The Birth and History of Artificial Intelligence

Artificial intelligence did not begin with one invention.

It emerged from several older questions:

Can reasoning be formalized?

Can machines compute?

Can nervous systems be modeled?

Can intelligent behavior be engineered?

By the middle of the twentieth century, these questions converged.

Before AI Had a Name

Long before electronic computers, thinkers imagined artificial reasoning.

Leibniz dreamed of a formal calculus in which disputes might be settled by calculation.

Logic slowly became mechanized.

Formal Logic

The nineteenth and early twentieth centuries transformed logic into symbolic systems.

Boole, Frege, Russell, and others helped make reasoning mathematically explicit.

This created conceptual machinery for automated inference.

Computability

In the 1930s, Alan Turing, Alonzo Church, and others formalized computation.

The Turing machine showed that mechanical symbol manipulation could be studied mathematically.

This gave AI a theory of what machines could compute.

Electronic Computers

During and after World War II, electronic computers made large-scale automatic computation practical.

Machines were no longer only mathematical abstractions.

They became programmable artifacts.

McCulloch and Pitts

In 1943, Warren McCulloch and Walter Pitts published a mathematical model of artificial neurons.

Their simplified neurons could implement logical functions.

This linked:

• brain,
• logic,
• computation.

Hebb

In 1949, Donald Hebb proposed a learning principle often summarized as:

cells that fire together strengthen their connection.

Hebbian learning influenced later neural-network ideas.

Turing’s Question

In 1950, Alan Turing published “Computing Machinery and Intelligence.”

Instead of trying to define thinking directly, he proposed an imitation game.

The question became behavioral.

Dartmouth 1956

The term artificial intelligence is usually associated with the Dartmouth Summer Research Project on Artificial Intelligence in 1956.

The proposal was organized by figures including:

• John McCarthy,
• Marvin Minsky,
• Claude Shannon,
• Nathaniel Rochester.

The field gained a name.

Early Optimism

Early researchers believed many aspects of intelligence might yield quickly to symbolic computation.

Computers solved:

• algebraic problems,
• logic puzzles,
• games.

Success encouraged ambitious predictions.

Logic Theorist

Allen Newell, Herbert Simon, and Cliff Shaw developed the Logic Theorist in the 1950s.

It proved mathematical theorems.

This was one of the earliest landmark AI programs.

General Problem Solver

Newell and Simon later developed the General Problem Solver.

It attempted to solve problems through general search strategies.

The dream was domain-independent intelligence.

Search as Intelligence

Early AI often framed intelligence as:

search through a space of possibilities.

The challenge was controlling combinatorial explosion.

Heuristics became essential.

LISP

John McCarthy developed LISP in the late 1950s.

Its symbolic data structures and recursive style made it highly influential in AI research.

For decades, LISP became strongly associated with AI.

Perceptron

Frank Rosenblatt developed the perceptron in the late 1950s.

It learned linear decision boundaries from data.

This was an important early neural-learning system.

Symbolic AI

By the 1960s, much AI focused on explicit:

• symbols,
• rules,
• search.

This became known as symbolic AI.

Knowledge was represented in human-readable forms.

SHRDLU

Terry Winograd’s SHRDLU, developed around 1968–1970, manipulated a simulated blocks world through natural-language commands.

It looked remarkably intelligent within its narrow domain.

Its success also revealed the importance of restricted environments.

ELIZA

Joseph Weizenbaum’s ELIZA simulated conversation through pattern matching.

Its apparent social fluency exceeded the depth of its underlying model.

The ELIZA effect became a lasting lesson.

Robotics

Early robotics projects attempted to integrate:

• perception,
• planning,
• action.

Shakey the Robot became a landmark system combining these capabilities.

The Frame Problem

As AI moved into richer environments, commonsense reasoning became difficult.

How should a system represent:

• what changes,
• what stays the same?

The frame problem exposed limits of naive symbolic approaches.

Combinatorial Explosion

Many early methods worked on small examples.

Scale them up and the number of possibilities exploded.

Search could become computationally impossible.

The First AI Winter

Expectations had grown faster than practical capability.

Funding and enthusiasm declined in parts of the field during the 1970s.

This period became known as an AI winter.

The phrase describes a broader pattern rather than one exact global date.

Expert Systems

AI revived strongly through expert systems.

These programs encoded specialist knowledge as rules.

They performed well in narrow domains.

DENDRAL

DENDRAL helped infer molecular structures from chemical data.

It demonstrated that domain knowledge could outperform generic problem-solving methods.

Expertise mattered.

MYCIN

MYCIN used rules to recommend treatments for certain bacterial infections.

It became a classic example of expert-system reasoning.

It was primarily a research system, not routine autonomous clinical practice.

Knowledge Engineering

Expert systems shifted attention from:

general reasoning

to:

capturing domain knowledge.

The bottleneck became:

How do we encode expertise?

Knowledge Acquisition Bottleneck

Experts often know more than they can state explicitly.

Turning tacit expertise into thousands of formal rules was difficult.

Systems were also brittle outside their encoded domains.

Commercial Boom

During the 1980s, companies invested heavily in expert systems.

Specialized AI tools became commercially attractive.

Optimism returned.

Second AI Winter

Maintenance costs, brittleness, unmet expectations, and changing hardware economics contributed to another major decline in enthusiasm in the late 1980s and early 1990s.

Again, the decline was uneven across research communities.

Probabilistic AI

Meanwhile, probabilistic approaches grew.

Bayesian networks provided tools for reasoning under uncertainty.

AI became less dependent on purely logical rules.

Machine Learning

A major shift gradually occurred:

Instead of hand-coding every rule,

let systems learn patterns from data.

