AI Algorithm Blocks (Search, Uncertainty, Learning)

Context: FIT1061_MOC · the unit’s replacement for the word “AI” (The AI Effect (Defining AI)) — every algorithm answers exactly one of three questions

Quick Revision

• 🎯 Objective: match an algorithm to its question (search · uncertainty · learning) ➔ recover its block’s recipe, flagship system and week.
• ⚠️ Key Constraint: the blocks are questions, not techniques ➔ a system spanning all three (an LLM) is classified component by component, never as a whole.

📝 Core

Block	The question	The recipe	Flagship milestone	Weeks
A · Search	the answer is hidden in a space of possibilities — find it without checking everything	network proposes promising moves → tree search explores consequences → self-play feeds the network better data	AlphaGo (DeepMind, 2016) beat Lee Sedol $4$–$1$ at Go	W2–5
B · Uncertainty	the right answer is probabilistic — estimate it from data and act on the estimate	estimate $P(\text{outcome}\mid \text{data})$ from past examples → combine features into one decision → audit who it gets right and wrong	Watson (IBM, 2011) beat the Jeopardy! champions; same machinery runs COMPAS	W6–9
C · Learning	the right rule is unknown — extract it from many examples	millions of labelled examples → define a loss → nudge the weights downhill billions of times	AlphaFold (DeepMind, 2024 Nobel) predicts any protein’s 3D shape	W10–12

• Block A tools ➔ BFS/DFS (W2) · greedy best-first (W3) · A* · hill climbing (W4) · minimax on game trees · consolidation (W5).
• Block B tools ➔ probability & Bayes’ rule (W6) · Naive Bayes, the spam-filter algorithm (W7) · weighted sums / perceptron, the unit cell of every neural net (W8) · fairness metrics (W9).
• Block C tools ➔ gradient descent in 1D and 2D (W10) · perceptron training · convolution, how machines see (W11) · deployment & model cards (W12).
• The same recipes scale ➔ A drives AlphaProof (IMO silver 2024), Gemini Deep Think (IMO gold 2025) and every planning agent (Cursor, Claude Code); C drives ChatGPT, Tesla Autopilot and GraphCast weather forecasting.
• LLMs use all three ➔ A agent wrappers search over actions and reasoning models over chains of thought · B next-token prediction is a conditional probability · C gradient descent trained the weights, RLHF polished the behaviour.
• Critique refrain ➔ ask of every algorithm built and every system met outside the unit: (1) how does it work? (2) when does it fail? (3) who does it affect?

⚠️ Common Mistakes

• 💡 Classifying the product, not the component ➔ “ChatGPT is Block C” drops the search and probability machinery that make it usable.
• 💡 Reading blocks as a difficulty ladder ➔ they are three independent questions, and Block A’s 1959 shortest-path algorithm still runs in production.

🧠 Active Recall

A spam filter and AlphaFold both "learn from data" — why is one Block B and the other Block C?

Answer

• Short answer: they differ in what is being computed, not in whether data is used.
• Why: Block B estimates a probability ➔ counts past examples to get $P(\text{spam}\mid \text{words})$, then acts on that estimate — the decision rule is given, the numbers are filled in. Block C searches for the rule itself ➔ a loss is defined over millions of labelled examples and the weights are nudged downhill until the mapping emerges.