Chapter 5 asked which existing contributions from the AI field we can recognise and place in the “living intelligence” framework. This chapter is that survey: seventy years, at altitude.
The golden age begins at Dartmouth College, New Hampshire, in the summer school of 1956. John McCarthy invites all scholars interested in computing towards intelligence. Among those present:
The term Artificial Intelligence was actually coined there, to describe the overall new field of research. And the deck flags something that will run through the entire history: the dualism between AI as intelligent behaviour and AI as intelligent process was already there, from the beginning.
That dualism is not a historical curiosity, it is the same fault line met twice already. Chapter 1 traced it in the Turing Test, a criterion that is behavioural by construction: independent of how the thing works, dependent on what it achieves. Chapter 5 traced it in the definition of intelligence itself: the bacterium exhibits intelligent behaviour with no intelligent process in any recognisable sense. Which of the two you are trying to build determines what counts as success.
Building general-purpose intelligent machines is the goal of General AI. The obstacle is stated with disarming honesty: we have a poor understanding of human intelligence, and of intelligence in general. Early AI therefore focused on intelligent components — decomposing the problem, in the absence of a theory of the whole.
Understanding the environment, through sensors of any sort, and interpreting the overall situation. The deck marks it explicitly: one of the most difficult tasks of AI. Chapter 4 has already told us why this matters structurally — perception is one half of the minimal agent, and the perception-versus-representation problem is one of the things an agent must encapsulate.
Learning from data; building models, for instance classification; making predictions. The example given is face recognition through training. Section 10 returns to it as the driver of the modern era.
Problem solving and planning: devising a course of actions towards a goal, based on a repertoire of actions. The example given is playing games. This component gets a whole chapter of its own, Chapter 10.
Representing knowledge; inferring new knowledge from available knowledge; and doing so in a consistent and robust way. Chapters 8 and 9 are the long version of this one tab.
The ability to understand human languages, either spoken or written, possibly engaging in conversations with humans. Chapter 1 has already recorded what happened to this component: it is the one that broke out, and Chapter 11 will argue that its dominance produced a logocentric bias in how we judge intelligence.
Hold this list of five next to Chapter 4. There, the question was what an intelligent agent should encapsulate, and the answer named knowledge representation and reasoning, practical reasoning and planning, machine learning, natural language processing, perception and action. It is the same list. The difference is the framing: in 1956 these were the subfields into which AI decomposed; in Chapter 4 they are the techniques an agent has to hold together. That shift from decomposition to composition is the whole content of the phrase “intelligent systems engineering”.
The first great theme of early AI. Searching for a solution means going from the current state of the world to a desired one, called the goal, based on a repertoire of available actions modelled in terms of:
From these, a search tree can be built: from the initial state as its root, giving forward search; or from the final state as its root, giving backward search.
Three things are worth noticing about this paragraph, because they are the skeleton of two later chapters. The action model — preconditions and effects — is exactly the one Chapter 10 formalises as feasibility preconditions and expected effects, and then as the PRE/ADD/DEL of STRIPS. The forward/backward duality is exactly the forward chaining and backward chaining of Chapter 8, where it becomes a question about how to traverse a proof tree. And the whole thing is the reason Chapter 9 can say that automated reasoning relies almost universally on the AI paradigm of search.
And then the trouble. Explosion of the space of solutions: search trees become too large, in a combinatorial explosion — the example given is the game of chess.
The possible approach: heuristics, as rules of thumb to direct the search and reduce computational complexity. The example: IBM’s Deep Blue [Campbell et al., 2002] against Gary Kasparov in 1996, combining computational power with heuristics.
Two celebrated systems of the golden age, each impressive and each carrying its own limitation.
SHRDLU [Ward, 2006] worked in the Blocks World: a virtual environment where a number of coloured objects — blocks, boxes, pyramids — are arranged on a surface, a table. SHRDLU was aimed at arranging the objects following the user’s instructions, using a simulated robot arm with heuristics. And it worked: both the user interaction via a sort of natural language, and the practical reasoning — reasoning towards action. The limit: language and environment were actually very simplistic, not representative of real-world interaction.
