Definitions

1. Artificial Intelligence (AI)

The study of creating machines and computer systems capable of performing tasks that typically require human intelligence.

Types of AI

  • Narrow AI: AI designed and trained for a specific task or narrow set of tasks.
  • General AI: A hypothetical future AI system with human-level intelligence.

Applications of AI

  • Chatbots: AI programs designed to engage in natural conversations with people.
  • Computer Vision: Enables computers to interpret and analyze visual information like images and videos.
  • Robotics: Focuses on designing, constructing, and operating robots.
  • AI Tools: Software programs designed to assist users in performing AI-related tasks.

2. Algorithms

Defined methods or processes employed to train models, generate predictions, and execute tasks using data.

Used in machine learning, AI models, and statistical analysis.

3. Machine Learning (ML)

A branch of AI that enables computers to improve their performance through experience without explicit programming.

Types of Machine Learning

  • Supervised Learning: Uses labeled data for training.
  • Unsupervised Learning: Uses unlabeled data without explicit guidance.
    • Uses advanced ML algorithm to process the unstructured data to train AI
  • Reinforcement Learning: Trains agents through rewards and penalties based on interactions with an environment.

Subfields of ML

  • Neural Networks: Computational models inspired by the human brain, used for recognizing patterns, classification, and predictions.
    • Deep Learning: A subset of ML using multi-layered neural networks to process large amounts of data.
    • Fine Tuning: the process of adjusting parameters/ weights/biases of a pre-existing neural network model so a new model does not have to be trained from scratch
  • Generative AI: AI systems capable of creating new content (e.g., text, images, music, code); often powered by LLMs
    • Temperature : the parameter that influences the level of randomness in generated outputs; lower temperatures produces conservative/deterministic outputs while high temperaures lead to more diverse results

4. AI Models

Computer programs designed to make predictions or decisions based on input data.

  • Large Language Models (LLMs): Trained on massive amounts of text to understand and generate human-like language.
  • Image Model: Trained on creating images
  • Diffusion Model: For editing and polishing images
  • Other model types: Image recognition, Speech recognition, Recommendation models

5. Natural Language Processing (NLP)

A subfield of AI focused on enabling computers to understand, interpret, and generate human language.

  • Applications: Chatbots, Language translation systems, Sentiment analysis, Text summarization
  • Related Models: Large Language Models (LLMs)

6. Statistical Analysis

Techniques involving data collection, organization, examination, and interpretation to identify patterns and make predictions.

Supports: Machine learning model training, AI model evaluation, Data-driven decision-making.

Uses of Artificial Intelligence

  • Text Extracting : get specific information from the text
    • Document Summarization : helps document codes or other contents
    • Extract keywords use for SEO
    • Classify the input and re-organize it
  • Optimization: Optimizes routes for deliveries, processes, and testing.
  • Research & Testing: Research variety of topics or tech and give quick analysis as a preliminary for a project.
  • Software Development: Assisting developers by autocompleting code, explaining errors, documentation, debugging, and optimizing programs.
    • Read code documentation and answer questions
    • Help make preliminary plans for software architecture
  • Data Handling: Finding insights/predcitions within large, complex datasets across finance, healthcare, research, and more.
    • Data cleaning and correction : identifies and correct errors in data; break down the cleaning prompts to column-by-column
    • Data classification : categorizing data into predefined classes or labels
    • Intelligent grouping and clustering : detects patterns, clusters similar data points, and identifies trends
    • Data visualization : transform datasets into intuitive visual formats
  • Language Processing: Rephrasing and rewriting text in different styles while preserving the core meaning, as well as understanding and generating human-like language.
  • Recommendation Systems: Powering engines that suggest relevant content, products, or services.
  • Work Automation: Helping schedule meetings, manage tasks, prioritize work, and improve productivity.
  • Computer Vision: Identifying objects, people, text, and other elements within images and videos.
  • Audio Processing: Transcribing recordings, identifying speakers, removing noise, and generating speech.
  • Creative Generation: Producing original images, music, stories, and other creative works from text prompts.

