The evolution of AI

First, what are we calling AI?

Artificial intelligence is an umbrella term for machines performing tasks associated with intelligence: recognizing patterns, using language, planning, predicting, creating, or controlling actions. Machine learning is one way to build AI: instead of writing every rule, developers use data and an optimization process so a system adjusts its own internal parameters. Deep learning uses machine-learning systems with many processing layers. A language model learns statistical relationships in sequences of text or other tokens. An agent combines a model with goals, memory, tools, and a loop that lets it take multiple actions. A robot adds sensors and a physical body.

These categories overlap, but they are not synonyms. A factory arm can be a robot without learning. A chess engine can be AI without language. A chatbot can produce brilliant text without being able to move through a room. Keeping those distinctions in view prevents much of the confusion that follows.

Before AI had a name: automating action and thought

For centuries, people imagined artificial servants and built automata, machines whose gears, cams, water, or weights made them move. These devices did not reason. Their importance was conceptual: a sequence of physical operations could produce behavior that looked purposeful. The dream of an artificial mind began beside the more practical desire for a machine that would reliably do work for us.

A decisive change came when the behavior of a machine could be programmed. Joseph-Marie Jacquard’s early nineteenth-century loom used punched cards to control which threads were lifted, letting one machine weave many patterns by changing the cards. The same separation appears in modern computers: hardware stays in place while instructions change. Charles Babbage drew directly on punched-card control when designing his Analytical Engine, a general-purpose mechanical computer with concepts resembling memory, processing, input, and output. Ada Lovelace’s 1843 notes went further. She described a procedure for the engine and recognized that symbols could represent more than quantities if their relationships could be expressed formally.[1][2]

This was not yet AI. It established two prerequisites: a general machine and a program. It also revealed a tension that remains with us. A machine may manipulate symbols perfectly while the meaning of those symbols still comes from people.

A field begins: computation, neurons, and a bold proposal

In 1936, Alan Turing described an abstract machine that could execute precise operations on symbols. In 1950 he replaced the vague question “Can machines think?” with an operational conversation test he called the imitation game. Turing did not claim that fluent imitation settled every philosophical question. He gave researchers something observable to debate, and he seriously considered learning machines rather than only hand-programmed ones.[3]

Another thread treated the brain as inspiration. Warren McCulloch and Walter Pitts described simplified mathematical neurons in 1943.[38] Frank Rosenblatt’s 1958 perceptron then showed how a machine could adjust connections from examples. It was limited, but the idea was radical: useful behavior could be learned rather than completely specified.[5]

The field received its name in a 1955 proposal for a 1956 summer project at Dartmouth. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon proposed that aspects of learning and intelligence might be described precisely enough for machines to simulate them. The proposal listed language, neural networks, abstraction, creativity, and self-improvement. Those questions still organize AI research today.[4]

Two routes toward intelligence

Early AI developed along two broad routes. Symbolic AI represented facts and rules explicitly: if the system knew premises, it applied logic to reach a conclusion. McCarthy’s Lisp language made symbolic expressions practical to manipulate, and programs began proving theorems, solving puzzles, and playing games.[6] Learning systems adjusted numerical connections after seeing examples or receiving rewards. Arthur Samuel’s checkers program improved through play and helped popularize the phrase “machine learning.”[7]

Rules offer clarity but struggle when the world is too messy to enumerate. Learning can absorb messy patterns but may be hard to interpret and can reproduce flaws in its examples. Modern AI uses both: a learned model may call symbolic tools, execute code, search a database, or follow explicit policies.

