The AI Journey
FreeMalta / A history of changing questionsFile 001 · 06 Oct 2026
The AI Journey

It began
with a question.

Seventy-six years later,
we are asking how much we should let it do.

Enter the journey
1950202676 years. 21 moments.
One question that kept changing.
Not a race to the smartest machine.A journey into the decisions we give away.

Read it in order.
Or enter where the question changed.

A curated history, not an inventory of every invention. Dates mark publications, releases or public events—not a claim that each idea appeared from nowhere.

All 21 moments / choose a date
  1. 1950A question enters the room.
  2. 1956The ambition gets a name.
  3. 1958Do not write every rule. Let it learn.
  4. 1965Intelligence narrows its brief.
  5. 1966It sounds like someone is listening.
  6. 1973The promises meet the bill.
  7. 1980AI gets a job description.
  8. 1986The error travels backwards.
  9. 1997The champion loses. The world watches.
  10. 2006Depth becomes worth trying again.
  11. 2009Someone has to label the world.
  12. 2012The machine learns what to look for.
  13. 2016A move nobody expected.
  14. 2017Pay attention. In parallel.
  15. 2020Describe the task. Show a few examples.
  16. 2022The prompt becomes a picture.
  17. 2022The research preview walks into everyday life.
  18. 2023There is no longer one front door.
  19. 2024More work before the answer.
  20. 2025The answer reaches for the controls.
  21. 2026Capability is not permission.
Chapter 01 / 1950–1966

Can a machine do what a mind does?

A name, a learning machine, an expert and a conversation. The ambitions are large. The systems are small.

02 / 2119566 years later

John McCarthy, Marvin Minsky, Nathaniel Rochester & Claude Shannon / Dartmouth

The ambition gets a name.

A proposal for a summer research project gathers several approaches under the name artificial intelligence. Its authors suggest that aspects of learning and intelligence might be described precisely enough for machines to simulate them. Dartmouth gives a research community a shared banner; it does not produce general intelligence in one summer.

A collection of ideas becomes a named research agenda.

The recordThe Dartmouth proposal, 1955
03 / 2119582 years later

Frank Rosenblatt / the perceptron

Do not write every rule. Let it learn.

Rosenblatt’s perceptron explores how a machine can learn to distinguish patterns by changing connection weights. It offers an alternative to spelling out every decision in advance. The publicity outruns the modest system, but the central idea survives: useful behaviour can emerge from training, not only from hand-written instructions.

Learning becomes an engineering route, not just an aspiration.

The recordCornell’s account of Rosenblatt’s work
04 / 2119657 years later

Stanford / DENDRAL

Intelligence narrows its brief.

A Stanford team including Edward Feigenbaum, Joshua Lederberg and Carl Djerassi develops DENDRAL to help infer molecular structures. Instead of trying to reproduce the whole human mind, it encodes specialist knowledge for a bounded scientific problem. This is an important change of ambition: a machine does not need to know everything to be useful somewhere.

Domain expertise offers a practical path to useful AI.

The recordComputer History Museum / DENDRAL
Chapter 02 / 1973–1986

What survives contact with reality?

Funding, specialist knowledge and a way to learn from error. Progress does not move in a straight line.

06 / 2119737 years later

James Lighthill / a contested review

The promises meet the bill.

Lighthill’s review for Britain’s Science Research Council questions whether AI’s laboratory successes can scale to harder real-world problems. Researchers dispute its judgement; John McCarthy publishes a pointed response. It becomes a marker of the first funding winter, not a verdict that all AI research everywhere has stopped.

Demonstrations must answer to scale, evidence and funding.

AI winters were uneven periods of reduced confidence and support. Their boundaries depend on the institution, country and approach being discussed.

The recordUCL archive / April 1973 symposium McCarthy’s response to Lighthill
07 / 2119807 years later

John McDermott / R1, later XCON / Digital Equipment Corporation

AI gets a job description.

R1 uses expert rules to configure computer systems for Digital Equipment Corporation. The attraction is concrete: specialist decisions become a repeatable business process. Commercial expert systems also reveal a burden—someone must acquire, update and maintain their knowledge. The later retreat in confidence does not erase the lesson that deployment is more than a clever demonstration.

The economics of a working system enter the frame.

The recordMcDermott / R1, AAAI 1980 Expert-system pioneers / CHM oral history
Chapter 03 / 1997–2016

What can it do better than us?

