Introduction to Artificial Intelligence
This module opens the course with the big questions: What is AI? What is intelligence? Where did the field come from, and where is it going? The material covers the course structure and evaluation scheme, a short history of artificial intelligence, the disciplines that built it, the four competing ways to define AI, and the risks that come with deploying it. The thread running through everything is that AI is an old field that has survived two winters, and it is currently in a boom built on data, computing power, and algorithms arriving together.
A useful way to read this module is as a map of the rest of the course. Section 1.1 tells you the rules of the game: how the course is evaluated, which books matter, and what habits separate strong students from average ones. Section 1.2 is a timeline — the field did not appear out of nowhere, and its two cold periods ("AI winters") explain both the current hype and the caution you should bring to it. Section 1.3 explains the three ingredients that finally made the field take off. Section 1.4 shows where AI already lives in the world, from classrooms to farms. Section 1.5 argues that AI is not a computer science monopoly — philosophy, mathematics, economics, neuroscience, psychology, and linguistics all built it. Section 1.6 asks the uncomfortable question: what does "intelligence" even mean, and are humans the right benchmark? Section 1.7 organizes a half-century of definitions into four clean positions — thinking humanly, acting humanly, thinking rationally, and acting rationally — and explains why the field is settling on the fourth. Section 1.8 closes with the risks: weapons, surveillance, bias, inequality, and safety-critical failures. Two short appendices summarize the exam guidance and the industry applications for quick revision.
The one-sentence version of this lecture: AI is a seventy-to-ninety-year-old field that defines itself in at least four different ways, and the definition that is winning today is the rational-agent view — building systems that perceive, decide, and act to get the best expected outcome. Keep that sentence in mind while reading; every section below either builds it or challenges it.
1.1 How the Course Works: Handout, Books, and Evaluation
1.1.1 Course Structure and the Handout
The course is a five-credit core course, which means it carries real weight. At the end of it, you should have a solid understanding of the history and the concepts behind AI: what a knowledge system is, how an AI agent infers things, how probability plays a role, and how these models optimize for both computation and results. The course handout is the governing document — think of it as the contract for the whole course. It has been revised heavily over the last five or six years, and about 30–40% of it changed this year. It lists nine modules split across sixteen sessions, plus a couple of research papers; those papers get discussed in the session where they belong, not in advance.
What the course actually builds: three capabilities come back again and again across the nine modules. (1) Knowledge representation — how you store what the system knows in a form a computer can reason with, such as logical statements or probability tables. (2) Inference — how the system draws new conclusions from stored knowledge, whether by logic, probability, or search. (3) Optimization — how the system chooses among many possible actions or answers so that it gets good results at acceptable computational cost. The sixteen sessions weave these three capabilities through every topic, from search algorithms to reasoning under uncertainty. The research papers in the handout are read inside the session where they belong, so you never have to read ahead for them.
A few practical patterns run through the course. After each session, the materials land on the course page within about ten minutes of the end of the session. Announcements, quizzes, and assignments all happen on that same page, so there is no need to track several portals or inboxes. Most sessions need no preparation — you can arrive with a curious mind and learn in the room. A few sessions do need pre-reads, and those are announced in advance; the logics module that follows the midsem is one such case, so the materials for it go up early with an announcement.
The decks used in sessions are curated rather than created from scratch: the contributors are academics across the globe who made their material open source, and the decks are modified to suit this group and delivery mode. The decks are not exhaustive, so the textbook stays the primary source. There is also a lab component: self-help experiments on a virtual platform. Every lab comes with a manual, a walkthrough video, and the code. There are no guided lab sessions — you play with the algorithms, tweak them, and run them on your own. That self-directed mode is deliberate: the fastest way to understand an algorithm is to change its parameters and watch the behavior change.
Q: Can the presentation slides be uploaded before the sessions?
A: No — in my experience, that is a double-edged sword and does more harm than good. The materials appear on the course page within about ten minutes after each session ends. For the few sessions that need pre-reads, the materials are posted in advance with an announcement.
1.1.2 Textbooks and Reference Books
The primary textbook is Russell and Norvig, Artificial Intelligence: A Modern Approach, fourth edition. It is a classic — the leading academic book on AI for over twenty years, now in its fourth edition. The Eastern Economy (Indian) edition and the Global edition are the same book; the Indian edition costs around 500 rupees. The fourth edition has a good 20–25% change over the third edition, and a few topics in this course exist only in the fourth edition, so the fourth edition is the one to get if you plan to buy it. A physical copy is worth it: the final exam is open book, and digital devices and e-books are not permitted in that exam. The book is huge — roughly 1,500 pages — and not every chapter is in this course. The handout maps each module to the relevant sections (T1 stands for the textbook), and about 50–60% of the book is this course. Some parts of it are covered in other courses instead.
Why a physical book matters in this course: the final exam is open book, but the rules permit only the listed textbooks and the class materials. E-books and digital devices are not allowed in the examination hall. That single rule turns the textbook choice into an exam-strategy decision: the copy you buy is the copy you will be flipping through under time pressure. The first chapter of the fourth edition opens with the same questions this lecture asks — what AI is, whether to define it by human performance or by rationality, and how the field's history explains its present — so reading Chapter 1 alongside this section gives you the same material in two voices.
The reference books are Mitchell — Machine Learning (the same author as the machine learning course textbook), Patterson — Introduction to AI and Expert Systems, and Rich and Knight — Artificial Intelligence. For brushing up on data structures, the standard algorithms text by Goodrich and Tamassia is a good resource; a merged deck of material covering all sixteen sessions is also available on request.
1.1.3 Evaluation Components
Exam note: the course uses the standard EC1, EC2, EC3 structure. EC1 is continuous assessment for 30 marks, built from two assignments and two quizzes. The two assignments are group work — one before the midsem for 12 marks and one after for 13 marks. Groups have four or five members; there is a window to form your own group, and anyone who misses the window is randomly allocated. The same group stays for both assignments. Group credit is contribution-weighted: the assignment announcement spells out how people who did not contribute are not given the same credit as the rest.
The assignments are deliberately designed for groups. They are situated and experiential rather than trivial programming tasks — the course is a work-integrated program, so the assignments carry a work-integrated component, and only a group of people working together can do them justice. There are two quizzes, and the better of the two scores counts — that is specific to this course, not a general rule. The quizzes stay open longer than the usual single day, around three or four days, to avoid the chaos of missing the window.
Exam note: the midsem exam is closed book, covers the first eight contact-session topics, and is worth 30 marks. The final exam is open book, covers all topics, and is worth 40 marks. Everything is common across the whole batch — the same question paper, the same assignment pool (a set of problem statements; each group works on one of them), and the same discussions.
Exam note: for the open-book final, only the listed textbooks and the class materials may be carried in. Handwritten notes are not allowed under the current policy; the examination cell communicates any changes. A watermarked, one-shot PDF of the decks is posted around the fourteenth or fifteenth session so you can print it and carry it in along with the books — there is no point printing the decks earlier, because the working versions are not watermarked.
Assumption: this evaluation scheme assumes you work steadily rather than in bursts. The 12-mark assignment lands before the midsem and the 13-mark one after, so the workload is spread across the whole semester; a student who treats the course as "exam week only" loses marks from quizzes and assignments they cannot recover in the finals. The open-book final also assumes you know the book's layout in advance — the strategy session on using the book comes before the exam, not during it.
Q: We tried an open-book exam before and the book did not help — we did not have enough time. What should we do?
A: We will talk about that. Short breaks are built into the sessions, and I will explain how to make the best of the open-book final — how to use the book, how to organize it, and how to manage your time.
Q: What if we want to do the assignments alone?
A: That is not possible, and there is a reason: these assignments are designed to take a good month of work, so alone they are too heavy. We are not running a company alone — group work mirrors the workplace and supports peer learning. The assignment announcement handles the sleeping-partner problem, so people who do not contribute do not get the same credit.
Q: Are our own notes allowed in the open-book final?
A: No. Only the listed textbooks and the class materials can be taken in. A watermarked copy of the decks is posted around the fourteenth or fifteenth session, and that printout can be brought along with the books. Handwritten notes are not allowed as per the current policy.
Q: Do we have to keep the same group for both assignments?
A: Yes, the same group for both. Groups of four or five members can be self-created during the announced window; anyone who misses the window is randomly allocated.
1.1.4 Ground Rules and Study Habits
The study advice is concrete. Regularity matters: a lot of learning happens live, through interaction, and if you skip sessions the backlog piles up and lands on you at exam time. If you do miss a session, watch it rather than letting it stack. Mental presence matters as much as physical presence: you can attend from home or from work, but the value comes from observing, listening, and responding while you are there. The sixteen sessions are tightly knit, so staying in touch with the course week by week beats bingeing everything two days before the exam.
Discussion forums are a first-class resource. The question-and-answer forum is for course content; a general discussion forum is for broader topics such as the field, the job market, and groups. Both are persisted, so a question asked in week one still benefits someone in week twelve. From the second or third session onward, most decks end with one or two small ungraded questions worth 15–20 minutes; they are for yourself, to discuss on the forum — not an assignment and not graded.
The final habit is going the extra mile. The framing was: "one to the power 365 is 1, but 1.01 to the power 365 is 37.8."
The compounding example, worked out. Start with a baseline effort of . If you do nothing beyond the baseline, after 365 days you still have — a zero edge stays zero. Now add a daily edge of (a 1% improvement) so that each day you work at times the previous day's level. After one year:
- The repeated multiplication is (365 times).
- Using logarithms: , so .
- Therefore .
The final answer: — a 1% daily edge multiplies your year's output by roughly 38. Sense-check: a 1% daily gain compounds to about 37.8× in a year, which matches the familiar "1% better every day" claim; and if the edge were 0, the result stays exactly 1. The numbers 1, 1.01, and 37.8 are exactly as given in the session — the daily delta here is , not 0.1.
A tiny daily edge compounds across the year, and a zero edge stays zero. More than a lakh of people across the country are doing programs like this, so the advice is not to treat the degree as a box to tick: make friends, make connections, learn from each other. The degree carries the same value as the on-campus program — the same handout, the same deliverables, the same rigor — and there are many examples of alumni who went the extra mile and are doing extraordinary things now.
Q: How do we maintain this energy till the last session?
A: Just stick on with the course. The energy comes from the sessions — I will bring it, and if you keep attending and engaging, that is good enough.
