Part 1: Course Overview and General Introduction

1. Terminology: AI, machine learning, deep learning, data science (p.6)

The deck opens by fixing four terms that are used interchangeably in industry and are not interchangeable here. Slide p.5 gives the definition of AI the course starts from, attributed to Herbert Simon: "Machines that exhibit behaviors that would be regarded intelligent if they were exhibited by human beings". Hold on to the shape of that sentence, because by p.56 the course will have replaced it with something much more precise, and the replacement is the examinable one.

The four terms, as p.6 defines them:

• Artificial intelligence: the Simon definition above, so behaviour that would count as intelligent if a human produced it.
• Machine learning: "A machine learns when it improves its performance on a specific task with experience". The slide adds an arrow: "Central to AI (no intelligence without learning)".
• Deep learning: a subset of machine learning algorithms that model high-level abstractions in data through neural networks.
• Data science: "Inter-disciplinary field that uses scientific methods, processes, algorithms and systems to extract knowledge and insights from structured and unstructured data".

The geometry of the figure carries the argument. Deep learning is drawn strictly inside machine learning, which is drawn strictly inside AI. Data science is a separate circle that overlaps all three without being contained by any of them. So the picture says: every deep learning system is a machine learning system, and every machine learning system is an AI system, but the converse fails in both directions. The region of AI outside the machine learning circle is not empty, and it is where this course lives: search, constraint satisfaction, game trees, logic, planning and Markov decision processes are all AI and none of them is learning.

2. The course as an object: objectives, contents, position (pp.8-12)

The four objectives (p.8)

Read these as a description of the exam, because that is what they are.

1. Deep knowledge and insight in a limited number of fundamental techniques. The phrase "a limited number" is doing real work: the course prefers depth on about a dozen algorithms over a survey.
2. Simulate these techniques with pen and paper on small new examples, and develop solutions for other problem settings using the provided software tools. This is the exercise exam in one sentence: pen and paper, small scale, new examples.
3. Insight in the relations between these techniques and in their relevance for real-world applications. This is the "synthetic knowledge" the theory exam tests.
4. Basic insight into the ethical implications of AI, which is Part 14.

Course contents: the 13-lecture map (p.9)

Six blocks. The table below maps each item on slide p.9 to the chapter of this pack that covers it, because the deck's block order and the deck numbering do not agree everywhere.

Two structural facts follow from this slide. First, block 5 tells you what "machine learning" means inside FAI: version spaces and pattern mining, both symbolic, both driven by generality orderings (relations that say which hypothesis or pattern is more general than which). Second, the slide's last two lines are the programme split: "+ Exercise sessions" for everyone, "+ Project (for Bachelor students, see later)". On the postgraduate track there is no project.

Where the course sits (p.10)

Blended learning and workload (p.12)

Slide p.12 lists three things done before the contact hour: watch prerecorded lecture parts, read or recap upfront, and prepare the exercise session upfront. In the lecture the third was singled out: the exercise slot exists to walk through the model solution and answer questions, not to attempt the exercises cold. The slide closes with a workload arrow: reserve time in between lectures, this is a 5 or 6 ECTS course. The slide gives no hour figure. Applying the general European convention of 25 to 30 hours per credit, which is not stated anywhere in the course material, 5 ECTS works out at 125 to 150 hours in total.

3. Evaluation, and what it implies for how you answer (pp.13-15)

Slide p.13 governs your exam on 18 August 2026 (postgraduate and advanced master, 5 ECTS):

• Open book exercise examination, 10/20.
• Closed book theory examination, 10/20.
• Arrow: "Tests for factual and synthetic knowledge".

Slide p.14 is the Bachelor variant (6 ECTS): 8/20 exercises, 8/20 theory, 4/20 programming project in three parts. It does not apply to you. If a past-paper discussion mentions the Pacman project, that is the p.14 track.

4. A brief history, in four eras (pp.17-21)

In this pack's judgment this is breadth material: know the eras and roughly what happened in each, know the two or three names the deck attaches to each, and move on. The one date worth fixing exactly is 1956, the Dartmouth meeting where the term was adopted, because the deck returns to it on p.43.

The "TODAY" quiz (pp.20-21). The same twelve-item list appears twice, but p.21 is not a plain repeat: it prints the verdicts as colours. The twelve items are: table tennis, quiz participation, driving a curving mountain road, driving in the streets of Paris, buying groceries on the web, buying groceries in a supermarket, discovering and proving a new mathematical theorem, conversing successfully for an hour, performing a surgical operation, putting away the dishes and folding the laundry, translating spoken Chinese into spoken English in real time, writing an intentionally funny story.

