Read Time: 6 minutes │ Subscribe to the FREE Newsletter │Join the Premium Patch

Let’s address the elephant in the room.

Ammi has been MIA; it has been over one month since our last newsletter!

But I have never, in all my days, met anyone who lamented fewer emails in their inbox. So I guess you’re welcome 😁 

In all seriousness, it has been too long. It is great to see you.

I’m excited to get this thing back on track!

🌿 Let’s grow!

This week’s letter is brought to you by Growth School!

Become an AI & ChatGPT Genius in just 3 hours for FREE!  (Early Easter Sale)

Join ChatGPT & AI Workshop (worth $199) at no cost (Offer valid for first 100 people only) 🎁

Planting the Seed: Explore AI topics and headlines in simple language.

Tomorrow (April 19) is National AI Literacy Day! AI literacy refers to the knowledge and skills needed to understand and use Artificial Intelligence technologies as they become increasingly present in our work and personal lives.

If you’ve been with us since our first newsletter, you'll know we’ve covered decent ground in this Planting the Seed section (pun intended 🌱).

If not, no problem! I compiled a quick cheat sheet of key AI topics covered in the last 19 issues of ammi.ai. We have much more to cover, but this AI literacy garden is blooming beautifully. Thanks for nourishing it with your thoughtful conversation, questions, and insights.

Ammi’s AI Literacy Cheat Sheet

What is Artificial Intelligence (AI)? 

AI is technology that enables machines or computers to perform tasks that typically require human intelligence such as processing natural language or distinguishing between a sloth and pain au chocolate. How is this done? By some combination of mathematical model(s) + data to train the model(s) + computer(s) to power it all. (Issue #1)

Why is this ChatGPT variant of AI causing such a ruckus?

In contrast to ‘classical’ AI, as some call it, which can do things like predict (this user is likely to cancel their subscription) or classify (this is an image of a turtle), Generative AI can…generate. Generate images, text, and audio. Generate ideas and content. It has immediate & vast applications in amplifying human productivity and creativity. (Issue #1)

Why does AI need to be regulated?

Like humans, AI systems are imperfect. AI can be wrong, AI can be biased, and AI can be exploited. There may be instances where AI is considered capable of fully replacing a human in a job when it should not be. AI regulation is crucial to ensure the development and use of AI is ethical, safe, and fair. Hopefully, AI regulations will protect citizens’ safety and privacy and hold people and organizations providing AI technologies accountable. (Issue #2)

What is a decision tree?

A decision tree is a popular, easy-to-understand machine learning model. It is essentially math that models decision-making based on historical data to help a machine (computer) make predictions on new data. After seeing a decision tree represented in code, I understood how machines could be trained to make decisions like us. (Issue #3)

What are deepfakes?

Deepfakes are synthetic images, video, or audio generated via - you guessed it - artificial intelligence. Specifically, they’re built with an AI technology called Generative Adversarial Networks (GANs).

Deepfakes are primarily associated with creating realistic-looking but entirely fabricated videos or images of people, often portraying them saying or doing things they never actually did. (Issue #4)

Personal AI: Science Fiction to Reality

One AI innovation is personal AI, or AI that aims to understand an individual and provide personalized support as a good friend, assistant, or coach. Thus far, we’ve primarily seen personal AI in Hollywood films, like Samantha in Her or J.A.R.V.I.S. in Iron Man. Well, personal AI has transitioned from the realm of science fiction to reality. (Issue #5)

What are some potential challenges presented by AI-generated content?

As AI-generated content looks & sounds increasingly real, it’s getting harder for people to determine whether they’re looking at something created by humans or machines. This content can be used to spread false and misleading information. There are AI tools available that can be used by people with bad intentions to create scams and lies quickly and easily.

Examples of misuse of AI-generated content include doctored media about world events, fake advertisements using the likenesses of celebrities to peddle products, or phone calls that replicate the voices of family members claiming to be kidnapped and needing money. (Issue #6)

How will AI Labeling legislation help to address the challenges posed by AI-generated content?

The AI Labeling legislation will require companies and AI developers to label AI-made content clearly so people know what they're looking at, watching, or listening to. This will help prevent rampant misinformation and fraud, as people will not need to investigate deeply to find out whether something was made by AI or not. (Issue #6)

What does it mean to ‘train’ an AI model?

An AI model is like a complex math equation with many variables. To get accurate results, it's trained with numerous examples, called training data. The model adjusts its variables based on this data until it produces acceptable answers.

What are some of the data sources that were used to train GPT models?

The purpose of the GPT models is to generate human-like text in response to a given prompt. Here is an ingredient list, or the training corpus, for the GPT models. The latest GPT model, GPT-4, is trained with data until April 2023.

• Common Crawl: A dataset that includes a wide range of web pages.

• WebText2: A dataset consisting of web pages with a focus on diverse and high-quality text.

• Wikipedia: A crowd-sourced online encyclopedia that provides a wealth of information on various topics.