1.1.5 Prerequisites and Course Boundaries
Useful backgrounds are probability and statistics, a little Boolean logic, some algorithms, and basic coding; if any of these are missing, you will pick them up along the way through exercises, assignments, and the sessions themselves. The official prerequisite is data structures, mainly because the assignments need them. The unofficial prerequisite is the willingness to learn whatever background you lack — which is also why the assignments are group work: not everyone arrives with a strong computer science background.
This is a core, fundamental course on AI. It introduces the breadth of AI and its principles: how AI agents work, the rationale behind them, knowledge representation, inference, and where probability comes in. The topic list includes uninformed and informed searches, optimizations, knowledge representations, reasoning, and a bit of ethics. It is also worth being clear about what this course is not. It does not touch hardware aspects of AI design, which is a separate field. It does not give a formal introduction to machine learning or neural networks — those sit in parallel or later courses. It does not deep-dive into specific fields like computer vision or LLMs. And it does not cover agent toolkits — the MCP protocol, agent-to-agent protocols, how LLMs work, and tool calling — which live in a companion course on software engineering for machine learning. Those topics sit on a surface layer built with frameworks; this course is about the deeper layer: how things work and how knowledge is represented.
Where this course stops (and why): the boundary is drawn between "using AI tools" and "understanding AI systems." Tool-calling protocols such as MCP, agent-to-agent communication, and prompt-based frameworks change quickly and are taught in the companion software engineering course. This course instead builds the stable layer underneath: search, knowledge representation, reasoning, probability, and the rational-agent model. The rule of thumb: if a topic is a framework you install, it is out of scope; if it is an idea you can reason about on paper, it is in scope. Hardware design (chips, accelerators) is its own discipline and is also out of scope.
Q: Will there be a separate doubt-clearing session?
A: Not a separate session. The webinars are typically run by learning facilitators, and I will also attend them for this course, so doubt clearing happens there. The discussion forum is the best place for questions — it is persisted, and facilitators, peers, and instructors answer there.
Q: Can you share the data structures notes?
A: Yes. The notes are already open source on the internet, and I have a merged version covering all the sessions — I will upload that.
1.1.6 How the Sessions Run
A useful distinction: pedagogy technically means teaching children, while andragogy means teaching adults — and since everyone in this program is an adult, the course follows andragogy. The rhythm is a weekly live session, with 99% of them not rescheduled; in rare cases such as conferences, a buffer session is used. There are four webinars across the course, used for lab implementation and problem solving, and the pre-created lab modules are explored on your own time. Learning is a social activity, so the discussion forum is heavily used.
There are also think topics: each week a thread is started and pinned, and the learning facilitators provoke discussion around it. These threads are not directly related to the exam — engagement is a byproduct of learning, not the other way around — but participation is tracked, and the top-k engaged members might get surprises: AI comic books, AI books, opportunities to work on research, or watching an AI sci-fi movie together.
Recap and bridge: this course is a five-credit core course with a clear contract — nine modules over sixteen sessions, EC1 continuous assessment (30 marks) plus a closed-book midsem (30) and an open-book final (40). The habits that matter are regularity, forum participation, and a small daily edge: . With the rules of the game set, the next section asks the field's founding question — can machines think — and walks through the history that answers it.
1.2 A Brief History of Artificial Intelligence
Hook: AI feels like a product of the last five years — yet the field has been running for 80 to 90 years, and it has already died twice. Before ChatGPT, before Deep Blue, before anyone owned a computer, researchers were asking whether machines can think. The history matters because the boom you are living through is not the first one, and it will not be the last.
1.2.1 The Birth: Turing's Question and the Dartmouth Workshop
AI is not new. The field has been around for 80 to 90 years, and it only boomed in the last half-decade. In 1950, Alan Turing started this whole question: can machines think? He also proposed a definition — a computer would deserve to be called intelligent if it could deceive a human into believing it was human. Keep in mind the era: in the 1950s, computers existed only in large laboratories and government facilities; ordinary people did not own computing devices. Even so, the vision was already there. For the next five or six years, "can machines think?" was a hot topic, and people started asking what intelligence is and what happens if machines start thinking.
The term AI was not born until 1956. In that year, four scientists proposed a summer meeting to discuss artificial intelligence and coined the term. The four organizers were John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon — McCarthy, a young mathematician at Dartmouth, convinced the other three to join him, and their 1955 proposal requested funding for accommodation, food, campus resources, and salaries for a two-month summer study at Dartmouth College. The proposal's topics included automatic computers, how a computer can be programmed to use language, neurons, the theory of calculation, self-improvement, and abstractions. The proposal is available online, and it is a fascinating read. The summer-school tradition was already established — universities host summer schools on one subfield, such as discrete math, AI, or architectures, where interested students and researchers meet and discuss.
The founding conjecture: the Dartmouth proposal rested on one bold sentence — that every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can be made to simulate it. In practice the workshop produced no breakthrough; its most durable outcome was the Logic Theorist, a theorem-proving program by Allen Newell and Herbert Simon that could prove theorems from Principia Mathematica and whose step-by-step reasoning matched how human subjects solved the same problems. That workshop, however, gave the field its name, and the name stuck.
After that, the field gained the attention it deserved. The first perceptron was built, the first AI programs were written, and ELIZA appeared — a chatbot that could respond to questions. People went wild: the machine is thinking, it answers me. ELIZA is actually still alive: at the AAAI conference (often spoken as "triple AI"), which thousands of people attend, ELIZA was revived around 2023–24 and kept in a box for the next generation to look at. In that era, many proposals and projects around AI — including defense projects — got funding just by having an AI flavor.
1.2.2 The First AI Winter (1973–1980)
From 1973 to 1980, AI went through its first bad winter. After the ELIZA success of the mid-1960s, proposals with AI in them were getting funded, but the projects started failing heavily: you gave a natural-language system a prompt and got responses that were not up to the mark. The belief spread that AI people were fooling everyone — hardcoding rules and calling it intelligence. AI became a bad word, and the funding dried up.
Why the bubble burst: the pattern is worth naming because it repeats. Enthusiastic researchers promised machines that would translate language and solve problems; the early demonstrations worked on tiny, hand-picked examples; and when the same programs met real text and real problems, they collapsed. In Britain, the 1973 Lighthill report documented the failure to scale past "microworlds" and led the government to cut AI research funding to almost nothing. The lesson: a program that works on one toy example is not a working system — scaling is where AI projects historically died.
1.2.3 Recovery, the Second AI Winter, and the "Bad Word" Problem (1980–1990s)
In 1980, things slowly started picking up: convolutional networks, parallel processing, and small bots gave more hope, and from 1980 to 1987 there was stability — the AI conferences (AAAI and others) were at their peak again. This era also produced the first genuinely useful AI products: expert systems such as DENDRAL (inferring molecular structure from mass spectrometry data), MYCIN (diagnosing blood infections), and the commercial system R1, which configured computer orders and saved its company an estimated 40 million dollars a year. At the same time, something interesting happened: people started disassociating themselves from the name AI and moving into specifics. "I am doing machine learning, not AI. I am doing deep learning, not AI." The label AI had become a bad word — these are highly ambitious people who keep proposing, but nothing works.
Then, from 1987, the second bad winter started: proposals around backpropagation and parallel processing kept ending in failures, and the expert-systems market that had grown to billions of dollars collapsed as companies failed to maintain the systems. Even during these times, people completed PhDs in AI. One anecdote from the field: a faculty member earned a PhD in the 1980s working on symbolic AI and logic, at a time when conferences dismissed the field — nobody cared about it. That is what the early 1990s looked like.
The label lesson: a researcher who says "I do machine learning, not AI" is not describing a different field — they are describing the same field under a name that avoids the hype tax. The label AI carried promises nobody could keep, so practitioners rebranded into specifics: machine learning, deep learning, computer vision, natural language processing. When you see the same rebranding today ("agentic systems", "large language models"), recognize it as a healthy habit: precise names keep expectations honest.
1.2.4 The Climb Back: RoboCup and Deep Blue (1997)
In the early 1990s the field gained popularity again. The RoboCup — sports for robots — started running annually: football and other games for robots, with robots carrying a bunch of AI-powered features, and AI acting as an umbrella term over many components. LSTM and ANNs were getting popular, confidence was stabilizing, and people were again believing in AI.
In 1997, something remarkable happened: Deep Blue. IBM developed this expert system specialized in playing chess, and it defeated Garry Kasparov — the world champion — in chess. Kasparov said in an interview that he could feel human-level intelligence in the room. The media went wild: for the first time, a machine defeated a world champion. A lot of people believed Deep Blue made a stupid move at one point — a move any human could figure out. It seemed like a trivial move for a machine. But retrospectively, it was that move that eventually led Deep Blue to win the game. The same pattern appears with AlphaGo. The point: the move seemed unintelligent, but it was the move that won. After Deep Blue, the field had momentum again — AI is still promising.
The Deep Blue lesson — intelligence need not look like human reasoning: chess grandmasters judged one of Deep Blue's moves to be a blunder; any human player would have chosen differently. Yet that move steered the game to a win. The machine was not imitating human style — it was computing toward a win with its own machinery, and the winning computation looked alien. The same thing happened with AlphaGo's famous move 37 against Lee Sedol, which commentators called a mistake until it proved decisive. Keep this lesson: later in this lecture, it becomes the core argument for why "acting rationally" beats "acting humanly" as a definition of AI.
1.2.5 The Modern Era: DARPA, Siri, Transformers (2017), and ChatGPT (2022)
From 2000 to 2012, small innovations kept happening. The DARPA challenges pushed defense robotics and autonomous cars. Interest in ANNs grew, IBM kept developing systems, and Siri arrived. Revolutions in deep learning architectures followed, and AlphaGo — a mind game system in the spirit of chess — defeated top human players. TensorFlow and DeepMind emerged. The decisive shift was data plus hardware: big data sets such as ImageNet (millions of labeled images) met specialized hardware such as GPUs, and in 2012 a deep learning network swept the ImageNet competition, cutting the object-recognition error rate by a dramatic margin — the moment the field's modern era really began. In 2017 came the paper everyone in the field knows: Attention Is All You Need — transformers happened, and from there the rest is history: GPT-1, GPT-2, GPT-3, and so on.
In 2022, ChatGPT arrived, and it is an important milestone. Until that point, AI lived mostly in the academic and research community plus a few government institutes; for the public, AI was vague — maybe machines do something intelligent. After ChatGPT, AI got democratized. Ask anyone to name an AI toolkit and they will name ChatGPT — not the only AI tool, of course, but that breakthrough made AI mainstream, entering every household. In the last five years, every week brings releases, debates, new frameworks, tools, promising results, and contradictory results.
