Skip to content
Guides

October 23, 2026

The Story of AI: From Turing's Question to Autonomous Agents

Probir Sarkar
14 minute read

Eighty years of AI in one brief guide: from Turing's imitation game to ChatGPT, agents, and an honest look at where AGI actually stands. Every figure comes from a primary source.

The Story of AI: From Turing's Question to Autonomous Agents — BrowserStay blog card

The question that started it

The mathematical blueprint came first. In 1943, Warren McCulloch and Walter Pitts published "A logical calculus of the ideas immanent in nervous activity," modeling neural events with propositional logic — the first description of what researchers would later call a neural network.

Timeline of AI milestones, 1950 to 2026
[object Object]

Seven years later, Alan Turing reframed the field's oldest question. His 1950 paper "Computing Machinery and Intelligence," published in the journal Mind, opened with "Can machines think?" — and then replaced that vague question with an operational one he called the "imitation game," the basis of the Turing test: can a machine's conversation be made indistinguishable from a human's? Turing also made a prediction: in about fifty years, computers with storage of about 10^9 units would play the imitation game so well that an average interrogator would have no more than a 70 per cent chance of correctly identifying the machine after five minutes of questioning.

The name came in 1955. On 31 August 1955, John McCarthy wrote a funding proposal to the Rockefeller Foundation, co-signed by Marvin Minsky, Nathaniel Rochester and Claude Shannon, proposing "a 2 month, 10 man study of artificial intelligence" for the summer of 1956 at Dartmouth College — the earliest documented appearance of the phrase. The Rockefeller Foundation backed it with $7,500, a little over half of the $13,500 requested. The workshop ran that summer, from roughly 18 June to 17 August, and is widely considered the founding event of AI as a field. By 1956 the field had its first running program: Logic Theorist, by Newell, Simon and Shaw, proved 38 of the first 52 theorems in chapter two of Whitehead and Russell's Principia Mathematica — and found a shorter proof for one of them.

Hype, collapse, and the two AI winters

The pattern that would repeat for decades — bold demo, bigger headlines, brutal reckoning — appeared immediately. The Georgetown–IBM machine-translation demonstration of 1954 generated headlines like "Robot brain translates Russian into King's English," but had actually translated a curated set of just 49 sentences with a 250-word vocabulary. Frank Rosenblatt's perceptron, published in 1958 and first implemented in software for the IBM 704, showed machines could learn from examples — until Minsky and Papert's 1969 book Perceptrons proved single-layer networks could not even learn XOR, contributing to a roughly ten-year decline in neural network research.

ELIZA, Joseph Weizenbaum's conversation program published in January 1966, simulated a Rogerian psychotherapist — the first chatbot. The same year, the ALPAC report concluded machine translation was more expensive, less accurate and slower than humans; after roughly $20 million spent, US support ended. In 1973, Sir James Lighthill's report to the UK Science Research Council criticized AI's failure to meet its "grandiose objectives," leading to the near-complete dismantling of AI research in Britain. Between 1974 and 1980, the US and UK governments cut funding for undirected AI research — the first "AI winter," a term first used publicly at AAAI's 1984 annual meeting, when Roger Schank and Marvin Minsky warned that 1980s enthusiasm had spiraled out of control and disappointment would follow.

The 1980s boom failed the same way. Expert systems were the commercial success story — XCON, written at Carnegie Mellon for Digital Equipment Corporation, had processed 80,000 orders with 95–98% accuracy by 1986, saving an estimated $25 million a year. But in 1987 the market for specialized AI hardware collapsed when desktop computers from Apple and IBM outgrew expensive Lisp machines, demolishing an industry worth half a billion dollars and opening the second AI winter, which lasted into the 1990s.

Machines beat humans at their own games

The comeback came as brute force. On 11 May 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov 3½–2½ in their six-game rematch, with Kasparov resigning game 6 after only 19 moves — the first match loss by a reigning world champion to a computer under standard tournament time controls. A year earlier, Deep Blue had won a single game before losing the Philadelphia match 4–2. The machine's strength was not learning but search: 32 processors evaluating 200 million chess positions per second.

