Artificial Intelligence, A Modern Approach - Artificial Intelligence - Introduction

Artificial Intelligence: A Modern Approach, 3rd Edition, by Stuart Russell, Peter Norvig

Chapter 1 - Introduction

1.1 Define in your own words: (a) intelligence, (b) artificial intelligence, (c) agent, (d) rationality, (e) logical reasoning.

Intelligence is the capacity to perceive, understand, predict, and manipulate the world. Artificial intelligence is the field of study concerned with building intelligent systems. An agent is an actor that operates autonomously, perceives its environment, adapts to changes, creates and pursues goals over time. Rationality is a quality that measures the success of outcomes relative to ideal performance, considering the available information. Logical reasoning is a key part of thinking rationally, one of the four approaches to AI that adhere to irrefutable laws of thought, originally developed by Aristotle; the other approaches are acting rationally, thinking humanly, acting humanly.

1.2 Read Turing's original paper on AI (Turing, 1950). In the paper, he discusses several objections to his proposed enterprise and his test for intelligence. Which objections still carry weight? Are his refutations valid? Can you think of new objections arising from developments since he wrote the paper? In the paper, he predicts that, by the year 2000, a computer will have a 30% chance of passing a five-minute Turing Test with an unskilled interrogator. What chance do you think a computer would have today? In another 50 years?

The paper Computing Machinery and Intelligence (Turing, 1950) asks "Can machines think?" and proposes it can be answered by an imitation game as a test, essentially reframing the question as "Can a machine interact with a person well enough such that the person can't tell whether it's a machine or another person?" Turing places limits on the communication medium between the person and the unknown entity to account for obvious shortcomings, e.g. looks, touch, handwriting. Today’s artificial intelligence, especially as evidenced by large language models from Google, ChapGPT, Claude, etc., have shown some objections to carry little weight. However, the Theological Objection is still fundamentally valid because it debates whether thinking includes elements of spirituality, but Turing adequately reframes the definition of thinking back to interactions. The Consciousness Objection is also fundamentally valid to the question and definition of thinking, but again made moot by the proposed test. The Disabilities Objection still holds weight, because the imitation test is too conveniently designed to eliminate natural human abilities form impacting the assessment to the advantage of the computer for the purposes of equalizing the method of expression from machine and person. The Extrasensory Perception Objection still holds great weight because no machine has shown psychokinesis.

I propose another objection to Turing: the Connection Objection. Recent studies in childhood and parenting have shown that emotional connection is a critical aspect of relationships and evolves with every interaction. This is inadequately hinted at by Turing’s Informality of Behavior Objection. Today’s artificial intelligence systems learn separately from analysis/response. The difference between a machine and person would become obvious with sufficient time and interactions.

A key issue with predicting whether a machine will pass a five-minute Turing Test is a matter of economics. The impressive capabilities of today’s large language models show a very high chance that the state of artificial intelligence could pass the Turing Test. However, the investment in hundreds of billions of dollars would not prioritize passing a test without any financial reward. Investment demands return. Today’s artificially intelligent systems are prioritizing solutions to economically valuable problems, not a philosophical problem.

1.3 Are reflex actions (such as flinching from a hot stove) rational? Are they intelligent?

Reflex actions such as flinching from a hot stove are rational because it is logical to quickly take action to prevent burning oneself. Reflex actions are not intelligent because intelligence typically implies careful consideration and slow deliberation.

1.4 Suppose we extend Evans's ANALOGY program so that it can score 200 on a standard IQ test. Would we then have a program more intelligent than a human? Explain.

Suppose we extend Tom Evan’s ANALOGY program so that it can score 200 on a standard IQ test. This program would still not be more intelligent than a human because its domain of expertise would be limited to domain of IQ tests. One metric of intelligence is insufficient to broadly claim higher intelligence.

1.5 The neural structure of the sea slug Aplysia has been widely studied (first by Nobel Laureate Eric Kandel) because it has only about 20,000 neurons, most of them large and easily manipulated. Assuming that the cycle time for an Aplysia neuron is roughly the same as for a human neuron, how does the computational power, in terms of memory updates per second, compare with the high-end computer described in Figure 1.3?

