33.07.02 · history-of-science / digital-revolution

Computing history: Turing, von Neumann architecture, transistor, and the internet

stub3 tiersLean: nonepending prereqs

Anchor (Master): Turing, A. M. — On Computable Numbers (1936)

Intuition Beginner

The computer was invented twice — once in theory and once in practice. In 1936, Alan Turing (1912-1954) described a theoretical "universal machine" that could compute anything that is computable. He proved that some problems can never be solved by any computer (the halting problem). During WWII, Turing built real machines at Bletchley Park to crack Nazi Germany's Enigma cipher — work that shortened the war and saved millions of lives.

The ENIAC (1945) was the first general-purpose electronic computer — 30 tons, 18,000 vacuum tubes. John von Neumann designed the "stored program" architecture, in which computers hold their own instructions in memory. Nearly every computer built since uses this design. The earliest programmers were often women, including the six mathematicians who configured ENIAC for its first runs.

The transistor (1947) replaced fragile vacuum tubes — smaller, cheaper, more reliable. Integrated circuits (1958) placed thousands of transistors on a single chip. Gordon Moore noticed that transistor density doubled roughly every two years ("Moore's Law"), a trend that has driven the digital revolution for six decades.

ARPANET (1969) connected distant computers by packet switching — the birth of the internet. Tim Berners-Lee invented the World Wide Web (1989) — making the internet usable for everyone. The personal computer and then the smartphone placed a global network into billions of pockets.

Visual Beginner

Milestone	Year	Innovation	Significance
Turing machine	1936	Universal computation, halting problem	Defined what computers can and cannot do
Colossus	1943-44	First programmable electronic computer	Wartime codebreaking at Bletchley Park
ENIAC	1945	First general-purpose electronic computer	Ballistics, weather, hydrogen bomb
Von Neumann architecture	1945	Stored-program design	Template for all modern computers
Transistor	1947	Solid-state switch	Replaced vacuum tubes
Integrated circuit	1958	Many transistors per chip	Began miniaturisation
ARPANET	1969	Packet-switched network	Foundation of the internet
TCP/IP	1974	Universal networking protocol	A network of networks
World Wide Web	1989	Hypertext over the internet	Brought the net to the public
Smartphone	2007	Pocket, always-on computing	Computing becomes ubiquitous

Worked example Beginner

Moore's Law is one of the most striking quantitative regularities in the history of technology. Gordon Moore predicted in 1965 that the number of transistors on an affordable chip would double about every two years. To see what that implies, compare the Intel 4004 (1971) with a modern chip.

The Intel 4004 held 2,300 transistors. The years from 1971 to 2023 span 52 years, which is $52/2 = 26$ doubling periods. So the prediction gives a 2023 count of:

2, 300 \times 2^{26} \approx 2, 300 \times 67, 108, 864 \approx 154 billion transistors .

A 2022 Apple M1 Ultra chip contains about 114 billion transistors. The figures are in the same range, which is remarkable for a 1965 forecast.

Doubling produces growth that is hard to picture directly. A quantity that doubles every two years grows a thousandfold in twenty years and a millionfold in forty. That is why a chip that once held thousands of transistors now holds billions, and why a pocket phone outperforms a room-sized machine from the 1950s.

The trend cannot last forever. Transistors are approaching the size of single atoms, where the usual physics stops applying. The industry is now exploring multi-chip designs, specialised accelerators, and entirely new substrates.

Check your understanding Beginner

Exercise (easy, multiple choice).

What was Alan Turing's 1936 contribution to the foundations of computing?

A. He built the first electronic computer at Bletchley Park B. He defined an abstract machine capturing what "computable" means and proved some problems cannot be solved by it C. He invented the transistor at Bell Labs D. He designed the TCP/IP protocol for ARPANET

Hint

Turing's 1936 paper addressed a mathematical question about what a mechanical procedure can decide.

Answer

Option B.

Turing introduced an abstract device (now called a Turing machine) to make the notion of "effective procedure" precise. He showed that well-defined problems exist that no such machine can settle, most famously the halting problem. This set the absolute limits on what any computer can do, before any electronic computer had been built.

Exercise (easy, multiple choice).

What does the "stored program" idea in the von Neumann architecture mean?