Machine learning moved toward the center of AI.

Statistical Learning

Methods such as:

• decision trees,
• support vector machines,
• probabilistic models

became important.

AI increasingly became empirical.

Reinforcement Learning

Reinforcement learning developed formal methods for agents learning through reward.

Ideas from:

• psychology,
• control theory,
• dynamic programming

converged.

Neural Networks Return

Neural networks had never disappeared completely.

But improved algorithms, larger datasets, and more computing power made them increasingly competitive.

Backpropagation

Backpropagation became widely influential in the 1980s as a practical method for training multilayer neural networks.

It allowed hidden layers to learn useful representations.

Deep Learning

In the 2000s and especially the 2010s, deeper neural networks began achieving major successes.

Three ingredients mattered greatly:

• data,
• computation,
• algorithms.

ImageNet Moment

In 2012, AlexNet dramatically improved performance in the ImageNet image-classification competition.

This became a symbolic turning point for deep learning.

Computer vision changed rapidly afterward.

GPUs

Graphics processing units were well suited to the matrix operations used in neural networks.

Hardware shaped algorithmic possibility.

AI progress became closely tied to computational infrastructure.

Speech Recognition

Deep learning also transformed:

• speech recognition,
• machine translation,
• vision.

Tasks once dominated by hand-engineered pipelines shifted toward learned representations.

AlphaGo

In 2016, AlphaGo defeated Lee Sedol in Go.

The system combined:

• neural networks,
• search,
• reinforcement learning.

Go had long been viewed as resistant to brute-force search.

The victory became a public landmark.

Representation Learning

Deep learning reduced the need to hand-design features.

Systems learned useful internal representations directly from data.

This shifted the role of the engineer.

Transformers

In 2017, the transformer architecture introduced attention-based sequence processing without recurrence as its central mechanism.

Transformers later became foundational for large language models.

Pretraining

A powerful strategy emerged:

train large models on broad datasets,

then adapt them to many downstream tasks.

This improved transfer dramatically.

Foundation Models

Large pretrained models began functioning as general-purpose bases for many applications.

Instead of training one model per task, one model could support many tasks.

This changed the economics and architecture of AI development.

Large Language Models

Large language models learned statistical structure from massive text corpora.

Scaling produced surprising abilities in:

• language generation,
• question answering,
• code.

The relationship between scale and capability became a major research topic.

Generative AI

AI systems increasingly produced:

• text,
• images,
• audio,
• code.

Generative AI shifted public attention from classification to creation.

The interface to AI became conversational and creative.

Multimodal Systems

Models began integrating:

• text,
• images,
• audio,
• video.

This moved AI closer to systems operating across multiple representational domains.

Retrieval and Tools

Modern AI systems increasingly combine learned models with:

• search,
• databases,
• external tools,
• code execution.

Intelligence becomes architectural rather than located in one model alone.

Agents Return

The old idea of an intelligent agent has reappeared in new engineering forms.

Systems can:

• plan,
• call tools,
• maintain state,
• execute multi-step tasks.

Old concepts return under new technology.

AI History Is Not Linear

The story is not:

symbolic AI failed, neural AI won.

Many ideas coexist.

Modern systems use combinations of:

• search,
• optimization,
• symbolic tools,
• statistical learning.

Progress is hybrid.

Hardware Matters

AI history cannot be separated from hardware.

Changes in:

• memory,
• processors,
• GPUs,
• networking

made previously impractical methods feasible.

Algorithms live inside infrastructure.

Data Matters

Modern learning systems depend heavily on data.

The internet created massive collections of:

• language,
• images,
• code.

AI capability partly reflects accumulated human culture.

Benchmarks Matter

Research follows what can be measured.

Chess, ImageNet, language benchmarks, and coding tests shaped priorities.

Benchmarks accelerate progress.

They can also narrow it.

Public Expectations

AI has repeatedly oscillated between:

• hype,
• disappointment.

This reflects both genuine technical progress and exaggerated forecasts.

Historical memory is useful.

Intelligence Keeps Moving

When machines master a task, people sometimes stop calling the task intelligence.

Chess once seemed a deep test.

Now it is treated as computation.

The cultural definition of intelligence shifts.

The AI Effect

This tendency is sometimes called the AI effect:

once AI solves something, it no longer seems like AI.

The boundary moves with technological familiarity.

Ethics Enters the Center

As systems affect real institutions, questions increasingly concern:

• bias,
• privacy,
• accountability,
• labor,
• safety.

AI is no longer only a laboratory problem.

It is a social technology.

Alignment

More capable systems raise a deeper issue:

How do we ensure that optimized behavior matches human intentions and values?

This is the alignment problem.

It will return later in Part XV.

History as Warning

AI history teaches several recurring lessons:

• narrow success can be mistaken for general intelligence,
• scaling can change what is possible,
• brittle systems can look impressive,
• infrastructure matters,
• expectations can outrun evidence.

History as Continuity

Modern AI still carries old questions:

Can symbols represent meaning?

Can learning replace hand-coded knowledge?

Can intelligence emerge from simple processes?

Can machines understand?

The technology changes.

The philosophical questions persist.

The Philosophical Lesson

Artificial intelligence grew from a convergence of:

• logic,
• computation,
• neuroscience,
• statistics,
• engineering.

Its history is not a straight march toward intelligence.

It is a sequence of changing representations of what intelligence itself might be.

The Next Question

One idea has followed AI from near its beginning:

Perhaps we should not ask whether a machine truly thinks.

Perhaps we should ask whether its behavior is indistinguishable from ours.

That proposal belongs to:

The Turing Test.