SHAKEY with STRIPS [Furcy and Koenig, 2006] was a remotely-controlled mobile robot, moving objects in an office-like environment by pushing them, planning with STRIPS. It somehow worked. But it had a limited ability to perceive and understand the real-world environment, so the environment had to be built around it to make SHAKEY work. The deck’s verdict is memorable: basically, it looked like a real-world robot, but never worked in the real world.
Both systems are worth remembering for what they bequeathed rather than for what they failed at. The Blocks World becomes the standard example domain of planning, and you will meet it again in Chapter 10 in its STRIPS formulation with stack, unstack, put and puttable. STRIPS itself becomes the ancestor of PDDL and of an entire competition series. Failure in the real world did not prevent them from defining the vocabulary.
Then the correction. The deck gives two reasons, one empirical and one theoretical.
Microworlds do not scale up to the real world. Solutions in microworlds apparently do not work in the real world — and the deck leaves the diagnostic question open: qualitative or quantitative issue?
NP-complete problems. Non-deterministic Polynomial time complete problems are easy to check for correctness of solutions, and almost impossible to find a solution for in a reasonable amount of time — the example given being the Travelling Salesman Problem [Papadimitriou, 1977]. And the exclamation that follows: almost all relevant problems in AI are (at least) NP-complete.
Together with hype, that was enough. AI started with way too much hype; the General Problem Solver never worked for real-world problems [Newell et al., 1959]; in a couple of decades nothing was really accomplished. Funding for AI research was cut everywhere, starting from the seventies.
Be able to give both halves of the AI winter diagnosis, since giving only one is the common mistake. The practical half: microworld results did not transfer, and the hype had promised that they would. The theoretical half: almost all relevant AI problems are at least NP-complete, so the failure was not a matter of waiting for faster machines. Chapter 9 will restate the same lesson in its own vocabulary: decidable does not mean tractable.
The way out came from an observation about people. Humans do not explore enormous search spaces to find solutions: they use (possibly extensive) domain knowledge to cut down the search space — chess players being the example. Hence the idea: exploiting explicit domain knowledge representation to make solutions computable.
A new class of AI systems emerged. They were no longer General AI, since they were tailored to specific domains and worked only there — yet they were finally facing real-world problems successfully. Expert systems were built by:
MYCIN [Shortliffe, 1976]: nearly 200 rules, about blood disease, working as a doctor’s assistant. Its issues were two, and they are the permanent issues of the approach: it is difficult to extract knowledge from humans and to precisely formalise it in terms of rules, and the system operates only in a very narrow area of expertise.
CYC [Lenat et al., 1985]: representing everything in the world in terms of rules — an expert system containing all common-sense human knowledge. Doug Lenat’s CYC started in 1984 and is still alive. Way too ambitious; some call it a failure; yet it now lives out of its commercial applications.
The trade the expert systems era made is the trade that structures everything after it. Give up generality, gain results. General AI was aiming at general-purpose intelligent machines and delivering microworlds; expert systems abandoned generality by construction and delivered working assistants. Note also the fourth bullet in the list above, which will look very valuable again in Chapter 13: expert systems produced an answer along with an explanation. Section 10 shows what happened when that bullet was dropped.
Logic programming is presented here as the rigorous version of the same knowledge-based idea: capturing knowledge and inferring new knowledge with the precision and rigour of logic. Computing with logic means the PROLOG programming language [Colmerauer and Roussel, 1996], nowadays usually written Prolog, where knowledge and programs are expressed in terms of logic axioms in the form of Horn clauses, altogether forming a logic theory against which goals are demonstrated — computation as goal demonstration [Lloyd, 1984].
Its issues, as listed: heavy computational load, and not fitting (alone) real-world representation and computation. Chapter 8 gives this paragraph its full treatment, including why the declarative and procedural readings of a clause coincide.
Coping with uncertainty, on a different basis. The deck states the premise bluntly: humans are not good decision makers. Rational decisions are better than human decisions in most fields, particularly where uncertain information is available — for instance when probabilistic reasoning is involved. Bayes developed the mathematics in the eighteenth century, and a couple of centuries were required to make Bayesian reasoning usable for AI [Barber, 2012] — many NLP systems, for instance, use Bayesian networks.