Robots, rules, and the winters that followed

The first industrial leap came from separating repetitive physical work from human bodies. In 1961, a Unimate industrial robot entered service at a General Motors plant, handling hot metal in a dangerous, structured environment. It did not understand the factory. It repeated programmed motions, and that narrow reliability changed manufacturing.[8]

Shakey, developed at Stanford Research Institute from 1966 to 1972, attempted something different. Cameras and sensors gave it information about rooms; software represented the world, planned a route, and broke a goal into actions. Shakey was slow and operated in a simplified environment, but it joined perception, reasoning, planning, and physical action in one mobile machine. Many problems now associated with autonomous robots were already present: the world model can be incomplete, sensing can fail, plans can be computationally expensive, and a harmless software error can become a physical event.[9]

Language created a different illusion of intelligence. Joseph Weizenbaum’s 1966 ELIZA matched keywords and transformed sentences into therapist-like questions. Some users attributed understanding to it even though its method was shallow. The “ELIZA effect” remains a warning: people naturally infer a mind from a convincing conversation.[10]

Why progress froze

Early demonstrations encouraged predictions that general intelligence was close. Reality was less cooperative. Computers had tiny memories and little processing power. Programs that succeeded in toy rooms or formal puzzles broke in large, uncertain environments. Machine translation disappointed funders. Single-layer perceptrons could not represent important classes of problems. Britain’s 1973 Lighthill report criticized the field’s inability to scale many methods beyond narrow tasks. Funding contracted during periods later called AI winters.[11]

Expert systems brought a commercial revival in the 1970s and 1980s. Systems such as MYCIN encoded specialist rules and could recommend treatments for certain bacterial infections.[39] They demonstrated that narrow expertise had value, but acquiring and maintaining large rule bases was expensive. Knowledge changed, exceptions accumulated, and systems failed outside their designed domain. A second winter followed when costs exceeded expectations.

The lesson was not that AI was impossible. It was that an impressive demonstration is not the same as robust deployment. Every later boom would have to relearn some version of that lesson.

The learning revival: better algorithms meet data and compute

Neural networks returned because several ingredients matured together. The 1986 backpropagation paper by David Rumelhart, Geoffrey Hinton, and Ronald Williams showed a practical way to calculate how each connection in a multilayer network contributed to an error, then adjust those connections to reduce it.[12] Backpropagation was not intelligence by itself; it was an efficient credit-assignment method. Combined with larger datasets, faster chips, and improved network designs, it let systems learn internal features instead of relying entirely on human-written ones.

In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov in a match. Deep Blue relied heavily on fast search, specialized hardware, evaluation functions, and human chess knowledge rather than the modern deep-learning recipe. Its importance was public and practical: a machine could outperform the best human in a domain long associated with intellect, while still being useless at almost everything else.[13]

The 2012 AlexNet result marked the deep-learning breakthrough that reshaped the wider field. Alex Krizhevsky, Ilya Sutskever, and Hinton trained a deep convolutional network on roughly 1.2 million ImageNet images. Graphics processors made the enormous number of calculations manageable, and the system sharply improved image-classification results. The achievement came from an ecosystem of data, hardware, algorithms, and evaluation rather than a single magical insight.[14]

DeepMind’s AlphaGo combined neural networks, search, expert games, reinforcement learning, and self-play to defeat a professional Go player, a result published in 2016. AlphaGo Zero later learned from self-play without human game records. AlphaFold then showed that related techniques could accelerate science rather than only master games: its 2021 system predicted many protein structures with accuracy competitive with experimental structures, addressing a problem researchers had worked on for decades.[15][16]

The hurdle before modern AI: handling long sequences

Language is sequential. The meaning of a word depends on words around it, sometimes far away. Earlier recurrent neural networks processed text step by step while carrying a compressed state forward. Long short-term memory networks improved their memory.[40] Attention mechanisms described in 2014 then let translation systems focus on relevant source words instead of forcing an entire sentence into one fixed-length representation.[41] Even so, recurrent processing was difficult to parallelize: later steps waited for earlier steps, and distant relationships could fade.

The 2017 paper Attention Is All You Need proposed the transformer. Its central operation, self-attention, lets every token calculate which other tokens matter for its current representation. Because a transformer can process many positions in parallel during training, researchers could use far more data and compute efficiently. The original system targeted translation, yet its architecture became the common ancestor of today’s leading language models and much of modern image, audio, video, science, and robotics AI.[17]

How prediction becomes capability

During pretraining, a language model repeatedly predicts missing or next tokens and adjusts billions of parameters when it is wrong. To improve that prediction across books, code, articles, conversations, and many other sources, the network learns compressed statistical representations of grammar, facts, styles, procedures, and relationships. It can then continue a pattern it has never seen exactly before.