A. The program is saved on a hard disk after it runs B. Instructions and data share the same memory, so a computer can be re-tasked by loading new instructions C. Programs are stored in the cloud and streamed on demand D. Each program is hard-wired into a separate chip

Hint

Think about why one piece of hardware can run a word processor, a game, and a simulation.

Answer

Option B.

In the von Neumann design the central processing unit, memory, and input/output are separate, but instructions live in the same memory as the data they act on. Loading a new program changes what the machine does without changing the hardware. This universality is what makes a computer general-purpose rather than special-purpose.

Formal definition Intermediate+

A Turing machine is the formal model Turing introduced in "On Computable Numbers" (1936) to make "computable by a mechanical procedure" precise. It is a 7-tuple $M = (Q, Σ, Γ, δ, q_{0}, q_{acc}, q_{rej})$ where $Q$ is a finite set of states, $Σ$ is the input alphabet, $Γ \supseteq Σ$ is the tape alphabet (including a blank symbol), $δ : Q \times Γ \to Q \times Γ \times {L, R}$ is the transition function, $q_{0}$ is the start state, and $q_{acc}, q_{rej}$ are distinct halting states. The machine operates on an unbounded tape divided into cells, reading and writing one symbol at a time under a mobile head.

A function $f$ is Turing-computable if some Turing machine halts on every input $x$ with $f (x)$ on the tape. The Church-Turing thesis asserts that every function computable by any effective procedure is Turing-computable. The thesis is not a theorem — "effective procedure" is informal — but every alternative model (lambda calculus, recursive functions, register machines) has proved equivalent in power, which is the evidence for it.

A universal Turing machine $U$ takes as input a description $⟨ M ⟩$ of an arbitrary machine $M$ together with input $w$ and simulates $M$ on $w$ . Universality is the theoretical seed of the stored-program computer: one physical device, given different instructions, carries out any computation. This is the precise sense in which Turing's 1936 paper contained the idea later engineered as the von Neumann architecture.

The von Neumann architecture is the engineering realisation of universality. It comprises (i) a central processing unit containing an arithmetic-logic unit and a control unit, (ii) a memory holding both data and instructions, (iii) input and output, and (iv) a bus connecting them. The fetch-decode-execute cycle reads an instruction from memory, decodes it, carries it out, and advances the program counter. Because instructions are ordinary data in memory, programs can be loaded, modified, and even generated at runtime — exactly what a universal machine prescribes.

A semiconductor transistor is a three-terminal solid-state device in which a small voltage or current at one terminal (the gate or base) controls the current between the other two. Manufactured in silicon by doping regions to be $n$ -type or $p$ -type, it operates as a switch and an amplifier. An integrated circuit fabricates many transistors (and their interconnections) on a single semiconductor substrate; a microprocessor is an integrated circuit implementing a complete CPU.

The internet is, formally, a federation of networks interconnected by the TCP/IP protocol suite. Packet switching breaks data into labelled packets routed independently across shared links and reassembled at the destination, in contrast to the dedicated circuits of the telephone network. The World Wide Web is an application-layer system of interlinked documents identified by URLs, transferred by HTTP, and authored in HTML.

Key result with derivation Intermediate+

Key theorem (Turing, 1936 — undecidability of the halting problem). There is no Turing machine that, given an arbitrary machine $M$ and input $w$ , decides whether $M$ eventually halts on $w$ .

Derivation (diagonal argument). Suppose a decider $H$ exists. On input $⟨ M ⟩$ and $w$ , $H$ halts and outputs "halt" or "loop." Build a machine $D$ that takes a description $⟨ M ⟩$ and calls $H$ on $⟨ M ⟩$ and $⟨ M ⟩$ itself; if $H$ says "halt," $D$ loops forever, and if $H$ says "loop," $D$ halts. Now run $D$ on its own description $⟨ D ⟩$ . If $D$ halts on $⟨ D ⟩$ , then $H$ reports "halt," forcing $D$ to loop — contradiction. If $D$ does not halt on $⟨ D ⟩$ , then $H$ reports "loop," forcing $D$ to halt — contradiction. So $H$ cannot exist.

The proof is a self-referential diagonal construction in the spirit of Cantor and Gödel. It shows a precise, permanent limit on what computers can do: there are well-posed yes-or-no questions no algorithm can settle.

