[0:00] Learn how to get chat GPT and other LLMs
[0:02] to give you the perfect responses
[0:04] by mastering prompt engineering strategies.
[0:07] Anu Kubo is one of our most popular instructors.
[0:10] And in this course, she will teach you the latest techniques
[0:13] to maximize your productivity with large language models.
[0:17] Hi everyone and welcome to this course on prompt engineering.
[0:21] My name is Anu Kubo and I'm a software developer
[0:24] as well as course cruiser here on FreeCoCamp
[0:26] as well as on my own channel.
[0:28] This is going to be a unique course for me
[0:30] as there is going to be a lot less coding going on
[0:33] but a lot more understanding
[0:35] about the topic of prompt engineering
[0:37] and why some companies are paying up to $335,000 a year
[0:42] according to Bloomberg for people in this profession.
[0:44] And no, no coding background is necessarily required.
[0:48] So what are we waiting for?
[0:50] Let's do it.
[0:51] In this course, we will learn what prompt engineering
[0:54] is exactly, get a brief introduction to AI,
[0:58] a look at large language models or LLMs
[1:01] such as chat GPT, a look at text to image models
[1:04] such as mid journey, a look at emerging models.
[1:07] This would include text to speech,
[1:09] text to audio or speech to text
[1:11] as well as prompt engineering mindset, best practices,
[1:15] zero-shot prompting, few-shot prompting, chain of thought,
[1:19] AI hallucinations, vexes, text embeddings
[1:22] and also end with a quick intro to chat GPT.
[1:26] So let's start off with looking at what exactly
[1:29] prompt engineering is in the first place.
[1:31] Prompt engineering in a nutshell is a career
[1:34] that came about of the back of the rise
[1:36] of artificial intelligence.
[1:38] It involves human writing, refining and optimizing prompts
[1:41] in a structured way.
[1:43] This is done with the intention of perfecting the interaction
[1:46] between humans and AI to the highest degree possible.
[1:49] Not only that, however, a prompt engineer is then also
[1:53] required to continuously monitor those prompts
[1:55] ensure their effectiveness with time as AI progresses.
[1:59] Maintaining an up-to-date prompt library will be a requirement
[2:02] that is placed onto the prompt engineer
[2:04] as well as reporting on findings and in general,
[2:07] being a thought leader in this space.
[2:09] But why do we need it?
[2:11] And how did it come about from AI?
[2:13] Before moving on, let's actually make sure
[2:15] we are on the same page about what exactly AI is.
[2:19] Artificial intelligence is the simulation
[2:22] of human intelligence processes by machines.
[2:25] I say simulation as artificial intelligence
[2:28] is not sentient, at least not yet anyways,
[2:31] meaning it cannot think for itself
[2:33] as much as it may seem it does.
[2:36] Often, and this is certainly the case with tools
[2:38] such as chat GPT, for example, when we say AI,
[2:43] we are simply referring to a term called machine learning.
[2:46] Machine learning works by using large amounts
[2:48] of training data that is then analyzed
[2:51] for correlations and patterns.
[2:53] These patterns are then used to predict outcomes
[2:56] based on the training data provided.
[2:58] So for example, here we are feeding data
[3:01] saying that if a paragraph looks like this
[3:03] with this type of title, then it should be categorized
[3:07] as international finance.
[3:09] The second paragraph should be put in the category
[3:11] of earning reports and so on.
[3:14] With some code, we should be able to train our AI model
[3:18] to correctly guess what future paragraphs are about.
[3:21] And that's it.
[3:22] Of course, that is a super basic example
[3:25] and we would need way more data
[3:27] than these five paragraphs right here,
[3:29] but you get the idea.
[3:30] If you want to build your own AI model
[3:32] and understand the concept of machine learning better
[3:34] as a total beginner, please check out my video on it
[3:37] on my channel, Code with Anya Kubo.
[3:39] As of now, rapidly improving general AI techniques
[3:42] can create realistic text responses
[3:44] and even images, music, and other media
[3:47] thanks to the huge amounts of training data
[3:49] and talented developers working on it today.
[3:53] Why is prompt engineering useful?
[3:56] With the quick and exponential growing rise of AI,
[3:59] even the architects of it themselves struggle
[4:01] to control it and its outputs.
[4:04] This might be a bit hard to understand,
[4:06] but think of it this way.
[4:07] If you were to ask an AI chatbot what is four plus four,
[4:11] you would expect it to say eight, right?
[4:13] The result of eight is pretty indisputable.
[4:16] However, imagine you are a young student
[4:19] trying to learn the English language.
[4:21] I'm going to show you just how different responses can be
[4:24] based on the prompts you feed
[4:27] and in turn your learning experience.
[4:29] For this example, I'm going to be using
[4:31] Chat GPT's GPT-4 model.
[4:34] So let's start with the basics.
[4:36] If you were to type, correct my paragraph
[4:39] and then paste it a badly written paragraph.
[4:43] So just like this, today was great in the world for me.
[4:45] I went to a Disneyland with my mom.
