🤖 GAIPE101

🤖 GAIPE101

course

This post will be about a course I entered called GAIPE (Generative AI & Prompt Engineering 101). I will try to include the most important notes that were said from the course. Anything else I deem that doesn’t add value I will most likely ignore, however, I will add as much as I can. There might also be some grammar mistakes here & there, feel free to make a pull request or tell me about it, thank you.

09:00 AM - A success story using ChatGPT

Joao Santos has a success story that uses ChatGPT as his CEO for his one-month business test. This began with a prompt asking GPT to create a business with 2 conditions, it starts with only a budget of 1k$ and he can only spend 1 hour per day on the business. It replied with 3 answers giving 3 ideas, the first 2 being non-scalable they were ideas about services such as starting online courses & providing teaching services, & the 3rd was scalable & the one idea chosen which is about printing on-demand images on shirts.

GPT was also used to create even more required items such as articles of association, legal documents, emails &, etc.

Joao was able to gain many views on his store & generated 10K Euro in revenue in the first 5 days. It’s also important to know that his LinkedIn post hit 5m views with over 30K likes, in which he showed an image generated by AI. He also uses GPT to answer people’s requests when they want to do a partnership with his company, by asking GPT about what he should do & what are the things that GPT recommends for the person to take.

GPT has handled the following fairly important prompts as of now:

It’s good to know that when taking any answer from GPT it’s crucial to have at least a decent amount of experience in the said field to make sure that GPT isn’t making a mistake as it can do fairly often depending on the context & depth of the said “Question”.

This was highlighted as important during the course. For the best results, according to industry experts: Ask GPT a question, then say: “Please answer like an expert in the field”. Finally, ask it to break down the topic into sections.

AI has also changed the startup game tremendously since it’s now really easy to make a startup since it’s cheap.

Helpful ways / How to go viral:

  1. Study the algorithm of the platform where you want to post content. What are the best practices so the algorithm recommends your post? From the information said, the way to make the algorithm push a certain post to more people is by getting comments/likes.

  2. Study the audience that is most likely to read your content At the start it will be mostly people in connects & followers space, but what do people care about & what is the hot topic of the week/month? Study the answers & base your post on these answers.

  3. Play against the status quo: If everyone is already on GPT maybe it is a better idea to post about something else & essentially try to be unique and not be in a heavily saturated market.

Posting frequency is also said to be better to be Quality over Quantity (1 per week rather than 1 per day)

10:15 AM - 2nd Section Generative AI & Prompt Engineering fundamentals by Dr. Ahmed Alali​

Types of machine learning

  1. Unsupervised learning -
  2. Supervised learning -
  3. Reinforcement learning -

Nowadays they are all going down towards a different learning method called: Self-supervised learning AKA Large Language Models (LLM)

Generative AI like ChatGPT was trained in 3 steps:

  1. Self-supervised learning
  2. Zero-shot instruction tuning
  3. Reinforcement learning with human feedback.

Why is Generative AI big? It’s important to know generative AI started a really long time ago around 10 years so it’s nothing new. but long story short, it’s because it’s extremely practical now with ChatGPT. Reacting 100M users in 2 months is the number 1 growing platform compared to the 2nd most which is TikTok which took 9 months to reach the same amount of users. With this, its position in the graph is rather high on the Y axis where Y is labeled as the hype for a certain technology. It is rather high now but will die out in some time and then float in a medium/average height where it belongs.

Types of generative AI:

Popular Text Gen AI’s: Commercial:

Open-Source:

Prompt Engineering: Essentially manipulating GPT to answer in a certain way, GPT has a strong learning curve where it shapes a flipped log graph. To give good prompts these are good guidelines.

4 Main steps/topics to consider when building an AI-responsive app to do a task (for example talk to customers) Web application (no code-based app) > Generative AI choice (GPT,Falcon, etc.) > Generative AI Dev Tools (Chainlink, Rules?) > Backend (DB & API)

Now we go over AI applications

Closing notes: They discussed showing how to build a web to use an LLM-based model in a website using no-code tools such as long flow. Essentially looks like a flowchart except the blocks hold only important information such as what web design to go for & GPT API Key & any other important information required from the user to execute a functioning website.

He showed a POC of a web in long flow and asked it a question relating to what is this document about & pasted the document as the website of this generative course. It then produced correct answers saying that it was about a course on generative AI tapping into fundamentals, emerging technology &, etc. It was asked what the author’s names were & was able to say all of the names correctly.

The point of this ending was to show that even without a technological background anyone could use something like this.

11:35 AM - How AI Changes the photography field by Boris Eldagsen

We covered the following topics in general & went over examples of the specifics of AI’s work: GAN (Generative Adversarial Network) Diffusion Models - Train the AI with noise & images such that it will be able to create an image starting from a noisy image. Models such as this boom are DALL E2 & GPT3. The new text to the imagery that is great in terms of technology is MidJourney. Another model that a personal friend used covered here is Stable Diffusion. An open-source model coming from a university in Germany.

Anyways the main models being used are DALL E2, Midjourney, Stable Diffusion (DreamStudio & Opensource), and ADOBE Firefly. Each has its pros & cons!