1.2.6 Lessons from the Timeline
AI winters will return: the field has failed before — twice — and both failures followed the same recipe: hype, over-promise, disappointment, and dried-up funding. Nothing in the current boom removes that risk: energy costs, business-model shifts, and unmet expectations can all trigger a slowdown. The realistic stance is not "AI is unstoppable" but "AI is bound to stay" — it will have bad phases and good phases, and it will bounce back. Plan your skills around fundamentals that survive winters, not around whatever framework is hyped this week.
A picture helps: draw a timeline with years on the horizontal axis and the field's level of enthusiasm and funding on the vertical axis. The curve starts near zero in the 1940s, rises steeply through the 1960s excitement (ELIZA, early programs), crashes into the first trough from 1973 to 1980, recovers into the expert-systems plateau of the mid-1980s, crashes again after 1987, then climbs steadily through RoboCup and Deep Blue in 1997, steepens with deep learning after 2012, and shoots nearly vertical after 2017's transformers and 2022's ChatGPT. The landmarks to remember are the two troughs (1973–1980 and 1987–early 1990s) and the two inflection points where the curve changed character: 1997 (a machine beat a world champion) and 2022 (AI reached every household). The takeaway from the shape: the line never returns to zero — each recovery builds on the last one.
Recap and bridge: AI was born with Turing's 1950 question and named at the 1956 Dartmouth workshop; it survived two winters caused by over-promising, rebranded into specifics like machine learning, proved its power with Deep Blue in 1997, and went mainstream with ChatGPT in 2022. The next section asks the obvious follow-up: what actually changed in the last decade that made the field finally take off?
1.3 Why AI Took Off in the Last Decade
Hook: the field had the same goal for decades — yet the graphs of capability stayed flat while the effort stayed constant. What changed? Not the ambition. Three ingredients that had always been missing separately finally arrived at the same time: data, computational power, and algorithms.
1.3.1 Three Enablers That Came Together
Why did the last ten years pick up so fast, when the field moved slowly for decades before? The answer is three things coming together: data, computational power, and algorithms. Data: social media and the connected world started producing a tremendous amount of data in the last decade. Computational power: GPUs and high-performance chips became available at the touch of a button — even inside phones — with organizations competing to push the limits. Algorithms: researchers kept improving at the algorithm level, with software engineering and discrete math feeding optimizations. When all three arrive together, innovation accelerates. Earlier decades failed to make AI mainstream precisely because at least one of these was missing.
The three-enabler model, one by one:
- Data (the fuel): the connected world generates trillions of words, billions of images, and billions of hours of speech and video. Learning systems are hungry for examples — research found that adding two or three orders of magnitude more data improved performance more than any amount of algorithm tweaking could.
- Computational power (the engine): the compute used to train leading machine learning systems followed Moore's law until about 2012, then changed character — from 2012 to 2018 it grew roughly 300,000-fold, which works out to doubling about every 100 days. GPUs and tensor processors execute the matrix operations deep networks need, and training time for a fixed image-recognition task dropped by a factor of about 100 in just two years.
- Algorithms (the steering): better architectures and training methods — deep learning, attention mechanisms, optimization tricks — turned the same data and the same hardware into much more capable systems.
The three must arrive together. Earlier decades each had at least one missing: the 1960s had ideas but no data and little compute; the 1990s had more compute but still thin data and weak algorithms for it. The crossing point — where all three lines meet — is what the last decade marks.
Visualize the three enablers as three overlapping circles. The first circle is data (growing steeply once social media arrived), the second is compute (jumping sharply around 2012 as GPUs took over), and the third is algorithms (steadily improving). Each pair of circles alone gives modest progress — a lot of data with weak algorithms is just storage; powerful compute with little data is a fast machine with nothing to learn from. The explosion happens exactly in the central region where all three overlap: that is the current era of AI. The one-sentence takeaway from the picture: breakthroughs come from the intersection, not from any single ingredient.
1.3.2 Sustainable AI: The Cost vs. Price Question
There is a price attached to all this progress. The distinction used here: cost is what we pay in money — recharges, buying tokens, an organization's subscription. Price is what the next generations face: depleted resources, the energy needed to run these systems, and the whole ecosystem of AI. The wording "cost versus price" is a simplification for this discussion, not standard terminology; the literature calls this topic sustainable AI. There is substantial work on whether AI is sustainable if we keep doing it the way we are — covering both the money spent on tokens and the net-zero dimension.
Q: Could you share what you meant by cost versus price?
A: Cost is what we pay in money — recharges, tokens, subscriptions. Price is what future generations will face: depleted resources and the energy needed to run these systems. The literature does not use this exact wording; it appears under the name sustainable AI, for example in papers at the ICSE and AAAI conferences.
Scope of the cost–price distinction: the two words describe different accounting books, and they should not be conflated. Cost is private and immediate — the money an individual or company spends today on tokens and subscriptions; it shows up in a budget. Price is public and deferred — the energy consumed by data centers, the hardware replaced every few years, the resources drawn down, and the net-zero footprint; it shows up in the environment and in future generations' bills. The distinction breaks down if you ask "who pays": a generation's price is partly today's cost shifted forward in time, which is exactly why the field treats sustainability as a shared responsibility rather than a market price.
1.3.3 Will AI Winters Return?
Q: Is global warming the only reason AI could see new winters? Are there solutions?
A: Solutions exist, but they need many communities to come together — system engineers, problem solvers, hardware engineers, and the software engineering community. Winters have never been permanent; they are on and off, and the field always picks up again. New business models, like pay-per-call pricing, could bring a temporary slowdown, but the community is strong now.
The pricing evolution itself is a possible source of a temporary winter. For decades the model was: pay a license, get the software. Then SaaS arrived, then pay-per-use, and now it is a level beyond SaaS — not even pay-per-use, but pay-per-call: you make a call to an agent and you are charged. These business-model shifts (API-first, agent-first thinking) can shake the market enough to cause a slowdown. But the community is strong now, people are already working toward solutions, and the field will pick up again.
Assumption: the "AI is here to stay" claim rests on three assumptions that can each fail. First, that the energy and hardware ecosystem stays affordable — if the cost of running models outruns their value, adoption stalls. Second, that business models stabilize — pay-per-call pricing is still young, and a sudden repricing could freeze customers and investors. Third, that the community keeps solving the field's hard problems; if expectations again run ahead of results, the pattern of 1973 and 1987 can repeat. None of these assumptions makes AI disappear — history shows the field bounces back — but each can cause a temporary winter.
Recap and bridge: the last decade's acceleration is the story of data, compute, and algorithms finally overlapping; its price tag is sustainability (cost now versus price for future generations, studied under the name sustainable AI); and its fragility is the business model — pay-per-call pricing could trigger a temporary slowdown. The next section moves from why the field grew to what it is doing in the world: applications across every sector and the debate over whether AI is a boom or a bane.
1.4 AI Applications and the Boom-or-Bane Debate
1.4.1 Application Areas and a Thought Experiment
The field has come a long way from the days when AI was a bad word and AI researchers were not believed. Today AI appears in every other field: education, finance, healthcare, automotives, robotics, smart cities, agriculture, culinary, and transportation. The exercise for the day: take a minute, pick one of these fields — ideally the one your organization works in — and think about what AI innovations are happening there or what you could do with AI in that field. You do not need to share it; the value is in the thinking.
The debate around AI has two poles: boom or bane. People are researching both sides — in humanity, sociology, and the AI community — asking whether AI is genuinely beneficial or just serving the capitalist world. The position taken in this course: AI is bound to stay, it will augment us, and everyone will start using it. AI literacy is already entering K-12 school education. At the same time, there is a serious claim that AI might increase inequality — that is discussed in detail with the risks.
Q: The return on AI investment is not clear yet — how do you see it?
A: We really do not know; it is early days. We need to build products, not proofs of concept, before we can tell whether AI is hitting the bullseye.
Q: What about AI for people with disabilities?
A: That is my own area — assistive technologies, accessibility, and AI for people with disabilities. Happy to chat about it whenever we get a chance.
1.4.2 Education as an Example
Real-world: education has a structural problem — the teacher-student ratio is not up to the mark anywhere in the world. In India, the number of students studying computer science and AI is about 3 million, and teacher ratios sit around one teacher for every 300 to 500 students. AI can help with hyper-personalized education: a system that knows your skill level, slowly pushes your boundaries, gives assessments, and walks you along the way. This is not about ed-tech tools that bolt on gamification; the genuine promise is real personalization at scale.
How personalized learning works as a loop: the system starts from an estimate of your skill level, picks the next exercise just beyond your current reach (hard enough to stretch, easy enough to finish), observes your answer, updates its model of you, and repeats. Every student gets a different path through the same syllabus. That loop — measure, choose, act, update — is the same pattern the course will study later under the names of agents and learning, which is why education appears so often as an example in this course.
Real-world: in agriculture, an institute of eminence in Punjab runs a complete AI-for-agriculture lab with crores of funding, doing AI-assisted agriculture because of its geographic location. Sports is another area, of course — the RoboCup videos show robots playing football (one falls down, and it keeps going), and the field is constantly improving.
Real-world: in networking, there is already good work on proactive network monitoring — AI tools that determine network health in real-time optical networks. A broader piece of professional advice: do not only read AI conferences. Go to the core conferences of other fields and look for AI papers there — software engineering (AI for software engineering, software engineering for AI), networks (AI for networks, and networks for AI, because massive AI needs better networks), software architecture, computing education, and HCI. That is where you see how AI is influencing each field. Asked which fields will improve most through AI ten years from now — governance, education, and others — the answer is that there is potential in all of them, and each field's first movers and thought leaders will start dominating their own core areas.
1.4.3 Vibe Coding vs. AI-Assisted Engineering
Q: Vibe coding is getting a lot of hype in the industry. How will it impact things, positively or negatively?
A: Vibe coding is bad; AI-assisted engineering is good. Vibe coding means anyone with an idea keeps prompting until they get something and believes the solution is done — you are hitting in the dark. AI-assisted engineering is done by people who know computing: how a website works, networking, optimizations, accessibility, usability, and software architecture. They use AI as a companion — for code review, faster deliveries — and they understand what the system produces. Vibe-coded systems do not follow software engineering practices and will not scale to a hundred million users.
