can see how powerful this is, right? Like he kept the background. Perfect. Here. And Just going to see what this comes up with. Yeah, here we go. 

It's working.  You can imagine there's a lot of power here. You can generate quite a lot of stuff and you can make a lot of changes if you need. And because it's so perfect. The segmentation, it just really works well. If you wanted to segment more stuff, I would recommend doing them one at a time and then just saving them in a file and bring them together. 

The masks. But just because it doesn't work that well, if you click on multiple things the other thing you can do though, is you can use this ankle grounding, dyno.  If you click enable grounding, Dano and. Just put in a word. So like people. Then just hit preview. It will attempt to find the people in the image and it's using a different kind of model for that. 

But it will only find the people in the image and create masks and remove them out as well. Which is very magical and powerful.  We'll see how that works. But this is really useful. I think. In general, it just makes it very precise. At what ever it is. You're trying to mask out. 

Yeah. And here we go. So it's detected like her at least as a person and that's a pretty good mask.  If you try this and different images, you get different results, you can also dial up the threshold as well at like maybe if we stick a higher a lower threshold, it might identify this guy as well. But I really powerful stuff. Really useful. It's more advanced, obviously, because you have to use automatic 11, 11. But I'm sure that's a barrier you can get across. If it's going to save you a bunch of time masking. All right. Hopefully that was useful.

How do you install the segmentation extension?

What type of model is used for the Segment Anything extension?

What is the purpose of the segmentation extension?

What should be done if the segmentation extension does not work well?

Inpainting is a powerful tool, but it can be a real pain to neatly cut out parts of an image by hand using a brush tool.

How do we cut out and replace parts of an image?

You want to produce longer form stories with AI, but you find you can't fit enough info into the prompt to guide it to reliable results.

It's a tradeoff
Don't put enough context and the AI hallucinates
Put too much and you hit the token limit
There is a paper I read which helps you recursively build a coherent story using reprompting
Let me dig that out for you...

Recursive reprompting and revision is a technique for writing coherent stories that goes beyond the context length of traditional natural language processing (NLP) models. This technique was developed by a team of researchers to enable artificial intelligence (AI) to generate longer stories, and it involves using a plan and draft module, combined with a rewriting module. 

In the plan module, the premise of the story is given. For example, a new law graduate who returns home to start a career while facing a broken justice system. This is then used to generate the setting and characters. For example, the two characters in the story could be Lela Smith and Toby Johnson. The main plot points of the story can then be outlined. In this case, the story follows the hero's journey template, with Lillyn and Toby discovering the broken justice system, fighting against injustices, and ultimately becoming respected advocates. 

The draft module is used to generate the story. It works by first providing the previous summary and the upcoming passage. For example, if the previous summary is "Lula moves back to her hometown", the upcoming passage could be "She is dismayed to discover...". The draft module also uses the least to most technique, starting with the least amount of information (the premise) and gradually expanding to more information (the plan).

The rewriting module is used to correct any issues with the text, such as hallucinations, or facts that don't make sense. For example, if the story says that Lela and Toby have been friends for a long time, even though the previous text said that they only just met, the rewriting module would identify this inconsistency and correct it. It also works by progressively extracting facts from the text, building a list of facts as it goes. This allows it to check for inconsistencies between the old facts and the new facts in the text. 

Overall, the recursive reprompting and revision technique is a great way for AI to generate longer, more coherent stories. It allows AI to be more creative, as it can generate stories that reflect the structure of successful stories, such as the hero's journey. It also prevents issues like hallucinations, as the rewriting module can identify any inconsistencies it finds.

