AI Foundations for Everyone
A beginner-friendly guide to how modern generative AI is trained, how transformer-based systems generate outputs, where reliability breaks down, and how to prompt, verify, retrieve, delegate, and use agents responsibly. Dated studies and product policies are labeled as snapshots so durable principles stay separate from fast-changing facts.
20 sections · 85 lessons
Course outline
How Language Models Generate Text
- How Chatbots Generate Text
- How Prompts Become Context
- Why Responses Vary
- Temperature and Sampling
Learning Patterns from Training Data
- Training Data and Its Gaps
- Generalizing Beyond Examples
- Parameters: What a Model Learns
Training Models at Scale
- Loss, Gradients, and Parameter Updates
- The Pre-training Loop
- Training Progress and Evaluation
- How Much Compute Training Takes
- Training and Inference Are Different
- Why AI Uses GPUs
- Why Transformers Train in Parallel
Turning a Base Model into an Assistant
- Pre-training and Post-training
- Supervised Fine-Tuning
- Learning from Human Preferences
- The Modern Training Pipeline
- Knowledge Cutoffs and Current Information
Inside Transformers: Building Contextual Representations
- Why Transformers Changed AI
- Tokenization: From Text to Vectors
- Embeddings and Contextual Representations
- Attention: Mixing Context
- Queries, Keys, Values, and Attention Heads
Inside Transformers: Producing the Next Token
- Feed-Forward Networks
- Why Transformers Stack Layers
- From Hidden Vectors to Token Probabilities
- What “Attention Is All You Need” Actually Showed
- The Full Generation Pipeline
Capabilities, Reliability, and System Boundaries
- What Language Models Do Well
- Why Verifiable Tasks Improve Faster
- Capability Is Not Reliability
- The Model Is Only One Part of the System
Build Your First Reliable AI Workflow
- Write a Clear Task Brief
- Use Examples, Constraints, and Checks
- Engage Critically and Create Iteratively
- Manage AI and Shape Its Use
- Run a Small, Low-Risk Pilot
Context Windows and Product Memory
- Context Windows Are Working Memory
- Why Models Miss Information in Long Contexts
- Context Engineering: Curate What the Model Sees
Reasoning Models and Answer-Time Compute
- What Test-Time Compute Does
- When Extra Thinking Helps
- How Reasoning Models Are Trained
Why AI Capability Is Uneven
- Why AI Capability Is Jagged
- Map AI to Your Own Tasks
- What the BCG Study Found
Why AI Produces False Claims
- Why Models Produce False Claims
- Where Fabrication Risk Is Highest
- What the 2025 Hallucination Paper Argues
- Sycophancy: Why Models Agree with You
Checking AI Output
- Why Language Models Struggle with Arithmetic
- Use Tools for Exact Computation
- Separate Fluency from Correctness
- Match Verification to the Stakes
- Ask for Critique, Not Agreement
Working with Documents and Retrieval
- Grounding Answers in Sources
- How Retrieval Can Miss the Answer
- Why Citations Do Not Guarantee Accuracy
- Specify the Output Format
- Structure Output to Make Errors Visible
Choosing Models and Product Surfaces
- Match Model Capability to the Task
- Choose the Right Product Surface
- Evaluate Models on Your Own Tasks
- Open Weights vs. Open Source
Images, Audio, and Video
- Transcription and Vision Can Hallucinate
- Verify Synthetic Media Out of Band
- How Image and Video Generation Work
AI Agents and Tool Use
- The Agent Loop
- How Tool Calls and MCP Work
- Why Agent Errors Compound
- Put Human Checks Before Irreversible Actions
Delegating and Collaborating with AI
- Delegate Steps, Not Responsibility
- Keep Accountability with the Human
- What the Workslop Survey Found
- What the 2025 METR Study Did—and Did Not—Show
Security, Privacy, Bias, and Rules
- How Prompt Injection Works
- Why Risk Grows with System Access
- Check the Data Terms for Your Plan
- Separate Training, Retention, and Human Review
- How Unspecified Defaults Create Bias
- Disclosure, Records, and Existing Rules
Staying Current
- Separate Durable Principles from Product Facts
- Build a Small Personal Evaluation Set
- Make a Personal AI Watchlist