[4:47] It could have been better though if it wasn't raining.
[4:50] Great, the young English learner has a better sentence,
[4:53] but it kind of stops there
[4:55] and the learner is just left to their own devices.
[4:57] And honestly, the sentence really isn't that great anyway.
[5:02] What if the learner could get the best sentences possible
[5:05] from a teacher who understands their interests
[5:08] to keep them engaged?
[5:10] With the correct prompts,
[5:11] we can actually create that with AI.
[5:14] So let's give it a go and let's write a prompt to do this.
[5:17] So here's the prompt I'm going to give it.
[5:20] I'm going to write,
[5:21] I want you to act as a spoken English teacher.
[5:24] I will speak to you in English
[5:25] and you'll reply to me in English
[5:27] to practice my spoken English.
[5:30] I want you to keep my reply neat,
[5:32] limiting the reply to 100 words.
[5:35] I also want you to strictly correct
[5:37] my grammar mistakes and typos.
[5:39] And I want you to ask me a question in your reply.
[5:43] Now, let's start practicing.
[5:45] You could ask me a question first.
[5:48] Remember, I want you to strictly correct
[5:50] my grammar mistakes and typos.
[5:52] So there's my prompt.
[5:54] You can also go a step further
[5:56] and ask it to correct your factual errors too,
[5:59] which I think would be an excellent addition to the prompt
[6:01] that will benefit the young learner.
[6:03] Okay, so let's go ahead and add that to the prompt.
[6:06] Okay, and let us do its thing.
[6:09] And great, so now this is way more interactive.
[6:13] As you will see, it's asking you a question
[6:15] and telling you what to do
[6:17] and will provide you with corrections if needed.
[6:19] So in a way, you're communicating with the AI.
[6:23] It's giving you suggestions
[6:24] and you're learning along the way.
[6:26] It's a completely different experience
[6:28] thanks to the prompt that we wrote.
[6:30] Pretty cool, right?
[6:31] We're gonna be diving into a bunch of these concepts soon,
[6:34] but first let's start with the basics.
[6:37] Linguistics.
[6:38] Linguistics is the study of language.
[6:41] It focuses on everything from phonetics,
[6:44] so the study of how speech sounds
[6:46] are produced and perceived.
[6:47] Phonology, so the study of sound patterns and changes.
[6:51] Morphology, the study of word structure.
[6:54] Syntax, so the study of sentence structure.
[6:57] Semantics, so the study of linguistic meaning.
[7:01] Pragmatics, so in other words,
[7:02] the study of how language is used in context.
[7:05] Historical linguistics, or the study of language change.
[7:09] Sociolinguistics, or in other words,
[7:12] the study of the relation between language and society.
[7:16] Computational linguistics,
[7:18] so the study of how computers can process human language.
[7:21] And physiolinguistics,
[7:23] or the study of how humans acquire and use language.
[7:27] So a lot.
[7:29] Linguistics are the key to prompt engineering.
[7:32] Why?
[7:33] Understanding the nuances of language
[7:35] and how it is used in different contexts
[7:37] is crucial for crafting effective prompts.
[7:40] Not only that,
[7:41] but knowing how to use a grammar or language structure
[7:44] that is universally used
[7:45] will result in the AI system
[7:47] returning back the most accurate results.
[7:50] As you can imagine,
[7:51] the sheer amount of data that it is trained on
[7:54] is most likely to mostly use the standard grammar
[7:58] and language structure that is universally used.
[8:00] So sticking to the standardization is key.
[8:04] Language models.
[8:06] Imagine a world where computers possess the power
[8:08] to understand and generate human language.
[8:11] A world where machines can chat, write stories,
[8:14] and even compose poetry.
[8:16] In this magical realm,
[8:18] language models can come into play.
[8:21] They are like the wizards of the digital realm
[8:23] capable of understanding and creating human-like text.
[8:27] A language model is a clever computer program
[8:31] that learns from a vast collection of written text.
[8:34] It takes in books, articles, websites,
[8:36] and all sorts of written resources,
[8:38] allowing it to gather knowledge
[8:40] about how humans use language.
[8:42] Just like a master linguist,
[8:44] it becomes an expert in the art of conversation,
[8:46] grammar, and style.
[8:48] But how does this all work?
[8:50] Well, imagine you feed a sentence.
[8:53] The language model will then analyze the sentence,
[8:55] examine the order of words, their meanings,
[8:58] and the way they fit together.
[8:59] Then the language model would generate a prediction
[9:03] or a continuation of the sentence
[9:05] that makes sense based on its understanding of the language.
[9:08] It will weave words together one by one,
[9:11] creating a response that seems like it was crafted
[9:13] by a human being.
[9:15] Imagine having a conversation with the language model
[9:18] as if you were exchanging ideas with a digital friend.
[9:20] You ask a question,
[9:22] and it responds with a well-crafted answer.
[9:24] You tell a joke, and it counters with a witty remark.
[9:27] It's like having a language expert by your side
[9:29] always ready to assist and engage in conversation.
[9:33] Now, you may wonder where these language models are used.