Public contests meet less visible foundations: training methods, shared datasets and the people who build them.

10 / 2120069 years later

Geoffrey Hinton, Simon Osindero & Yee-Whye Teh / deep belief networks

Depth becomes worth trying again.

A paper presents a way to train deep belief networks layer by layer, with further fine-tuning. It helps renew interest in learning through multiple layers when training such networks is difficult. This is one route in a broader revival, not the invention of deep learning or the final architecture that later dominates it.

Training deeper models becomes a more credible research programme.

The recordA fast learning algorithm for deep belief nets
11 / 2120093 years later

Jia Deng and collaborators / ImageNet

Someone has to label the world.

ImageNet introduces a large, organised image database for visual recognition research. Shared data makes systems easier to train and compare; later competitions give the field a common test. Behind the model is another kind of work: collecting, organising and labelling the examples from which it learns.

Data infrastructure becomes a driver of progress.

The recordImageNet / the 2009 research paper
13 / 2120164 years later

Google DeepMind / AlphaGo / Lee Sedol

A move nobody expected.

AlphaGo defeats Lee Sedol four games to one in Seoul. Its combination of neural networks, learning and search tackles a game long considered a formidable challenge for computers. Go becomes another reminder that expertise can be learned and combined with computation in ways that do not simply imitate a person’s familiar choices.

Learning and search show their power together.

The recordDeepMind / AlphaGo
Chapter 04 / 2017–2022

What happens when everyone can ask?

An architecture becomes a platform. A prompt becomes an interface. The laboratory door opens.

15 / 2120203 years later

Tom Brown and collaborators / GPT-3

Describe the task. Show a few examples.

GPT-3 research tests a large language model on tasks presented through instructions and a few examples, without task-specific weight updates for each test. The prompt starts to look like a flexible interface rather than just an input sentence. Useful results coexist with limitations: fluent continuation is not the same as dependable truth or judgement.

One model can be steered towards many tasks through context.

The recordLanguage Models are Few-Shot Learners
16 / 2120222 years later

Stable Diffusion / Stability AI and research collaborators

The prompt becomes a picture.

Stable Diffusion’s public release makes model weights and code available for text-to-image generation under a use-restricted licence. It is not the first image generator, but the ability to run and adapt released models changes access for developers and creators. Questions of provenance, permission and misuse become inseparable from the creative possibilities.

Generative imagery becomes something people can build with.

The recordStable Diffusion / public release announcement
17 / 212022Later that year

OpenAI / ChatGPT

The research preview walks into everyday life.

OpenAI releases ChatGPT as a free research preview with a conversational interface. Follow-up questions and ordinary language make a language model approachable without requiring people to design a technical integration. The interface changes the audience—and puts plausible but incorrect answers in front of it, a limitation the launch itself acknowledges.

The conversation becomes a mass-facing way to use AI.

The recordOpenAI / Introducing ChatGPT
Chapter 05 / 2023–2026

What should we let it do?

Models see more, spend more effort and use tools. Capability expands. Responsibility does not disappear.

18 / 2120231 year later

Claude, GPT-4 & Gemini / competing model families

There is no longer one front door.

Anthropic introduces Claude in March; GPT-4’s technical report describes image-and-text input; Google announces Gemini in December with a multimodal design. Different model families compete for the same work while expanding beyond text alone. An announcement is not identical to universal access: features, modalities and rollout schedules differ.

Choosing and evaluating a model becomes part of the work.

The recordAnthropic / Introducing Claude, 14 March GPT-4 technical report, March 2023 Google / Gemini, 6 December
19 / 2120241 year later

OpenAI / o1-preview

More work before the answer.

OpenAI introduces o1-preview, trained to spend more computation working through problems before responding. This makes the inference stage—the work performed when a model is used—a more visible lever for performance on difficult tasks. Calling it reasoning describes an approach and evaluated behaviour, not proof of human-like thought or consciousness.

The cost and time of producing an answer become adjustable levers.

The recordOpenAI / introducing o1-preview
The story does not end here

The question changed.
The responsibility stayed.

1950 / Turing’s opening question

“Can machines think?”

2026 / Our editorial question

What should we
let them do?

There was never one purpose behind AI. Researchers pursued formal reasoning, learning, language and useful work along different paths. Understanding that history makes it harder to mistake a convincing demo for a settled future.

Understand the presentModels. Agents. Your next move.Put it to workBuild a system, not a promise.
Return to the first question ↑