Recursive Reprompting and Revision

00:01 Okay, so this is recursive, reprompting and revision. So this is a technique for writing coherent stories way beyond the context length.
00:11 This comes from a paper that you told me to hear generating longer stories with, with a recursive, reprompting revision. And the general idea here is that you can't just say, like, write me a story.
00:24 I mean, novel to the AI because the context window is relatively small. So you can get like a few thousand words in and then you know, then you have to kind of generate the next section, but it's not aware of what it wrote in the previous section, right?
00:40 So you know, chat GPT makes this a little bit easier in that it can kind of see past messages, but again, you know, you can kind of run your tissues with hallucinations.
00:49 So what these guys came up with is if you kind of recursively build the story using these kind of plan and then draft and rewriting modules.
01:01 So the way this works is first you generate a plan. So what's the premise of this? You know, what are the characters you know, where's the location.
01:10 And then you use the plan to generate each draft the plot points. And then for each plot point kind of generate the draft text.
01:19 And then. And then rewrite the draft text in case it doesn't make sense with any of the facts that are coming in so far.
01:26 So let me just dive in here and show you how this works. And so this is a you know, Google Colab notebook.
01:35 Just, you know, you can run this and you're gonna share this as long as you have your own API key.
01:40 If you click here and go the API keys, I've also linked to the paper here. But you know, if you run the cell, if you just click play, if you press like shift and enter, command and enter then it's gonna prompt you for your key.
01:52 You put that in there, install open AI. So run that. And then this is just some code to run the completion.
02:00 So like to actually message open AI and get the text back. Cool. And if, if all this is working, you say, you know, you'll get a response from open AI.
02:10 All right. So here's the plan module. This is the first part you put in the premise and you can type it in here.
02:16 Actually. This is just like a feature you can do with Google notebook. A new grad law a new law grad returns him.
02:26 Start a career, but start also with a broken justice system. And you know, we we feed the premise in to open AI and then we get the story.
02:39 Setting back, right? So that's the first thing we get back. And this is why it's called recursive re-prompting. Cause it's kind of, you know, you put in the premise and you get the, the st- setting.
02:49 And then we're going to put in the premise and the setting and get the characters. So you can see here, we put in the premise to the prompt and the setting.
02:58 And then we give it like a character format. So we've asked it you know, give it like name and then context, given some examples.
03:07 And then told it not to reuse the same character names because it has a habit of doing that. Cool. So you run that and then you're going to stop when it gets to the full stop.
03:19 So make sure that it's only one complete sentence, otherwise you get really long descriptions sometimes. And and then it's going to kind of generate the characters.
03:28 So here I've just given it just asked if it's just going to get two characters here. It came up with Lela Smith.
03:40 So now we're just asking to outline the main plot points of the story. So again, it's recursive re-prompting. But now we're in.
03:48 We use the premise to get the setting, but now we use the premise and the setting to get the characters.
03:52 And we're using all three to get the outline of the plot points. So the plot is basically Lillyn Muftham. She discovers a broken g- the justice system, she meets Toby.
04:04 Together they fight against injustices. And despite facing obstacles, they find success. And in the end they become respected advocates. If you're paying attention here, this is basically the hero's journey.
04:15 So, it's really interesting that you know, we didn't tell her to use the hero's journey template, but it's obviously, you know, derived that from the data.
04:23 So, this is the same story behind Harry Potter, The Matrix and. Lion King you know, any Marvel movie really, like every, every successful movie uses this template.
04:35 So, pretty interesting. It comes up with that. Cool. So, now, this is what we have so far. We have the premise, the setting, characters, and then.
04:43 And this is the planning module altogether, right? So, this is kind of planning what the story's going to be. And this is also using another technique that you might be familiar with, which is, is in general.
04:57 Purpose technique with AI which is the least to most technique. Right? We started with the least, like the smallest thing, which is just the premise.
05:05 And then we're kind of expanding out to the most, right? Like, we've, now we have a whole plan. We can kind of feed that.
05:11 And to, you know, generating more stuff. Cool. So now we're asking for the plot points. So I'm just going to split.
05:17 So we have like a list of these without the numbers. So that's useful. And then we're going to iterate through.
05:25 And with every plot point, we're going to give it the previous summary story. We're going to say immediately before this passage.
05:32 This is what you wrote. And then in the upcoming passage, this is the new plot point. Here's some relevant. Context as well.
05:41 And you can kind of see that coming together here. So if you have if you start with say plot point two you know, it's given the context of the characters.
05:53 Previous summary story is that Lula moves back to her hometown as this is the text that it generated just before.
05:59 And then in the upcoming passage, she's dismayed to discover So what we're doing here is, you know, really based off this draft module in this paper.
06:10 And, you know, it's giving the relevant context to, you know, upcoming passage, etc. So. We've just kind of translated that into our prompt template here.
06:20 Cool. So that actually generates the whole story and you can, you can join it all together here, so like, the Smith graduated from college, you know, you should move back.
06:30 She knows the area well, you know, so this is like the whole story. And then introduces Toby Johnson. Also returned to his hometown.
06:39 Cool. Let's leave that in. Toby were excellent lawyers, right? Yeah. So this is like the whole story. It's put the whole story together and it's much more coherent than if you just kind of prompted these things in isolation because it has the continuity.
06:55 Of each draft module feeding off the previous one. But there's sometimes issues with this. One thing I found when I run it with this prompt in particular, like the premise is that sometimes like Leela is, you know the same, you know, graduated at the same time as Toby, which doesn't
07:16 necessarily make sense, right? Sometimes it says that they move, move back and meet each other, but then it. The next bit it says that they had been friends for a long time or that they've been practicing a lot together for a long time, which doesn't make any sense.
07:30 Right. So they still have some issues with you know, hallucination and. And this is like a pretty useful module, even on its own, this revision module to, to start hallucination.
07:44 You can basically just ask it, you know, are you making things up, right? Like, like what's wrong? With this text, you can get it to correct itself.
07:52 So the way that we've done this I've departed a little bit from this paper. They have this whole coherence and relevance thing where they generate multiple drafts and then.
08:06 You know, different algorithm. I didn't do that. I think it's overkill potentially for this. And that's not necessarily prompt engineering, right?
08:16 But you can kind of think about. This is like a multi-vote. So you know, it generates multiple drafts and then it ranks them based on coherence and then chooses the top one.
08:25 That's one way to do it. You know, there's also the edit model, but they kind of keep this two different steps of these inferred tax, inferred facts.
08:35 They keep like an attribute dictionary. So like Lisa Turner is the younger sister, you know, gender is female. You know, workplaces, the restaurant, right.
08:45 So they, they gone a little bit deeper with this. I just want to demonstrate the whole revision kind of prompt concept.
08:52 So what I've done is just focus on the Vax. So what facts can be inferred from this text? And you can see here, I start actually on index one meaning you know, we, we skipped the first text.
09:08 So we're going iteratively through the te- the drafts, and we're taking the first draft as a gospel. So anything that's in the first draft is gonna be inferred as, as facts, right?
09:20 Because there's nothing previous to disagree with. But we're going through like the reader would and then we're checking whether the next paragraph you know, agrees with the old facts.
09:31 So we have this faxed fax array, and we have this edits array. And what we're doing is we're checking the facts for the old text.
09:42 We're getting the response. And then and then we're adding that to the facts to this array. And then we're getting.
09:50 Checking in new facts. So what could be a friend from this text? And then we're joining this is kind of getting the old facts like as, as we go through this, we're going to have the old facts from the first text and then we'll get to the second text.
10:04 We're going to have old facts from two texts, right? The two previous texts. So we're going to have like a longer list.
10:09 This kind of joins all that together with spaces and then, or with the new line. And then we're asking it for consistency.
10:18 So saying, are any of the new facts inconsistent with the old facts? And if you remember the old facts will continue to grow, right?
10:26 Like every time we add a new edit it's going to be checking the old texts that came before. And adding those facts.
10:33 So by the time we get to the sixth paragraph, we're going to have, say, five sets of old facts. So keep up with it.
10:41 This is a concept I call progressive extraction because we're kind of. Progressive summaries, the summarization is where you go through and you summarize each text and then you maybe like summarize the summaries.
10:54 Progressive extraction is where you go through each text progressively. But you extract the facts out or you extract quotes out or whatever it is you're extracting information.
11:05 So we're going through kind of building this list of facts and then over time we've got, you know, more facts to agree with essentially.
11:13 So we're checking the consistency and then here we're just simply asking it like edit so any consistencies are addressed. And we've got the rewrite and rewritten text, which is what actually goes into the edits and the edits.
11:26 The final kind of version of the text. So I'm just putting it out as you go through so you can see you know, Lula Smith have been pressing the law, the old facts.
11:39 You know, we're here. They, they moved back from their hometowns. So they've made a formidable legal team, blah, blah, blah.
11:49 Here it says in terms of consistency it's saying that they've been practicing the law in their hometown for a few years while Tony jumps and only just moved back to his hometown.
12:00 So they both graduated from the law school, both moved back to their hometowns. Was it but then they're saying. That they had worked basically like he had only just come back and she had been back for a while.
12:15 So so basically it rewrote the text based on this issue. Right. Here we found no consistency. So in this module, I've kind of kept it relatively simple.
12:28 I, I just rewrite it anyway, even if even if it's there are no consistency, inconsistencies found. But like if you were improving this, you might want to kind of pass the response and say like, if it says no inconsistencies found then, you know, don't, don't rewrite it.
12:46 Like that's one way you can kind of save on tokens. But you know, for, for. For me, it's kind of trivial, I think in this case.
12:52 Cool. So you know, most of the time it doesn't find any consistencies, which is nice. You're knowing this is here.
13:00 It sees here. So we have all that together. And. And really, if you know, if you want to get like all of the whole story together, we can just join the edits.
13:15 And that's basically the final. And you can kind of read that yourself and see if it's consistent. The, the few other things you can do to improve this that I'm really just kind of demoing this concept of recursive repurting revision, right?
13:31 Like. But you know, you could improve each one of these modules as you can see, right? Like we could you know, not rewrite it if if there are no inconsistencies found.
13:41 We could check for more than just inconsistencies. Sorry, we can check whether it's interesting. Or check whether you know whether, whether there's like enough drama.
13:51 You could actually when you're generating the plot points, you can give it a specific example and say like, I want a plot that's like synced- similar to this, right?
14:00 You know, or, or I want to use this framework. I want to use Joseph Campbell's monomyth, right? But, you know, as long as you kind of understand how this works, the general architecture, you can improve each individual module.
14:12 And, and get different responses depending on what your job is, right? This is specifically to write a story but you can imagine you can use this for anything, right?
14:21 Like, you use this for a blog post, a lot of fun blog posts you can use it for creating children.
14:27 You can use it for creating ebooks, whatever. And it doesn't have to be necessarily encode. You could, it would be tiresome, but you could do this with chat GPT, you know, just by kind of.
14:40 You're diligently repasting these different prompts until you have the whole book written. Right. So really you saw to understand this.
14:48 I think it's a pretty advanced topic, but you know, it's pretty powerful concept once you. Figure out the architecture.