The neural structure of the sea slug Aplysia has been widely studied because it has only 20,000 neurons, most of them large and easily manipulated. Assuming that the cycle time for an Aplysia neuron is roughly the same as for a human neuron, then the computational power, in terms of memory updates per second, is 2x10^10. This is on par with a personal computer with 4 CPUs.

1.6 How could introspection – reporting on one's inner thoughts — be inaccurate? Could I be wrong about what I'm thinking? Discuss.

Introspection – reporting on one’s inner thoughts – could be inaccurate for a variety of reasons. According to cognitive science and psychology, one’s inner thoughts depend on incoming information and this information could be incorrect due to misunderstanding or poor perception like bad eyesight or poor hearing. The representation of the incoming information could also be inaccurate because of emotional events of the past causing incorrect assessments. And, in translating thoughts into action, such as communicating my thoughts, I could be wrong about what I’m thinking because of hidden motives or poor communication. Ultimately, we have no way to compare the state of mental information processing with external objective, observable truths.

1.7 To what extent are the following computer systems instances of artificial intelligence:Supermarket bar code scanners, Web search engines, Voice-activated telephone menus, Internet routing algorithms that respond dynamically to the state of the network.

Supermarket bar code scanners contain do not contain artificial intelligence because they algorithmically decode a series of lines and spaces arriving on an optical sensor and output a matching sequence of numerals according to a clearly defined set of rules. This is equivalent to a reflex, not intelligence.

Web search engines today are highly advanced instances of artificial intelligence. Search engines can accept complex search queries, pass the query through a large language model, and lead with the output of that model. Early search engines from 2000, however, were reflexive and not intelligent because they worked with hash tables to translate search queries to URLs.

Voice-activated telephone menus, where a person’s voice drives the navigation of a pre-defined map, is a moderate instance of artificial intelligence. It takes a human-level of thinking to interpret an audio stream into words or intent because of the complexity of human voices and accents. Once the intent is analyzed, then the machine takes logical action consistent with the intent.

Internet routing algorithms that respond dynamically to the state of the network are reflexive and not intelligent because they follow clearly defined sets of rules that don’t require sophisticated perception and analysis. However, the development of the original internet routing algorithms was highly intelligent.

1.8 Many of the computational models of cognitive activities that have been proposed involve quite complex mathematical operations, such as convolving an image with a Gaussian or finding a minimum of the entropy function. Most humans (and certainly all animals) never learn this kind of mathematics at all, almost no one learns it before college, and almost no one can compute the convolution of a function with a Gaussian in their head. What sense does it make to say that the “vision system” is doing this kind of mathematics, whereas the actual person has no idea how to do it?

The sense and purpose of saying that the “vision system” is doing complex mathematical operations is to state a hypothesis that describes our best understanding of how the human brain and intelligence develops from a set of neurons and synapses that connect them. The mathematics propose a theory of operation that model the vision system. The person sees without knowing how they see, just as a person moves without knowing how to move.

1.9 Why would evolution tend to result in systems that act rationally? What goals are such systems designed to achieve?

Evolution tends to result in systems that act rationally because rationality is determined by the degree to which systems behave relative to ideal performance. If a system were to behave completely opposite to ideal performance, then its resources would be wasted in ways that do not serve its purposes and the system would expire. A simple example of an irrational system is a rabbit that hops 100 feet away for a single beautiful green blade of grass, rather than eating the acceptable blade beside the one it ate last. The rabbit would perish and not reproduce to pass on its irrational habits. The rational rabbit succeeds.

Whether such systems are designed is beside the point. Rational systems survive and give an opportunity to evolve because they behave ideally. This is true in life, business, and AI.

1.10 Is Al a science, or is it engineering? Or neither or both? Explain.

Artificial Intelligence is a field of science and engineering. AI is a field of science because deep study, hypothesis, and testing to model human intelligence in theory. AI is also a field of engineering because people and businesses are applying the scientific and mathematical principles to design, build, and maintain systems, structures, and technologies that solve problems and improve society.

1.11 "Surely computers cannot be intelligent-they can do only what their programmers tell them." Is the latter statement true, and does it imply the former?