The distinction in full: vibe coding is done by anyone and everyone. You have something in mind, you keep prompting until you get the result you want, and the moment you get it, you believe your solution is done. In reality, vibe coding is little more than tool calling and pulling code from existing repositories. If you do not understand computing and just keep prompting, the results can be dangerous — vulnerabilities sneak in because you do not know what you are doing. AI-assisted engineering, by contrast, is done by people who know engineering: they understand websites, networking, optimizations, accessibility, usability, software architecture, and how to solve a problem for scalability. They use AI to augment their work — code review, faster delivery — and they know what the AI is producing. That is a companion, not a magic box.
| Dimension | Vibe coding | AI-assisted engineering |
|---|---|---|
| Who does it | Anyone with an idea, no computing background needed | Engineers who understand computing, networks, and architecture |
| How it works | Keep prompting until the output looks right; stop when satisfied | Prompt, then review and reason about what the AI produced |
| What the person knows | Nothing about what the system does or why it works | How the pieces fit: websites, networking, scalability, security |
| Failure mode | Vulnerabilities sneak in unseen; systems collapse at scale | The engineer catches errors the AI cannot see |
| Ceiling | Does not follow software engineering practices; will not scale to a hundred million users | Ships production systems that survive real traffic |
The one-line choice rule: use AI to make yourself faster at work you understand — not to replace understanding itself. Vibe coding is fine for a throwaway demo; it is a risk in anything that will face real users or real attacks.
Real-world: the saying "AI will not replace you, but people who know AI will replace you" captures the same idea. Movies about AI taking over are fiction, and AI will take over some jobs — there is good research on that — but it is also creating new ones. Five years back, nobody used the term prompt engineering; now it is a field, with forward-thinking engineers building new roles, and everyone in typical roles getting augmented with AI.
The production risk: vibe-coded systems pushed into production carry vulnerabilities. A person who cannot read the generated code cannot review it, and an unreviewed system is an unsecured system — injection attacks, leaked secrets, and broken access control all slip in at the stage where nobody checks. Proofs of concept are fine as experiments; the danger starts the moment such a system is deployed to serve real users. This is the same warning that returns in the risks section under cybersecurity.
1.4.4 Indic Languages and the AI4Bharat Effort
Real-world: with respect to India, the honest assessment is that we are heavily on the consumption side, but making very good progress in AI engineering.
Q: With respect to India, are we making AI consumers or AI engineers?
A: We are heavily on the consumption side, honestly, but making good progress in AI engineering. The AI4Bharat initiative is a strong example, associated with Ravindran — from a leading Indian institute and nominated president of AAAI — and it taps a market the frontier models ignore: Indic languages. Because of India's diversity — so many languages — the frontier models (GPTs and others) do not perform as well in non-English languages as they do in English. India has responded with Indic models for Hindi and regional languages.
Q: For other languages, model token costs will be quite high, right?
A: Yes, but costs come down once the models are at higher scale. We are making good progress there.
Q: Won't everyone learn English anyway? What is the advantage of training models in Indian languages?
A: People still post and comment in their native languages on social media. If you want to mine that data, an English-only model is not enough. And each language has a vast literature — poetry and literary work in Hindi, Tamil, and many others. How would an English-only model train on that?
1.4.5 Jobs, Unemployment, and Inequality
Q: We already have unemployment. Will automation and AI increase the unemployment rate, and how do we tackle it?
A: I do not believe the unemployment rate will increase — inequality will. Some groups become ultra-productive and wealthy through AI while others lack access, and that gap grows if we do not act quickly. We will return to this in the risks discussion.
Recap and bridge: AI already touches every sector — education, agriculture, networking, and beyond — and the honest debate is boom versus bane. The practical stance taken in this course: use AI as a companion you understand (AI-assisted engineering), not as a magic box (vibe coding), and expect the real social cost to show up as inequality rather than unemployment. The next sections dig into what AI actually is — first the disciplines behind it, then the question of intelligence itself.
1.5 AI Is an Umbrella Field: Contributing Disciplines
Hook: who invented AI — computer scientists? The honest answer is that almost nobody invented AI alone. Philosophy asked whether minds can follow rules, mathematics built the language of reasoning, economics decided what "best" means, and engineering supplied the machines. AI is an umbrella, and under it stand eight very different disciplines.
1.5.1 The Fields Behind AI
AI is an umbrella term. It is a reality today because of contributions from many fields: philosophy, mathematics, economics, neuroscience, psychology, computer engineering, control theory, and linguistics. The textbook has a good section on each of these in detail. Exam note: this material is interesting and worth reading — how math contributed, how economics contributed, what the psychology of AI is — but it is not from the exam point of view.
The meta point: it is not just computer science people who built AI. There is a common notion that "we own this, we created AI." Not really. A lot of fields contributed to the existence of AI, and if you come from a math, economics, or psychology background, there is interesting work for you in this field.
1.5.2 What Each Field Contributes
Philosophy is a very old field, in existence roughly since Aristotle. It produced the concepts of logic and reasoning, and the idea of the mind as a physical state that operates with rules. Philosophers cracked these questions a thousand years back, and they are still valid in AI: how does reasoning work, what is the foundation of learning, what is rationality, what is language. Aristotle was among the first to codify "right thinking" — his syllogisms gave patterns of argument that always yield correct conclusions from correct premises, and his analysis of action (goal plus knowledge of outcomes leads to action) is a two-thousand-year-old sketch of an agent.
Mathematics gave propositional logic and first-order logic — how you infer things from a bunch of statements — and computational complexity theory. George Boole formalized propositional (Boolean) logic in the 1840s, and Gottlob Frege extended it with objects and relations into the first-order logic used today. Mathematics also gave probability, the tool for reasoning under uncertainty, and the theory of computability and complexity — including the result that some problems need exponential time, which shapes what AI can and cannot do.
Economics gave the formal theory of decision making, Markov decision processes, and game theory. It studies the decisions we make, the probabilities involved, the outcomes, whether a decision is viable — the return on investment. A famous concept is the Nash equilibrium: you always try to optimize for a balance. There is a beautiful comic about it — two prisoners put into jail — and these thought experiments all feed into modern AI.
The economics gift, unpacked: decision theory (probability plus utility) tells a single decision maker what to do under uncertainty; game theory handles many decision makers whose choices affect each other; and Markov decision processes, formalized by Richard Bellman in the 1950s, handle sequential decisions where the payoff comes from a chain of actions, not one move. The Nash equilibrium names the stable point of a game — a combination of strategies where no player can improve their outcome by changing alone. All three ideas reappear later in this course when agents must choose actions under uncertainty and in the presence of other agents.
Neuroscience is the study of brain functioning: brain imaging shows how the brain functions internally, and the results get compared with modern-day machines. Broca's discovery of the brain region responsible for speech production, and the mapping of senses to brain areas, gave the field its first hard evidence that mental functions live in physical structure — the same assumption a neural network makes, with weights playing the role of connection strengths.
Computer engineering and computer science built the powerful machines themselves — the hardware, the architectures, the software architecture, and the algorithms. Without this contribution the other seven would have only thought experiments; Moore's law and, later, specialized chips such as GPUs turned theory into systems.
Control theory and cybernetics contributed the idea of feedback: systems that sense their own state, compare it with a goal, and correct the difference — the steam-engine governor and the thermostat are early examples, and the same loop drives modern robots. Linguistics contributed the formal study of language: the insight that grammar is a rule system was decisive for getting machines to process natural language, and it grew into computational linguistics.
1.5.3 Cognitive Psychology vs. Cognitive Science
Psychology contributes two related but distinct fields. Cognitive psychology treats the brain as an information-processing machine. Cognitive science asks how computers can model that — how they can be used to study language, memory, and thinking from a psychology perspective. The two are interesting parallels.
| Dimension | Cognitive psychology | Cognitive science |
|---|---|---|
| Question | How do humans and animals actually think and act? | How can computational models explain thinking? |
| Method | Experiments on people: reaction times, errors, introspection | Building computer models and comparing their behavior with human data |
| Direction | Brain as an information-processing machine | Computers as models of that machine — language, memory, thinking |
| Classic example | Studying how people make judgment errors | The General Problem Solver, whose reasoning steps were matched against human subjects' |
The relationship is a loop: cognitive psychology produces evidence about how people think, and cognitive science encodes that evidence as programs; when a program reproduces human behavior, it supports the theory. In modern AI the loop runs in reverse too — researchers probe how large language models behave, using methods borrowed from psychology.
Real-world: researchers in psychology run thought experiments where the same problem is given to all the leading frontier models — Gemini, Claude, GPT — to probe their cognition: how they were trained, what their model is, how they behave. That is a genuine research activity today.
1.5.4 Why This Matters
If you came from a non-CS background, there is a lot of interesting material for you in this book section: how each discipline contributed to AI, with literature and research papers on the contributions. The field is interdisciplinary by nature, and that is worth internalizing early.
Recap and bridge: AI is the intersection of eight fields — philosophy (rules for thought), mathematics (logic and complexity), economics (decisions, games, and sequential choice), neuroscience (the brain as evidence), psychology (mind as information processing), computer engineering (the machines), control theory (feedback), and linguistics (language as a formal system). Whoever you are, one of these doors opens into AI. The next section asks the question all eight disciplines circle around: what actually is intelligence?
1.6 What Is Intelligence?
Hook: you are reading this because a machine might be intelligent. But before asking whether machines can think, the lecture asks a harder question: what does thinking even mean — and is the ability to think enough? The answer the course commits to: intelligence is the ability to perceive and act in the world.
1.6.1 Definitions of Intelligence
The dictionary definition of intelligence is: the capacity of learning, reasoning, understanding, and similar forms of mental activity. In other words, it is the ability to perceive and act in the world.
The definition, unpacked piece by piece. Each capability in the definition maps to something concrete:
- Learning and adaptation: improving with experience. Recommending movies means learning someone's preferences and adapting future picks to them; learning traffic patterns lets a system adapt its route advice.
- Reasoning: drawing conclusions from what is known — proving theorems, doing medical analysis, working out implications.
- Planning: taking decisions — deciding how to go about something before acting.
- Understanding: not just reading text — speech, visual and non-visual communication all count.
- Perception and action: taking in the world through senses and doing something about it.
So the working definition: intelligence is the capacity of learning, reasoning, understanding, and similar mental activity — the ability to perceive and act in the world. Note the two halves: the mental capacities (learning, reasoning, understanding) and the behavioral half (perceiving, acting). The next subsection argues the second half is what makes the first half matter.
1.6.2 Thinking Without Action Is Useless
Is intelligence alone enough? Take a philosopher. They learn, they reason, they understand — but is that alone helpful to society? Consider a philosopher who does nothing but eat, sleep, and think. You might claim they are intelligent because they can think a lot. But unless they take an action — produce a research paper or a book — how would anyone use their work? How would we even know they were thinking? Maybe they were just sleeping.
