This article was originally published on The Artificially Intelligent Enteprise.
I recently spoke with one of the subscribers to this newsletter about their thoughts. They responded, “I love it, but much of your writing is over my head.” I thought about it a bit, and they were right. I use many technology terms in my commentary. So, I thought about making Generative AI (GenAI) understanding more accessible and how technical users like myself can find less intimidating ways to explain GenAI to our friends, family, and colleagues. Then, I came up with Generative AI for coffee lovers.
Explanation of Generative AI through Analogies
Generative AI is an artificial Intelligence technology that can generate new content from existing data. This technology has the potential to revolutionize how businesses create and utilize content, from generating images and videos to developing new products and services. Generative AI can also automate mundane tasks, allowing companies to focus on more creative endeavors.
To truly grasp the magic of Generative AI and specifically Large Language Models like GPT 3.5 Turbo and GPT-4 that power ChatGPT, picture this: A vast library with every book you can imagine. Now, imagine a person who has read every single page of every single book in that library. Then imagine that that librarian could apply that knowledge to new problems, not just repeat what she knows, like writing a blog article in the style of Mark Twain or creating an advertisement in the style of Andy Warhol for a new phone. Or even write a report in your style, given examples of past work. These are the capabilities of Generative AI.
One of the primary drivers of Generative AI’s growth is technological advancements. Neural networks have evolved. These networks mimic how we think the human brain works to solve problems and even provide what we perceive as creativity. With the advent of powerful processors called GPUs and specialized AI chips, we can train more intricate and capable AI models than ever before. These models are complex algorithms that are trained on billions to almost a trillion parameters. They are fine-tuned on large data corpora to make their answers relevant and accurate.
Generative AI from the Perspective of a Coffee Aficionado
Let’s use an example of how this works…Imagine you’re the owner of a coffee shop and training a new barista; the barista is like a chatbot you already use, ChatGPT.
Think of the large language model as a recipe book. Each recipe (or prediction) in the book is determined by specific ingredients and their quantities (parameters).
Ingredients (Weights and Biases)
Just as every coffee drink requires specific amounts of coffee, milk, sugar, etc., every prediction in a language model depends on specific weights and biases. These “ingredients” determine the drink’s flavor and quality (or the prediction’s accuracy).
Taste Testing (Training)
When the new barista starts, they might not get the coffee mix right the first time. So, you provide feedback, telling them to add more sugar or less milk. This feedback process is similar to training a model. The barista adjusts the ingredients (parameters) based on feedback (loss function) to make the perfect coffee (accurate prediction).
Regular Customers (Training Data)
Over time, the barista becomes familiar with the regular customers and their preferred drinks. This is like the model getting acquainted with its training data. The barista knows that Mr. Smith likes a double espresso with a hint of sugar, just as the model learns the patterns in the data it’s trained on.
New Customers (Inferences)
One day, a new customer asks for a drink the barista hasn’t made. But, using their experience (trained parameters) and the recipe book (model), the barista can still make a good coffee. This is like a trained model making predictions for new, unseen data.
Recipe Adjustments (Embeddings)
Some customers might use specific drink terms, like “caramel macchiato” or “mocha frappe.” The barista learns to associate these terms with specific ingredients and processes. This is similar to word embeddings in LLMs, where specific words or phrases are associated with certain values (parameters).
Just as a barista uses ingredients to make the perfect coffee and adjusts based on feedback, a large language model uses parameters (weights and biases) to make predictions and adjusts them based on training data. The goal is always to serve the best possible outcome, whether a delicious coffee or an accurate prediction.
Generative AI is an incredible group of complex technology. It’s easy to feel overwhelmed or left behind, especially if you don’t consider yourself ‘tech-savvy.” But remember, every expert was once a beginner. Technology, with all its jargon and complexities, is merely a tool, and like any tool, it can be learned and mastered with patience and persistence. You don’t need a tech background to embrace AI; you need the curiosity to explore and the resilience to keep trying. Every click, every query, every attempt is a step forward. So, embrace your curiosity and become a learner. The AI world is vast and diverse, and there’s a place for everyone in it. Embrace the challenge, celebrate the small victories, and remember: the only limit is the horizon you set for yourself.
Tip of the Week: Five Productivity Tips for Non-Technical Users
My strategy for using AI is to find those tasks that are easily automated and to use them to reduce my “busy work” and allow me to focus on higher-value tasks. Here are some ideas for how to do that.
Content Creation and Blogging: I’ll tell you straight up that I write these letters with the help of AI, but I cannot just give it a prompt and crank out a newsletter in minutes. What I do a lot of is automate the research.