Verified from the text colours in the deck, p.21 splits the twelve items three ways. Two are printed in red: buying a week's worth of groceries in the supermarket, and performing a surgical operation. Five are printed in amber: driving safely in the streets of Paris, discovering and proving a new mathematical theorem, conversing successfully for an hour, putting away the dishes and folding the laundry, and writing an intentionally funny story. The remaining five stay in the deck's default dark blue: table tennis, quiz participation, the curving mountain road, groceries on the web, and real-time Chinese to English translation. The deck prints no legend, so the reading has to be inferred from the question "Which of the following can (not) be done at present?": dark for done, amber for partly or in progress, red for not done.

The pattern is what to retain, not the individual verdicts. Informational tasks (quiz answering, translation, web shopping) sit in the done group. Tasks requiring physical manipulation under uncertainty are the hard ones: the lecture singled out folding a towel as a standing benchmark problem in robotics, because the fabric is unknown in advance, the towel can slip from the gripper, and an early fold that is slightly wrong ruins every later one. Note one place where slide and lecture differ in emphasis: surgery is red on the slide, while in the lecture he added that some procedure types, the ones where the required view and the required actions are clearly delimited, are already possible. That dependency structure, task properties decide feasibility, is exactly the taxonomy Part 2 formalises.

5. AI and its successes: what to name, what to skip (pp.22-41)

Twenty pages of breakthroughs. Two of them feed directly into later chapters. In this pack's judgment the rest are context: name them and move on.

The two that carry forward

Shakey the robot (p.24). Developed in 1965 at SRI, driven by a remote-controlled computer that formulated plans for moving and acting. The slide names two people: Nilsson, associated with the STRIPS planning formalism, and Hart, associated with the A* search algorithm. Those two names return as the two central algorithms of the course: A* in Part 5 and STRIPS in Part 11. Shakey is also the first full instance of the agent loop you will meet on p.63: the photo is annotated with a television camera and bump detectors (sensors) and with drive wheels, a caster wheel and their motors (actuators). For the decision process, follow the slide's own bullet rather than the photo: the plans for moving and acting were formulated by a remote-controlled computer, so the deliberation sat off the robot even though the photo also carries an onboard logic callout. In the lecture one detail was added for scale: a single action took roughly half an hour of computation before the robot moved for a few seconds.

Foundational subdomains of AI (p.26). Eight of them: robotics, computer vision, language and speech processing, planning and scheduling, theorem proving, human computer interaction, declarative programming, machine learning. This course touches planning and scheduling (Part 11), theorem proving (Part 10) and machine learning (Parts 12 and 13) directly.

The timeline, in one table

Application slides pp.37-41 group further examples: the (bio)medical domain (combining image and language processing to support patient diagnosis; predicting protein structures for drug development), robotics (reinforcement learning teaching a robot dog to walk in an hour), computational creativity (NVIDIA's GauGAN2, OpenAI's DALL-E, a bot generating tweets, a deepfake video), and a long list of industrial applications on p.41 running from gas and electricity consumption modelling to fraud detection and smart grids. Note that reinforcement learning itself is out of scope: the lecturer says of it at the end of Part 9 that "we don't cover here", so in this pack's judgment the robot-dog slide is context rather than examinable content.

Several pages in this stretch are built around video or screenshots, and each still prints its own caption. Checked page by page: p.25 is the only page in the deck's success stretch with no text at all (Shakey footage); p.28 prints "2007: Boss wins Darpa Urban Challenge" together with a YouTube link; p.31 prints the full Watson caption, naming the champions Brad Rutter and Ken Jennings and the 1 million dollar prize. The clips were played in class; the captions are the part that stays on the page.

6. What is AI today: the four-quadrant matrix (pp.42-54)

This is the examinable core of the chapter. Everything before it is orientation; everything after it is a consequence.

The starting point (p.43)

The Dartmouth proposal, McCarthy and colleagues: "Every aspect of learning or any other features of intelligence can in principle be so precisely described that a machine can be made to simulate it." That is a hypothesis about describability, and it is deliberately silent about whether the machine should think or merely behave, and about whether it should copy a human or do better. Those two silences are exactly the two axes of the matrix.

The two axes (p.45)

Slide p.45 introduces the axes as two different research motivations:

• "The attempt to make computers more 'intelligent'" leads to the question: is it about thinking or acting?
• "The attempt to better understand human intelligence" leads to: is it about modeling human behavior (with all its defects) or normative behavior (how should a rational being think/act)?

Stated as a test you can apply to any sentence:

The horizontal axis is the one students misread. "Rational" is not a synonym for "clever" or "logical" here. It means: judged against a norm rather than against human performance. That is why a definition can mention logic and still be in the human column (if it is about how people reason), and can avoid logic entirely and still be in the rational column.

