Lessons

lesson

- Intro to Prompt Engineering and Why It Shapes Every LLM Response - How Prompts Steer the Probability Space of an LLM - Context Engineering for Landing in the Right “Galaxy” of Meaning - Normal Prompts vs Engineered Prompts and Why Specificity Wins - Components of a High-Quality Prompt: Instruction, Style, Output Format - Role-Based Prompting for Business, Coding, Marketing, and Analysis Tasks - Few-Shot Examples for Teaching Models How to Behave - Synthetic Data for Scaling Better Prompts and Personalization - Choosing the Right Model Using Model Cards and Targeted Testing - When to Prompt First vs When to Reach for RAG or Fine-Tuning - Zero-Shot, Few-Shot, and Chain-of-Thought Prompting Techniques - PAL and Code-Assisted Prompting for Higher Accuracy - Multi-Prompt Reasoning: Self-Consistency, Prompt Chaining, and Divide-and-Conquer - Tree-of-Thought and Branching Reasoning for Hard Problems - Tool-Assisted Prompting and External Function-Calling - DSPy for Automatic Prompt Optimization With Reward Functions - Understanding LLM Limitations: Hallucinations, Fragile Reasoning, Memory Gaps - Temperature, Randomness, and How to Control Output Stability - Defensive Prompting to Resist Prompt Injection and Attacks - Blocklists, Allowlists, and Instruction Defense for Safer Outputs - Sandwiching and Random Enclosure for Better Security - XML and Structured Tagging for Reliable, Parseable AI Output - Jailbreak Prompts and How Attackers Trick Models - Production-Grade Prompts for Consistency, Stability, and Deployment - LLM-as-Judge for Evaluating Prompt Quality and Safety - Cost Optimization: How Better Prompts Reduce Token Usage

lesson

- Intro to Synthetic Data and Why It Matters in Modern AI - What Synthetic Data Really Is vs Common Misconceptions - How Synthetic Data Fills Gaps When Real Data Is Limited or Unsafe - The Synthetic Data Flywheel: Generate → Evaluate → Iterate - Using Synthetic Data Across Pretraining, Finetuning, and Evaluation - Synthetic Data for RAG: How It Stress-Tests Retrieval Systems - Fine-Tuning with Synthetic Examples to Update Model Behavior - When to Use RAG vs Fine-Tuning for Changing Information - Building RAG Systems Like Lego: LLM + Vector DB + Retrieval - How Vector Databases Reduce Hallucinations and Improve Accuracy - Generating Edge Cases, Adversarial Queries, and Hard Negatives - Control Knobs for Diversity: Intent, Persona, Difficulty, Style - Guardrails and Bias Control Using Prompt Engineering and DPO - Privacy Engineering with Synthetic Data for Safe Testing - Debugging AI Apps Using Synthetic Data Like a Developer Debugs Code - LLM-as-Judge for Fast, Cheap, Scalable Data Quality Checks - Axial Coding: Turning Model Failures Into Actionable Error Clusters - Evaluation-First Loops: The Only Way to Improve Synthetic Data Quality - Components of High-Quality Prompts for Synthetic Data Generation - User Query Generators for Realistic Customer Support Scenarios - Chatbot Response Generators for Complete and Partial Solutions - Error Analysis to Catch Hallucinations, Bias, and Structure Failures - Human + LLM Evaluation: Combining Experts With Automated Judges - Model Cards and Benchmarks for Understanding Model Capabilities

zaoyang

lesson

- Feedforward networks as transformer core - Linear layers for learned projections - Nonlinear activations enable expressiveness - SwiGLU powering modern FFN blocks - MLPs refine token representations - LayerNorm stabilizes deep training - Dropout prevents co-adaptation overfitting - Skip connections preserve information flow - Positional encoding injects word order - NLL loss guides probability learning - Encoder vs decoder architectures explained - FFNN + attention form transformer blocks

zaoyang

lesson

- Intuition for decoder-only LLMs - Tokens, embeddings, transformer pipeline - Autoregressive next-token generation - Generative AI modalities overview - Diffusion vs transformer model families - Inference flow and prompt processing - Build a real LLM inference API - Architecture: attention, context, decoding - Training phases: pretrain to RLHF - Vertical vs generic LLM design - Distillation, quantization, efficient scaling - Reasoning models: Chain of Thought and Test Time Compute - Hands on Exercises

zaoyang

lesson

- How AI Thinks in Numbers: Dot Products and Matrix Logic - NumPy Power-Tools: The Math Engine Behind Modern AI - Introduction To Machine Learning Libraries - Two and Three Dimensional Arrays - Data as Fuel: Cleaning, Structuring, and Transforming with Pandas - Normalization in Data Processing: Teaching Models to Compare Apples to Apples - Probability Foundations: How Models Reason About the Unknown - The Bell Curve in AI: Detecting Outliers and Anomalies - Evaluating Models Like a Scientist: Bootstrapping, T-Tests, Confidence Intervals - Transformers: The Architecture That Gave AI Its Brain - Diffusion Models: How AI Creates Images, Video, and Sound - Activation Functions: Teaching Models to Make Decisions - Vectors and Tensors: The Language of Deep Learning - GPUs, Cloud, and APIs: How AI Runs in the Real World

zaoyang

lesson

- Why context is fundamental in LLMs - Limits of n-grams, RNNs, embeddings - Self-attention solves long-range context - QKV: query–key–value mechanics - Dynamic contextual embeddings per token - Attention weights determine word relevance - Multi-head attention = parallel perspectives - GQA reduces attention compute cost - Mixture-of-experts for specialized attention - Editing and modifying transformer layers - Decoder-only vs encoder–decoder framing - Building context-aware prediction systems

zaoyang

lesson

- Foundations of multimodal representation learning - Text, image, audio, video embeddings - Contrastive learning for cross-modal alignment - Shared latent spaces across modalities - Vision encoders and patch tokenization - Transformer encoders for text meaning - Audio preprocessing and spectral features - Time-series tokenization via SAX or VQ - Fusion modules for modality alignment - Cross-attention for integrated reasoning - Zero-shot retrieval and multimodal search - Real-world multimodal applications overview

zaoyang

lesson

- Tokenization as dictionary for model input - Tokens → IDs → contextual embeddings - Semantic meaning emerges only in embeddings - Transformer layers reshape embeddings by context - Pretrained embeddings accelerate domain understanding - Good tokenization reduces loss, improves learning - Tokenizer choice impacts RAG chunking - Compression tradeoffs differ by domain needs - Tokenization affects inference cost and speed - Compare BPE, SentencePiece, custom tokenizers - Emerging trend: byte-level latent transformers - Generations of embeddings add deeper semantics - Similarity measured via dot products, distance - Embeddings enable search, clustering, retrieval systems

zaoyang

lesson

Victor

lesson

Victor

lesson

Victor

lesson

Victor