Key result (Rice, 1953). Every nontrivial semantic property of Turing machines (does it halt on empty input? does it ever print a 1? does it accept the empty language?) is undecidable. Here "nontrivial" means the property holds of some machines and fails of others.

Sketch. If property $P$ is nontrivial, choose a machine $M_{yes}$ that has $P$ and a machine $M_{no}$ that does not. Given any $M$ , build $M^{'}$ that first simulates $M$ on its input and, if $M$ halts, behaves like $M_{yes}$ . If $P$ were decidable, we could decide whether $M^{'}$ has $P$ , and thereby decide whether $M$ halts — contradicting Turing's result. Hence $P$ is undecidable.

The historical consequence is sharp. No fully automatic program verifier, optimiser, or security analyser can answer every meaningful question about every program. The digital revolution enlarged what could be mechanised, but Turing's ceiling bounds what mechanisation can ever reach.

Exercises Intermediate+

Exercise (medium, short answer).

Explain why the universal Turing machine can be regarded as the theoretical ancestor of the von Neumann architecture. What does each contribute that the other does not?

Answer

A universal Turing machine shows that a single device can simulate any specialised machine, provided its instructions are given as data on the tape. That is the conceptual content of "stored program": instructions are data, and one physical machine serves every purpose. The von Neumann architecture is the engineering realisation — a concrete layout of CPU, memory, bus, and input/output, plus the fetch-decode-execute cycle, that makes the universal-machine idea physically buildable and fast.

What Turing contributes is the existence proof and the limitative result (some problems are uncomputable). What von Neumann contributes is the practical blueprint that turned an abstract machine into working hardware. The first without the second is pure theory; the second without the first is a design without justification for its universality.

Exercise (medium, short answer).

Packet switching was developed independently by Paul Baran (RAND) and Donald Davies (NPL). Explain what packet switching is and why it was preferred over circuit switching for a survivable military-research network.

Answer

Packet switching divides a message into small, individually addressed packets that are routed independently across shared links and reassembled at the destination. Circuit switching, as used by the telephone network, dedicates an end-to-end path for the duration of a call.

Packet switching was preferred because no single link or node carries the whole conversation: if a route is destroyed or congested, packets can be rerouted around it. This redundancy made the network resilient to partial failure — a central design goal for a network funded by the Defence Department during the Cold War — and also let many conversations share the same infrastructure efficiently.

Exercise (hard, essay).

Turing's proof that the halting problem is undecidable appeared in 1936, before any electronic computer existed. Why did a result about the limits of computation matter so much for the positive history that followed — the building of actual computers?

Hint

Consider what the universal machine establishes about the scope of a single physical device, and what the limitative result rules out as a goal for automation.

Answer

The positive force of Turing's paper lies in the universal machine. By exhibiting one machine that can simulate any other, Turing showed that a single physical device could in principle carry out any computation that any special-purpose device could. That legitimated the ambitious goal of building general-purpose hardware rather than yet another calculator tied to one task. ENIAC, EDVAC, and the von Neumann architecture inherit this universality as their core design principle.

The negative result — undecidability of the halting problem — set expectations honestly. It ruled out the fantasy of a machine that could settle every mathematical question, and later, via Rice's theorem, the fantasy of a fully automatic program verifier. Engineers could therefore aim at the achievable (powerful general-purpose computation) without chasing the provably impossible.

So the same paper supplied both the licence to build universal computers and the boundary on what they could ever be asked to do. The practical history of computing is, in a real sense, the working-out of both halves of Turing's 1936 argument.

Advanced results Master

The long arc of artificial intelligence

Artificial intelligence began as a formal programme at the 1956 Dartmouth Conference, where John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester coined the term and predicted human-level machine intelligence within a generation. The early decades alternated between optimism and collapse. The General Problem Solver of Newell and Simon, Arthur Samuel's checkers player, and early theorem provers suggested rapid progress; the Lighthill Report (1973) and the practical limits exposed by Minsky and Papert's Perceptrons (1969) produced the first AI winter, a deep contraction of funding and confidence.