The eight definitions, and where each one lives (pp.46, 50, 52, 54)

The deck shows the same eight-definition table four times, highlighting one cell each time. That build order is itself the answer key, and it is worth recording:

7. The four cells, one at a time (pp.47-55)

Acting humanly: the Turing test (pp.47-49)

Slide p.47 pairs the cell with Alan Turing and the Turing test: an interrogator in one room questions an unseen respondent in another and must decide whether it is a human or a machine. The test is behavioural by construction, which is exactly why it sits in the acting row: it inspects the transcript, never the mechanism. In the lecture one nuance was added: Turing's original formulation was quantified, predicting that by 2000 an interrogator would be fooled in roughly 70 percent of cases after five minutes of questioning, so the criterion has a time budget and a success rate, not a binary verdict.

Slides p.48 and p.49 supply the critique. The Loebner Prize is an annual competition in the format of a standard Turing test; the slide notes it was "Won three times by Alice, which is unable though to pass the Turing test", and that Marvin Minsky (of Dartmouth) "promised a prize for the person that can stop the Loebner Prize competition". A.L.I.C.E. (Artificial Linguistic Internet Computer Entity) is described on p.49 as inspired by Joseph Weizenbaum's classical ELIZA program, "with still a lot of pattern matching". The point of these three slides together: winning a behavioural imitation contest is not evidence of the mechanism, because pattern matching on the input and reflecting it back can carry a short conversation. That is the weakness of the whole cell, and the reason the course does not adopt it.

Thinking humanly: cognitive science (p.51)

The question is "What cognitive capabilities are necessary to produce intelligent performance?". The slide sets the priorities of this cell bluntly:

• Important: being able to solve problems like a human would. Fields: cognitive science and cognitive psychology. Also important for human-machine interaction.
• Not important: being able to solve problems correctly.
• Arrow: "Will not be discussed in this course".

That arrow is a printed scope limit, so you can rely on it. The reason given in the lecture is the same one that motivates the entire right column: human thinking is not always correct, people are demonstrably poor at logic and probability problems, and reasoning is partly emotional, so a system that faithfully imitates it inherits its defects.

Thinking rationally: the laws of thought (p.53)

The question is "What are the laws of thought? How should we think?". Two frameworks are named:

• The logical approach, which returns as propositional and first order logic in Part 10.
• The probabilistic approach, which returns as Markov decision processes in Part 9.

The slide also lists two problems with the cell: representing problem descriptions in a formal notation, and computational aspects. Both are course-length themes rather than passing remarks. The first is the problem-formulation skill of Part 3 and the knowledge-representation exercises of Part 10; the second is why Part 5 spends a whole lecture on heuristics and why SAT solving is hard.

So the course does not discard this cell. It borrows its machinery. What it declines is the cell's criterion: correct inference is a means here, not the definition of success.

Acting rationally: the rational agent (p.55)

Slide p.55 states the position the course adopts:

• "A rational agent acts so as to achieve its given goals, under the assumption that its impressions of the world and its convictions are correct".
• "Rational thinking is a prerequisite for rational acting, although it is not a necessary condition".
• Arrow: "What to do, for example, when we must make a decision faced with insufficient information?"

The second bullet needs care, because "prerequisite" and "not a necessary condition" pull against each other if you read "prerequisite" strictly. The reading that fits the rest of the deck, and the lecture, is: rational thinking is the normal route to rational action and the course will use it heavily, but it is not required, because rationality is a property of the decisions, not of the process that produced them. Slide p.56 says exactly that in its own second bullet. Practical consequence: a lookup table, a reflex, or a learned policy (a rule fixing which action to take in each state) that happens to select the utility-maximising action counts as rational in this course's sense, with no reasoning step anywhere.

The third bullet is the bridge to the next slide. Once you admit that information is incomplete, "achieve your goals" is not yet a decision rule, because you cannot guarantee outcomes. The fix is to take an expectation.

8. Rational decisions: maximize your expected utility (p.56)

Slide p.56 is the sentence to memorise, in the sense that you should be able to produce all four bullets under exam conditions:

• Rational: maximally achieving pre-defined goals. The term is used "in a very specific, technical way", in the slide's own words.
• Rationality only concerns what decisions are made (not the thought process behind them).
• Goals are expressed in terms of the utility of outcomes.
• Being rational means maximizing your expected utility.

The slogan is worth taking apart word by word, and the deck itself does that on pp.58-61. The order below is the deck's build order, which is also the order the lecture spoke: Maximize (p.58), Utility (p.59), Expected (p.60), Your (p.61). Each word is a different piece of machinery in a different later chapter.