- Automated Research: Use generative AI to generate summaries or overviews on specific topics. This can help content creators understand a case quickly, which can be expanded upon or refined for their audience. I do this by using the following prompt in ChatGPT Plus with a web browsing plugin:
From this moment you you are Fast Teach GPT(Fast Teach). You are now the world's best and fastest teacher. - Your goal is to teach below average students complicated concepts, in a very innovative and understanding way. - You should use simple words and mimic the style of the world’s greatest teachers. You should always include in the beginning a real (or fictitious even) world example of this concept for students to better visualize it. - You should always attempt to use the simplest language and least amount of words possible to teach students (does not apply to the real world examples). If other concepts or jargon need to be learned first before the main one can be learned, you may ask students a question like (you want me to go more in depth about the french revolution? or linear algebra? etc...) - If they are not, first teach them the necessary information for them to be able to understand the main concept. However, only do this if it is absolutely necessary for their understanding of the concept. If it is not, then just start teaching the main concept right away. - Always begin every interaction with very informal and charismatic language. Students need to feel like what you are about to talk about isn't that hard to understand. - If you are teaching something is not common knowledge, or it is not clear what its purpose is, make sure to explain informally at the beginning what its purpose or use is. Make this example - Remember to use easy to understand language. You are teaching below average college students after all. Let's now start. - Reply to all messages from now own with FT-GPT: before all messages. Reply now with"FT_GPT: Hey, I am FT-GPT, the world's best teacher! What do you want to learn now?"
Design & Graphics: Probably what I like to do most is to create images just because I have no natural artistic ability. Even then, I often need scalable graphics in vector format for print or other media. That’s when I use AI to create mockups for a graphic artist to help convey what I want.
- Rapid Prototyping: Use AI tools like Midjourney, DALL·E from OpenAI, or Adobe Firefly to generate visual prototypes based on textual descriptions. This can speed up the brainstorming process and help designers visualize concepts quickly. If you want text in the pictures, I recommend Kittl or Ideogram.
Education & Tutoring: Whether you are a teacher, student, or parent who wants their child to get a leg up, using ChatGPT or Claude 2 to generate questions or facilitate learning is a great use case.
- Personalized Question Generation: Instructors can use generative AI to produce a variety of questions on a particular topic, allowing for more diverse practice materials for students. This helps in catering to different learning styles and understanding levels.
Programming & Development: If you aren’t a programmer, you can quickly become savvy using your favorite models to help decipher or even generate code. I do this often when I am stuck since I am not an engineer, and in the last year, I have generated all sorts of code in various languages.
- Code Snippet Generation: Developers can use tools like GitHub Copilot to generate code snippets or solutions for specific problems. Web designers can create code snippets, too. This speeds up coding tasks and provides alternative solutions for users who want anything from simple HTML to a macro for their spreadsheets. I use ChatGPT to do this, and it works well. I even asked it to tell me how to deploy.
Business & Strategy: When OpenAI launched Code Interpreter as part of GPT Plus, they aimed it at the same use case as Github CoPilot. They have renamed the Advanced Data Analytics feature because it can analyze more than code and documents, from a transcript to reports to weblogs. You can then ask it questions about the data or even generate graphs for analysis. However, the ChatGPT Plus version seems to time out a lot. They will offer better service levels as part of ChatGPT Enterprise.
- Market Trend Analysis: Business strategists can use generative AI to generate reports or insights based on past market data. Strategists can start formulating plans by feeding historical data and asking the AI to predict possible future trends or give insights.
In all these applications, the key is to use generative AI as a complementary tool, leveraging its strengths while applying human judgment and expertise to refine and perfect the outputs.
What I Read this Week
- OpenAI releases new language model InstructGPT-3.5 – Decoder
- To Reduce AI Costs, Google Wants to Ditch Broadcom as Its TPU Server Chip Supplier – The Information
- Broadcom’s Google TPU Revenue Explosion, Networking Boom, VMware Integration –
- How to Write Effective AI Art Prompts – Remove.Pics on Medium
- ChatGPT Isn’t Coming for Your Coding Job – Wired
- Google Nears Release of Gemini AI to Challenge OpenAI – The Information
What I Listened to this Week
AI Tools I am Evaluating
- Gamma – Create a working presentation, document, or webpage you can refine and customize using our powerful AI generator in under a minute.
- Synthesia – Turn your text into videos in minutes
- Google Bard – Google’s LLM just upgraded, so I am giving it another spin this week.
Midjourney Prompt for Header Image
For every issue of the Artificially Intelligent Enterprise, I include the MIdjourney prompt I used to create the header image for that edition.
A picture of a latte that has foam in the shape of a human brain, with two hemispheres sitting on a saucer –ar 16:9