The latter statement, computers “can do only what their programmers tell them”, is true. Computers are composed of centralized processing units (CPUs) that literally execute instructions, which is what their programmers and programs tell them. However, this does not imply the former statement, “computers cannot be intelligent”. Computers can be intelligent, as demonstrated by today’s large language models by GPT (OpenAI), Claude (Anthropic), Gemini (Google), LLaMA (Meta), DeepSeek (DeepSeek), Grock (xAI).

1.12 "Surely animals cannot be intelligent—they can do only what their genes tell them." Is the latter statement true, and does it imply the former?

The latter statement, animals “can do only what their genes tell them”, is true. Animals fulfil their DNA expression. The latter statement does not imply the former statement because there are animals that are reflexive and not intelligent, e.g. jellyfish.

1.13 "Surely animals, humans, and computers cannot be intelligent-they can do only what their constituent atoms are told to do by the laws of physics." Is the latter statement true, and does it imply the former?

The latter statement “animals, humans, and computers can do only what their constituent atoms are told to do by the laws of physics” is true. We are all acting in accordance to the actions of our atoms. The latter statement does not imply the former statement “animals, humans, and computers cannot be intelligent” because humans demonstrate intelligence in perceiving, understanding, predicting, and manipulating the world.

1.14 Examine the AI literature to discover whether the following tasks can currently be solved by computers: Playing a decent game of table tennis (Ping-Pong). Driving in the center of Cairo, Egypt. Driving in Victorville, California. Buying a week's worth of groceries at the market. Buying a week's worth of groceries on the Web. Playing a decent game of bridge at a competitive level. Discovering and proving new mathematical theorems. Writing an intentionally funny story. Giving competent legal advice in a specialized area of law. Translating spoken English into spoken Swedish in real time. Performing a complex surgical operation.

Today’s large language models offer a wonderful opportunity to efficiently/quickly research literature. Prompt: “Examine the AI literature to discover whether the following task can currently be solved by computers, and summarize in 3 sentences:”

Grok (xAI) says: As of 2025, AI-powered robots can play a decent game of table tennis, sustaining long rallies and defeating beginner to intermediate human players in real matches. Leading examples include Google DeepMind’s system (winning 55% against intermediate opponents), MIT’s high-speed robot (88% rally success), and UC Berkeley’s humanoid that achieved 106 consecutive hits against humans. While these systems already perform at amateur-to-intermediate human level with adaptive spin handling and precise placement, they remain limited by hardware reach, extreme spin recovery, and lack of full-court mobility compared with professional players.

Grok (xAI) says: As of November 2025, AI systems cannot currently solve the task of fully autonomous driving in the center of Cairo, Egypt, due to persistent challenges including chaotic, lane-less traffic with high pedestrian and vendor density, inadequate infrastructure (e.g., missing markings and signage), data scarcity for training models, and underdeveloped regulations. While simulations like multi-agent models for Cairo's Ring Road demonstrate feasible traffic modeling, and prototypes such as Brightskies' 2020 Level 4 highway system have been authorized for testing, no real-world deployments exist that handle central Cairo's unconventional urban dynamics. Recent progress includes IoT approvals for smart vehicles and national AI strategies, but literature highlights that these "unconventional odds" in the Arab region, including Egypt, limit AI to controlled or simulated environments rather than operational success in dense city centers.

Grok stopped responding at this point. I tried GPT but was greeted by a paywall. Then, I tried Gemini:

Gemini (Google) says: Based on AI literature, driving in Victorville, California, is a task that can be solved by today's advanced autonomous systems, although it is not a "solved" problem in terms of being commercially available to the public. Victorville was famously the site of the 2007 DARPA Urban Challenge, which established the foundational capability for AI to navigate a staged urban environment, and current technology is significantly more advanced. While robotaxi services are geofenced to other cities, the California DMV permits dozens of companies to test autonomous vehicles (with safety drivers) on any public road in the state, making the challenge of driving in a standard grid-like suburban city like Victorville well within the capabilities of current test-level AI.