What is the purpose of the revision module in the tutorial?

What is the purpose of the progressive extraction process?

The context window is a killer when it comes to AI: you can only stuff so much relevant info into the prompt.

Is it possible to write longer stories without the AI hallucinating?

GOOLYBIB HEADQUARTERS – AI STRATEGY DISCUSSION

Alright, team, gather 'round! We've got an exciting experiment lined up today. We're going to be testing different prompts to see if they impact the length of our AI-generated responses. I know, I know, it sounds a bit technical, but trust me, it's going to be fun!

So, here's the plan. We'll be using Google Sheets and ChatGPT to conduct this prompt testing. We'll input various prompts and observe how the AI responds in terms of length. It's like a little game of wordplay, but with a purpose.

Now, the results might surprise us. We'll find out that specifying a word length doesn't consistently affect the response length. But here's the interesting part: using emotional prompts in all caps does have a noticeable impact. Fascinating, right?

So, grab your keyboards and let's dive into the world of prompt testing! We're eager to see what our little experiment reveals. Remember, this is just the first step towards making our generated blog posts longer and better. 

Let's get started!

Alright, team, gather 'round! We've got an exciting experiment lined up today. We're going to be testing different prompts to see if they impact the length of our AI-generated responses. I know, I know, it sounds a bit technical, but trust me, it's going to be fun!\nSo, here's the plan. We'll be using Google Sheets and ChatGPT to conduct this prompt testing. We'll input various prompts and observe how the AI responds in terms of length. It's like a little game of wordplay, but with a purpose.\nNow, the results might surprise us. We'll find out that specifying a word length doesn't consistently affect the response length. But here's the interesting part: using emotional prompts in all caps does have a noticeable impact. Fascinating, right?\nSo, grab your keyboards and let's dive into the world of prompt testing! We're eager to see what our little experiment reveals. Remember, this is just the first step towards making our generated blog posts longer and better.\nLet's get started!\nGustav Gieger at CEO.

The Importance of Prompt Testing in AI Experiments

Prompt testing is a crucial aspect of AI experiments, and the good news is that you don't need coding skills to perform it. In the provided transcript, the focus is on testing the effectiveness of different prompts in increasing the word length of blog posts. The goal is to explore whether specifying a word length in the prompt can influence the output generated by the language model.

The transcript highlights the use of emotional stimuli to enhance the performance of large language models. AI, being a simulation of the human brain, can respond to emotional cues in a similar way to humans. This understanding forms the basis of the hypothesis being tested in this experiment.

The experiment involves comparing different variations of prompts, labeled as Variation A, Variation B, and Variation C. The control prompt serves as the starting point for each variation. The experiment aims to determine if specifying a word length in the prompt will result in the desired outcome. The length of the responses is measured by counting the number of words in the generated text.

The experiment is conducted by pasting the prompts into a Google Sheet and using the OpenAI API to generate the responses. The AI model used is GPT 4, and the responses are copied back into the sheet for analysis. The experiment is run multiple times to account for the variability in AI model responses.

The results of the experiment show that specifying a word length in the prompt does not consistently yield longer responses. In some cases, the responses are shorter than expected. This finding suggests that the AI model does not effectively follow word length instructions.

The experiment demonstrates the importance of prompt testing in AI experiments. By systematically testing different prompts and measuring the outcomes, researchers can determine which prompts are effective and which ones are not. This information can be used to refine the prompts and improve the performance of AI models.

Moreover, the experiment highlights the value of documenting and sharing the results of prompt testing. This fosters a culture of testing within the organization and enables collaboration and validation of findings. The data collected from prompt testing can be used to guide further experiments and inform decision-making processes.

In conclusion, prompt testing is a vital step in AI experiments, allowing researchers to evaluate the impact of different prompts on the output of language models. By conducting systematic experiments, documenting results, and sharing findings, researchers can improve the performance and effectiveness of AI models.

Prompt Testing

Here you can see the outcome of a test.

I'm going to teach you about prompt testing, and the good news is you don't have to know how to code to do this.

This is variation A versus variation B. And specifically I was trying to test something that would increase the word length of blog posts. Even if you ask it for quite a long blog post, it doesn't listen all the time.

There's all sorts of tricks to getting it to listen. And the trick that I tried here, if you look into the datasheet  I specifically just said, make it really long or I lose my job

The reason why it works, if you look for the emotion prompting paper is that large language models can actually understand and can be enhanced by emotional stimuli.

And that's pretty interesting because, AI is like a simulation of a human brain, if you think about it that way. So it does make sense that they would react in similar ways to the way that we work, right?

So that's, the technical explanation to why this works. All right, so what you can see here, and I'll just outline these, is that, these responses are actually longer, but not consistently in some cases they're shorter. In this case we actually got like a pretty short response back, in a few other cases we've gotten below a thousand words.

The way I'm calculating word word length here is I'm just counting the result of splitting the text by space. You can see that this doesn't have to be really complicated. Evals is a whole topic, right? There's a lot that you can learn about evals and most of them are geared towards programmers who can set up like a eval framework, but in this case, you just need something simple, and Google Sheets, so you just need to count the number of words.

All right We have an existing experiment, and we've tested, adding it in all caps. But we want to add another variation here. And that should be as simple as just creating, Variation C. And then we're going to run it ten times.  Just count ten of these.  There we go.

So we have 10. And the hypothesis I have is that if we take this the hypothesis I have is that if I specify a word length then it might Listen to that, right? We'll see, and this is an experiment, right? Because we don't actually know if it's capable of following word length at this point.

I'm just going to call this specify  word length.  And I'm going to take this existing prompt, the control, and it's important that you start from the control each time.  Because, you might be getting different results based on, the different changes that you made. I'm just going to paste in the template here.

Get rid of these  these quote marks. And then I'm going to say write an outline for a blog post. Each section should be a minimum of paragraphs, two paragraphs long. And the overall article should be more than 2, 000 words. I'm going to pass that in there and then, so this is all, this is going to be the same every time,  and the actual  actual post here.

I'm just going to paste this in. So this is when the template's filled in. And if you see here we have these, in curly brackets. We have writing style, we have topic. These are variables that you need to fill in. And if we see an example here,  this is where it's  filled in. So you can see that, we filled in the topic, productivity with time blocking, and we filled in the writing style that we want, and we filled in the example post as well. So we've given it one example already, which is a post here about the Pomodoro technique.

And and then we're just gonna wait for it to come back with the actual response. We need to adapt this this filled in, we'll call it the formatted prompt as well. Because that's what we're actually going to copy and paste into chat GPT.  So now we have it in here. We need to just make that change that we made to the template. We said minimum of two paragraphs long, and there should be more than two thousand words.

Okay, so that's the only change that we're testing here. And we're just isolating that one variable. And we can just copy and paste this down.  So if you're running a more advanced experiment, you might want to try multiple things. So maybe you try multiple topics, maybe try different examples whatever it is you can test that.

And, the reason why, by the way, I'm running the same thing 10 times, as you can see. Just how often the results vary. And this is the thing with LLMs. They don't always give you the same response back. So it's really important that you actually take a look at what's coming back and you actually run experiments like this.