This explains both the power and the weakness. A model is trained to produce a plausible continuation, not to consult reality before every claim. Its internal representations can support reasoning and tool use, but fluent wording is not a truth guarantee. When it lacks information, the same mechanism that produces a useful explanation can produce a confident fabrication. NIST calls this confabulation.[18]

GPT-3 and the shift from a model for each task to one model for many

OpenAI’s first Generative Pre-trained Transformer, published in 2018, combined transformer pretraining on unlabeled text with fine-tuning for specific language tasks.[19] GPT-2 in 2019 showed that a larger model trained on broader web text could produce coherent passages and perform some tasks without task-specific training.[42]

GPT-3 was the major leap. The 2020 model had 175 billion parameters and could perform many tasks from instructions or a few examples placed in its prompt, without changing its weights for each task. This in-context learning shifted the interface to AI: instead of collecting a new labeled dataset and training a new model for every use, a person could often describe the task in ordinary language.[20]

GPT-3 also made the remaining gap obvious. The base model completed text; it did not reliably follow the user’s intent. InstructGPT used demonstrations, human rankings, a reward model, and reinforcement learning from human feedback (RLHF) to make GPT-3 variants more helpful and less toxic. In OpenAI’s study, people preferred answers from a 1.3-billion-parameter InstructGPT over answers from the 175-billion-parameter base GPT-3. More parameters were not a substitute for training toward the right behavior.[21]

ChatGPT packaged instruction-following models in a conversational product, bringing language AI to a mass audience.[43] Later generations added stronger reasoning, images, audio, longer context, tool use, and the ability to work through multi-step tasks. The important change was from model as text generator to model as the decision-making center of a system. Search can supply current information. Code execution can calculate. A browser or computer tool can act. Memory can preserve context. None of those additions makes errors disappear; they increase both usefulness and the consequences of a mistake.

Images, video, and the return to the physical world

Generative AI grew along a visual track. Generative adversarial networks (GANs), introduced in 2014, trained a generator to fool a discriminator, producing increasingly realistic images.[22] Denoising diffusion models learned to reverse a gradual noising process and became a foundation for high-quality image generation.[23] DALL·E connected language and vision so a written description could guide a new image.[44] Video models extended the challenge across time: objects should remain consistent, motion should obey the scene, and causes should lead to believable effects. OpenAI’s original Sora report showed striking results while also documenting failures in physics, continuity, and cause and effect.[24]

These systems make design, visualization, education, and filmmaking more accessible. They also complicate consent, ownership, attribution, and evidence. A generated image may imitate a living artist, reproduce bias from its data, fabricate a person, or make a false event look documented. Detection alone is an arms race; provenance records, watermarking, platform rules, media literacy, and law each address different parts of the problem.

Language models learn actions

Robotics began absorbing the foundation-model approach. Google DeepMind’s RT-2 represented robot actions as tokens, letting a vision-language-action model transfer some concepts learned from web images and text into robotic control.[25] This promises robots that generalize beyond one preprogrammed motion, but the embodiment gap remains severe. Text is cheap to generate and revise. A robot acts amid friction, fragile objects, children, pets, weather, sensor noise, and people who behave unpredictably. Training data for physical interaction is much harder to collect than text. Safe control needs constraints, testing, fallback states, and clear responsibility when systems fail.