The 1980s brought expert systems — MYCIN for diagnosis, DENDRAL for chemistry — encoding human expertise as large rule sets. They achieved commercial success in narrow domains but proved brittle, expensive to maintain, and unable to learn. A second winter followed from 1987 to 1993. The neural-network lineage, dormant since the perceptron critique, was revived by Rumelhart, Hinton, and Williams's 1986 backpropagation algorithm, which made it practical to train networks with hidden layers.

The current boom dates to AlexNet's 2012 win on the ImageNet benchmark, where a deep convolutional network cut the error rate dramatically. The convergence of massive internet datasets, GPUs originally built for graphics, and improved architectures (convolutional networks, transformers, diffusion models) produced breakthroughs in vision, language, and protein structure. AlphaGo's 2016 defeat of Lee Sedol signalled that AI could master domains once thought to require human intuition, and contemporary large language models extend this to open-ended text generation, code, and multimodal reasoning. Each of these strands rests on the theoretical foundation Turing laid in 1936: every one of these systems is, in the end, a program executing on a universal machine.

Computing, mathematics, and limits

Computability and complexity form a continuous thread from Turing to the present. The Cook-Levin theorem (1971) established NP-completeness of Boolean satisfiability, and Karp's 1972 paper catalogued twenty-one further NP-complete problems, revealing a hidden unity among apparently unrelated optimisation tasks. The P versus NP question — whether every problem whose answer can be checked quickly can also be solved quickly — is now a Clay Millennium Problem and the most consequential open question in the theory of computation. Cryptography depends on its presumed answer: RSA, Diffie-Hellman, and elliptic-curve systems are secure only because factoring, discrete logarithm, and related problems are believed to be hard.

Physical limits of computation are now studied with the same rigour as mathematical ones. Landauer's principle fixes a minimum energy cost $k T ln 2$ for erasing one bit of information, tying computation irreversibly to thermodynamics. The Bekenstein bound limits the information storable in a finite region of finite energy, and black-hole thermodynamics turns these bounds into sharp statements about information and gravity. Quantum computing, proposed by Feynman in 1982 and formalised by Deutsch in 1985, does not evade these limits but exploits superposition to solve certain problems (Shor's factoring, Grover's search) asymptotically faster than any known classical algorithm.

Computing, biology, and the sciences

Computational methods have rewritten the practice of the natural sciences. The Human Genome Project assembled three billion base pairs using algorithms for sequence alignment, assembly, and annotation; tools such as BLAST made sequence search routine. Molecular dynamics, climate models, and the data pipeline of the Large Hadron Collider each depend on massive computation, as does contemporary structural biology: AlphaFold's prediction of protein structures from sequence rests entirely on learned models running at scale.

Computational neuroscience models neural systems at levels from individual ion channels to whole-brain connectomes, with projects such as the Blue Brain Project and the Human Brain Project attempting large-scale simulation. Artificial life — Langton's loops, Ray's Tierra, evolutionary computation more generally — probes whether lifelike behaviour can arise from simple computational rules, paralleling the long-running philosophical question of whether minds could be machines.

Computing and society

The societal consequences of computing are now inseparable from its technical history. Mass surveillance is technically straightforward on infrastructure designed for packet routing; the Snowden revelations of 2013 made plain the scale on which states had adopted it. Shoshana Zuboff's "surveillance capitalism" names the commercial counterpart — the conversion of human experience into behavioural data that can be predicted, sold, and used to shape further behaviour. Algorithmic curation of social media has been linked to filter bubbles, polarisation, and the rapid spread of disinformation, while algorithmic decision-making in lending, policing, and hiring raises acute questions of bias and accountability.

Cybersecurity and cyber warfare have become domains of statecraft in their own right; Stuxnet (2010), which damaged Iranian uranium-enrichment centrifuges, marked the first widely-recognised use of malware as a weapon of physical sabotage. The digital divide — roughly 2.7 billion people still lack internet access — intersects with global inequality, gender, and geography, and determines who can participate in the economic, educational, and civic life the network now mediates. None of these issues is an external side-effect; each follows from design choices made during the decades traced in this unit.

The future of computing

Three frontiers define the present edge of the field. First, the slowing of Moore's Law pushes the industry toward multi-chip packaging, domain-specific accelerators (GPUs, TPUs, NPUs), and alternative substrates — photonics, spintronics, memristors, and exploratory molecular computing. Second, quantum computing has moved from speculation to small noisy machines; quantum supremacy claims (Google 2019) remain contested in practical scope, but the trajectory is real. Third, neuromorphic chips that mimic aspects of biological neural processing raise the old question of substrate independence in new form: could a sufficiently structured physical system realise conscious experience?