A different reaction to the same frustration, and a much more radical one. Against reasoning, against representation [Brooks, 1991a; Brooks, 1991b]: frustration from the absence of real-world results led to a focus on the real world itself, against knowledge-based and logic-based AI, with perceive and react as the main cycle of robots in the real world, rather than symbolic reasoning.
Hence the robotics revolution [Brooks, 1986]: focus on the expected specific intelligent behaviours; just build them, according to how they relate to each other — in particular, which ones take precedence over the others. The example given is the Roomba vacuum cleaner.
Read nouvelle AI through Chapter 5 and it stops looking like a heresy. Brooks is building non-explicit intelligence: competence with no representation, behaviour that works beautifully without the system having any idea of what it is doing. The bacterium is the existence proof that this is a legitimate form of intelligence rather than a shortcut — and Chapter 4 records the same lineage from the agent side, where the robotic notion of agent brings with it the non-symbolic approach and the subsumption architecture.
The Grand DARPA Challenges: a race for autonomous vehicles in the desert of Nevada, 2005, won by STANLEY [Thrun et al., 2006] — a converted Volkswagen Touareg equipped with seven onboard computers, interpreting sensor data from GPS, laser rangefinders, radar and video feed.
Machine learning aims at computers that learn knowledge they have not been explicitly told [Michalski et al., 1983]; they are typically trained. Two paradigms are named:
| Paradigm | How it works |
|---|---|
| Supervised learning | Training by examples of the thing the computer is trying to learn — the training data. Example: face recognition. |
| Reinforcement learning | A system is able to experiment by making decisions, receives feedback on those decisions, and learns from that. |
And the diagnosis of why now, in three factors:
Neural nets implement machine learning by connecting many artificial neurons in complex networks [Lippmann, 1987]. Each neuron gets input from neighbours and produces an output based on some weights; learning occurs by adjusting weights. The research came in two phases: first, simple neural networks show potential and are proven to have limits; then more complex networks were proven able to overcome these problems. The example given is character recognition.
And the issue the deck marks with an exclamation: they are opaque, no explanation made available — knowledge hidden in weights.
Put this next to the expert-systems list in section 7, where providing an answer along with an explanation was a defining capability. The modern era did not merely fail to provide explanations; it adopted a representation in which the knowledge is not anywhere inspectable. That is the exact problem Chapter 13 takes up, and it is also why Chapter 11 insists that an agentic system is more than the model at its core: everything that surrounds the model — state, goals, plans, tool calls, traces — is legible in a way the weights are not.
AlphaGo [Silver et al., 2016]. In 2014 DeepMind demonstrated a system learning how to play arcade games just by looking at the video and accessing the scores, using the same controls as humans. Acquired by Google, they built AlphaGo, which beat Go champion Lee Sedol 4 to 1 in 2016, exploiting deep neural networks along with self-training.
Why it mattered: the Go search space is so huge that brute force just does not work, so it was considered impossible for a machine to beat a human. The consequence the deck draws is a statement about expectations rather than about Go: this also made everybody know that there were no known limits to the ability that machine intelligence could reach.
The closing list [Wooldridge, 2018] is deliberately banal, and that is its argument. AI everywhere: digital assistants like Siri, Alexa and Cortana are classic AI; face recognition software in social media is AI; the software in a car’s satellite navigation system is AI; product recommendations from online stores are made by AI; cruise control systems in modern cars are AI; AI-based automated translation systems are routinely used across the world; LLMs pervade almost any application domain; and agentic AI is the most hyped technology today. What is next?
AI is just everywhere already.
The historical narrative is examinable as a shape, not as a list of dates. The shape: general ambition (1956, components of intelligence) → success in microworlds that did not transfer → theoretical and financial winter (NP-completeness, hype, funding cuts) → two escapes, one through explicit knowledge (expert systems, logic programming, Bayesian reasoning) and one through no knowledge at all (nouvelle and behavioural AI) → the modern era where a third escape, learning, wins on data and compute → and today, where the winner is powerful and opaque, and agentic AI is the current hype. If you can narrate that, everything in this deck hangs off it.