Okay, cool. Now we have Just to summarize, we have the specified word length, we have a hypothesis of what we think might work, we have the actual prompt template, which we're recording here for posterity, and we could use that template to fill in different topics or different writing styles, whatever it is.

Then we have the formatted prompt for this run, right? And this is each one of these rows is a run of chat GPT. And this is exact text that we've pasted in, that we will paste into chat GPT. So it's the prompt with the variables filled in. That's the formatted prompt. And then we just need to put in here which model we're going to use.

In this case we're going to use GPT 4, which is ChatGPT and then this is where we're going to paste in the response. We can also copy this formula down. And I'm just going to drag that down there. So that's going to count the number of words.  Cool. All right. So now I'm not going to use ChatGPT because ChatGPT I've found is relatively inconsistent.

It's somewhat personalized to people's results. If you have custom instructions, it's going to change things, right? I would say whenever you're templating something, you want to use the API, but obviously I, I promised you, you wouldn't have to code here. So I'm not going to make you but if you sign up for a developer account at OpenAI,

the main thing that I use here is this playground

this is the system message. This is like the custom instructions. We're not going to play with that here. We're just going to emulate talking to ChatGPT. We have the model over here, so we can use Turbo Preview. We can use GPT 4 whatever it is.

And then you can set the temperature as well. I just leave it at one right now. And that's relatively balanced. If you want it to be more deterministic, so you want it to give the same result every time, you could lower it. Or if you want it to be a bit more crazy or creative, you could increase it.

The main thing you want to make sure is you set the maximum length here. I've set it for 4, 000 because that should be enough. One token like it's saying here is roughly four characters of English text, so about third three fourths of a word. 4, 000 tokens would be about 3, 000 words.

There should be plenty for what we, what we need. And then don't worry about these other settings. I'm going to paste in the prompt here.  And we can see actually let me just  delete the quote marks. This is something that Google Sheets adds. But yeah, we have, this is our formatted prompt.

Remember, don't paste the template in there. So we have here, the productivity with time blocking topic. We have. The writing style.  And then we have this example post of another post I've written in the past. And it's going to go off that. Okay, cool. So let's hit submit.  And it's going to generate this response.

And, this takes a little bit of time with GPT 4. It is much faster with GPT 3. 5. But the nice thing about this, by the way, is you're also not going to hit your limits. Because GPT 4, you can only send 50 messages a day. Or, in three hours. Which is super annoying, right? And if you have this situation where you're trying to test a prompt, and you need to run it ten times, that means I can only test Five things and that sucks.

It's better to do it through the open AI API playground instead and one thing I should also mention is that there is some cost to this GPT 4 is pretty cheap But you can see the pricing and the usage in here, I'm  just going to try and open that in another tab so you can see.

So you can see I only spent three bucks this month, but last month I spent  a fair bit, 130 bucks. Just depends on what you're doing. This actual, these API calls, so like this one I just made here is like five cents.  So it's not a big amount so don't worry too much about it.

Cool. Alright here we go. We have  response here.  And I'm just gonna  copy.  I'm not even gonna read it. I'm just gonna paste in here. And here we go. We've only got 933 words. So it did ignore.  Or admonition to make the word land long. You have a couple of choices here. Either one, you can iterate on the prompt and just say, It did so badly on the first try that I'm, not even going to try it multiple times.

I'm going to try something else. Or you could keep running it ten more times and pasting these in, just to see if there's some variation. I'm going to do that, and then I'll restart.  Okay on the last run now you can see that I've just been copying and pasting in here and it's not looking good.

Actually, we're getting worse. results so far. Yeah, with an experiment like this, typically I would run it a couple of times, and if it's not even close, I would stop it. But this is interesting data, right? Because here we go, we've got the last one here. If we just paste that one in, what this tells us is that it does not respect word count.

It has basically completely ignored what we said here. That it should be more than 2, 000 words. That is, a good result when it comes to an experiment you'd be surprised because one, it tells us that technique doesn't work, saves us a lot of time and effort, like worrying about whether it's going to work or not and we're not going to try that one next.

We're going to try a different technique. The other thing that's really helpful here is that when we see other people using this technique we can tell them of this experiment and actually give them the data and say, look, it doesn't look like it works but you test it, and it fosters a culture of testing within the organization.

The really nice thing as well is that, this is just a small scale test with, you doing it in Google sheets, but like you could pass this to a developer and say, hey can you test this at a larger scale? Or maybe we could. If you find a result that does work like we did with make it really long in caps the emotion prompting then that's something that you could validate with a larger scale test.

And you could try it with lots of different topics or lots of different example blog posts or lots of different writing styles and see which of those things impacts the overall length.  You could also then scale this up and do, other types of evaluations. So maybe the word length isn't as good when we use one type of prompt but maybe the quality is high.

So you could have maybe outsource this have someone read each of these blog posts and then put like. A rating out of 10. The reason why it's useful to save this stuff here is that rather than just running it in chat gbt and then you lose the output, you can never find it again because it's in your history somewhere then you can actually give this data to someone else.

You actually have some real data now. And you can use that for whatever it is you need. You can go back and review it in different ways. You can add new tests to the experiment. Really, whatever it is you want. That's a really powerful thing you can do here. And just on the final thing, let me just update this this pivot table.

Okay, so to change the range, you just select the pivot table, then you click edit here, and then if you scroll up you'll see the range here. Right now it goes to G21,  which is just the bottom here, and we want it to go all the way down to G31. So we're just going to change that.  Or if you want to make this automatic, you could just say to G21.

And then it's gonna grab everything if you want, but  I'm just gonna do this manually. Cool. Alright, so we can see that we actually had a lower  response than I'llLoseMyJobs in caps. But we did have a slightly higher response to the control. The way we figured that out, I just copy and pasted the formula.

Actually let me just do it here, so  it would be this divided by this minus one so it is 7. 4 percent better in terms of length across, 10 observations, so it's only pretty small, but it did make a difference, it just didn't make as much of a difference as the all caps emotion response.

You could also test other things here, so you could take away the caps and just use the root emotion, or you could put, the word length in caps as well if you want. You can get pretty fine grained with this, but now you know how to do prompt testing, you can test whatever you want. So feel free to make a copy of this sheet, but also do make your own sheets, or add your own stuff if you find that necessary for the purposes of your test, or the type of evidence you need to be able to give to the other people on your team.

How did the speaker try to increase the word length of blog posts in the experiment?

Why did the speaker mention emotional stimuli in relation to AI models?

How was the word length calculated in the experiment?

Why did the experimenter run the same prompt multiple times with ChatGPT?

The biggest difference between prompting ChatGPT and prompt engineering is running each response 10x instead of trusting the first result.

Make our generated blog posts longer

LOCAL COFFEE SHOP – WEDNESDAY AFTERNOON

You and your boss Gustav Gieger, CEO of GoolyBib, are sitting in a local coffee shop on a Wednesday afternoon. You've been tasked with writing a whole story with AI, and you're discussing the best way to use recursive reprompting and revision, based on the hero's journey template and least to most technique. You also need to ensure accuracy and consistency in the story.

We need to make sure that our AI story generator has enough context to make a coherent story. It needs to know the beginning, middle and end of the story.
We should also look into the hero’s journey framework and the least-to-most technique to ensure accuracy and consistency.
Let me do a little research, and then we can come up with a plan for how to approach the task.\n
We can use recursive reprompting and revision to make sure that the story is as good as it can be.