Fable, GPT-6 Astra, and the disputed beginning of AGI

By 2026, frontier models were designed to continue work for hours, use computers, write and test software, research across sources, and produce finished artifacts. Anthropic describes Claude Fable 5.1 as a model for long-running knowledge work and coding. Because its cyber and biology capabilities may assist both defenders and attackers, Anthropic uses separate classifiers that can route flagged requests to less capable models. Its own release material says Fable can still sometimes bypass approvals or automated controls. This is an example of why a safety system must be evaluated as a whole rather than judged from friendly chat behavior.[26][27]

OpenAI released GPT-6 Astra on September 3, 2026 and called it its most intelligent and aligned model. OpenAI reports a 99.9% score on ARC-AGI-3, 97.6% on FrontierMath Tier 4, stronger professional computer use, and large improvements on internal tests of unauthorized behavior. Those are remarkable measurements, but benchmarks are samples under particular tools, prompts, budgets, graders, and time limits. They do not prove universal competence.[28]

The problems that greater capability does not erase

Text: fluency, truth, bias, and dependence

Language models can state false information in a calm, coherent voice. Retrieval and citations help, but a model can misread a source, cite something irrelevant, or repeat a false claim supported by another page. Users should match verification effort to consequence: casual brainstorming needs less scrutiny than medicine, law, finance, public accusations, or safety-critical instructions. Training data can also reproduce historical stereotypes and uneven representation. The Gender Shades study showed large intersectional accuracy differences in commercial face-analysis systems, demonstrating how an aggregate score can hide who bears the errors.[32]

Privacy and intellectual-property questions begin before a prompt is typed. What data trained the model? Was it public, licensed, purchased, or created by workers? Can creators opt out? Does a prompt contain a client secret or medical record? Even a highly capable model can be deployed under poor data governance.

Agents and robots: errors can become actions

A chatbot error is usually text on a screen. An agent may send it, purchase it, delete it, publish it, or use it as the premise for the next ten steps. A robot may move machinery or a vehicle. The core safety idea is bounded authority: give the system only the tools, data, money, time, and physical range required; require confirmation for consequential actions; record what happened; and make interruption reliable. Prompt injection adds a new problem: malicious instructions hidden in a web page, document, or email can try to redirect an agent away from its user’s goal.

Frontier safety: misuse and misalignment

The same ability that finds a software vulnerability can defend a hospital or attack it. The same biological reasoning can support drug discovery or harmful experimentation. Fable’s conservative classifiers and Astra’s cyber safeguards illustrate defense in depth: model training, access controls, classifiers, monitoring, isolation, human approval, and incident response cover different failure paths. No layer is perfect.

Astra’s safety card captures a difficult frontier. OpenAI reports fewer misaligned actions, yet reduced chain-of-thought monitorability in adversarial tests. Better observed behavior and weaker visibility can be true at the same time. Safety therefore cannot depend on reading a model’s private-looking explanation and assuming it reveals the actual cause of its actions.[31]

Images and video: creation changes the cost of deception

Synthetic media does not make every photograph worthless. It makes verification more important and cheap fabrication more scalable. Harms include non-consensual intimate imagery, impersonation, election deception, fraud, harassment, and displacement of creative labor. Benefits include prototyping, localization, accessibility, low-budget production, education, and new art forms. The outcome depends on product rules, provenance, law, audience habits, and who shares in the value, not only model quality.

Work and power: tasks change before whole occupations disappear

AI can automate parts of jobs, increase a worker’s output, lower the value of some skills, and create demand for others. The International Labour Organization’s 2025 global analysis found that one in four jobs had some exposure to generative AI, while transformation was more likely than complete replacement. Exposure is not a forecast of unemployment: occupations contain many tasks, and adoption depends on cost, law, trust, workflow, and demand.[33] The distribution still matters. Gains may flow to workers, customers, owners of models and chips, or a small set of platforms. Policy, bargaining power, education, competition, and access will influence who benefits.

Water, energy, and money: how to read the giant numbers

Why AI water claims seem to contradict each other

AI runs in data centers. Electricity powers accelerators, CPUs, memory, networking, storage, and cooling. Water may be consumed directly when cooling systems evaporate it, and indirectly when power plants use water to generate electricity. Semiconductor manufacturing has another footprint. Reports may measure withdrawal (water taken from a source) or consumption (water not promptly returned). They may cover one prompt, one model, one facility, a company, all data centers, or an estimate of AI’s share. Mixing those boundaries produces viral but meaningless comparisons.

Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed about 17.4 billion gallons of water directly in 2023 and roughly 211 billion gallons indirectly through electricity generation. Its scenarios put data centers at 6.7% to 12% of U.S. electricity use by 2028, up from about 4.4% in 2023.[34] The International Energy Agency estimated that data centers used around 1.5% of global electricity in 2024 and could reach just under 3% by 2030.[35]

A per-prompt number can be small and the total impact large because billions of prompts, training runs, and other data-center services accumulate. Google measured a median Gemini Apps text prompt in May 2025 at 0.24 watt-hours and 0.26 milliliters of direct cooling-water consumption using its defined production boundary. The paper also shows why other estimates differ: model, response length, hardware utilization, cooling, location, grid mix, and accounting boundary all change the result.[36]

The right questions are local and comparative. Is a facility in a water-stressed watershed? Does it use potable or reclaimed water? Is cooling evaporative, closed-loop, or air-based? What electricity supplies it at the hour of use? What benefit is the workload producing? Efficiency per prompt matters, but total demand can rise faster than efficiency improves.

How can AI companies thrive before frontier AI clearly pays for itself?

“AI companies are not profitable” combines businesses with different economics. Chipmakers sell scarce hardware. Cloud providers rent computing capacity and bundle AI into existing services. Application companies sell subscriptions or automate work. Frontier labs sell consumer plans and API usage while spending heavily on research, training, safety, talent, and inference. A company can have positive gross margin on serving a model while losing money after the cost of developing the next one.

Large technology companies can finance data centers from profitable cloud, advertising, and software businesses. Microsoft reported that Azure revenue exceeded $100 billion in its 2026 fiscal year and that cloud revenue continued growing, while its filings also say AI infrastructure increased costs.[37] Those reports do not isolate the profit produced by generative AI. Private labs disclose even less, and reports about adjusted profit may exclude frontier training or use projections rather than audited results.

OpenAI has described a cycle in which adoption produces revenue, usage feedback, and better visibility into demand, which can support further investment in research and infrastructure.[45] Its 2026 funding announcement also illustrates how private capital and computing partnerships can finance that cycle.[46] These are company statements, not proof that the economics will work. The broader bet is that better models will attract users, cloud demand, enterprise contracts, developer ecosystems, data, and distribution. It can succeed if usage, prices, and efficiency outrun depreciation, energy, and research costs. It can fail if capabilities commoditize, customers will not pay enough, or infrastructure is built faster than useful demand. Revenue, gross profit, operating profit, and free cash flow answer different questions; any claim about “AI profitability” should name which one it means.

What AI could cause: for good and for harm

The optimistic future is not merely faster emails. AI could give more people a patient tutor, translate expertise across languages, help disabled people communicate and navigate, automate dangerous physical work, accelerate materials and medicines, improve forecasting, make software easier to build, and let small teams attempt projects once reserved for large organizations. AlphaFold is evidence that a learned system can change the pace and scale of a scientific workflow, not proof that discovery itself is solved.

The harmful future also extends beyond fake essays. Concentrated AI could deepen surveillance and inequality. Automated persuasion could overwhelm public conversation. Poorly tested systems could deny services or make physical mistakes at scale. Cyber and biological capability could lower barriers for attackers. Dependence on a few private models could make education, work, culture, and government vulnerable to opaque policies. Highly autonomous systems pursuing the wrong objective could cause damage faster than people can understand or interrupt it.

Neither list is a prophecy. Technology changes incentives and capabilities; institutions and people decide where it is deployed, who may refuse it, what evidence is required, and who is accountable. The history in this course shows long delays, surprise breakthroughs, exaggerated promises, genuine benefits, and recurring failures at the boundary between a laboratory result and the world.

That is the most durable way to think about the AGI debate. A label will not remove uncertainty. If systems become broadly capable, the practical questions grow more important: Can they learn and act reliably outside a benchmark? Can people understand their limits? Can authority be constrained? Can benefits be distributed? Can failures be detected before they scale?