[9:36] They can be found in various places,
[9:39] from your smartphone's virtual assistants
[9:41] to customer service chatbots,
[9:43] and even in the creative realm of writing.
[9:46] They help us find information,
[9:48] offer suggestions, and create content.
[9:50] But remember, while language models
[9:52] possess incredible abilities,
[9:54] they still rely on the human to create and train them.
[9:58] They are, in fact, a fusion of human ingenuity
[10:00] and the power of algorithms,
[10:02] blending the best of both worlds.
[10:05] Let's start off by looking at the history of language models,
[10:08] starting with the first AI, Eliza, back in the 60s.
[10:12] Eliza is an early natural language processing computer
[10:15] program created from 1964 to 1966 at MIT by Joseph Wiesenbaum.
[10:22] Eliza was designed to simulate a conversation
[10:25] with a human being.
[10:26] Eliza had a special knack for mimicking
[10:28] a Rogerian psychotherapist,
[10:30] someone who essentially listens attentively
[10:33] and asks probing questions
[10:35] to help people explore their thoughts and feelings.
[10:37] Eliza's secret weapon was its mastery of pattern matching.
[10:42] It had a treasure trove of predefined patterns,
[10:44] each associated with specific responses.
[10:47] These patterns were like magical spells
[10:50] that allowed Eliza to understand
[10:51] and respond to human language.
[10:54] When you engage in a conversation with Eliza,
[10:56] it would carefully analyze your input,
[10:58] seeking patterns and keywords.
[11:00] It would then transform your words
[11:01] into a series of symbols,
[11:03] searching for patterns that match those symbols
[11:06] in its repertoire.
[11:07] Once a pattern was detected,
[11:09] Eliza would work its magic by transforming your words
[11:12] into a question or statement
[11:14] that aimed to explore your thoughts and emotions.
[11:17] It was as if Eliza was holding up a metaphorical mirror,
[11:20] encouraging you to delve deeper into your own thinking.
[11:24] For example, if you said something like, I feel sad,
[11:27] Eliza would detect the pattern
[11:29] and respond with a question like,
[11:31] why do you think you feel sad?
[11:33] It encouraged reflection and introspection,
[11:36] just like a caring therapist would.
[11:39] But here's the delightful twist.
[11:41] Eliza didn't truly understand what you were saying.
[11:44] It was just a clever illusion.
[11:46] It used pattern matching
[11:48] and some creative programming tricks
[11:50] to create the illusion of understanding
[11:53] while in reality,
[11:54] it was just following a set of predefined rules.
[11:57] Yet, even though Eliza was a simple program,
[12:00] people were often captivated
[12:02] by its conversational abilities.
[12:04] They felt heard and understood,
[12:06] even if they knew they were talking to a machine.
[12:09] It was like having a digital confident
[12:11] who was always ready to listen and offer gentle guidance.
[12:15] Eliza's creator, Weisenbaum,
[12:17] intended the program as a method to explore communication
[12:20] between humans and machines.
[12:22] He was surprised and shocked that individuals,
[12:24] including his secretary,
[12:26] attributed human-like feelings to the computer program.
[12:29] But that is a whole other topic of conversation in itself.
[12:34] Eliza's impact was profound,
[12:35] sparking interest and research
[12:37] in the field of natural language processing.
[12:39] It paved the way for more advanced systems
[12:41] that could truly understand and generate human language.
[12:45] It was the humble beginning of a grand adventure
[12:48] in the world of conversational AI.
[12:51] Fast forward to the 1970s,
[12:52] when a program named Shudlu appeared.
[12:56] It could understand simple commands
[12:57] and interact with a virtual world of blocks.
[13:00] Although Shudlu wasn't a language model per se,
[13:04] it laid the foundation for the idea of machines
[13:06] comprehending human language.
[13:08] But the true language models began around 2010,
[13:12] when the power of deep learning
[13:13] and neural networks came into play.
[13:16] Enter the mighty GPT,
[13:19] short for generative pre-trained transformer,
[13:22] ready to conquer the world of language.
[13:24] In the year 2018,
[13:26] the fast iteration of GPT emerged,
[13:28] created by the company OpenAI.
[13:31] It was trained on a large amount of text data,
[13:34] absorbing knowledge from books, articles,
[13:36] and a large chunk of the internet.
[13:38] GPT-1 was a taste of things to come,
[13:41] an impressive language model,
[13:42] but small compared to its descendants that we use today.
[13:46] As time went on,
[13:47] the saga continued with the arrival of GPT-2 in 2019,
[13:51] followed by GPT-3 in 2020.
[13:54] This was a titan among language models
[13:57] equipped with a large number of parameters,
[13:59] over 175 billion to be precise.
[14:02] GPT-3 dazzled the world
[14:04] with its unparalleled ability to understand,
[14:06] respond, and even generate creative pieces of writing.
[14:09] The arrival of GPT-3 marked a real turning point
[14:12] in terms of language models and AI.
[14:15] At the time of writing,
[14:16] we now also have GPT-4,
[14:18] trained on pretty much the whole internet,
[14:20] rather than outdated large data sets,
[14:22] as well as BERT from Google and so much more.