The safety of advanced AI has become a serious research field. Bostrom's Superintelligence (2014) and a growing technical literature on alignment ask how to specify objectives that remain benign as capabilities grow. These questions reconnect with the philosophical problems Turing raised in 1950: what would count for a machine's "thinking," and what would that mean for the beings who built it? The history of computing has not closed the questions it opened; it has widened them.

Connections Master

The deepest link is to formal computation itself. Turing machines, computability, and complexity theory are the technical core of the present unit, and their full treatment lives in the computing domain (50.*). Likewise the foundations of mathematics (20.09.02) supply the backdrop — Hilbert's Entscheidungsproblem, Gödel's incompleteness theorems, and the Church-Turing thesis jointly define the limitative landscape within which Turing's 1936 result sits.

The transistor and integrated circuit rest on quantum mechanics and solid-state physics. Band theory, doping, and the physics of semiconductors are the substance of the quantum revolution (33.05.) and of chemistry (14.). The industrial consequences — Shockley Semiconductor, Fairchild, Intel, and the rise of Silicon Valley — connect to the longer history of industrial transformation (32.18.*) and the emergence of entirely new economic sectors.

Wartime funding shaped the technology at every stage. Colossus and ENIAC grew out of WWII computing needs, ARPANET out of Cold War concerns about survivable command-and-control, and much of the early semiconductor industry out of military procurement. This places the unit within the history of big science and big-government research funding developed in 33.08., and within the wider history of twentieth-century conflict (32.22.).

The biological applications of computing reach back to molecular biology (33.06., 17.). The Human Genome Project was as much a computational as a biochemical achievement, and contemporary protein-structure prediction is a direct descendant of both. Computational neuroscience and the modelling of consciousness connect forward to neuroscience (29.02.) and the philosophy of mind (20.06.), including the long debate over whether minds could be machines.

The societal consequences intersect with several further chapters. Surveillance capitalism, platform power, algorithmic bias, and the digital divide connect to media literacy (36.) and to sociology (30.02.03, 30.07.03). Cryptography depends on number theory (37.) and underwrites the privacy questions raised here and in the AI-ethics material at 20.02.06. Gender history enters through the often-erased contributions of women in early computing (30.04.04) and through the life and persecution of Turing himself.

The successor unit (33.08.02) takes up "big science" and global collaboration — the regime in which computing, genomics, particle physics, and climate science are practised as large-scale collective enterprises that the digital revolution made possible.

Historical and philosophical context Master

The digital revolution is unusual in the history of science because its central artifact — the universal computer — was specified in theory before it existed in practice. Turing's 1936 paper answered a question in mathematical logic (Hilbert's Entscheidungsproblem) and, almost in passing, described a machine that twenty years of engineering would make real. The history of computing therefore reads in reverse: the boundary of the possible was drawn before the first machine was built, and the engineers who followed were, in effect, catching up to a limit that had already been proven.

The Second World War was the decisive accelerator. Bletchley Park's codebreaking effort (the Bombe, Colossus) and the American ballistics programme (ENIAC) drew together mathematicians, engineers, and women "computors" in institutions freed, by military urgency, from peacetime constraints on cost and risk. The Cold War sustained the investment: ARPA funded time-sharing, computer graphics, and the network that became the internet; the Pentagon bought the first integrated circuits; the nuclear programme drove simulation at Los Alamos. As in 33.05.*, the wartime state was the patient, deep-pocketed investor that private markets could not have been.

Turing's personal history haunts the field. Prosecuted for homosexuality in 1952, subjected to chemical castration, dead by cyanide in 1954 at forty-one, he was given a royal pardon only in 2013. The mismatch between what the British state owed him and what it did to him is one of the clearest cases in the history of science of social prejudice destroying a life that had shortened a war. It also connects the unit to the history of gender and sexuality (30.04.04) and to the longer reckoning over who gets credit for scientific work — a reckoning that has restored the ENIAC programmers, the NASA "human computers," and the women of Bletchley Park to a narrative that long omitted them.