Gemini (Google) says: Based on AI literature, no, a computer cannot currently solve the complete task of physically buying a week's worth of groceries at a market. While "AI agents" can autonomously order groceries online for delivery, the physical act of navigating a crowded store, identifying varied products (e.g., a specific apple vs. a can), and manipulating diverse items (from eggs to heavy jugs) remains an unsolved robotics challenge. Current in-store robots are limited to tasks like scanning shelves for inventory, not the end-to-end process of shopping and purchasing.

Gemini (Google) says: Yes, based on AI literature, a computer can currently solve the task of buying a week's worth of groceries on the Web. This capability is being driven by the development of "agentic AI," which moves beyond simple chatbots to act as autonomous, goal-driven systems that can plan, manage budgets, and execute multi-step web tasks. These AI agents can compare prices across different retailers, integrate with loyalty programs, select items based on user preferences, and navigate the online checkout process from start to finish. While not yet a mainstream consumer product, the underlying technology is operational, and major retailers are actively developing and deploying these systems to automate the entire online shopping process.

Gemini (Google) says: Yes, based on AI literature, a computer can play a decent game of bridge at a competitive level, and has even defeated top-tier human world champions. An AI named "NooK" famously beat eight world champions in a 2022 tournament, demonstrating superhuman skill in the card-playing phase of the game (known as declarer play). However, this victory came with a significant caveat: the tournament omitted the bidding phase, which is widely considered the most complex part of bridge due to its reliance on partnership, communication, and reasoning with incomplete information. Therefore, while AI has mastered the "solved" mathematical aspects of card play, it has not yet solved the full, nuanced game of competitive duplicate bridge, particularly the art of cooperative bidding.

Gemini (Google) says: Yes, based on AI literature, computers are now capable of both discovering new mathematical conjectures and, in some cases, finding proofs for them, though this is not a fully autonomous "solved" problem. AI systems, particularly in collaborations with human mathematicians, have successfully guided the discovery of new theorems in fields like knot theory by identifying novel patterns that humans had missed. Separately, modern AI systems like Google's "FunSearch" have demonstrated the ability to find new solutions—which are a form of proof—for long-standing, unsolved problems, such as the "cap set problem." However, the current state-of-the-art involves AI as a powerful assistant; a system that independently intuits a new, interesting field of study, formulates a deep conjecture, and carries out the creative proof from scratch does not yet exist.

Gemini (Google) says: Based on AI literature, no, a computer cannot currently solve the task of writing an intentionally funny story, as it lacks the consciousness, lived experience, and emotional intelligence required for genuine intent. While modern AI can successfully generate formulaic jokes, puns, and even satirical headlines by recognizing and mimicking patterns from vast datasets, it does not actually "understand" why something is funny. The literature concludes that sophisticated comedic storytelling—which relies on cultural nuance, authentic vulnerability, and a deep understanding of the human condition—remains a uniquely human capability that AI cannot yet replicate.

Gemini (Google) says: Based on AI literature, no, a computer cannot currently solve the task of giving competent legal advice in a specialized area of law. The consensus in the literature is that while AI is a powerful tool for assisting lawyers with tasks like research, document drafting, and contract analysis, it cannot replace the required human judgment, ethical reasoning, and accountability that constitute "competent advice." Bar associations and legal scholars warn that AI models are prone to "hallucinations" (fabricating legal cases and facts) and lack a true understanding of legal nuance, making it ethically impermissible and professionally dangerous to rely on them without rigorous human oversight. Therefore, AI is treated as a sophisticated assistant that enhances a lawyer's competence, not as an autonomous advisor that possesses competence of its own.

Gemini (Google) says: Yes, based on AI literature, computers can currently solve the task of translating spoken English into spoken Swedish in real time. Advanced AI models can now perform direct speech-to-speech translation, a process that captures the input audio, translates it, and generates synthetic speech in the target language, all within a few seconds. This technology is no longer purely experimental; it is already integrated into major consumer products like Google's and Samsung's smartphones and specialized commercial platforms used for international conferences. Current research is now focused less on the basic possibility of this task and more on perfecting it by reducing latency and better preserving the original speaker's emotional tone and prosody.