• Books1: A collection of books covering a wide range of genres and subjects.

• Books2: Another collection of books that adds to the diversity of the training data.

This helps us understand what a given AI model knows and what it doesn’t. For example, ChatGPT should give me a decent summary of Taylor Swift’s rise to fame, but it can’t tell me what she’s wearing to Coachella! (Issue #7)

What is ChatGPT?

ChatGPT is a chatbot. A chatbot is a computer program that humans can interact with in natural language, like English, instead of clicking or coding.

ChatGPT is powered by AI to process human input and provide human-like outputs. Specifically, ChatGPT is powered by an AI model called GPT.

What is GPT?

GPT, or Generative Pre-Trained Transformer, is an AI model developed by OpenAI, designed to generate human-like text. Let’s break down the G, P, and T:

• Transformer models enable computers to understand the relative importance of words in sentences, making them efficient at learning and responding.

• By Pre-training these Transformers with data from the internet, the model learns language.

• Finally, provided human input, the model can Generate the next most likely word or sentence based on the context it's given.

In short, the T is P, so it can G!

Every so often, OpenAI re-trains GPT to update it with fresh data and improve its safety and performance. The model has evolved from GPT-1, released in June 2018, to the latest version, GPT-4, released in March 2023. (Issue #11)

What are LLMs?

ChatGPT is powered by GPT, an artificial intelligence model that can understand, process, and generate human language. This kind of AI model is known as a Large Language Model, or LLM, and GPT is among several notable LLMs.

What makes LLMs different from each other?

Large Language Models (LLMs) can be different from each other based on a few factors:

1. Training Data: The information that LLMs learn from can vary. Some learn from a wide range of topics on the internet, while others may focus on specific types of text, like scientific research papers, for example.

2. Size and Complexity: LLMs come in different sizes. Bigger models can understand language more deeply but need more power to run.

3. Purpose and Training: LLMs are trained for different uses. Some are better at answering questions, others at translating languages or writing creatively.

These attributes help determine what each LLM is best at and how it is used.

What can LLMs be used for?

LLMs are great for text in, text out tasks including:

• Conversation: here’s a question, greeting, or comment, respond to it.

• Summarization: here’s a very long blog post, summarize it.

• Content generation: here’s a message, write an email to convey it.

LLMs are great at processing and generating natural language, but what about image generation or audio generation? There’s AI for that, too. But we'll save those letters for another meal 🙂 (Issue #12)

What are Foundation Models?

Foundation Models are a class of large-scale AI models. Here are the primary types of Foundation Models:

• Language: Models like GPT and Claude that process and generate text.

• Vision: Models like DALL·E and Midjourney that process and generate images.

• Audio: Models like Whisper and WaveNet that process and generate audio.

• Multimodal: Models like Gemini that can process and generate data in multiple formats. (Issue #13)

Fruitful Harvest: Fresh AI fruits from the community garden.

Summarize Pesky Privacy Policies with AI

I try to be a responsible digital citizen and ensure that the apps and technologies my family and I use don’t pose major privacy or security concerns. But I must admit, as soon as a Privacy Policy or Terms and Conditions window pops up on an app I’m using, I tell myself, “eh, I’ve got nothing to hide!” and proceed to clickety, click-click my way into selling my soul to the Technology powers that be.

AI can help distill lengthy legal text and present summaries and key points in easy-to-understand language. Here’s a ChatGPT prompt to help decode those pesky legal texts. This prompt instructs the AI model underlying ChatGPT, GPT-3.5 or GPT-4, to act as a legal text analyst, summarize the text, and highlight any high-risk factors users should know about.

#Directive:
You are an AI tasked with processing and analyzing legal texts.
Please output the best results in accordance with the following constraints.

#Constraints:
Identify specific language that suggests data sharing with third parties, indefinite data retention, automatic renewal clauses, and other high-risk factors.

#Input Text:
A section of a privacy policy or terms and conditions.

#Output Text Requirements:
Provide a concise summary and highlight any clauses that involve high-risk factors or unusual terms.

#Example:
Input: "The company reserves the right to share personal data with global partners."
Output: "Summary: The company can share your personal data globally. Highlight: This clause allows for extensive sharing of personal data which might affect your privacy."

#First Response
After processing this prompt, ask the user to enter the legal text, privacy policy, or terms and conditions in the chat. 

To use this prompt:

1. Copy and paste the prompt in the black box into a new chat in ChatGPT. Press enter.

2. ChatGPT will prompt you to enter the legal text. An easy way to copy and paste Privacy Policies or legal text is to open the window with the text and hit Ctrl+A on your keyboard. This will highlight all the text. Copy and paste into ChatGPT.

3. Hit Enter, and let ChatGPT get to work for you!

Cultivating Conversation: your insights nourish the ammi.ai garden.

Thanks for spending a few of your precious, precious minutes with us.

See y’all in 2 weeks!
Ruqaiya
Ammi by day, Ammi by night