[14:25] It would seem we are only just at the start
[14:27] when it comes to language models and AI.
[14:30] So learning how to harness this data with prompt engineering
[14:32] is a smart move for anyone today.
[14:36] The prompt engineering mindset.
[14:38] When thinking of good prompts,
[14:39] it is always best to get in the correct mindset.
[14:42] Essentially, you want to just write one prompt, right?
[14:45] And not have to waste time and tokens,
[14:48] writing lots of different prompts
[14:50] until you get the result you desire.
[14:52] So essentially, kind of the same
[14:54] as when you Google stuff, right?
[14:56] How good are your Googling skills now,
[14:58] as opposed to five years ago?
[15:00] I'm assuming a lot better.
[15:02] We have grown to intuitively know
[15:04] what to type into Google the first time round
[15:07] as to not waste time.
[15:09] Having the same mindset for prompt engineering
[15:11] can also be applied.
[15:13] Mahail Eric of the Infinite Machine Learning Podcast
[15:16] says it well when he says,
[15:18] I personally like the analogy of prompting
[15:20] to designing effective Google searches.
[15:23] There are clearly better and worse ways to write queries
[15:26] against the Google search engine that solve your task.
[15:29] This variance exists because of the opaqueness
[15:32] of what Google is doing under the hood.
[15:35] We are gonna keep this in mind
[15:36] for the remainder of the course.
[15:39] A quick intro to using chat GPT by OpenAI.
[15:43] So as I said in this course for the examples,
[15:46] I will be using chat GPT's GPT for model.
[15:51] In order to follow along or just to understand
[15:55] how we are going to be using the platform,
[15:58] please head over to openai.com
[16:00] and just go ahead and sign up.
[16:03] I've already signed up,
[16:04] so I'm just gonna go ahead and log in
[16:07] and that will take me to the platform
[16:09] in which I can choose what I want to interact with.
[16:13] For this tutorial,
[16:15] we are going to be interacting with chat GPT for.
[16:19] So please go ahead and click on here
[16:22] and that will take you to the platform.
[16:24] And then I'm just gonna go ahead
[16:26] and switch to the GPT for model,
[16:29] which is the latest one.
[16:31] Okay, so great.
[16:33] You will see here all the previous chats that I've had.
[16:37] I'm just going to minimize this.
[16:39] And if we want to create a new chat,
[16:41] all we'd have to do is click the new chat button.
[16:44] Okay, so here, for example,
[16:47] what I can do is just go ahead and ask any questions.
[16:51] So what is four plus four?
[16:54] And then hit send.
[16:55] And that will essentially give me a response.
[16:58] So I'm now interacting with chat GPT for.
[17:02] On this occasion,
[17:03] I can actually build on the previous conversation.
[17:07] So what I can do is say, great.
[17:11] Now, can you add another five to that?
[17:19] What is the answer?
[17:21] And it will take into account everything
[17:24] that I have previously.
[17:25] Okay, I am building on top of the knowledge
[17:28] that it already has.
[17:30] Great.
[17:31] So again, this is just a very quick introduction
[17:35] to how to use this,
[17:36] we will be doing a deeper dive into this
[17:39] throughout this course.
[17:41] So once again, to create a new chat,
[17:43] all you'd have to do is create new chat.
[17:45] And if you want to delete the old one,
[17:46] just go ahead and click that delete button.
[17:48] And that will delete it.
[17:50] And then a new chat is brought up automatically.
[17:53] Wonderful.
[17:55] Now you might have seen my course on open AI
[17:58] in which we can also use the API.
[18:02] And to use the API,
[18:03] just go over to the API references.
[18:06] And all you would have to do is get an API key.
[18:10] And the API key can be viewed here.
[18:13] And then you just go ahead and create an API key.
[18:16] And that will allow you to interact with the open AI API
[18:21] in order for you to build out your own platforms
[18:24] just as the one we saw before.
[18:26] So if you're interested in that,
[18:27] please do head over to my tutorial on this,
[18:31] again, on the free COCAM channel.
[18:33] Otherwise, let's continue using chat GPT.
[18:38] So once again, I'm just going to log in
[18:40] and head back to the platform here.
[18:43] And that is it, we are ready to go.
[18:47] Another thing I want to discuss is tokens.
[18:51] As you might find that you have run out of the free tokens
[18:55] that you have in order to interact with chat GPT.
[18:58] So in order to do that, I'm just going to show you this,
[19:02] I'm going to head over to the documentation
[19:05] and talk to you a little bit about tokens.
[19:09] So GPT-4 essentially processes all texts
[19:13] in chunks called tokens.
[19:16] And this token is approximately four characters
[19:20] or 0.75 words for English text.
[19:24] And we essentially get charged by token.
[19:27] If you want to know exactly how many tokens you are using,
[19:31] you can check it out here,
[19:33] you can check out the tokenizer tool,
[19:35] and it will give you a rough example.
[19:37] So for example, I can say, what is four plus four.