Recursive reprompting and revision is an AI technique which seeks to generate longer stories beyond the limitations of the context window. By using a plan-and-draft-and-rewrite module, AI can be given a premise and then use that to generate a setting, characters, plot points, and ultimately a coherent story.

This technique starts with the premise, which is used to generate a setting. Characters are then generated based on the premise and setting. From there, the plot points are outlined, and the draft text is generated for each plot point. Finally, the draft text is rewritten in case it doesn't make sense with the facts that have been established.

The plan-and-draft-and-rewrite module can also be used to address the issue of hallucination, which is when AI makes up facts that don't actually exist. To do this, it checks for consistency between the new facts and the old facts. For example, if an AI writes that two characters have been friends for a long time, but it was established earlier that they just met, the AI will rewrite the text to match the facts.

This technique is based on the hero's journey, which is a template used in many successful movies such as Harry Potter, The Matrix, and the Lion King. It is also a form of progressive extraction, where the AI progressively extracts facts from each text and builds up a list of facts.

In conclusion, recursive reprompting and revision is a powerful technique which allows AI to generate longer, coherent stories. By using plan-and-draft-and-rewrite modules, hallucination can be addressed, and the AI can generate stories based on the hero's journey template. Through progressive extraction, AI can build up a list of facts as it goes through each text.

00:01 Okay, so this is recursive, reprompting and revision. So this is a technique for writing coherent stories way beyond the context length.
00:11 This comes from a paper that you told me to hear generating longer stories with, with a recursive, reprompting revision. And the general idea here is that you can't just say, like, write me a story.
00:24 I mean, novel to the AI because the context window is relatively small. So you can get like a few thousand words in and then you know, then you have to kind of generate the next section, but it's not aware of what it wrote in the previous section, right?
00:40 So you know, chat GPT makes this a little bit easier in that it can kind of see past messages, but again, you know, you can kind of run your tissues with hallucinations.
00:49 So what these guys came up with is if you kind of recursively build the story using these kind of plan and then draft and rewriting modules.
01:01 So the way this works is first you generate a plan. So what's the premise of this? You know, what are the characters you know, where's the location.
01:10 And then you use the plan to generate each draft the plot points. And then for each plot point kind of generate the draft text.
01:19 And then. And then rewrite the draft text in case it doesn't make sense with any of the facts that are coming in so far.
01:26 So let me just dive in here and show you how this works. And so this is a you know, Google Colab notebook.
01:35 Just, you know, you can run this and you're gonna share this as long as you have your own API key.
01:40 If you click here and go the API keys, I've also linked to the paper here. But you know, if you run the cell, if you just click play, if you press like shift and enter, command and enter then it's gonna prompt you for your key.
01:52 You put that in there, install open AI. So run that. And then this is just some code to run the completion.
02:00 So like to actually message open AI and get the text back. Cool. And if, if all this is working, you say, you know, you'll get a response from open AI.
02:10 All right. So here's the plan module. This is the first part you put in the premise and you can type it in here.
02:16 Actually. This is just like a feature you can do with Google notebook. A new grad law a new law grad returns him.
02:26 Start a career, but start also with a broken justice system. And you know, we we feed the premise in to open AI and then we get the story.
02:39 Setting back, right? So that's the first thing we get back. And this is why it's called recursive re-prompting. Cause it's kind of, you know, you put in the premise and you get the, the st- setting.
02:49 And then we're going to put in the premise and the setting and get the characters. So you can see here, we put in the premise to the prompt and the setting.
02:58 And then we give it like a character format. So we've asked it you know, give it like name and then context, given some examples.
03:07 And then told it not to reuse the same character names because it has a habit of doing that. Cool. So you run that and then you're going to stop when it gets to the full stop.
03:19 So make sure that it's only one complete sentence, otherwise you get really long descriptions sometimes. And and then it's going to kind of generate the characters.
03:28 So here I've just given it just asked if it's just going to get two characters here. It came up with Lela Smith.
03:40 So now we're just asking to outline the main plot points of the story. So again, it's recursive re-prompting. But now we're in.
03:48 We use the premise to get the setting, but now we use the premise and the setting to get the characters.
03:52 And we're using all three to get the outline of the plot points. So the plot is basically Lillyn Muftham. She discovers a broken g- the justice system, she meets Toby.
04:04 Together they fight against injustices. And despite facing obstacles, they find success. And in the end they become respected advocates. If you're paying attention here, this is basically the hero's journey.
04:15 So, it's really interesting that you know, we didn't tell her to use the hero's journey template, but it's obviously, you know, derived that from the data.
04:23 So, this is the same story behind Harry Potter, The Matrix and. Lion King you know, any Marvel movie really, like every, every successful movie uses this template.
04:35 So, pretty interesting. It comes up with that. Cool. So, now, this is what we have so far. We have the premise, the setting, characters, and then.
04:43 And this is the planning module altogether, right? So, this is kind of planning what the story's going to be. And this is also using another technique that you might be familiar with, which is, is in general.
04:57 Purpose technique with AI which is the least to most technique. Right? We started with the least, like the smallest thing, which is just the premise.
05:05 And then we're kind of expanding out to the most, right? Like, we've, now we have a whole plan. We can kind of feed that.
05:11 And to, you know, generating more stuff. Cool. So now we're asking for the plot points. So I'm just going to split.
05:17 So we have like a list of these without the numbers. So that's useful. And then we're going to iterate through.
05:25 And with every plot point, we're going to give it the previous summary story. We're going to say immediately before this passage.
05:32 This is what you wrote. And then in the upcoming passage, this is the new plot point. Here's some relevant. Context as well.
05:41 And you can kind of see that coming together here. So if you have if you start with say plot point two you know, it's given the context of the characters.
05:53 Previous summary story is that Lula moves back to her hometown as this is the text that it generated just before.
05:59 And then in the upcoming passage, she's dismayed to discover So what we're doing here is, you know, really based off this draft module in this paper.
06:10 And, you know, it's giving the relevant context to, you know, upcoming passage, etc. So. We've just kind of translated that into our prompt template here.
06:20 Cool. So that actually generates the whole story and you can, you can join it all together here, so like, the Smith graduated from college, you know, you should move back.
06:30 She knows the area well, you know, so this is like the whole story. And then introduces Toby Johnson. Also returned to his hometown.
06:39 Cool. Let's leave that in. Toby were excellent lawyers, right? Yeah. So this is like the whole story. It's put the whole story together and it's much more coherent than if you just kind of prompted these things in isolation because it has the continuity.
06:55 Of each draft module feeding off the previous one. But there's sometimes issues with this. One thing I found when I run it with this prompt in particular, like the premise is that sometimes like Leela is, you know the same, you know, graduated at the same time as Toby, which doesn't necessarily make
07:17 sense, right? Sometimes it says that they move, move back and meet each other, but then it. The next bit it says that they had been friends for a long time or that they've been practicing a lot together for a long time, which doesn't make any sense.
07:30 Right. So they still have some issues with you know, hallucination and. And this is like a pretty useful module, even on its own, this revision module to, to start hallucination.
07:44 You can basically just ask it, you know, are you making things up, right? Like, like what's wrong? With this text, you can get it to correct itself.
07:52 So the way that we've done this I've departed a little bit from this paper. They have this whole coherence and relevance thing where they generate multiple drafts and then.
08:06 You know, different algorithm. I didn't do that. I think it's overkill potentially for this. And that's not necessarily prompt engineering, right?
08:16 But you can kind of think about. This is like a multi-vote. So you know, it generates multiple drafts and then it ranks them based on coherence and then chooses the top one.
08:25 That's one way to do it. You know, there's also the edit model, but they kind of keep this two different steps of these inferred tax, inferred facts.
08:35 They keep like an attribute dictionary. So like Lisa Turner is the younger sister, you know, gender is female. You know, workplaces, the restaurant, right.
08:45 So they, they gone a little bit deeper with this. I just want to demonstrate the whole revision kind of prompt concept.
08:52 So what I've done is just focus on the Vax. So what facts can be inferred from this text? And you can see here, I start actually on index one meaning you know, we, we skipped the first text.
09:08 So we're going iteratively through the te- the drafts, and we're taking the first draft as a gospel. So anything that's in the first draft is gonna be inferred as, as facts, right?
09:20 Because there's nothing previous to disagree with. But we're going through like the reader would and then we're checking whether the next paragraph you know, agrees with the old facts.
09:31 So we have this faxed fax array, and we have this edits array. And what we're doing is we're checking the facts for the old text.
09:42 We're getting the response. And then and then we're adding that to the facts to this array. And then we're getting.
09:50 Checking in new facts. So what could be a friend from this text? And then we're joining this is kind of getting the old facts like as, as we go through this, we're going to have the old facts from the first text and then we'll get to the second text.
10:04 We're going to have old facts from two texts, right? The two previous texts. So we're going to have like a longer list.
10:09 This kind of joins all that together with spaces and then, or with the new line. And then we're asking it for consistency.
10:18 So saying, are any of the new facts inconsistent with the old facts? And if you remember the old facts will continue to grow, right?
10:26 Like every time we add a new edit it's going to be checking the old texts that came before. And adding those facts.
10:33 So by the time we get to the sixth paragraph, we're going to have, say, five sets of old facts. So keep up with it.
10:41 This is a concept I call progressive extraction because we're kind of. Progressive summaries, the summarization is where you go through and you summarize each text and then you maybe like summarize the summaries.
10:54 Progressive extraction is where you go through each text progressively. But you extract the facts out or you extract quotes out or whatever it is you're extracting information.
11:05 So we're going through kind of building this list of facts and then over time we've got, you know, more facts to agree with essentially.
11:13 So we're checking the consistency and then here we're just simply asking it like edit so any consistencies are addressed. And we've got the rewrite and rewritten text, which is what actually goes into the edits and the edits.
11:26 The final kind of version of the text. So I'm just putting it out as you go through so you can see you know, Lula Smith have been pressing the law, the old facts.
11:39 You know, we're here. They, they moved back from their hometowns. So they've made a formidable legal team, blah, blah, blah.
11:49 Here it says in terms of consistency it's saying that they've been practicing the law in their hometown for a few years while Tony jumps and only just moved back to his hometown.
12:00 So they both graduated from the law school, both moved back to their hometowns. Was it but then they're saying. That they had worked basically like he had only just come back and she had been back for a while.
12:15 So so basically it rewrote the text based on this issue. Right. Here we found no consistency. So in this module, I've kind of kept it relatively simple.