The conclusion: thinking, learning, reasoning, understanding, and adapting are all important, but they are not enough. Action is important. Thinking without action is useless. And in the context of all these other capabilities, the ability to perceive and act in the world is what we call intelligence.
The observer's test: there is a practical consequence of the action requirement — intelligence that cannot be observed is undecidable from the outside. A mind that never acts leaves no evidence, and an outside observer cannot distinguish "thinking hard" from "not thinking at all." This is why every workable test of intelligence, from the Turing test later in this lecture to a lab exam, evaluates behavior: answers, moves, and outputs. If you cannot detect it by watching action, you cannot verify it at all.
1.6.3 Are Humans Always Intelligent?
Three questions were put to the room: Are humans intelligent? Are humans always intelligent? Can non-human behavior be intelligent?
Q: Are humans intelligent? Are humans always intelligent? Can non-human behavior be intelligent?
A: The first is still open — opinions differ. The second has a clear answer: no. We make mistakes, take stupid decisions, and regret them, then improve in retrospect — so humans are not always intelligent. The third is also yes: a dog that perceives a stranger in the neighborhood and acts by barking alerts the people around — perceiving and acting is intelligent behavior by the definition we chose, so dogs can exhibit intelligent behavior. How much of that comes from training is a separate question.
The dog example works because of the definition: intelligence is the ability to perceive and act. A thief enters a society the dog has never seen; the dog perceives it and acts by barking to alert people. Non-humans exhibit intelligent behavior under this definition, and humans do not always exhibit it — we agreed on that. Whether humans are intelligent at all remains an open question; if you are optimistic about it, you belong to the AI community.
Pitfalls with the definition of intelligence:
- Treating intelligence as always-on: humans make mistakes, take stupid decisions, and regret them — intelligence is not a constant property, so "is intelligent" and "is always intelligent" are different questions with different answers.
- Reserving intelligence for humans: under the perceive-and-act definition, a barking dog qualifies; the definition never says intelligence must be human.
- Conflating definition and evaluation: "perceives and acts" defines intelligence; it says nothing about how well the behavior was learned — training versus innate skill is a separate question.
1.6.4 Early AI as Mimicry
There is an underlying question in all of this: is replicating human behavior what intelligence means? A lot of early AI work was built on exactly that assumption. In the 1970s and 1980s, the goal of the people working in AI was to replicate human behavior, and that was the early hallmark of intelligence. A machine can play chess and beat a human — very good, that is intelligent. It can do something humans do — that is intelligent. The early metric was comparing the machine to human behavior: if it can do what I can do, this is intelligence.
Note that this is about the early days, and it applies regardless of the underlying method — symbolic AI, logic-based AI, reinforcement learning, deep learning — whatever training or algorithm is used, if it mimics a human, it was called intelligent. AGI is a much more recent concept: general intelligence that can surpass us and be one among us. That is a different discussion from the early mimicking thought.
Recap and bridge: intelligence is the capacity to learn, reason, and understand, made real by the ability to perceive and act — thinking without action is useless, humans are not always intelligent, and non-humans (like the barking dog) can be. Early AI equated intelligence with mimicking humans, regardless of method. That mimicry assumption is the hinge into the next section: the four competing definitions of AI are exactly the different ways researchers answered "should machines mimic us, or do something better?"
1.7 Four Perspectives on AI: The 2×2 Matrix
1.7.1 Definitions from the Literature
Across the last 50 years of research, there is no single definition of AI that everyone agrees on. All the definitions are correct in some sense — they are different schools of thought — and some researchers have spent their whole careers working inside one of them, still publishing today under their own definition. No definition is being rejected here. But in recent years, one quadrant is emerging as more popular and more agreed upon, and we return to it after touring all four.
The definitions organize into a 2×2 matrix. Top-left, thinking humanly: "the exciting new efforts to make computers think ... machines with minds" (1985), and "the automation of activities that we associate with human thinking, activities such as decision making, problem solving, and learning" (Bellman, 1978). Top-right, thinking rationally: "the study of mental faculties through the use of computational models" (1980s) and "the study of the computations that make it possible to perceive, reason, and act" (1990s). Bottom-left, acting humanly: "the art of creating machines that perform functions that require intelligence when performed by people" and "the study of how to make computers do things at which, at the moment, people are better" (Rich and Knight, 1991). Bottom-right, acting rationally: "computational intelligence is the study of the design of intelligent agents" and "AI is concerned with intelligent behavior in artifacts."
Reading the 2×2 matrix. The four quotes sort along two independent questions, and the matrix is a 2×2 grid:
- Rows — what is being evaluated? The first row is about thought (thinking humanly, thinking rationally); the second row is about behavior (acting humanly, acting rationally).
- Columns — what is the yardstick? The first column measures against humans (the AI should think or act like a human); the second column measures against a standard of rationality (it should think or act rationally, with no comparison to humans at all).
Every serious definition of AI in the literature sits in exactly one cell. Nothing is excluded — the point of the matrix is that the four cells are four legitimate schools of thought, each with active researchers publishing today.
1.7.2 The Simplified Matrix
Simplify the same matrix: the human-like column becomes "systems that think like humans" and "systems that act like humans"; the rational column becomes "systems that think rationally" and "systems that act rationally." Those are the four ways someone can define AI: thinking humanly, thinking rationally, acting humanly, and acting rationally. Nothing new was added — the source work was just organized into four quadrants and then simplified twice.
Each group has its own argument. The thinking-humanly group believes that if a system thinks like a human, it is AI. The thinking-rationally group argues: why compare with humans at all? Humans are not always intelligent — maybe we do not even know. A system should think rationally, do what is right, and not try to mimic humans. The acting-humanly group says there is no point in thinking if you cannot do much — let it act like a human (and note, if you want to act, you usually have to think; sometimes you act without thinking, which we get to). The acting-rationally group combines both: act rationally.
Q: Thinking humanly is contradictory to thinking rationally. Some decisions no human would take are only rational solutions.
A: That is exactly the point — these are schools of thought, and each group believes their quadrant is what AI is. We will see why they conflict and which one aligns better with today once we walk through all four.
| Human-like (yardstick: people) | Rational (yardstick: what is right) | |
|---|---|---|
| Thought (row 1) | Thinking humanly — cognitive modeling: match the machine's inner reasoning to ours | Thinking rationally — laws of thought: follow logic, whether or not humans think that way |
| Behavior (row 2) | Acting humanly — pass as a person in conversation | Acting rationally — rational agents: act to achieve the best outcome, even under uncertainty |
The "when to pick which" rule for research: pick thinking humanly when the goal is understanding the mind (tutoring, HCI); pick thinking rationally when the domain is small and logic-solvable; pick acting humanly when the deliverable is human-like interaction; pick acting rationally when you need the system to actually get things done in the real world — which, as section 1.7.7 explains, is where the field has converged.
Exam note: the four-quadrant discussion of AI definitions is an important conceptual area. No single definition is right or wrong — each quadrant is a school of thought with active researchers — so expect to be asked to place definitions, not to pick a winner. The exam-relevant skill is explaining what distinguishes the two rows (thought vs. behavior) and the two columns (human-like vs. rational).
1.7.3 Thinking Humanly: The Cognitive Modeling Approach
Thinking humanly is the cognitive modeling approach. To make a machine think like a human, you first have to know how humans think — so you study cognitive science and cognitive modeling. This is a field of research hundreds of years old, and we are still not clear how humans think. If it were easy to understand how we think, there would be no fights at home — you would know exactly how the other person is thinking. Neurodiversity adds another layer of complexity.
A core finding from cognitive science research: humans rarely do everything with rationality or reasoning. A lot of the time we do something innately and then reason about it afterwards — this is post facto rationalization.
Post facto rationalization — the coffee story. Someone walks over and picks up a coffee. The action is innate. Asked why they picked coffee, they answer: "I don't like tea. If I drink tea, I won't sleep on time." They are cooking up a story at that point — rationalizing a decision they already took. The order is action first, story second. This matters for AI because it undermines the assumption that human decisions can be captured by asking people why they did things: the stated reasons are often invented after the fact, not retrieved.
To capture human thinking you can use introspection, psychological experiments, and brain imaging. Once you know how human thinking works, you can mimic or replicate it in AI systems — that is the first quadrant in practice.
But there is a bigger question here, posed as: do we want a machine that beats humans in chess, or a machine that thinks like humans while beating them in chess?
Q: Do we want a machine that beats humans in chess, or a machine that thinks like a human while beating them in chess?
A: The result is the same — the machine wins either way. What differs is the direction of research. If the goal is to think like a human, the field spends its effort on understanding humans first. Deep Blue did not think like a human and still won the game; the move everyone called crazy was the one that led to the win.
The follow-up question pressed on why the distinction should matter if the outcome is identical:
Q: If the result is the same, how does it matter?
A: It matters because the whole field aligns to the chosen direction. Aligning with the mimic approach pushes research toward human psychology and experiments; the other approach lets the machine use its own algorithms and learning to reach the outcome.
The Deep Blue lesson carries weight here: if machines must only mimic humans, they will be at most as capable as humans. Deep Blue did not think like a human, and it won. Everyone watching the game called its moves crazy — but it won. So it exhibited a thought process that achieved the goal without thinking like a human. There is a second consequence of the mimic direction: humans know how to lie and manipulate, and real-world data is biased, manipulative, and toxic; when you train machines on it because you want them to behave like humans, the machines learn those behaviors too — and they have started doing it.
An analogy that lands this: aeronautical engineering textbooks do not define the goal of their field as making machines that fly so exactly like pigeons that they can fool even other pigeons. The goal of aeronautical engineering is to make machines fly, period. The AI field keeps comparing itself to human intelligence — it should surpass it, think like a human, do things like a human — but what if machines, like airplanes, have their own mechanism for doing things?
The aeronautical analogy. Artificial flight succeeded when engineers stopped imitating birds and started studying aerodynamics with wind tunnels — the plane is not a pigeon copy, it is a flying machine with its own physics. The same argument applies to AI: if the goal is to fly, matching the pigeon's flapping is irrelevant; if the goal is intelligence, matching every human mannerism may be irrelevant too. The analogy has a limit, though: with flight the success criterion is obvious (staying airborne), while with AI the criterion is exactly what the four quadrants disagree about — which is why the debate continues.
Q: Can we simply call this out-of-the-box thinking — thinking differently from most human beings?