John McCarthy invited all scholars interested in computing towards intelligence to a summer school at Dartmouth College, New Hampshire. Among the participants were Marvin Minsky, later co-founder of the MIT AI Lab; Alan Newell and Herb Simon, authors of Logic Theorist, an automatic theorem prover and likely the first AI program; and McCarthy himself, inventor of LISP, the first programming language for AI. The term Artificial Intelligence was coined there — and the dualism between AI as intelligent behaviour and AI as intelligent process was already present.
Perception (understanding the environment through sensors, interpreting the situation — flagged as one of the most difficult tasks of AI); machine learning (learning from data, building models such as classifiers, making predictions); problem solving and planning (devising a course of actions towards a goal from a repertoire of actions); reasoning (representing knowledge and inferring new knowledge consistently and robustly); natural language understanding (understanding spoken or written human language, possibly conversing).
Searching for a solution means going from the current state of the world to a desired one, the goal, based on a repertoire of available actions modelled by their pre-conditions (the state of the world in which they can be applied) and their effects (the state they bring about). A search tree can then be built from the initial state as root — forward search — or from the final state as root — backward search.
The explosion of the space of solutions: search trees become too large, in a combinatorial explosion — the game of chess being the example. The mitigation is heuristics, rules of thumb to direct the search and reduce computational complexity, as in IBM’s Deep Blue against Gary Kasparov in 1996, which combined computational power with heuristics.
Working in the Blocks World — coloured blocks, boxes and pyramids on a table — it arranged objects following user instructions, using a simulated robot arm with heuristics. It worked, both for user interaction through a sort of natural language and for practical reasoning, that is reasoning towards action. Its limit: language and environment were actually very simplistic, and not representative of real-world interaction.
A remotely-controlled mobile robot moving objects in an office-like environment by pushing them, planning with STRIPS. It somehow worked, but it had a limited ability to perceive and understand the real-world environment, so the environment had to be built around it. In the deck’s summary: it looked like a real-world robot, but never worked in the real world.
Practical: microworld solutions apparently do not work in the real world; AI started with way too much hype; the General Problem Solver never worked for real-world problems; in a couple of decades nothing was really accomplished, and funding was cut everywhere from the seventies. Theoretical: NP-complete problems are easy to check but almost impossible to solve in reasonable time — and almost all relevant problems in AI are at least NP-complete.
The idea: humans do not explore enormous search spaces, they use extensive domain knowledge to cut the search space down — chess players being the example — so exploit explicit domain knowledge representation to make solutions computable. Expert systems were built by eliciting knowledge from domain experts, representing it as rules, answering human questions over a domain, providing an explanation along with the answer, and often coping with uncertainty. They were no longer General AI: tailored to specific domains, working only there, but finally facing real-world problems successfully.
MYCIN: nearly 200 rules about blood disease, working as a doctor’s assistant; difficult to extract knowledge from humans and to formalise it precisely as rules, and operating only in a very narrow area of expertise. CYC: representing everything in the world as rules, an expert system containing all common-sense human knowledge, started by Doug Lenat in 1984 and still alive; way too ambitious, called a failure by some, yet living on through its commercial applications.
Nouvelle AI is against reasoning and against representation: frustration at the absence of real-world results led to focusing on the real world itself, against knowledge-based and logic-based AI, taking perceive and react as the main cycle of robots rather than symbolic reasoning. Behavioural AI is the robotics revolution that follows: focus on the expected specific intelligent behaviours, just build them according to how they relate to each other, and in particular which take precedence over which — the Roomba being the example.
Scientific breakthroughs, with deep learning dealing with complex problems; the fact that training requires lots of data, and data are nowadays hugely available; and the fact that training requires computational power, which is more and more available.
Because learning occurs by adjusting weights, and the result is opaque: no explanation is made available, and the knowledge is hidden in the weights. Set against expert systems, which delivered an answer together with an explanation, this is a regression in one specific dimension — the one that Chapter 13 addresses under the name of explainability, and that Chapter 1 listed among the standing requirements on artificial systems as trustworthiness and accountability.
Because the Go search space is so huge that brute force simply does not work, so beating a human champion was considered impossible for a machine. DeepMind had shown in 2014 a system learning arcade games just from video and scores, using the same controls as humans; AlphaGo then beat Lee Sedol 4 to 1 in 2016 using deep neural networks with self-training. The consequence drawn is about expectations: it made everybody aware that there were no known limits to what machine intelligence could reach.