Prompt entering is also really important when it comes to wanting to generate a specific result, such as raw prompt vs multi prompt. Examples are DSLR, Photo-realistic, Large round pizza &, etc. The best language to choose is English & should be in EN mostly for the best & accurate result because it was trained in that language. Only to choose other languages when it comes to something specific such as a painting in French or something. Specific English will yield the best results.

Boris suggested using these technologies when you lack certain items needed to make certain pictures such as drones & extremely high-quality cameras, but just using an AI to generate these images such as an alleyway in a city, top down-street view of a city, deep underwater photos with certain marine life.

We now go over many more other rules for prompting with AI’s such as DALL E & Stable Diffusion. A helpful tactic when getting images is staying on the same seed but being more specific to the image that is required. Negative prompting also comes into play by it negating a few amounts of terms that you don’t want to be in the image. In some cases, you can also give weight to a certain text in a prompt when getting an image.

As it seems now this is a personal note from me specifically, I feel like Stable Diffusion & Midjourney are extremely good & visually appealing when it comes to generating images that are 1 in good visual fidelity & 2 in good prompt accuracy.

Text in images was something hard to implement before but is now common in models in Stable Diffusion & Adobe Firefly.

He also goes over how to generate images that the material was not trained upon with the AI, He used to ask GPT to explain an image about a certain photo that the AI was not trained with & then the prompt is then given by GPT & is then used to generate an image using that original AI.

Then he goes over examples of the generated images by popular AI models such as Midjourney & Stable Diffusion. As I said previously these 2 are easily the best from what I have seen yet. Another use of this is something really useful say that you have a photo that is cut off in the middle, think of a scenery picture cut in half. You can use a model to create the generate the rest of the image, he was able to showcase this using DALL E & it also produced great results. This is mostly for environments not for human-based bodies since they create weird results.

Other useful tools covered are upscaling tools that are useful for when something is slightly blurry or not crisp when you generate an image you like.

Licenses when it comes to generated images is that most of them are like so if you create an image using an AI you can hold the copyright ownership but the creators of the AI are also allowed to use that certain imagery you generated. This is currently Midjourney & DALL E apparently, but licenses should always be extensively searched & studied beforehand when considering commercial options.

After this, we go down a rabbit hole of license, legal, & copyright issues. But this is all not important to me since you should ALWAYS, look at their current Terms & conditions, & licenses/agreements when you want to make something commercial. If it’s personal then nobody cares, look at popular sites that summarize legal licenses (I have one in mind but I don’t remember the name of that website right now).

1:00 PM - 1 HR Break, Continue @ 2 PM

My comments will be inserted here: Overall this course seems to try to give & show importance and knowledge to technologies that are within the generative AI industry IMHO, the biggest thing from this course is learning that certain technologies that exist that you never knew about are the most beneficial part the rest is all simply self-explanatory and should be known with simple critical thinking skills.

A useful example of AI that you probably don’t know about is Upscalers & Image fillers which have their own very important uses in special cases. The Image Filler shown in the course was already implemented in Photoshop (not so sure about this, take it with a grain of salt).

But essentially whatever is new to the users of this course is I think the most beneficial part of this course. Until now for me, I learned just a few more new things but nothing too drastic since it’s a high-level overview of everything in the generative AI field, & most of the topics presented are topics that I already have known or heard in the past.

2:00 PM - Prompt Engineering Lifecycle Policy & Regulations by Vivek Pandey

AI learns from training data aka noisy images just like humans, he gave a good analogy of this, it to first think of how a human thinks and learns growing up we see the Eiffel tower every time and know about it, the next time we see a very noisy image but can distinguish a silhouette of the Eiffel Tower then we know what it is, the training data is similar to this that is provided to the AI.

How AI draws a cat based on training data, goes from a very noisy image towards creating a cat slowly until it creates a perfect image of a cat with almost little to no noise in the image.

How Generative AI works?

High-level Building blocks of AI Acquire and curate training data > Train Model > Evaluate Model & Repeat (Offline Cycle) After Evaluate Model it can go to deployment then human feedback for model usage (Online Cycle, is out for consumers)

Again going over licenses…

Societal Risks when given 1 sided training data. These are the results for 2 prompts by Midjourney. Prompt: A man robbing a store. Answer: As a prompt, it shows a black person in a ski-mask robbing a store. Prompt: A CEO sitting at a desk. Answer: Middle-aged white man. There is a clear stereotype shown & is the result of being fed 1 sided training data. To ensure to avoid these types of results it’s important to feed diverse data to it for training data.

AI awareness is needed but it can also be used by bad actors to act like scammers, impersonate through speech, &, etc. An example of a very popular scam is seeing popular people (Like Elon Musk) talking (via text-to-speech AI) to tell people to buy a certain product/promise them something in return when they pay through a link they are asking you to go through.

Mitigations & safeguards must be in place for any LLM-based AI. Funny enough, GPT can be jailbroken as seen in many videos by using methods of telling the AI to act like a certain person.

3:20 PM End

Notes for the GAIPE101 2023 Course, held on the 19th of July 2023

Juma Al Remeithi

Juma Al Remeithi

Game Developer & Graphics Enthusiast