What is the purpose of the recursive reprompting and revision technique?

What is the hero's journey template mentioned in the tutorial?

What is the difference between the paper's approach and the approach taken in the tutorial?

It's easy to create short form content that's consistent, but AI tends to hallucinate and introduce inconsistencies when you go outside the amount of text that fits in a single prompt.

I want to write a whole story with AI

LeftMedia Headquarters - Pre-Meeting Banter

You're in the LeftMedia headquarters for the pre-meeting banter. You're a part of the team that will be using LangChain to summarize large documents. You've heard of the power of LangChain, PyPDF, and recursive text splitter, and you're excited to learn more.

Have you guys heard of LangChain?
It's a powerful tool for summarizing large documents quickly and efficiently.
With the recursive text splitter and PyPDF, we can get the job done in no time.
Let's get started and see what this software can do!

Delve deep into the world of LangChain, learning how to leverage its capabilities to condense large documents into easy-to-digest summaries.
One of the key skills you will acquire is the use of the "map_reduce" chain type for summarizing voluminous documents.
Also it's possible to customize the summarization chain to suit different languages and specific documents. We'll teach you how to create a custom "map_reduce" chain for generating summaries in different languages, and customize map and reduce patterns as per the needs of specific documents. 

Automatically Summarizing with LangChain!

 So in this video, we're gonna cover how you can summarize large documents using Lang chain. So let's start by installing Lang chain And PI P W F if you don't already have it. 

  After installing those packages, what we're then gonna do is import a range of different packages from Lang chain and requests. So we're gonna import the chat models, OpenAI. We're gonna also import a document loader called PI PDF loader. that's going to allow us to load PDF documents directly into Lang chain documents. 

We'll also use one of Lang chain's most common text splitters called the recursive text splitter, And we'll load in something called a load summarized chain function, as well as that we're going to import requests and we'll be going and getting a marketing book called principles of marketing. 

PDF. we're gonna just download that and store it on the local file system with the requests. As after we have downloaded the PDF, what you are then going to do is you're going to use a recursive character, tech splitter. Which will recursively split across a range of different separators. 

and it will try and keep within a specific size. So there are a couple of different parameters that you can choose with the recursive text splitter. One that's quite common is using something called the chunk size. And the chunk size is where you're defining how large your chunks are. And you can basically. determine what size, the chunks that you would like to have. and then, by default, all of the lo the documents that you create will have that size in terms of character length. 

You'll see here that we're loading in the PDF and we're then splitting it. so we're doing loader load and split. And then if we have just a brief look here at the pages, you will see. That we now have these pages, which are basically line chain documents, and it's got some metadata about what the source of it is. And the page of it is. 

also notice how we've got some of the page content here, so you can see the first page. is, doesn't have much content. And then if we start to look at some of the other ones, we can see that this book is produced by the university of Minnesota. what we want to do is we want to load in our chat model as an L L M. and this is important because we're going to use this in the various functions. So what you're gonna do is you're going to import that chat open AI. Remember that if you do. need to set an API key. you will have to set that here. so the way that you can do that and I'll leave this in the code would be to use the OS package and then you could do os Enron. and then we'll put the open AI. API key. And you will add your API key here. And then that would basically mean that this model is then authenticated, uh, so that you can easily call open AI. You'll see that the chat model by default uses GT 3.5 turbo, and this will be you know, upgraded as more, more and more models. Roll out. 

Um, we're gonna use a function today called load summarized chain. And essentially what this does is it takes in an L L M a chain type. So that could be stuff. Map map reduce. there's ones that are quite important. I personally find that map reduces very good for large documents. And essentially what happens in MapReduce is you can think of MapReduce being, you will do an operation on each individual document, and then once you've got all those documents, 

you will then use a reduce pattern on the, all of those documents. So you could have a hundred documents you would create a hundred summaries, and then you would have a reducing pattern that would then recursively reduce those a hundred documents into a final summary. the way that we can load this up is by using load summarize chain, the l lm And the chain type. and in this case, we're passing in map underscore reduce.  this one in particular to do the whole 400 pages of a book. that took about 30 minutes. 

so it does take a lot of time to run a large summarization chain. Um, however, this is kind of the summary that you can see, and it says this passage covers a range of topics related to marketing including defining marketing, strategic planning, consumer behavior. So we've got a really nice high level summary after running that summarization chain on the entire p d f of the book. 

