"Any sufficiently advanced technology is simply undistinguishable from magic." Arthur C. Clark

When I saw many of my friends completing Career Essentials in Generative AI by Microsoft and LinkedIn, it piqued my interest. It must be really good. And with Generative AI being the latest and ongoing trend, I gotta set time aside to give this a go.

It’s pretty informative and covers more than just Generative AI. I learnt concepts on Prompt Engineering, Microsoft Copilot, Ethical Artificial Intelligence, Machine Learning Algorithms, Artificial Neural Networks and more.

As usual, I took some notes to capture my key learnings for easier recap, and to share them with you. Here’s what I learnt from the 4 hours course in a 8 mins recap:

1. What is Artificial Intelligence (AI)

Artificial Intelligence (AI) Applications
Reactive Machines Self-driving Cars
Limited Memory AI Forecasts weather
Theory of Mind Virtual customer assistance
Narrow AI Customized product suggestions
Supervised Learning Object identification from image/video
Unsupervised Learning Fraudulent banking transactions
Reinforcement Learning Teach machine how to play game

2. What is Generative AI (Generates new content)

Generative AI Applications
Natural language models ChatGPT, GitHub Copilot
Text to Image Application Midjourney, DALL-E & Stable Diffusion
Generative Adversarial Networks From art design, visual effects to fraud detection
Variational Autoencoders Anomaly Detection


"Generative AI will be used to automate repetitive tasks and improve efficiency in a wide variety of industries."

3. Prompt Engineering Tips in Reasoning Engines
      a. Be specific
      b. Provide context (Examples of type of answers)
      c. Break things down (Reduce complexity)
      d. Use clear language with proper grammar
      e. Experiment (Trial & error)
      f. Role-play scenario (Indicate persona)
      g. Request for analogies
      h. Debate-style questions (Argument for & against)
      i. Creative exercises (Ideate & brainstorm)

4. Additional Resources on Prompt Engineering (One/Few Shot Engineering)
      a. OpenAI documentation
      b. ChatGPT Discord server
      c. Prompt Engineering Guide
      d. PromptVine
      e. Learn Prompting
      f. PromptPapers
      g. PromptHub

"Any new technology is only as ethical as the underlying data that it's trained on."

5. Risks & Challenges
      a. Deep Fakes (Fraud)
      b. Inaccurate chatbots (False advice)
      c. Legal confusion (Ownership)
      d. Biases (Hiring)

6. Ethical AI Framework
      a. Responsible Data Practices (Training data source & reduce bias)
      b. Boundaries on safe & appropriate use (Organization and target audience goals)
      c. Robust Transparency (Traceable outcomes for human audit of ethical accountability & fairness)

7. Ethical Data Organization
      a. Prioritizing privacy (Protect sensitive data, privacy policy audit, employee training)
      b. Reducing bias (Bias audit & analysis on data collection & data use)
      c. Promoting transparency (Be open with how data is collected & used)

"Ethical analysis needs to be intertwined with the initial design of new products and at every phase of deployment."

8. Machine Learning Algorithms
      a. Supervised Learning (Starts with labeled data)
            i.   K Nearest Neighbor KNN (Euclidean distance between data points)
            ii.  Regression Analysis (Predictors vs outcomes trend lines)
            iii. Naive Bayes (Class predictor probability based on features in data, assuming all predictors are independent)
      b. Unsupervised Learning (Systems see patterns that’s impossible for humans to create)
            i. K-Means Clustering (Iterate & redistribute until best centroids are identified)
      c. Q-Learning (Reinforcement learning incentivize the system to find new patterns)

9. Fit the Algorithm
      a. Ensemble modeling
            i.  Bagging (Several versions of the same machine learning algorithm)
            ii. Stacking (Different machine learning algorithms)
      b. Bias-variance trade off
            i.  Bias is the number assigned to each neuron (Gap between the predicted value and the actual outcome)
            ii. Variance is adjusted by adding weights to connections (Predicted values are scattered all over the place)
      c. Overfitting (Complexity/many predictors)
      d. Underfitting (Simple/little rules)

10. Artificial neural networks (Feedforward neural network supervised learning)
      a. Layers of neurons: Input, hidden, output layers
      b. Deep learning = many hidden layers = easier to identify complex patterns
      c. Activation function (Allows the hidden layer neuron to decide whether to feed the data forward)
      d. Advantage: Self-tuning to understand data by adding weights to the connections in the hidden layer
      e. Caution: Artificial neural networks tend to overfit the data by overlearning lessons about the data
      f. Cost Function (Measure output against correct answer. The better the output, the smaller the number)
      g. Use gradient descent to determine wrongness and backprop (backpropagation of errors in the network) to adjust weights and biases

11. Artificial intelligence systems
      a. Only as good as the data they’re given (Accurate & representative of the real world)
      b. Learn by trying different things (Needs time for trial & error)
      c. Can do things human can’t (Find patterns in large amount of data with supervision and direction)

"Remember that machine learning systems feed on data to learn new things."