[19:44] And with that piece of text,
[19:47] the total count of tokens is going to be six.
[19:51] Okay, so there we go.
[19:54] That is exactly how many tokens will be used
[19:56] in order to produce a answer for this request.
[20:02] Great, so here is a URL for that.
[20:04] You can play around with it.
[20:06] I hope you have fun.
[20:08] If you want to see your usage,
[20:09] you can head over to account
[20:11] and then you can manage your account.
[20:14] And this should be able to show you your usage.
[20:17] So it will show you your usage right here per month.
[20:21] And then you can, of course, also add billing
[20:24] in order to carry on using chat GPT
[20:28] in case you've used up all your tokens
[20:30] and you can't use it anymore.
[20:32] So once again, this is under account billing overview.
[20:36] So just go check it out.
[20:38] Okay, and now back to the platform.
[20:42] Best practices.
[20:44] The biggest misconception
[20:46] when it comes to prompt engineering
[20:48] is that it's an easy job with no science to it.
[20:51] I imagine a lot of people think
[20:53] it's just about constructing a one-off sentence
[20:56] such as correct my paragraph
[20:58] that we saw in the previous example.
[21:00] When you start to look at it,
[21:01] creating effective prompts
[21:03] relies on a bunch of different factors.
[21:06] Here are some things to consider
[21:08] when writing a good prompt.
[21:10] Consider writing clear instructions
[21:12] with details in your query.
[21:14] Consider adopting a persona
[21:16] as well as specifying the format
[21:19] using iterative prompting,
[21:21] meaning if you have a multi-part question
[21:24] or if the first response wasn't sufficient,
[21:26] you can continue by asking follow-up questions
[21:28] or asking the model to elaborate
[21:31] and avoid leading the answer.
[21:33] Try not to make your prompts so leading
[21:35] that it inadvertently tells the model
[21:37] what answer you're expecting.
[21:39] This might bias the response unduly.
[21:42] And finally, limit the scope for long topics.
[21:45] If you're asking about a broad topic,
[21:47] it's helpful to break it down
[21:48] or limit the scope to get a more focused answer.
[21:51] Let's look at some of these now.
[21:53] In order to write clearer instructions,
[21:55] we can adopt writing more details in our queries.
[21:59] And to get the best results,
[22:00] don't assume the AI knows what you are talking about.
[22:04] Writing something like, when is the election,
[22:07] implies that you are expecting the AI to know
[22:10] what election you are talking about
[22:12] and what country you mean.
[22:14] This may result in you asking a few follow-up questions
[22:17] to finally get the result you want,
[22:19] resulting in time loss and frankly,
[22:21] perhaps some frustration.
[22:23] Consider taking the time to write a prompt
[22:25] with clear instructions.
[22:27] So, for example, instead of writing,
[22:30] when is the election, you could write,
[22:32] when is the next presidential election for Poland?
[22:36] So let's go ahead and run this.
[22:39] And this will be much more precise
[22:41] and knows exactly what we are asking about.
[22:44] It's not gonna go guessing and waste our time
[22:48] as well as waste our resources.
[22:50] So in other words, tokens that we are using,
[22:52] so in other words, money,
[22:54] in order to get the right answer the first time.
[22:57] Here are some other examples
[22:59] of how you could write clearer prompts.
[23:01] So for example, we have this prompt here,
[23:03] which says write code to filter out the ages from data.
[23:07] And if you run it,
[23:09] you don't really know what language
[23:10] it's gonna come back with, let's see.
[23:12] So for example, here it's using Python.
[23:15] I actually didn't wanna use Python, okay?
[23:18] So now we've lost some tokens asking this.
[23:21] We've also lost some time
[23:23] and we just haven't got the right response.
[23:25] This could have been so easily avoided.
[23:28] So I'm gonna stop this from generating
[23:30] and let's try again.
[23:32] So this time, let's be more specific by writing,
[23:36] write a JavaScript function
[23:38] that will take an array of objects
[23:40] and filter out the value of age property
[23:42] and put them in a new array.
[23:44] Please explain what each code snippet does.
[23:47] In this example, I am not assuming the AI knows
[23:50] what computer language I like to use
[23:52] and I am being more specific
[23:54] about what my data actually looks like.
[23:56] On this occasion, it's an array of objects.
[23:59] Not only that, I'm also asking the AI
[24:02] to explain why it's doing each step
[24:05] so that I in turn can understand
[24:07] and not just copy paste the code
[24:08] without gaining any knowledge from it.
[24:11] Okay, so here you can see a live example
[24:13] of what's coming back to us from GPT-4.
[24:17] It's given us the correct code.
[24:20] So I have checked that
[24:21] and I was also giving us an example
[24:24] of how you would use the function,
[24:25] which is something I didn't ask,
[24:27] which is super useful.
[24:28] It's kind of gone above and beyond
[24:30] for helping me out in understanding what is going on.
[24:35] Let's look at another example.
[24:37] We can write, tell me what this essay is about.
[24:41] So I'm going to just type this
[24:44] and then just paste in an essay and hit go.