Gemini (Google) says: Based on AI literature, no, a computer cannot currently solve the task of performing a complex surgical operation fully autonomously. While surgical robots are now used as high-precision tools under a surgeon's direct control, the AI literature makes a clear distinction between this "AI-assisted" surgery and full autonomy. Groundbreaking research has shown robots can autonomously perform specific, pre-programmed tasks—like suturing or a gallbladder removal on a pig—in a lab, but no system can yet manage the dynamic, unpredictable, and high-stakes environment of a complete, complex operation on a human. The consensus is that AI will continue to evolve as a "co-pilot" that enhances a surgeon's capabilities, but the "solved" goal of an independent, thinking robot surgeon remains a distant prospect.

1.15 Various subfields of Al have held contests by defining a standard task and inviting researchers to do their best. Examples include the DARPA Grand Challenge for robotic cars, the International Planning Competition, the Robocup robotic soccer league, the TREC information retrieval event, and contests in machine translation and speech recognition. Investigate five of these contests and describe the progress made over the years. To what degree have the contests advanced the state of the art in Al? To what degree do they hurt the field by drawing energy away from new ideas?

The contests have advanced the state of the art in AI in the problem areas of translation, planning, and autonomous driving. There is no apparent evidence that the contests hurt the field by drawing energy away from new ideas. Rather, assuming a prize purse and reward, the contests provide incentive to accelerate research and development in the field.

Research is an ideal opportunity to use AI. Prompt: “Various subfields of Al have held contests by defining a standard task and inviting researchers to do their best. Examples include the DARPA Grand Challenge for robotic cars, the International Planning Competition, the Robocup robotic soccer league, the TREC information retrieval event, and contests in machine translation and speech recognition. Investigate five of these contests and describe the progress made over the years. Limit the response to progress to 3 sentences per contest.”

Gemini (Google) says: Here is an investigation of five major AI contests and the progress made in each.

🤖 DARPA Grand Challenge (Autonomous Driving)

The 2004 Grand Challenge was a landmark failure where no robotic vehicle finished the 150-mile desert course, with the best traveling only 7.3 miles. Just one year later in 2005, this failure catalyzed massive progress, resulting in five vehicles successfully completing an even more difficult 132-mile course. This competition is credited with kick-starting the modern autonomous driving industry, proving the viability of using LIDAR, sensor fusion, and machine learning for real-world navigation.

🏆 RoboCup (Robotic Soccer)

Since its first official competition in 1997, RoboCup has driven rapid progress in multi-agent robotics, coordination, and computer vision. Early humanoid league games featured slow, unstable robots struggling to find the ball, whereas modern competitions showcase teams of robots that can walk, run, and kick with agility. The challenges have evolved from basic 2-vs-2 matches to complex, coordinated team play, advancing the ultimate goal of defeating a human World Cup champion team by 2050.

📊 International Planning Competition (IPC)

The IPC, running since 1998, has been the primary driver for progress in automated planning and scheduling, the AI subfield focused on finding a sequence of actions to reach a goal. The competition's evolving tracks—from classical (deterministic) problems to more complex numeric, temporal, and probabilistic challenges—have pushed planners to become exponentially faster and more capable. This has resulted in AI moving from solving simple "toy" problems to successfully tackling complex, real-world logistics, manufacturing, and robotics tasks.

📚 TREC (Text REtrieval Conference)

Since the early 1990s, the TREC has benchmarked information retrieval systems, providing the standard data that directly fueled the development of modern search engines. Early "ad-hoc" tasks focused on finding relevant documents for simple queries, but the competition's progress is marked by its introduction of increasingly complex tracks like question answering, cross-language retrieval, and spam filtering. This evolution has pushed the field beyond simple keyword matching to understanding user intent, context, and nuance, as seen in today's sophisticated search tools.

🗣️ WMT (Workshop on Machine Translation)

The WMT competitions have chronicled the most dramatic performance leap in modern AI: the shift from statistical machine translation (SMT) to neural machine translation (NMT). For years, progress in SMT was slow and incremental, but the introduction of NMT models (like the Transformer) around 2016 caused quality scores to skyrocket, producing translations that were far more fluent and accurate. The competition's focus has now shifted from mainstream language pairs to solving harder problems, such as "low-resource" languages, speech-to-speech translation, and multilingual models.