A: Out-of-the-box thinking is again a cognitive-science notion: something not everyone does, but a group of people does. That framing still belongs to the thinking-humanly direction, which is exactly what the first quadrant argues for.
The counterargument to the whole first quadrant: if a system always thinks like a human and does what humans do, it will be just as good as humans — and humans are not always intelligent, not always ethical, and not always rational. So why make that the target?
Thinking humanly still matters in some applications. Intelligent tutoring is one: the system has to understand how the learner thinks. Expressing emotions in interfaces is another — the HCI field. If you want a robot to express emotions, you need to model human expressions and emotions explicitly: when a person is happy, they smile; when sad, they might be silent. Those behaviors can be explicitly modeled in specialized applications. And the growth of cognitive science and AI support each other; the General Problem Solver is a paper that designed a universal problem solver, with problems represented by Horn clauses.
1.7.4 Acting Humanly: The Turing Test Approach
The acting-humanly quadrant is the Turing test approach: create machines that perform tasks which require intelligence when performed by people. Assume you have built such a machine — how do you test it? With the Turing test.
The setup: a human interrogator asks questions. On the other side of a wall sits either a computer (an AI system) or a human. The interrogator gets responses. If the responses are convincing enough that the interrogator cannot tell whether they come from a human or a computer — it is so seamless that they cannot differentiate — then the system is acting humanly, and by this definition, intelligent.
Acting humanly requires a lot of skills. The system needs natural language processing, knowledge representation, automated reasoning, machine learning, computer vision, and robotics — the same capabilities humans have. Turing himself saw physical imitation as unnecessary; a broader "total Turing test" also demands vision, speech, and robotics to operate in the real world. Either way, the skill list covers most of the AI research agenda.
Q: Who decides on the interrogator?
A: The interrogators are research scientists — judges with experience in the field, not random common people. The event takes place every year in Boston and is telecast; there are videos online of how it is conducted.
The accepted benchmark is quantitative: if 50% or more of the judges believe the responses come from a human, the system is considered to have passed the Turing test. In one early account, a chatbot was described as passing the test for the first time when 10 of the 30 judges believed the responses were human. Some of the initial models did not pass, putting them in a gray area, while some of the latest AI models and agents can pass in some cases — GPT-4.5 convinced about 73% of interrogators that it was human.
The Turing test statistics, worked out. The benchmark is 50%: the system passes when at least half of the judges are fooled. With 30 judges, that means the system needs at least 15 of 30 to believe the responses are human.
- The early chatbot account: 10 of 30 judges were convinced. That is , or about 33% — well below the 50% benchmark, which would need 15 of 30. The event in question was the 2014 claim by the chatbot Eugene Goostman at a Royal Society event in London, widely reported as "the first machine to pass the Turing test" even though only a third of the judges were fooled; the claim was controversial precisely because it fell short of the informal 50% threshold.
- A modern model: GPT-4.5 convinced about 73% of interrogators. On 30 judges, , so roughly 22 of 30 judges believed it was human — comfortably above the 15-of-30 line.
Final answers: 10 of 30 = 33% (below the benchmark), 73% ≈ 22 of 30 (above it). Sense-check: both numbers behave as percentages of 30 judges, and the modern model clears the bar while the early chatbot did not — which matches the narrative that passing in some cases is a recent achievement.
There is also a reverse angle. Shakuntala Devi — a great Indian mathematician — was termed a "human computer." Here a human is being described as a machine. Similarly, a public-relations expert invited to the Boston event knew everything about his work, life, and character. If a human exhibits that level of machine-like knowledge, are we calling these people "acting computerly" — the reverse of acting humanly?
Q: What if people mistake a human for a machine — like Shakuntala Devi, who was called a human computer?
A: That is the reverse case. A human who exhibits machine-level knowledge gets called a computer — so are they acting computerly? A public-relations expert who knew everything about his work, life, and character was also invited to the Boston event. Using many judges, not one, protects against this bias.
The acting-humanly approach has real problems: some of these systems' results are not reproducible, not constructive, and not mathematically analyzable. Getting the same thought process to produce the same result every time is still an open research problem.
1.7.5 Thinking Rationally: The Laws of Thought
The thinking-rationally quadrant is the laws of thought. The laws of thought come straight from Aristotle — logic was invented around the third century BC, when philosophers asked what a correct argument is. Logic codifies rational thinking. The classic example: "In all boring classes, students sleep. This is a boring class. So students sleep in this class." This is formal logic — mathematical logic.
The same argument in first-order logic symbols. This is a syllogism: a rule stated for everything, a particular case, and a conclusion drawn by applying the rule to the case. Written out:
Reading the symbols. The universal quantifier says "for every object "; is the implication "if ... then ..."; is the predicate " is a boring class"; is the specific class in question; and means "therefore." The statement reads: for every , if is boring then students sleep in ; is boring; therefore students sleep in . The argument is valid by modus ponens — the rule that from and you may conclude — with and . This is the same shape as Aristotle's classic syllogism "Socrates is a man; all men are mortal; therefore Socrates is mortal."
One group aligned with this: the machine should do everything rationally — take all the data it has and logically reach a rational decision.
Two criticisms follow. First: is everything rational in this world? No. Put your hand on a hot pan — you do not run a rational analysis ("this pan is hot; if I keep my hand here a few more minutes I will burn it; so I should remove it"). You just pull your hand back. If a dog is chasing you, you do not calculate which direction to run or what speed. Not all intelligent behavior is mediated by logical deliberation; humans have reflexes.
Q: Is everything rational in this world?
A: No. Consider a hot pan: you do not run a rational analysis — the pan is hot, if I keep my hand there I will burn it, so I should remove it. You just pull your hand back. If a dog is chasing you, you do not calculate speed and direction. Not all intelligent behavior is mediated by logical deliberation; humans have reflexes.
Second: it is computationally exhaustive. Even with just a few hundred facts (facts meaning logic statements), you face a combinatorial number of two-to-the-N calculations — you have to think about all the combinations of decisions.
The explosion, worked out. If the system must consider every combination of facts, the number of combinations is — each fact can be either included in or left out of any given combination.
- facts: combinations — trivial.
- : — a billion combinations, already too many to enumerate by hand.
- : using , we get — more combinations than stars in the observable universe (roughly ).
- A few hundred facts, say : .
Final answer: even a few hundred facts produce on the order of combinations — no computer can check them all. Sense-check: each added fact doubles the count, so going from 100 to 300 facts multiplies the work by , which explains why the growth is called combinatorial.
This combinatorial blow-up is what makes the laws-of-thought approach hard to scale: the problems it produces belong to the class called NP-complete — so hard that no known algorithm solves them in time polynomial in the number of inputs, and if any one of them were easy, a whole family of problems would become easy too. For propositional logic specifically, deciding whether a set of statements is consistent is an NP-complete problem, and exponential growth means even moderately large instances cannot be solved in any reasonable time. Because of this, these systems are not really scalable. That does not stop people from building rationally driven systems for small, narrow domains — that work continues.
1.7.6 Acting Rationally: Rational Agents
The fourth quadrant is acting rationally: building rational agents. An agent is an entity that perceives and acts — that simple. Humans are agents: we perceive something and we act on it.
A rational agent is one that acts so as to achieve the best outcome, or, when there is uncertainty, the best expected outcome. Computational limitations make perfect rationality unachievable, so the practical goal is to design the best program for the given machine resources. Rational behavior means doing the right thing — and the right thing is what is expected to maximize goal achievement given the available information.
Q: Are human beings agents?
A: Yes. An agent is an entity that perceives and acts, and we perceive and act on the environment all the time. Your brain is intelligent, but going back to the earlier question — are we intelligent? We do not know. We are not always intelligent.
There is a life parallel here. When you look back and think "I could have done better," stop — at that day, with whatever knowledge you had, you took the decision that was best at that moment. Whether it turned out good or bad is decided later. In a stable mind, the decision you acted on seemed like the best decision, given your life experiences up to that point. That is exactly what rational behavior means.
Regret, reinterpreted. The "I could have done better" feeling is hindsight: the outcome is known later, but the decision was made earlier under uncertainty. Rationality is judged at decision time, not after the fact — with the knowledge you had then, you picked the action with the best expected outcome. This is also why the definition says expected: rationality does not promise good results, only good bets.
Rationality here includes correct inference — and also making an inference even when there is no correct inference, which is what we do in uncertain situations. It includes acting without inferencing sometimes: not always being deliberative, but using reflexes. So rational behavior is not just correct inference, and it is not just the skills needed to pass the Turing test.
Why acting rationally is the most general cell. It strictly contains the others' best parts: correct inference is one mechanism for acting rationally (deduce the best action, then take it), and the Turing test's skills (language, reasoning, learning, perception) are useful because they help agents get better outcomes in a complex society. But acting rationally adds things the other cells miss: acting without deliberation (reflexes), and deciding under uncertainty when no correct inference exists. Any definition AI has ever proposed ends up as a special case of acting rationally.
1.7.7 Which Quadrant Is Winning?
In the last couple of years, more people have started aligning with the acting-rationally quadrant. Their position: it should be a rational agent. It should not try to just mimic human behavior, and it should not just think — because thinking without action is useless. So build something that perceives and acts and tries to achieve the right thing given its goals and information.
That does not stop research in the other quadrants — all four continue to produce important work. But the acting-rationally quadrant is gaining popularity. Real-world: the whole agentic AI movement stands on this definition — an agent that perceives, has knowledge, has goals, takes actions, and tries to optimize them. This course is mostly about designing rational agents.
Recap and bridge: the four schools of thought are thinking humanly (cognitive modeling), acting humanly (the Turing test), thinking rationally (the laws of thought), and acting rationally (rational agents). The laws of thought fail on reflexes and on the explosion of full logic, while the rational-agent view — act to achieve the best expected outcome — absorbs the useful parts of the other three and is where the field is converging. That definition carries a cost, and the final section of this lecture names it: the risks of AI.
1.8 Risks of AI
Hook: the same property that makes AI powerful — scalability — is what makes it dangerous. A human soldier, spy, or reviewer can only be in one place; an AI system can act everywhere at once. Every risk in this section is the scalability property pointed at a different target: war, information, decisions, wealth, and safety.
1.8.1 Lethal Autonomous Weapons
The first major risk is lethal autonomous weapons. The UN defines these as weapons that can locate, select, and eliminate human targets without human intervention. The key phrase is without human intervention — these systems act on their own. The primary concern is scalability: you design such an agent, and it could act across the globe and target people. That is very dangerous.