[24:46] And then chat GPT will do its thing.
[24:49] It's going to give me a summarization
[24:51] as it thinks best.
[24:53] So on this occasion,
[24:55] it is essentially giving me numbered points
[24:57] about what this essay is about.
[25:00] They're really long.
[25:01] I really didn't want to read this much.
[25:03] It's pretty much looking to be the same
[25:06] as the original essay.
[25:08] So this is not something I wanted.
[25:11] I should have been way more specific
[25:13] in telling it what I need.
[25:16] So what I'm going to do is just add to this conversation.
[25:19] So it's going to learn on what I wrote previously.
[25:22] And I'm just going to specify to use bullet points
[25:24] to explain what this essay is about,
[25:27] making sure each point is no longer the 10 words long.
[25:30] So I am being super specific
[25:33] in providing the instructions of what I want
[25:37] and let's hit go.
[25:39] So now this is a lot shorter.
[25:40] Okay, as you can see,
[25:42] each point is no longer than 10 words long.
[25:44] And then I'm going to get a summary
[25:47] that is a bit longer
[25:48] and will give me a summary of the essay
[25:51] that I pasted in above.
[25:54] So great, of course you can do this
[25:57] on any essay or piece of text that you wish
[26:01] and you can set your own clear
[26:03] and specific instructions as well.
[26:07] Okay, so I hope you can see
[26:08] why the second one is better.
[26:11] It's because I am not assuming the AI knows
[26:13] what kind of format I want the summarization
[26:15] of the essay to be in.
[26:17] I am being specific that I want very short notes
[26:20] on the essay in bullet point format
[26:22] with a short conclusion at the end.
[26:25] If I did not put this,
[26:26] the summary could have been just as long as the essay itself
[26:30] and the prompt could be considered useless in my eyes.
[26:34] Next up, we can also adopt a persona.
[26:37] When writing prompts, it is sometimes helpful
[26:39] to create a persona.
[26:40] This means you're asking the AI to respond to you
[26:43] and a certain character.
[26:45] So exactly like the English language teacher example
[26:47] we saw earlier, using a persona in prompt engineering
[26:51] can help ensure that the language model's output
[26:53] is relevant, useful and consistent
[26:56] with the needs and preferences of the target audience,
[26:59] making it a powerful tool
[27:01] for developing effective language models
[27:03] that meet the needs of users.
[27:05] Let's look at some examples of adopting a persona.
[27:08] So for example, you're gonna have this prompt right here.
[27:12] Write a poem for a sister's high school graduation
[27:15] that will be read out to a family and close friends.
[27:19] So let's go ahead and run this
[27:21] and let's see what comes back.
[27:24] So there we go.
[27:25] It is quite good in a room filled with kin and close ties.
[27:29] We gathered to honor the mist in our eyes
[27:31] for a journey has ended another begun
[27:34] as our dear sister steps into the sun.
[27:36] Okay, so it's coming out with a poem I can see here.
[27:41] It is quite a good one, I guess,
[27:43] probably better than anything that I would have written.
[27:46] And it is maybe a little bit generic.
[27:50] Maybe that's what you wanted.
[27:52] I think we can do better than this.
[27:54] So let's try this.
[27:56] This time I'm gonna write a prompt with a persona.
[28:00] So this time I'm gonna specify who I'm writing as.
[28:03] I'm gonna do write a poem as Helena.
[28:09] Helena is 25 years old and an amazing writer.
[28:18] Her writing style is similar to famous
[28:24] 21st century poet Rupi Kaur.
[28:29] Writing as Helena, write a poem for her 18 year old sister
[28:44] to celebrate her sister's high school graduation.
[28:49] High school graduation, this will be read out to friends
[29:00] and family at the gathering.
[29:06] Okay, so let's check it out now.
[29:08] We're writing as Helena, she's 25, she's an amazing writer
[29:14] and we've also assigned a writing style.
[29:17] So let's check it out.
[29:19] Now Chat GPT should be using anything it knows about
[29:23] Rupi Kaur, hopefully from the internet,
[29:26] in order to apply that style to this poem.
[29:29] Okay, and as you can see, this is maybe
[29:32] a little bit more affectionate.
[29:34] We've said sister, it's obviously a younger sister,
[29:37] so the words little sister are being used.
[29:39] And in general, I think this is a much higher quality poem
[29:45] and if Helena truly does have the style of Rupi Kaur
[29:48] in writing, it will be almost indistinguishable
[29:52] who wrote this poem, Chat GPT or Helena.
[29:56] So here we go, here's the full thing.
[29:58] It starts off, in the garden of our youth,
[30:00] I watch you bloom from bud to blossom,
[30:03] from child to woman, 18 summers passed.
[30:06] Again, utilizing the fact that we fed it, she was 18
[30:10] and every winter's chill only made you stronger,
[30:13] a force of nature still.
[30:14] So yes, in my eyes, this poem is a lot more better
[30:18] it's much more refined, it's much more personal,
[30:21] thanks to the prompts that we wrote.