Philosophically, the digital revolution reopened the oldest question about mind and mechanism. Turing's 1950 paper "Computing Machinery and Intelligence" proposed the imitation game as a practical substitute for "can machines think?", and every major response — Searle's Chinese Room, the Lucas-Penrose argument from Gödel, functionalist theories of mind, integrated information theory — has been a reading of the consequences of universal computation for the nature of thought. The same question now returns with greater urgency through large language models, whose fluent output can be hard to distinguish from understanding and whose tendency to "hallucinate" is itself philosophically diagnostic.

The political economy of computing has shifted dramatically across the period. The early internet was built by a culture of open protocols, RFCs, and "rough consensus and running code," largely within universities and defence labs. The commercialisation of the 1990s, the platform consolidation of the 2000s, and the surveillance-advertising business model of the 2010s concentrated extraordinary power in a handful of firms. Whether the original architecture's openness was betrayed, merely outgrown, or is being reclaimed through open-source models and protocol-based alternatives is one of the live political questions the unit leaves open.

Two larger philosophical strands run through the whole. The first is the status of information itself: Shannon's 1948 theory made information a measurable quantity, and subsequent work (Wheeler's "it from bit," black-hole information debates, the simulation hypothesis) has pressed the question of whether information is more fundamental than matter. The second is the double character of computation as both liberating and constraining: the universal machine enlarged the possible, while undecidability proved permanent limits no engineering can lift. The history of computing is the history of a technology that, for the first time, came accompanied by a theorem about what it could never do.

Bibliography Master

Turing, A. M. "On Computable Numbers, with an Application to the Entscheidungsproblem." Proceedings of the London Mathematical Society 42 (1936): 230-265.
Turing, A. M. "Computing Machinery and Intelligence." Mind 59 (1950): 433-460.
Shannon, C. E. "A Mathematical Theory of Communication." Bell System Technical Journal 27 (1948): 379-423, 623-656.
von Neumann, J. "First Draft of a Report on the EDVAC." Moore School of Electrical Engineering, University of Pennsylvania, 1945. Reprinted in IEEE Annals of the History of Computing 15(4), 1993.
Bardeen, J. and Brattain, W. H. "The Transistor, a Semi-Conductor Triode." Physical Review 74 (1948): 230-231.
Shockley, W. "The Theory of p-n Junctions in Semiconductors and p-n Junction Transistors." Bell System Technical Journal 28 (1949): 435-489.
Kilby, J. S. "Invention of the Integrated Circuit." IEEE Transactions on Electron Devices 23 (1976): 648-654.
Moore, G. E. "Cramming More Components onto Integrated Circuits." Electronics 38(8), 1965.
Cerf, V. and Kahn, R. "A Protocol for Packet Network Intercommunication." IEEE Transactions on Communications 22(5), 1974: 627-641.
Berners-Lee, T. "Information Management: A Proposal." CERN, March 1989.
Cook, S. A. "The Complexity of Theorem-Proving Procedures." Proceedings of the 3rd Annual ACM Symposium on Theory of Computing, 1971: 151-158.
Karp, R. M. "Reducibility among Combinatorial Problems." In Complexity of Computer Computations (Miller and Thatcher, eds.), Plenum, 1972: 85-103.
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. "Learning Representations by Back-Propagating Errors." Nature 323 (1986): 533-536.
McCarthy, J., Minsky, M., Rochester, N., and Shannon, C. E. "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence." 1955.
Bowler, P. J. and Morus, I. R. Making Modern Science: A Historical Survey. 2nd ed. Chicago: University of Chicago Press, 2005. Ch. 11.
Ceruzzi, P. E. Computing: A Concise History. Cambridge, MA: MIT Press, 2012.
Dyson, G. Turing's Cathedral: The Origins of the Digital Universe. New York: Pantheon, 2012.
Isaacson, W. The Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution. New York: Simon and Schuster, 2014.
Abbate, J. Inventing the Internet. Cambridge, MA: MIT Press, 1999.
Russell, S. and Norvig, P. Artificial Intelligence: A Modern Approach. 4th ed. Hoboken: Pearson, 2021.
Zuboff, S. The Age of Surveillance Capitalism. New York: PublicAffairs, 2019.
Bostrom, N. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press, 2014.