Now I think it's worthwhile just seeing it in action. So if we just have a look at like, you know, the, the, 10 pages and running a summarization chain, You'll see it's taking a bit of time and what's going on underneath the hood is it's running that map, produced pattern across a range of different documents for you. 

And essentially getting summaries of each of the individual pieces. And then reducing that back into a single l l m response, which you'll then get back as a. summarized output. Uh, so we're just gonna see how long this takes here. We're currently at about half a minute. So let's actually just see what's come back now. So we've got, various marketing concepts as marketing segmentation, targeting planning, advertising. It's designed to help individuals effectively promote products and services. Cool. So it's very easy for you to, you know, do a summarization chain on any type of  document. and so essentially what we've got here is we've got this l l m, we have the chain type, and then we are also passing in, these pages, which is a list of documents. so now that you can sort of see that running in action, I think it's important for us to take a step back and say, well, maybe you want to do a summary in a different language, or, you know, maybe you've got something that you need to really specifically do at every stage of that map or reduce pattern. 

this is where you can create a map, reduce chain, which is a custom chain. And so what you'll see here is we've got this map pattern here that says, given the following pages of a marketing book, generate a summary in Spanish, and we've got the pages here. Uh, and then you say given the following summaries of pages of a marketing book, generate a high level description of the book in Spanish, we've set up a map prompt that will map over the pages. One at a time, we also set up a reduced prompt that will. 

go over the summaries, then we'll reduce those. We set up two L l m chains and then as well as that, we have a list of documents and combining that into a single string. we also have the map produce documents chain, which takes in the map and then associate all of the documents, the individual documents with pages. And we also have the combined document chain here. So just in case we need to. 

Combine these, uh, that's kind of how that's working underneath the hood. Uh, and then, so that's part of the reduced section when we're doing that combined. So that's a chain  to use to combine results of applying l l M chains to documents. We then set that up with a map, reduced chain and we feed it in the map. reduced documents, chain. 

And as well as that, you can also attach a text splitter. And then essentially you just passed in your input text. So in this scenario here, you can see, we call our map produce chain, which is called map produce. And then we pass in our input text. And what we've done is I've looked at the top. hundred documents to start with. And then what we've done is extracted the page content, joined those altogether and ran that through a map produced pattern. But notice now that the actual text that you're getting back is not necessarily an English summarization across all of these first hundred documents, actually a Spanish. translation. So, you know, being able to specifically. change what happens at the map section of your summarization chains or even just what's happening when all the documents are getting stuffed in to get reduced. can give you the flexibility to really optimize your summarization techniques.

What is the purpose of the recursive text splitter in LangChain?

What is the purpose of the map reduce chain?

Did you know it's possible to easily extract key insights from voluminous text, transforming an entire book into a digestible summary with LangChain?

Summarizing Large Documents

Charlie Cook, Vexnomics
It’s amazing, they can summarize whole books
They way they do it, is by breaking it down into little chunks
Then summarizing each chunk
Then summarizing the summaries
Pretty elegant actually
I wonder if we can use this to summarize some of our topics 

AI tools like GPT-3 have strict token or character limits to the text you can input, because the models are resource intensive to run. Of course GPT-3 is already good enough to not need a lot of context to complete tasks, but sometimes the limit is overly restrictive. For example if the task involved is to summarize a large block of text, say a book or anything over 1,500 words, you’re likely to run into their 4,000 token limit (prompt + completion). This can be a problem because many of the tasks we want to use GPT-3 for require multiple lines of text, or multiple examples fed to the AI so it knows what to do. The costs are negligible relative to the value of what you can create, so many people were interested in a workaround, where they could tradeoff more computing cost for longer text inputs.
To find a workaround, the OpenAI team experimented with progressive summarization, outlined in a paper they published (though they didn’t release the code or model publicly). The way it worked was quite elegant: break the large text into smaller chunks or sections, then summarize each section of text, before finally summarizing the summaries. The end result gives us a surprisingly good summary of all the text underneath. It still needs to be checked by a human for accuracy and small edits, but it saves a significant amount of time by letting you dig into interesting texts based on their summaries, rather than having to waste time reading the whole thing first. Editing is always a much cheaper and more efficient task than writing.

Progressive Summarization

What is the rough character limit of GPT-3?

Does GPT-3 always return the same response for the same input?

Why is it valuable to spend the time coding a GPT-3 summarization script when you could just summarize these articles manually?

What does the ‘temperature’ parameter control when using GPT-3?

You can read and summarize an article. No problem. It’ll take you 10 minutes. But how do you scale that to 1,000? You won’t believe me until you see it, but GPT-3 is not just ‘good enough’ at doing this, but the AI is actually *reliably better* than humans at the task!

Can we summarize at scale?

VEXNOMICS OFFICE – CONTENT SPRINT PLANNING MEETING

You've been tasked with scaling out the social media content. The social team are looking for innovative ways to improve their content velocity.

Hey, I'm looking for a way to re-write my Twitter content at scale. 
I heard GPT-3 is a great option. 
Can you tell me more about it?

Sometimes the process of re-writing competitor content is called 'Meme Jacking' - https://marketingmemetics.com/essays/meme-jacking

So what are the benefits of re-writing content using code?

1. Increased SEO visibility: Rewriting competitor content using GPT3 allows you to generate unique versions of the same content, meaning your website will be more likely to show up in search engine results for certain keywords and phrases.

2. Better customer engagement: By re-writing competitor content, you can update the content to better suit the needs of your customers. This can help keep them engaged and on your website for longer.

3. More cost-effective content production: Rewriting competitor content with GPT3 is far more cost-effective than creating new content from scratch. Therefore, it’s often a great way to increase the amount of content on your website without spending too much money.

4. Increased automation: GPT3 makes it easier to automate the process of rewriting competitor content at scale, meaning you can save time and resources that would otherwise be spent manually re-writing the same content over and over again.