[30:24] We've already had a brief look at specifying format
[30:27] when we limited the word count of our bullet points
[30:30] in a previous example.
[30:33] So that was a great example, limiting words
[30:35] is one that I use often.
[30:37] However, we can do a bunch of other things,
[30:41] including specifying if something is a summary,
[30:44] a list or a detailed explanation.
[30:47] Heck, you can even create checklists.
[30:49] So I'm going to go ahead and show you how to do this.
[30:52] Here is an example prompt and I'm just going to run it.
[30:56] And there we go, we have created a checklist.
[31:00] So many things you can do in Chat GPT.
[31:05] Just make sure to specify the type of format you want
[31:10] and it should be able to do it.
[31:13] Okay great.
[31:14] Now that we looked at some best practices,
[31:16] let's move on to some more advanced topics
[31:19] in the prompt engineering.
[31:21] In this section, I'm going to talk about two types
[31:23] of prompting we can do, zero-shot prompting
[31:27] and few-shot prompting.
[31:28] Zero-shot prompting leverages a pre-trained model's
[31:31] understanding of words and concept relationships
[31:34] without further training.
[31:36] And few-shot prompting enhances the model
[31:38] with training examples via the prompt avoiding retraining.
[31:42] So essentially in the context of the GPT-4 model,
[31:46] we don't really need to do much.
[31:48] We are already using all of the data that it has
[31:50] in order to ask when is Christmas in America?
[31:53] So let's go ahead and do some zero-shot prompting.
[31:57] When is Christmas in America?
[32:03] And here, go.
[32:05] Okay, so it clearly already has the data for this.
[32:10] We don't need to add any examples or anything like that.
[32:13] So as you can see, zero-shot prompting refers
[32:17] to a way of querying models like GPT
[32:21] without any explicit training examples
[32:23] for the task at hand.
[32:25] In the context of machine learning and not just GPT,
[32:29] zero-shot typically means that a model performs a task
[32:33] without having seen any examples
[32:35] of that task during its training.
[32:38] Okay great.
[32:40] So now let's look at few-shot prompting.
[32:43] So once again, with zero-shot prompting,
[32:46] we gave our language model a prompt and got a response.
[32:49] But sometimes that's just not enough
[32:52] and we need a bit more training.
[32:54] So let's use few-shot prompting
[32:57] and level up our language model by showing a few examples
[33:00] of the tasks we want it to perform.
[33:02] So instead of zero examples, we give it a tiny bit of data.
[33:07] So let's think, what would GPT for not know?
[33:12] I guess it would not know my favorite types of food.
[33:16] So for example, let's check,
[33:18] what is Ania's favorite type of food?
[33:26] Plural okay.
[33:28] And I mean, it can guess,
[33:30] but no, it's just telling me that it doesn't know.
[33:34] So that is absolutely fine, let's stop generating.
[33:38] So now let's feed it in some example data.
[33:40] I'm gonna feed it in a tiny bit of data.
[33:43] Ania's favorite type of food includes,
[33:51] sorry about my English.
[33:54] Let's go with burgers, fries, I love fries, pizza,
[34:01] and hit enter, okay?
[34:03] So we are essentially giving chat GPT some information.
[34:09] So now if I type what restaurant should I take Ania
[34:17] to in Dubai this weekend and hit here,
[34:26] it should hopefully understand
[34:28] that my favorite types of foods are burgers, fries,
[34:31] and pizza, and given that find me some restaurants
[34:35] in Dubai that I would like to go to.
[34:38] So here are some, this has been updated as of September,
[34:42] 2021, but you know, these are pretty good ones.
[34:45] So I would totally go to these.
[34:48] This is a great example of few shop prompting
[34:51] in which it wouldn't have been able to answer this question
[34:55] if I hadn't had given it some example data
[34:58] or just trained the model a little bit more
[35:02] in order to get the response that I want.
[35:05] Great, AI hallucinations.
[35:09] Now we're delving into something
[35:10] you probably never thought you'd hear in an AI context
[35:13] and that's hallucinations.
[35:16] So what exactly are AI hallucinations?
[35:19] And no, they're not when your AI assistants
[35:22] start seeing unicorns and rainbows.
[35:24] It's actually a time that refers to the unusual outputs
[35:28] that AI models can produce when they misinterpret data.
[35:32] A prime example of this is Google's Deep Dream.
[35:36] You know, that project that turns pictures of your dog
[35:38] into a nightmarish blend of dog faces
[35:41] and well, more dog faces.
[35:44] Deep Dream is an experiment that visualizes the patterns
[35:47] learned by a neural network.
[35:49] It is built to over interpret and enhance the patterns
[35:52] it sees in an image, or in other words,
[35:55] fill in the gaps with images.
[35:58] But some of the time those gaps
[36:00] can be filled with the wrong thing.
[36:02] This is an example of an unusual output
[36:05] that AI models can produce when they misinterpret data.
[36:10] Now, why do these hallucinations happen, you ask?
[36:13] They're trained on a huge amount of data
[36:16] and they make sense of new data
[36:18] based on what they've seen before.
