AI or not? You decide
A calculator app adding your CT marks
Google Maps rerouting you around Mirpur 10 traffic
Gmail pushing an email to spam
An AC that switches on above 26 °C
Deep Blue beating Kasparov at chess, 1997
Face unlock on your phone
An Excel sheet that turns marks into CGPA
A washing machine with a "fuzzy logic" sticker
A chatbot writing your assignment for you
Four ways to define AI
Two axes: do we care about thinking or acting, and do we measure against humans or an ideal of rationality?
Acting rationally
Do whatever maximises the expected outcome, given what you perceive and know. This is the approach of this course and of Russell & Norvig.
Examples: Self-driving cars, AlphaGo, route planners, insulin pumps.
On screen: JARVIS: never "human", always useful.
Rational ≠ emotionless robot. It means doing the right thing given what you know. A rational agent can still be wrong if its information was wrong.
The imitation game
“What did you have for breakfast?”
“Write one line about rain in Dhaka.”
“What is 7,493 × 28?”
1950: Turing swaps the fuzzy "Can machines think?" for a test. A judge chats by text with a hidden human and a hidden machine. Can't tell which is which? The machine passes.
Ex Machina is a Turing test with the stakes turned up. Its twist: passing the test says nothing about whether the machine shares your goals. Hold that thought for Week 12.
Passing isn't understanding? The Chinese Room
Today's chatbots arguably pass short Turing tests. Is an LLM just a very large Chinese Room?
Argue it in pairs for two minutes. There may be more than one defensible answer.
AI is built from eight other subjects
Philosophy
Can reasoning be written down as rules? Where does knowledge come from?
Logic & expert systems · W2Mathematics
Logic, probability, what is computable and how fast.
Search · W3–4, Bayes · W8Economics
Utility, decisions under uncertainty, game theory.
Utility agents · W1, Minimax · W5Neuroscience
How do networks of simple neurons produce thought?
Neural networks · W11Psychology
How do humans perceive, learn and decide?
Reinforcement learning · W9Computer engineering
Hardware to run it all: vacuum tubes to GPUs.
Why deep learning took off in 2012Control theory
Feedback loops that keep a system on target.
Agents · next classLinguistics
How does language carry meaning?
NLP & LLMs · W10, W1280 years in one scroll
AI wins Nobel prizes
Physics to Hopfield and Hinton for neural networks; Chemistry partly to Hassabis and Jumper for AlphaFold.
Narrow AI, general AI
Narrow AI
Superhuman at one task, useless outside it. A Go engine cannot fold laundry. Every deployed system you use today is narrow, including very capable chatbots (text in, text out).
General AI (AGI)
Human-level ability across almost any task, learning new ones quickly. Whether and when it arrives is one of the most argued questions in the field. We revisit it after we understand how LLMs work.
Worked example: is a smart traffic light AI?
Given
An adaptive traffic signal at Farmgate, Dhaka.
Find
Is it an AI system? Which of the four definitions fits? Use the four-question test.
Method: Russell & Norvig Ch. 1–2 (definitions, agents)
Solution
Predict the answer first. Then reveal one step at a time.
Your turn · 1: a Bangla voice assistant
A Bangla voice assistant on your phone sets alarms and answers questions.
Is it AI? Justify with the four-question test and name its quadrant.
Work it out first.
Press Solve when the class has an answer.
Your turn · 2: two quick verdicts
A washing machine that always runs a 40-minute cycle when you press Start.
Work it out first.
Press Solve when the class has an answer.
A diagnosis assistant that reads symptoms and lab values and ranks dengue, malaria and chikungunya by probability.
Work it out first.
Press Solve when the class has an answer.
Critical thinking · the AI effect
In 1997 beating the chess champion was ‘AI’. Today a free phone app does it and nobody calls it AI. Why does every solved problem stop counting?
Argue it in pairs for two minutes. There may be more than one defensible answer.
Your friend says: ‘If it doesn’t use machine learning, it isn’t real AI.’ Give two counter-examples from today.
Argue it in pairs for two minutes. There may be more than one defensible answer.
Critical thinking · design a Turing test for 2026
Chat-only tests are now easy to pass. Propose three changes to the Turing test so it measures something closer to intelligence in 2026, and one risk your new test creates.
Work it out first.
Press Solve when the class has an answer.
From our lab: AI for tropical-fever diagnosis
Dengue, malaria and chikungunya look alike in the first days: fever, aches, fatigue. In low-resource clinics lab tests are slow or unavailable. At BUET's mHealth Lab, your lecturer's team built mLabLLM: a fine-tuned LLaMA 3.2 language model combined with Bayesian reasoning, reaching 82.8% top-3 accuracy for these three diseases (AAAI 2025).
Why it matters today: a plain chatbot is "acting humanly" and may confidently invent answers. Adding explicit probability moves it toward "thinking rationally".
You will learn both halves in this course: Bayes in Week 8, LLMs in Week 12.
Pop quiz
Question 1 / 5
Which approach to AI does this course (and Russell & Norvig) mainly follow?
Three things to remember
Four definitions: thinking or acting, humanly or rationally. This course takes the rational-agent view.
Behaving intelligently ≠ understanding (Chinese Room). AI is bigger than ML: search, logic, probability, planning and learning all count.
History runs in hype cycles. Next class: what exactly an agent is, and PEAS.
Find one AI system used in Bangladesh (an app, a bank, a hospital, a farm). In five lines: what it perceives, what it chooses, its goal, and which quadrant it belongs to. Bring it: we use your examples next class.
Work it out first.
Press Solve when the class has an answer.