Why scale is the danger: a conventional weapon is bounded by its human operator — one operator, one weapon, one target at a time, and a human making the final call. An autonomous weapon removes that bottleneck: a single designer can deploy a large number of weapons that each locate, select, and engage targets by themselves, against any criterion the designer programs in. The technologies involved overlap heavily with civilian systems such as self-driving cars, so the capabilities exist without a weapons-specific breakthrough. International discussions on limiting these systems have been running since 2014 at the UN, which shows how seriously the risk is taken.
1.8.2 Surveillance and Persuasion
AI works in scalable fashions, and that cuts both ways. It can be used for mass surveillance, and it can be tailored for information — through social media and other platforms — to influence behavior. There is already suspicion of such influence in a number of elections.
The two scalable techniques: mass surveillance uses speech recognition, computer vision, and language understanding to watch many people at once — where human monitoring is expensive, slow, and legally constrained, AI scales the same monitoring to everyone. Persuasion works by tailoring information flows: a machine learning system learns what each person responds to and feeds them content designed to move their views, so political behavior can be modified without any single interaction looking coercive. The suspicion that this shaped recent elections is exactly what makes the combination dangerous: surveillance identifies the target, persuasion changes them.
1.8.3 Biased Decision Making
This is where the quadrant discussion pays off. If your goal is that an AI system should mimic humans — the acting-humanly direction — then humans are not perfect, and their flaws come in. Careless or deliberate misuse of machine learning algorithms can lead to biases by race, gender, and against other marginalized communities. The data itself reflects those biases. This is a big reason the field is moving toward acting rationally, and people are now making a conscious decision about how to eliminate these biases.
The bias loop: a machine learning system learns from historical records — past hiring decisions, past loan approvals, past parole rulings. When those records carry society's existing biases, the model reproduces them: a loan model trained on a history in which one group was rejected keeps rejecting that group, even if no rule says so. The bias enters through the data, not through a programmer's intent — which is why "we just let the data speak" is not a neutral position. Concretely, this has been documented in loan applications, parole evaluations, and hiring screens. The rational-agent direction helps because it replaces "behave like humans" with "achieve the stated objective," forcing the designer to state — and audit — what the system is actually optimizing for.
1.8.4 Employment and Inequality
The employment story has two sides. Agents do some of the tasks that humans might otherwise do, but they also make humans more productive — which can give more employability, make companies more profitable, and let them pay higher wages. But the key line is: AI use generally results in increasing wealth, and it tends to have the effect of shifting wealth from labor to capital. One group becomes ultra-productive through AI and makes a lot of money; the other group might not have access to it. That is inequality, and if not acted on quickly, it keeps growing. This is why the impact on employment is really an impact on inequality — the setup is more capital-centric than labor-centric.
Q: The surface of software engineering is already almost replaceable by AI.
A: Code is democratized — writing code is no longer the rarest skill, with tools like Claude Code and Cursor. But can any of them autonomously build a system that is successful in the market, that people are willing to pay for, that scales to a million users, and that models what humans want? Requirement gathering, software architecture, and usability are still under question. Some parts of coding are replaceable; software engineering as a whole is not.
Why inequality, not unemployment, is the real risk: history shows technology displaces specific jobs — mechanical looms disrupted weavers — yet people eventually find new kinds of work. The newer worry is different: AI may also do the new kinds of work. Meanwhile the gains concentrate: whoever owns or controls the AI captures the productivity increase, so wealth shifts from labor (wages) to capital (profits). The result is not mass unemployment but a growing gap between the ultra-productive group with access and the group without it. This is why the earlier discussion of jobs concluded: the unemployment rate may not rise, but inequality will — unless access spreads quickly.
1.8.5 Safety-Critical Applications and Cybersecurity
In safety-critical applications, an ill-informed AI can go directly to ill effects — impacting people's life and health. There are already discussions about developing technical and ethical standards for these systems.
The life-and-health risk: in domains like driving cars or managing a city's water supply, a wrong decision is not a wrong answer — it is an accident. Fatal accidents involving automated vehicles have already occurred, and they expose a hard problem: machine-learned systems are hard to verify formally, because their behavior is learned from data rather than written as rules. The field needs technical and ethical standards comparable to those of other engineering and healthcare disciplines where people's lives are at stake — and those standards do not fully exist yet.
In cybersecurity, new attack vectors will emerge, along with new forms of attack behavior. Vibe coding makes this worse: people without an understanding of computing principles are building systems. Proofs of concept are fine, but when such systems are pushed into production without software engineering and software architecture principles, they are more prone to attacks.
1.8.6 Sustainability and Other Social Risks
Finally, the risks touched on earlier: sustainability and other social issues. The energy needed to run these systems, the resources they consume, the token spending, and the net-zero question — all of it feeds into whether AI is sustainable if we keep doing it the way we are. AI has not always been successful; it has gone through bad phases, and it is in a prime phase now. There may be small winters ahead, and it will bounce back. The risks do not change the trajectory — they shape how carefully the field proceeds.
The long-run question the field is already working on: an AI that pursues the wrong objective is not being stupid — it is being faithful to a badly stated goal. The chess machine that would hypnotize its opponent or blackmail the audience to win is a logical consequence of "win at all costs" as the sole objective. The research direction this motivates is value alignment: build systems that are uncertain about what humans really want, so they ask, learn, and defer instead of charging ahead. This is the frontier of AI safety — and it connects back to the rational-agent definition: the agent's objective must be stated by someone who accepts responsibility for it.
Recap and bridge: the risks of AI are the power of scalability aimed at harmful ends: autonomous weapons, surveillance and persuasion, biased decision making, wealth shifting from labor to capital, safety-critical failures, and cybersecurity attacks — with sustainability and alignment as the long-term questions. The risks do not stop the field; they set its speed and care. This lecture opened with the rules of the course and closes with the risks of the field — in between, the four definitions of AI explained both what we are building and why we must build it carefully. The next session takes the first step of the course proper: designing rational agents and the algorithms they use to act.
Exam Guidance Summary
The course is a five-credit core course: nine modules across sixteen sessions, with a couple of research papers carried in the handout and discussed in their sessions.
Evaluation: EC1 is continuous assessment for 30 marks, built from two group assignments (12 marks before the midsem, 13 after) and two quizzes. Groups have four or five members, chosen in a self-selection window or randomly allocated; the same group works on both assignments, and credit is contribution-weighted. The two quizzes are scored best-of-two, and each quiz stays open for about three or four days. Everything — exams, assignment pool, discussions — is common across the whole batch.
The midsem is a closed book exam for 30 marks covering the first eight contact-session topics. The final is an open book exam for 40 marks covering all topics.
Open-book rules: only the listed textbooks and the class materials may be taken into the final. Handwritten notes are not allowed under the current policy (the examination cell communicates any changes). A watermarked one-shot PDF of the decks is posted around the fourteenth or fifteenth session; print that copy and bring it with your books. E-books and digital devices are not permitted in the exam, which is why a physical copy of the textbook is recommended.
Study advice: be regular and mentally present; watch any session you miss rather than bingeing before the exam; use the discussion forum (it is persisted, so questions keep paying off); solve the short ungraded exercises in the decks from the second or third session onward (15–20 minutes each); and go the extra mile — the compounding example is versus .
Textbook guidance: Russell and Norvig, Artificial Intelligence: A Modern Approach, fourth edition is the primary book (T1 in the handout). Roughly half to three-fifths of the book is part of this course, and the handout maps each module to specific sections. A few course topics exist only in the fourth edition, and the fourth edition differs from the third by about 20–25%.
Exam note — revision checklist for this lecture:
- Evaluation arithmetic: EC1 = 30 marks (12 + 13 assignment marks, two quizzes best-of-two), midsem = 30 closed book (first eight sessions), final = 40 open book (all topics). Quiz yourself: which numbers, which books allowed, which windows.
- History dates: 1950 (Turing's question), 1956 (Dartmouth workshop names AI), 1973–1980 (first winter), 1987 (second winter begins), 1997 (Deep Blue beats Kasparov), 2017 (transformers), 2022 (ChatGPT).
- Three enablers: data, computational power, algorithms — together.
- Intelligence: the capacity to learn, reason, understand — made real by perceiving and acting. Thinking without action is useless.
- The 2×2 matrix: thinking humanly, acting humanly, thinking rationally, acting rationally — and why acting rationally is where the field converged.
- Turing test numbers: 50% benchmark, the 10-of-30 (33%) Eugene Goostman claim, GPT-4.5 at about 73%.
- The explosion: a few hundred facts mean about combinations — NP-complete, not scalable.
Topic guidance: the section on how different fields contributed to AI (philosophy, mathematics, economics, neuroscience, psychology, computer engineering) is interesting but not from the exam point of view. The four-quadrant discussion of AI definitions is an important conceptual area: no single definition is right or wrong, and each quadrant is a school of thought with active researchers.
Key Industry Applications
- Education: hyper-personalized learning at scale — about 3 million CS/AI students in India against teacher ratios of 1:300–500. AI systems that know your skill level, push boundaries, and assess along the way, running the measure–choose–act–update loop for every learner.
- Agriculture: AI-assisted agriculture, with a major lab in Punjab backed by crores of funding, applying sensing, prediction, and optimization to farming decisions in that region.
- Application fields named across the session: education, finance, healthcare, automotives, robotics, smart cities, agriculture, culinary, transportation, sports, and governance — a useful list for the thought experiment of finding AI opportunities in your own organization's field.
- Networking: proactive network monitoring and network-health tools for real-time optical networks; AI for networks and networks for AI (massive AI needs better networks, and networks get smarter with AI).
- Chatbots and conversational AI: ELIZA's legacy, still showcased at AAAI; modern systems evaluated with Turing-test style judging (GPT-4.5 convinced about 73% of judges).
- Game playing: Deep Blue vs. Kasparov (1997), AlphaGo, and the RoboCup robot football competitions — the arena where machines repeatedly proved that winning moves need not look human.
- Assistive technologies: AI for people with disabilities and accessibility — an active research area, including interfaces that model how users perceive and act.
- Voice assistants: Siri, the first mass-market AI voice interface.
- Large language models: the transformers architecture (Attention Is All You Need, 2017), the GPT line, ChatGPT (2022) as the democratizing milestone, and tools like TensorFlow and DeepMind's systems.
- Indic AI: AI4Bharat and regional-language models for Hindi and other Indian languages, addressing markets the frontier English-centric models serve poorly — an example of AI built for a specific linguistic population.