Notes

  • Large, diverse, and veried datasets mitigate bias
  • Content filters and fact-checking mechanisms can help improve dataset and learning
  • Use human oversight in sensitive domains, like healthcare and law

LImits of Artificial Intelligence

  • Difficulty dealing with unknowns: AI is good at answering questions rather than discovering new ideas or solutions on their own
  • AI models rely on training patterns, not fact-checking which can lead to AI hallucination - when AI generates false, misleading, or fabricated info
  • Limited based on the scope of their training data
    • Can create nonsensical or unraelistic content
    • AI drift: overtime it can become less accurate compared to when it was first trained
  • Handling open-ended problems that require unconventional approaches
  • If the training data is biased in some way, AI will learn it and perpetuate it
    • Samping Bias : lack of diversity
    • Measurement Bias : errors in data or faulty tools
    • Algorithmic Bias : caused by bias decisions made in the training data or during algorithm design
    • Selection Bias : when training data is not randomly selected
    • Confirmation Bias : when dev only consider data that supports pre-existing beliefs
    • User & Prompting Bias : Repeatedly asking AI similar questions can skew responses, creating a feedback loop where AI reinforces certain perspectives
  • Predicting an individual human behavior
  • Lack of common sense.
  • Adversarial Attack: deliberate attempts to deceive or manipulate AI systems by introducing carefully crafted perturbations to the input data; like putting up false signs for AI to inpterpret

Designing Specific Prompts

  1. Identify the Topic
  2. Clarify the Objective: clearly define objective and desired output
  3. Specificity/ Scope: be specific
    • Add any relevant background/context information to improve the response
    • Make neutral and inclusive prompts to avoid biased outputs
  4. Revise and Reiterate: ask for revision if output is not optimal

Key Parts of a Prompt

  • Persona : define the AI's role or identy and shapes the tone/style of the response
  • Instructions : core purpose of the prompt; be clear and specific
  • Input Content : relevance info/background provided by the user
    • Separate the inputs from the rest of the instructions using delimiters like ### or --- or //
  • Output Format: : the desired structure, format, verbosity (detail-level), length, style, and tone of the output
    • Example: Companies: [list of companies]
  • Addtional Information : adding constraints/context will lead to more relevant, personalized responses
    • Give use cases/ scenarios to help AI understands the context
    • Context like programming language, frameworks, project structure
    • Example: "If you don't know the answer, then response with idk"
    • Example: "If you have question about the prompt, ask."

Advanced Prompting Techniques

  • Cognitive Verifier Pattern
    • Description: The AI asks clarifying questions before responding to ensure accuracy and relevance.
    • When to use: When working with ambiguous, open-ended, or multi-step queries.
  • Zero-Shot Prompting
    • Description: The model responds directly based on its existing knowledge without any examples.
    • When to use: For general queries where the model likely already has enough context.
  • Few-Shot Prompting
    • Description: Provides a few examples in the prompt to guide the AI toward more accurate results.
    • When to use: When precision is important or the task involves a specific structure or style.
  • Chain-of-Thought (CoT)
    • Description: Encourages the AI to explain its reasoning step by step, improving logical accuracy.
    • When to use: For multi-step problems like math, reasoning, or structured decision-making.
  • Least-to-Most Prompting
    • Description: Starts with simple prompts and gradually increases complexity as the AI builds understanding.
    • When to use: When solving complex problems where breaking them into smaller steps improves accuracy.
  • Tree-of-Thought (ToT)
    • Description: Explores multiple reasoning paths before selecting the best solution.
    • When to use: For problems with multiple possible answers or strategic planning scenarios.
  • Self-Consistency Prompting
    • Description: Generates multiple responses to the same prompt and selects the most consistent or accurate one.
    • When to use: When reliability and coherence are critical, such as research or reporting.
  • Generated Knowledge Prompting
    • Description: First generates relevant background information, then uses it to improve the final response.
    • When to use: For highly technical, niche, or domain-specific queries where additional context improves results.

Image and Video

  • DALL-E : generates images from text
  • Midjourney : allows users to refine images through iterative feedback
    • Uses Discord as an interface; Creates own server and add the Midjourney Bot into it
    • Use /imagine to start the prompt and add parameters like cmd
    • Produces images in varying styles
    • Library of Image Styles: midlibrary.io/
    • Documentation: link/
    • Parameter List: link/
  • InVideo : creates image and video
    • Create video from a script or title