Re-writing Content with GPT-3

0:00 All right. So we've been tasked with scaling out our social media content. A couple of ways that we could do this one in particular is we could automate some of our social media content by looking at what competitors have been doing and figuring out is this a personal story or is it not a personal story?
0:17 And if it isn't a personal story, then we could potentially rewrite that LinkedIn post or that Twitter post. So we've written a script to do this for you.
0:27 But what I'd like to do is run through it with you, but line by line. So hopefully it makes sense.
0:31 So at the top, we have a couple of Pippin installs with a package we'll be using called twk that allows us to interface with Twitter.
0:40 We've also got open AI that will be used for the GPT three and also a datetime package and Panders.  And then you'll see we've got some imports as well as some bits that you'll need to fill in.
0:52 So the open ai dot API key, you'll find in your open AI account. And these keys here, all the way from consumer key down to access token secret.
1:02 Those those you'll be able to get through setting up a Twitter app at the Twitter development. And you'll need to get these keys here, these API key and secret and the bar token and the access token secret and just click regenerate and get those.
1:16 And once you've got all of those, what you can then do is have a look at this code. So we're authenticating with Twitter.
1:24 We're also waiting on rate limits, which means if you do run into weight limits, then the API is gonna go to sleep for a couple of minutes.
1:32 And then you'll be able to then re-query inside the same function, which means you won't get errors if you do a large amount of querying.
1:40 So let's have a look at what it looks like if we're just querying for Python and getting a maximum count of 500 and the length of that.
1:49 And then you can see here we are able to essentially get a range of different tweets out you know, from the Python community.
1:58 So this is kind of like one thing that you could do. But what's quite interesting from our point of view is we've written some automatic Python functions to help you automate this.
2:08 So this one here, scrape tweets takes in a couple of different parameters. So like it takes in the Tweety object, takes in a keyword, which is a string and also it takes in the number of days which is an integer and also the number of pages, which is a integer.
2:24 And then what we're doing here is basically just finding how many number of days do you wanna look back, and how many pages do you want to see?
2:32 And notice here there's a count. So we're cursing through a hundred tweets at a time with a one second sleep.
2:38 And then we can go through any number of those pages and any number of those days. And then if we hit a time where we have gone back to the previous time then we break out of the for loop and then return the list.
2:50 So at some point we will break out of this loop if we go far back enough. So, and then the other thing that we've got is the ability to scrape user tweets.
3:01 So if we know a user's timeline then we can essentially go and get a bunch of their tweets based on chunks of a hundred and the number of pages that we want.
3:12 And however many pages you want going backwards in time. Cool. The other functions that I've written for us at the moment, you can see we've got this generate related keywords and we do a bit of text string manipulation to kind of clean up the data and return a Python list.
3:29 And so the prompt there is kind of like creating an order list sort of around up to 50 similar key keywords for seo given a a primary keyword.
3:39 And then basically what this does is it allows us to generate more keywords so that we could gather more different types of tweets to find more users, which you could, which we could then feed into this.
3:51 So the aim is we're kind of expanding out with a keyword research. So maybe start with, you know, one or two keywords and then you get a load of related keywords, and then you could then start branching out and finding all of the tweets and all of the users in a kind of semantically like relevant way.
4:08 This other prompt for GPT three is is it a personal story? So act as a social media strategic responsible for re reviewing competitor tweets with the sole aim to determine whether the capacity suite includes a personal story.
4:23 And then we do some simple classification and look inside that to see whether or not it is or not. And we've also got a separate prompt for rewriting the social media text.
4:34 So the first thing we're gonna do is we've had a look at like getting some tweets using just the, the native api with just Python.
4:42 So let's generate some recommended keywords first. So I'm gonna, I'm gonna comment out this code and we can just have a look at what this returns when I've when the, when the function gets back from GPT three.
4:58 And let's have a look at that. And you'll see that we've got a range of different types of words similar to Python, so like Python web framework, python data science.
5:07 So all of those could act as keywords that we may or may want to use. And then I'm not gonna run this piece of code cause it can take about a certain number of time, depending upon the number of days and the number of pages, parameters that you decide to choose.
5:21 But essentially what you're looking to get back is a bunch of tweets. And that should be a bunch of tweets for all of the different types of keywords that you've got in your recommendations.
5:30 At this time, I've only limited to five just for the sake of the tutorial. And then what we can then do is create a data frame based off that.
5:38 And then we've got a data frame of users and tweets. And you've got like the follow account frame count, listed count, et cetera.
5:44 And then what we want to do is we want to drop duplicates by the screen name. And then maybe, I don't know, maybe we wanna look at the ones who are actually quite popular and then maybe the look, look at the top 10 of those.
5:54 And now we've got a data frame of the top 10 of those with their screen names based on like their, you know, their follow account.
6:03 And, and so there we are. So cool. And then what we're gonna also do is then for all of these users, we're gonna scrape their tweets.
6:13 And I just wanna get 10 pages up to 10 pages of a hundred tweets for each user. And then what you'll end up with is we take that and save it into the user's tweet list.
6:22 And then we get all of those metrics out there. And then you'll see we get something called a user's tweets data frame.
6:29 And then inside of that we've then got a bunch of tweets against user IDs against text. And we've also got how popular it was in terms of was it retweeted, what's the favorite count and what's the retweet count as well.
6:42 And which is quite useful. So now we're gonna sort by the favorite count and you'll see if I now look at that data frame again, it completely changes and we start to see maybe some stuff of emojis, maybe some stuff that's a bit more interesting.
6:54 And so this is sort of by favorite count. And then what we can then do is, I've just limited it to the top 10 rows, but you could obviously do this for your entire data frame.
7:04 Is we then basically set a column called, is personal story called a false. And what we're doing here is we're looping over every row in the data frame, and we are then just updating that row ID with the T function to, to see whether it is a personal story or not.
7:21 And so if you have a look here, you'll see and I'll just show you this, that you can see that all of mine u users tweets, df.
7:30 So let's actually go and run this quickly. Cuz I, I actually just overwrite my data frame. So we actually do need to run this now.
7:38 So now it's going through and checking whether or not these are personal stories. And so if we now look at that, you'll see it doesn't believe that any of these tweets are personal stories, which means we could potentially rewrite all of these tweets and they probably still require some visual inspection from a human.
7:59 But the idea being is that we're just trying to, you know, throw out the ones that are like, you know, I went for a coffee last week with my friend and it was great kind of tweets.
8:09 But obviously you probably still have to look at them after they've been made. And then what we're doing is filtering out any of the ones that are a personal story.
8:18 So we want to say if the personal story is equal to false, then keep those ones. And if there aren't any of those, then we need to just bail on this point because we don't have any nonpersonal stories.
8:31 And then what we then do is we create a new column called Rewritten tweet. And then essentially we go through that data frame over again and essentially just write the text.
8:42 And so this one in action is then going through and rewriting each of these tweets to be unique and you know, completely unique.
8:51 And so this is something you could potentially do where, you know, a load of keywords or you know, a load of competitors.
8:58 And what you want to do is just put in, you know, a load of a load of their tweets and go, okay, are these personal stories?
9:05 Are these personal stories? Are these personal stories? And then the ones that aren't personal stories, we are then basically you know, creating new tweets from that.
9:15 And then after that you could give that to a low skilled worker potentially in, you know, in, you know, or, or maybe like a mid, mid-level worker who could then review those tweets and really clean out the junk and maybe even edit some of the tweets to make them make them viable.
9:31 And so this is something you would do as well. So I think the, the process doesn't stop here. We still need to check the, like, the validity of these tweets, but this definitely does give us the ability to scale and get to the point where we've, we've got a lot more of the tweets coming through.
9:45 So hopefully this was useful. And if you let need any help or, or anything, just let us know.

What are the benefits of re-writing existing content?

Why should you setup competitor tracking scripts on social media?

Scaling social media content can be a challenge, but it is a necessary part of any successful digital marketing strategy. Content scaling involves creating a consistent flow of social media posts across multiple platforms in order to reach a larger audience.

How can we scale our social media content?

We're in the Upsert headquarters the day before the end of the quarter and Sally Valentine, the Head of Marketing, has tasked us with finding a way to run a prompt locally instead of using OpenAI. She's looking for a way to test open source models, measure latency and response, and get reliable results at a lower cost.

There's no doubt that OpenAI's models are state-of-the-art, but you can quickly incur a lot of cost, be plagued by latency or reliability issues, and not every organization wants to send their data to OpenAI.

Vexpower Learning Certificate of Completion

Skill Path: Prompt Engineer