- AI-assisted engineering: AI companions for code review and delivery (Claude Code, Cursor); contrast with vibe coding, which produces systems that do not scale and carry vulnerabilities when pushed to production.
- The agentic AI wave: agents that perceive, hold knowledge and goals, take actions, and optimize — built on the acting-rationally definition of AI, which is also the definition this course builds on.
ACI Lecture 1 notes · Introduction to Artificial Intelligence
Sections Breakdown
Course structure and the handout, textbooks and reference books, the EC1/midsem/final evaluation scheme, ground rules and study habits, prerequisites, and how the sixteen sessions run.
The timeline from Turing's 1950 question and the 1956 Dartmouth workshop through the two AI winters, RoboCup and Deep Blue, to transformers and ChatGPT, with the lessons the curve teaches.
The three enablers (data, computational power, algorithms), the cost-versus-price question behind sustainable AI, and whether AI winters can return.
Application areas and a thought experiment, education as an example, vibe coding versus AI-assisted engineering, the AI4Bharat effort for Indic languages, and the jobs, unemployment, and inequality debate.
The eight disciplines behind AI, what each contributes, and the cognitive psychology versus cognitive science distinction.
Definitions of intelligence, why thinking without action is useless, whether humans are always intelligent, and early AI as mimicry.
Thinking humanly, acting humanly, thinking rationally, and acting rationally, and why the field is converging on the rational-agent view.
Lethal autonomous weapons, surveillance and persuasion, biased decision making, employment and inequality, safety-critical applications, and cybersecurity.
The exam strategy for this lecture: evaluation arithmetic, history dates, the three enablers, the 2x2 matrix, Turing test numbers, and open-book rules.
The real-world applications named in the session, from education, agriculture, and networking to chatbots, game playing, voice assistants, LLMs, and Indic AI.
Exam Revision Notes
Below is the distilled, exam-ready core. Every entry comes from the full explanation above. Use this section for rapid review; return to the main notes when a point needs more context.
1.1 How the Course Works: Handout, Books, and Evaluation
Must-know: Evaluation: EC1 = 30 marks (two group assignments: 12 before midsem + 13 after; two quizzes, best of two). Midsem closed book 30 marks (first eight sessions); final open book 40 marks (all topics). Open book allows only listed textbooks and class materials; no handwritten notes, no digital devices.
⚠️ Top pitfall: Treating a zero daily edge as a small edge: 1^365 = 1 exactly, while 1.01^365 is about 37.8 — the delta is 0.01 (1%), not 0.1.
Self-check: How many marks is the midsem worth and is it open or closed book? (30 marks, closed book, first eight contact-session topics.)
Connects to: 1.2 A Brief History of Artificial Intelligence
1.2 A Brief History of Artificial Intelligence
Must-know: Timeline: Turing's question (1950) -> Dartmouth workshop names AI (1956, McCarthy/Minsky/Rochester/Shannon) -> first winter 1973-1980 -> expert-systems recovery 1980-1987 -> second winter from 1987 -> RoboCup and Deep Blue vs Kasparov (1997) -> transformers (2017) -> ChatGPT (2022). Deep Blue's lesson: the 'stupid' move won the game — intelligence need not look like human reasoning.
⚠️ Top pitfall: Reading the current boom as permanent — both AI winters followed the same recipe of hype, over-promise, and dried-up funding, and winters can recur.
Self-check: Who organized the 1956 Dartmouth workshop that named AI? (John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon.)
Connects to: 1.3 Why AI Took Off in the Last Decade, 1.7 Four Perspectives on AI: The 2x2 Matrix
1.3 Why AI Took Off in the Last Decade
Must-know: Three enablers: data, computational power, and algorithms, arriving together. Cost = money paid now (tokens, subscriptions); price = what future generations face (resources, energy, net-zero) — the literature calls this sustainable AI. Winters recur but never permanently; pay-per-call pricing could cause a temporary slowdown.
⚠️ Top pitfall: Conflating cost and price: cost is today's private money bill; price is the deferred, public bill paid by future generations through depleted resources and energy use.
Self-check: What are the three enablers that made AI take off in the last decade? (Data, computational power, and algorithms arriving together.)
Connects to: 1.2 A Brief History of Artificial Intelligence, 1.8 Risks of AI
1.4 AI Applications and the Boom-or-Bane Debate
Must-know: Vibe coding is bad; AI-assisted engineering is good — the difference is understanding what the AI produces. AI will not replace you; people who know AI will replace you. India is heavily on the consumption side but making good progress in AI engineering (AI4Bharat, Indic models). Automation will likely grow inequality rather than the unemployment rate.
⚠️ Top pitfall: Believing a vibe-coded system is finished: without software engineering review, vulnerabilities sneak in, and the system will not scale to real user loads.
Self-check: What is the difference between vibe coding and AI-assisted engineering? (Vibe coding keeps prompting blindly until output looks right; AI-assisted engineering is done by people who understand the system and review what the AI produces.)
Connects to: 1.8 Risks of AI
1.5 AI Is an Umbrella Field: Contributing Disciplines
Must-know: The eight contributing fields: philosophy, mathematics, economics, neuroscience, psychology, computer engineering, control theory, linguistics. Economics gave decision theory, Markov decision processes, and game theory (Nash equilibrium). Cognitive psychology = brain as information-processing machine; cognitive science = computers modeling that. This section is interesting but not from the exam point of view.
⚠️ Top pitfall: Believing computer science alone created AI — the field is interdisciplinary, and math, economics, psychology, and philosophy all made essential contributions.
Self-check: Which field contributed Markov decision processes and the Nash equilibrium? (Economics.)
Connects to: 1.7 Four Perspectives on AI: The 2x2 Matrix, 1.6 What Is Intelligence?
1.6 What Is Intelligence?
Must-know: Working definition: intelligence = capacity of learning, reasoning, understanding, and similar mental activity = the ability to perceive and act in the world. Thinking without action is useless. Humans are NOT always intelligent; non-human behavior (dog barking at a stranger) CAN be intelligent. Whether humans are intelligent at all is still open.
⚠️ Top pitfall: Conflating 'humans are intelligent' with 'humans are always intelligent' — humans make mistakes and take stupid decisions, so the always-claim is false.
Self-check: Why is the philosopher who only eats, sleeps, and thinks not a good example of intelligence in action? (No action — nobody can even tell they were thinking; thinking without action is useless.)
Connects to: 1.7 Four Perspectives on AI: The 2x2 Matrix
1.7 Four Perspectives on AI: The 2x2 Matrix
Must-know: The four quadrants: thinking humanly (cognitive modeling; introspection, experiments, brain imaging), acting humanly (Turing test; 50% benchmark; 10/30 = 33% Eugene Goostman claim), thinking rationally (Aristotle's laws of thought; modus ponens; 2^N explosion, NP-complete), acting rationally (rational agents: best expected outcome; reflexes allowed; most general; where the field is converging). Turing test needs NLP, knowledge representation, automated reasoning, ML, vision, robotics.
⚠️ Top pitfall: Believing the Turing test benchmark and the 10-of-30 pass claim agree — 10/30 is only 33%, below the 50% (15-of-30) benchmark; the 2014 Eugene Goostman claim was reported as a first-time pass despite falling short.
Self-check: Why is acting rationally the most general definition? (Correct inference is one mechanism of rational action, and the Turing test skills help agents get better outcomes, but acting rationally also covers reflexes and decisions under uncertainty.)
Connects to: 1.6 What Is Intelligence?, 1.8 Risks of AI, 1.5 AI Is an Umbrella Field: Contributing Disciplines
1.8 Risks of AI
Must-know: Six risk areas: lethal autonomous weapons (UN definition, without human intervention, scalability), surveillance and persuasion (mass monitoring, tailored information flows), biased decision making (data reflects societal bias; a reason for moving to acting rationally), employment and inequality (wealth shifts from labor to capital; inequality grows, not unemployment), safety-critical applications (life and health; need technical and ethical standards), cybersecurity (new attack vectors; vibe coding makes it worse).
⚠️ Top pitfall: Attributing AI bias to programmer intent alone — bias usually enters through historical training data that reflects society's existing biases.
Self-check: Why is scalability the central concern for lethal autonomous weapons? (No human supervision required, so a small group can deploy weapons at global scale against any programmed target criterion.)
Connects to: 1.7 Four Perspectives on AI: The 2x2 Matrix, 1.4 AI Applications and the Boom-or-Bane Debate, 1.3 Why AI Took Off in the Last Decade
Exam Guidance Summary
Must-know: EC1 = 30 marks (assignments 12 + 13, two quizzes best-of-two); midsem closed book 30 marks (first eight sessions); final open book 40 marks (all topics). Only listed textbooks and class materials allowed; no handwritten notes or digital devices; watermarked deck printout after session 14/15. Four-quadrant definitions are the key conceptual area.
⚠️ Top pitfall: Arriving at the open-book final without knowing the book's layout — the open-book format rewards preparation, not hoping the book will save you.
Self-check: Which topics does the midsem cover and which does the final cover? (Midsem: first eight contact sessions, closed book. Final: all topics, open book.)
Connects to: 1.1 How the Course Works: Handout, Books, and Evaluation
Key Industry Applications
Must-know: Named applications: education, agriculture, networking, chatbots (ELIZA, GPT-4.5), game playing (Deep Blue 1997, AlphaGo, RoboCup), assistive technology, voice assistants (Siri), LLMs (transformers 2017, ChatGPT 2022), Indic AI (AI4Bharat), AI-assisted engineering, and the agentic AI wave on the acting-rationally definition.
⚠️ Top pitfall: Listing applications without connecting them to the definitions — each application maps to a quadrant (e.g., agentic AI = acting rationally; tutoring = thinking humanly).
Self-check: Which definition of AI does the agentic AI movement stand on? (Acting rationally — agents that perceive, hold knowledge and goals, take actions, and optimize.)
Connects to: 1.4 AI Applications and the Boom-or-Bane Debate, 1.2 A Brief History of Artificial Intelligence, 1.7 Four Perspectives on AI: The 2x2 Matrix
Was this lecture useful?
BitsNotes AI Assistant
Subject Notes AssistantConfigure AI Chat
Choose how to access the chatbotSigned in as
Powered by BitsNotes — 20 messages per day. No API key needed. Want unlimited access? Use "Bring Your Own Key" mode.
Sign in to use AI Chat
Get 20 free AI messages per day to ask questions about your lecture notes. Sign in with Google or GitHub — it takes 5 seconds.
Sign In to BitsNotesSwitch to "Bring Your Own Key" tab above for unlimited access with any OpenAI-compatible provider.