[36:20] Sometimes, however, they make connections that are,
[36:24] let's call it creative.
[36:26] And voila, an AI hallucination occurs.
[36:29] Despite their funny results,
[36:31] AI hallucinations aren't just entertaining.
[36:34] They're also quite enlightening.
[36:36] They show us how our AI models interpret
[36:38] and understand data.
[36:40] It's like a sneak peek into their thought processes.
[36:43] This example is actually an AI image hallucination,
[36:46] as I thought it would be a fun one to show.
[36:49] AI hallucinations can also happen with text models.
[36:52] An example of this is us asking a text model
[36:55] about a historical figure
[36:57] and the text model not having an answer
[36:59] and hallucinating one instead,
[37:00] resulting in an inaccurate response.
[37:04] Vectors and text embeddings.
[37:06] To finish off, I'm going to leave you
[37:08] with a slightly more complex subject.
[37:11] We're going to take a quick look
[37:12] at the topic of text embedding and vectors.
[37:15] In computer science,
[37:17] particularly in the realm of machine learning
[37:19] and natural language processing or NLP,
[37:22] text embedding is a popular technique
[37:24] to represent textual information in a format
[37:26] that can be easily processed by algorithms,
[37:29] especially deep learning models.
[37:31] In the context of prompt engineering,
[37:33] LLM embedding refers to representing prompts
[37:35] in a form that the model can understand and process.
[37:39] This involves converting the text prompt
[37:41] into a high dimensional vector
[37:43] that captures its semantic information.
[37:46] So essentially the word food is represented by this.
[37:49] If using the create embedding API from open AI.
[37:53] Okay, so you will see it's represented,
[37:56] the word was represented by an array
[37:58] of lots and lots of numbers.
[38:00] But why do this?
[38:02] Well, think about it this way.
[38:04] If you ask a computer to come back
[38:06] with a similar word to food,
[38:08] you wouldn't really expect it to come back
[38:10] with burger or pizza, right?
[38:12] That's what a human might do
[38:14] when thinking of similar words to food.
[38:16] A computer would more likely look
[38:18] at the word lexicographically,
[38:21] kind of like when you scroll through a dictionary
[38:23] and come back with foot, for example.
[38:26] This is kind of useless to us.
[38:28] We wanna capture a word's semantic meaning.
[38:31] So the meaning behind the word.
[38:34] Text embeddings do essentially that,
[38:36] thanks to the data captured in this super long array.
[38:40] Now, I can find words that are similar to food
[38:44] in a large corpus of text
[38:46] by comparing text embedding to text embedding
[38:49] and returning the most similar ones.
[38:51] So words such as burger instead of foot
[38:54] will be more similar.
[38:56] To create a text embedding of a word
[38:58] or even a whole sentence,
[39:00] check out the create embedding API here from OpenAI.
[39:05] So here's the URL that you should visit
[39:08] and to create your own text embedding,
[39:10] you're gonna have to create a post request
[39:13] to this endpoint.
[39:15] Okay, so that's what you can do.
[39:17] You can also take the code from here.
[39:20] So this is an example request written in Node.js
[39:25] and you would have to install the OpenAI package,
[39:30] take these two methods from it
[39:33] in order to pass through your OpenAI API key
[39:37] that I showed you how to get earlier.
[39:39] As a refresher, you should go here and view API keys
[39:42] to create your own API key.
[39:45] And once you pass that through,
[39:47] you can then use the configuration
[39:51] and pass it through here in order to use OpenAI
[39:55] and all of its wonderfulness,
[39:57] including the create embedding method
[40:00] in which you pass through an object.
[40:02] That object has to include the model
[40:05] and it also is gonna include the input.
[40:07] So in this case, it is the text
[40:10] that you want to turn into an embedding.
[40:13] So this is the code for doing that.
[40:16] And here is the response, which comes back as an object.
[40:21] And then you would use dot notation to get the data,
[40:24] go into the array in order to get the embedding.
[40:28] So this is gonna be the embedding that we saw earlier.
[40:31] It's an array made up of numbers
[40:34] and a lot of these numbers, okay?
[40:37] So not just these three,
[40:39] this does spill out
[40:40] and this will be a very, very long array of numbers
[40:44] that will give you this semantic meaning
[40:46] of whatever text you pass through, okay?
[40:50] So please have a go at playing around with this API
[40:54] in order to create your own text embeddings
[40:57] and compare them against other texts embeddings
[40:59] in order to find similar text.
[41:02] And that's it.
[41:04] Thank you so much for watching
[41:05] this course on prompt engineering
[41:08] in which we covered what it is.
[41:10] We covered an introduction to AI
[41:12] as well as looked at linguistics, language models,
[41:15] prompt engineering mindset, using GPT-4,
[41:19] how to look at best practices
[41:20] as well as zero shot and few shot prompting,
[41:23] as well as AI hallucinations, vectors or texts embeddings
[41:27] and this recap right here.
[41:29] So I hope you've enjoyed it
[41:31] and I'll see you again on the FreeCocam channel.