On the utility point, the lecture used route planning as the canonical case: for a route planner, the utility is the negative of travel time, so that maximising utility means minimising time. That sign flip is worth internalising now, because it is the standard way a cost-minimisation problem is written as a utility-maximisation problem.

Slides pp.57-61 are not five repeats of the same slide. They are the slogan itself, built one word at a time, and the build order is the order in which the lecture explained it. Verified from the text colours in the deck: p.57 shows "Maximize Your Expected Utility" with every word in the same dark blue, then each of the next four pages picks out exactly one word in the highlight colour, Maximize on p.58, Utility on p.59, Expected on p.60, Your on p.61. The lecture runs straight through those five pages without a break and continues into the aviation analogy on p.62. Slide p.68, at the very end of the deck, prints the same four words again with no highlight: it is the closing takeaway slide, and the lecture closed Part 1 by restating in one sentence that acting rationally means an agent tries to maximise the expected utility function.

9. The engineering perspective (p.62)

Slide p.62 makes the analogy that justifies the whole right column: the history of aviation. Early attempts to fly copied nature, strapping on constructed wings and jumping, and they failed. Flight was achieved when the goal changed from imitating birds to systematically engineering a machine that flies, which is the Wright brothers' picture on the slide. The slide's question is "Can we build machines that do X?", not "can we reproduce how humans do X?".

Transposed to AI: the human brain is the existence proof that intelligence is possible, not the specification. It makes non-rational decisions, and copying it would copy those too. So the course builds systems that perform a task correctly, and treats how a person would do it as out of scope. That is the same conclusion p.51 reached from the other direction.

10. Designing rational agents: the master figure (pp.63-64)

Slide p.63 is the most reused diagram in the course. Learn it now and you will meet it five more times: Part 2 pp.2, 3 and 11, Part 6 p.8, and Part 7 p.2.

The bullets on p.63:

• "An agent is an entity that perceives and acts." Nothing more is required: no mind, no learning, no representation.
• "A rational agent selects actions that maximize its (expected) utility." This is p.56 restated as a design rule.
• "Characteristics of the percepts, environment, and action space dictate techniques for selecting rational actions". This is the sentence that generates the syllabus. Change the environment's properties and you change which algorithm is appropriate.
• "This course is about:" followed by two sub-bullets, "General AI search techniques for a variety of problem types" and "Learning to recognize when and how a new problem can be solved with an existing technique".

The diagram itself is a closed loop with four labelled parts. The environment sends percepts to the agent's sensors; inside the agent a box marked "?" turns percepts into a choice; the choice leaves through the actuators as actions that change the environment; the changed environment produces the next percept. The "?" is deliberately blank. Every algorithm in this course is a candidate filling of that box.

Slide p.64 instantiates the loop with the course's running example, Pacman, which is also the environment used for the Bachelor programming project.

11. Current challenge: learning plus reasoning (pp.65-67)

Slide p.65 states the closing claim: "True intelligence requires integrating learning and reasoning!", and borrows Daniel Kahneman's two-system vocabulary to make the split concrete:

• System 1, 'Thinking fast': can perform tasks such as 2+2 and recognise objects in an image.
• System 2, 'Thinking slow': can reason about solving complex problems, and plan complex tasks.
• The slide lists the same split under other names: "data-driven vs. knowledge driven", "symbolic vs. subsymbolic", "solvers & learners". It closes with a list of researchers who argue the point: Marcus, Darwiche, Levesque, Tenenbaum, Geffner, Bengio, Le Cun, Kautz.

Slide p.67 closes the lecture with Leuven.AI, the KU Leuven interdisciplinary AI research network, spanning machine learning, computer vision, natural language processing, robotics, special hardware, probabilistic and logical reasoning, and also law, ethics, philosophy and cognition. In this pack's judgment it is context rather than examinable content.

The lecture's own closing summary of Part 1 was one sentence: acting rationally means that an agent tries to maximise the expected utility function. If you can produce the four-quadrant matrix, place a definition in it, and unpack that sentence word by word, this chapter has done its job.

12. Can you do these? (self-check)

Term box

Sources for this chapter: FAI_Part1_CourseOverviewAndIntroduction_25-26.pdf (69 pages, cited by PDF page index, which equals the printed slide number in this deck) and the Part 1 lecture recording. Verbatim lecturer quotes are drawn only from the verified quote reference and attributed (Lecturer, Tn). Slide text is quoted as printed. The route-planning lottery is a constructed example and is labelled as such; its arithmetic was checked by computation.