15. AI Cheat Sheet

The AI Hierarchy

┌─────────────────────────────────────────────┐
│                    AI                        │
│  (any machine simulating human intelligence) │
│  ┌───────────────────────────────────────┐  │
│  │          Machine Learning             │  │
│  │  (AI that learns from data)           │  │
│  │  ┌─────────────────────────────────┐  │  │
│  │  │        Deep Learning            │  │  │
│  │  │  (ML with neural networks)      │  │  │
│  │  │  ┌───────────────────────────┐  │  │  │
│  │  │  │      Transformers         │  │  │  │
│  │  │  │  (attention-based DL)     │  │  │  │
│  │  │  │  ┌─────────────────────┐  │  │  │  │
│  │  │  │  │       LLMs          │  │  │  │  │
│  │  │  │  └─────────────────────┘  │  │  │  │
│  │  │  └───────────────────────────┘  │  │  │
│  │  └─────────────────────────────────┘  │  │
│  └───────────────────────────────────────┘  │
└─────────────────────────────────────────────┘

Types of AI — Quick Reference

Classification	Types	Exists?
By Capability	Narrow AI	✓ Yes
By Capability	AGI	✗ No
By Capability	Super AI	✗ No
By Functionality	Reactive Machines	✓ Yes (Deep Blue)
By Functionality	Limited Memory	✓ Yes (ChatGPT, Tesla)
By Functionality	Theory of Mind	✗ No
By Functionality	Self-Aware	✗ No

How AI Works — One-Liner

Data → Train (minimize loss via backprop) → Model (frozen weights) → Inference

AI Lifecycle — 7 Stages

1. Problem Definition  → Is this an ML problem? What's the metric?
2. Data Collection     → Labeled/unlabeled, quality > quantity
3. Data Preparation    → Clean, label, engineer features, split (60-80% of time)
4. Model Development   → Choose arch, train, tune hyperparameters
5. Evaluation          → Test set, fairness, latency, cost
6. Deployment          → API, A/B test, rollback plan
7. Monitor & Maintain  → Watch for data drift, retrain as needed

Essential Terminology

Term	One-Line Definition
Model	Learned function: input → output
Parameters/Weights	Numbers adjusted during training
Training	Iterative weight updates to reduce loss
Inference	Using frozen model to predict
Loss	How wrong the prediction is
Gradient Descent	Optimizer that walks weights toward lower loss
Backpropagation	Computes gradient through the network
Epoch	One full pass through training data
Overfitting	Memorizes training, fails on new data
Underfitting	Too simple, fails on everything
Hallucination	LLM generates confident but false output
Token	Unit of text LLM processes
Context Window	Max tokens LLM can consider at once
RAG	Ground LLM answers in retrieved documents
Fine-tuning	Further train pretrained model on new data
Embedding	Dense vector representing a piece of content
Data Drift	Input distribution changes in production
RLHF	Align model via human preference feedback

AI Capabilities vs Limitations

Can Do	Cannot Do (reliably)
Generate fluent text	Verify facts
Classify images	Reason about physics
Translate languages	Maintain long-term memory
Detect patterns in data	Understand cause vs correlation
Code completion	Know its own knowledge cutoff
Summarize documents	Give consistent answers to edge cases

Key Milestones

Year	Event
1950	Turing Test proposed
1956	”AI” coined at Dartmouth
1997	Deep Blue beats Kasparov
2012	AlexNet — deep learning for vision
2017	Transformer architecture published
2018	BERT, GPT-1
2020	GPT-3 (175B params)
2022	ChatGPT — mainstream AI
2024	Reasoning models, AI agents, MCP

Ethics Quick Reference

Principle	What to Ask
Fairness	Equal performance across groups?
Transparency	Can decisions be explained?
Privacy	Minimum data needed? Consent obtained?
Accountability	Who owns failures?
Safety	Does it fail safely? Human override?
Beneficence	Who benefits? Who’s harmed?

Phase 1 → Phase 2 Bridge

You now understand:

What AI is and where it fits
The AI → ML → DL → Transformers → LLMs hierarchy
How learning happens (data, loss, backprop, inference)
The full AI lifecycle
Real-world applications and limitations
Ethical considerations

Phase 2: Machine Learning dives into the algorithms themselves — how models actually learn from data, what supervised vs unsupervised vs reinforcement learning looks like in code, and the math behind the patterns.