The courses I'm interested in so I could get going in AI, give the company a leg up:

• MIT course
  
  • Thomas Malone is the star here, author of Superminds, The Future of Work, etc.
  • Cost: $3200


• Texas McCombs UT Austin course. Talked to a course guide and it became apparent that Python is the language I should become conversant with.
• Cost: $3800
• Python based
• 6 mos duration

• Udemy AI for Business
• Going to start with this $13 course so I can get grounded.

3 AI Types

1. Artificial narrow intelligence (ANI) or Weak AI
   • Voice cloning
   • Self driving cars
   • Movie recommendations
   • Can't learn beyond specific tasks
   • Can't adapt to new tasks
   • Used for Siri and Alexa
   • Chatgpt is in between ANI and AGI
   
   
2. Artificial General Intelligence (AGI) or Strong AI, which is not yet ready for prime time
• Mimic human intelligence, can learn, reason, and adapt (HAL, from 2001)
• Autonomous improvement accompanied by technical advancement in the field
• Ethical concerns



3. Artificial Super Intelligence (ASI) is a future product
• The Terminator or God?
• Read Nick Bostrom's SUPERINTELLIGENCE

Machine learning

1. Machine learning is a subset of AI that involves developing algorithms that learn patterns from presented data. Instead of coding in every imaginable detail of Chewbaca feed multiple presentations of Chewbaca (in the sun, fighting, different sizes, etc) to the algorithms which fill in an ever-expanding set of Chewbaca details: this is machine learning based on experience and exposure to data. Humans learn this way as well.
2. Deep learning is a subset of Machine learning which uses artificial neural networks to model and solve complex problems by performing ever more detailed dives into data to recognize patterns. This allows for the formulation of informed guesses about which movie, book, or restaurant, given your history of past choices you might like to try next.

Four types of learning

1. Supervised learning means you have the data and you know what you want to achieve. Ex: extrapolate a price based on the amount of material used in a doll.
2. Classification groups data into a class.
3. Unsupervised machine learning works with unlabeled data and creates classes based on identified patterns.
4. Reinforcement learning: learning to make decisions by interacting w/environment and using a reward/penalty system to guide decision making.

Natural Language Processing

There are 5 categories:

1. Sentiment analysis
• Determine emotion (positive, negative, neutral) expressed in a piece of text. By monitoring social media, can identify trends, react proactively to negative reviews, capitalize on positive reviews, etc.
2. Text classification
• Grouping text into categories, such as spam, or resume screening
3. Named Entity Recognition (NER)
• Names of people, organizations, locations, dates
• Primary use is to extract key information out of an article - used by Wall St, by legal firms
4. Machine Learning Translation
• Foreign language support
5. Q&A
• ChatGPT - ask it a question, it answers
• FAQs, chatbots

ChatGPT and Prompt Engineering

Prompt engineering is an AI term. In addition to ChatGPT there is Google's Bard, Microsoft's Bing, etc. The 'T' in ChatGPT was developed by Google, and GPT = generative pre-training transformer. ChatGPT is also used by MSFT.

To make prompts most effective:

1. Directional/Instructional prompting: be discriptive as possible (see Output format #2).
2. Output format - Ex: List 5 programming languages focused on web development explaining primary purpose for each. Generate output with format: Programming language -> Utilization
3. Role-based/System prompting - Ex: You are an urban planner who has created many smart cities. 1. Create presentation slides etc... 2. No more than 3 bullet points per slide, 3. Put motivational quote on smart cities on second slide.
4. Few or zero-shot prompting is a technique that leverages the capabilities of Large Language Models (LLMs) to perform specific tasks. By providing a few examples of desired output, known as "shots," you can condition the model to generate similar output, be it text, code, or images.
5. Chain of thought prompting: 1) breaking down a complex subject into smaller chunks, 2) ask AI to explain the answer. Please see this pdf to see the fantastic results.
• Question #1: Currently, our company offers etc.
• Question #2: Our company has canvassed the existing marketplace etc.
• Question #3: What workflow design etc.
• Question #4: what technology stack etc.
• Question #5: what future technology etc.
• Question #6: what segment etc.
• Question #7: how would you etc.
• Question #*: what would you etc.
6. Tone prompting - ex: Rewrite the following email in the style of Rudyard Kipling and make it rhyme.
7. Style prompting - ex: Re-write any email that I paste in here using my style of writing. As a reference I will provide 2-3 samples I wrote: etc... ending with: To repeat, taking these 2-3 emails into account, rewrite any future text that I provide here in that style of writinhg. Got it?
8. Structure your data into a markdown table - ex: Marko is working in the IT industry as a QA. He was living in Belgrade, capital of Serbia, but after Covid he moved to Thailand to work remotely.
9. Text summarization - ex: Summarize the following text and create a title. Output format: SUMMARY: Text TITLE: Created Title
10. Classification text - ex: Classify provided text into three classes: Negative, Positive, Neutral. Output format: [SENTIMENT]: Text
11. Coding generation, test your code

ChatGPT Plugins

Plugin other sources of data which then get incorporated into the LLM. There are a lot of plugins, such as Kayak or Zillow, but you gotta have the professional version of ChatGPT. For example, you can use the Bing plugin and then prompt "using the internet as a source, present a list of the latest compositions for viola in the last 10 years."

I upgraded, used Kayak, and then thought I oughta figure out what it would take to develop a PRS plugin... used edX plugin and asked "given the fact that I have a basic understanding of programming in HTML, what are the educational steps I need to take to create a chatgpt plugin?" See answer here.

Excel and Machine Learning

AI Image generation

They are using Midjourney with Discord. No longer free. There are free image generators. We have, as part of Adobe, Firefly.

1. Do you want a photo or a painting?
2. What's the subject? Ex: painting of a cat
3. What camera angle? Lighting? Background? Ambience? Location?
4. You can blend images, etc in Photoshop
5. Uploiad an image and ask AI to generate a prompt on that image
6. Naming conventions (prefix, suffix)
7. Negative prompting (no clouds in background)
8. An aspect ratio is a proportional relationship between an image's width and height. For example, 1:1 means width and height are the same.
9. Prompt examples: --ar 5:4 --chaos 90 --stop 75
10. Dall-e is the chatgpt image generator

API Integration

Integrate Chatgpt directly into your app!

1. There is a pricing model, for both prompt and completion
2. Use python to send prompt to Chatgpt
3. Embed your own data source to act as a chatbot on your website
4. Fine-tuning: Teach the model how to answer a question (e.g. structure/format, personality, etc)
5. Embedding: Provide the model with new/specific information with which to answer questions.

Notes

1. Regression analysis is an attempt to define what's going to happen in the future by examining historical data.
2. TPU (tensor processing unit) and GPU definitions
3. AutoML (automatic model learning) uses machine learning models to view data using varying criteria, leveraging concepts such as hyper parameter optimization.
4. Transfer Learning: open source results
5. Have ChatGPT ask YOU clarifying questions!
6. Repitition of points in a prompt reinforce desired subject matter and result desired
7. Insertion or Jailbreak chats hack techniques
8. Use 'continue' if you want AI to continue...
9. Have it check it's own work by 1) copying initial answer and then 2) In a NEW chat, using/amplifying the context of the first chat, ask it to refine the text you copied!
10. LLMs are algorithms as much as they are neural networks - they rank the probability of words following an original word or set of words. For ex, start with 'I' and there's an 80% probability that the next word will be 'love' or 'am', etc. The larger the data set of words and responses, the more viable answers are likely to be.
11. King - man + woman = queen... the famous word vector equation. Check out Demystifying Word Vectors.