Autonomy in the physical world followed. On 8 October 2005, Stanford's Stanley won the DARPA Grand Challenge, completing a 212 km desert course in 6 hours 54 minutes — one of five vehicles to finish that day, a result that seeded the driverless-car era.

The deep-learning breakout

The modern era began quietly in 2006, when Geoffrey Hinton, Simon Osindero and Yee-Whye Teh published a fast learning algorithm for deep belief nets — the paper widely credited with reviving "deep learning." The fuel was ImageNet: introduced at CVPR 2009 with 3.2 million images, it grew to more than 14 million hand-annotated images across more than 20,000 categories.

The breakthrough moment came on 30 September 2012, when AlexNet — Krizhevsky, Sutskever and Hinton's convolutional network with 60 million parameters, trained on GPUs — won the ImageNet challenge with a top-5 error of 15.3%, more than 10.8 percentage points ahead of the runner-up's 26.2%. Deep learning had proved itself, and the field reorganized around it within a year or two.

Then machines learned to imagine and to play. In 2014, Ian Goodfellow and colleagues proposed Generative Adversarial Networks, training a generator and a discriminator against each other in a minimax game. In March 2016, DeepMind's AlphaGo defeated 18-time world Go champion Lee Sedol 4–1 before an audience of over 200 million — and in game two it played "Move 37," a shoulder hit its creators calculated had a 1 in 10,000 chance of being chosen by a human.

The architectural key arrived on 12 June 2017, when eight Google researchers posted "Attention Is All You Need," proposing the Transformer — a network based solely on attention mechanisms, dispensing with recurrence and convolutions, which set a then state of the art of 28.4 BLEU on English-to-German translation after training just 3.5 days on eight GPUs. Every frontier model today descends from that paper.

From language models to the ChatGPT moment

OpenAI applied the Transformer to text. GPT-1 arrived in June 2018, a 117-million-parameter model pre-trained on BooksCorpus; Google's BERT followed in October 2018 with state-of-the-art results on 11 NLP tasks. In February 2019, OpenAI announced GPT-2 — 1.5 billion parameters trained on 40GB of internet text — and withheld the full model "due to our concerns about malicious applications," releasing it only in November 2019 after staged rollouts found no strong evidence of misuse.

The scale-up crossed a threshold in May 2020 with GPT-3: 175 billion parameters, ten times more than any previous non-sparse language model, presented as a "few-shot learner" that could perform new tasks from a handful of examples. OpenAI opened the model family through a general-purpose API in June 2020 — "text in, text out" — which turned language models into a platform. In January 2021, the same stack learned to see and draw: DALL·E, a 12-billion-parameter GPT-3 variant generating images from text, and CLIP, which learned visual concepts from 400 million image-text pairs scraped from the internet.

Consumer AI arrived in 2022: Midjourney opened its Discord-based beta on 12 July, and Stable Diffusion's open release followed on 22 August, running text-to-image generation on consumer GPUs. Then, on 30 November 2022, OpenAI launched ChatGPT — a conversational model fine-tuned from GPT-3.5, free during the research preview. It reached an estimated 100 million monthly active users within two months, the fastest-growing consumer application in history according to a UBS study — versus nine months for TikTok and two and a half years for Instagram.

From chat to assistants (2023)

2023 turned the chatbot into an assistant. GPT-4, released 14 March 2023, "passes a simulated bar exam with a score around the top 10% of test takers," against the bottom 10% for GPT-3.5 — the first time a chatbot posted professional-exam-level scores. Anthropic — founded by ex-OpenAI researchers — launched Claude the same month, built on Constitutional AI, a training method where "the only human oversight is provided through a list of rules or principles" and AI feedback replaces human labels. Claude 2 opened to the public in July 2023 with a 100K-token context window — roughly a whole book per prompt.

Chat then gained eyes and ears. On 25 September 2023, OpenAI shipped GPT-4 with vision, framed as "ChatGPT can now see, hear, and speak." On 6 December 2023, Google announced Gemini, "built from the ground up to be multimodal" across text, code, audio, image and video. And the field started formalizing its destination: Google DeepMind's "Levels of AGI" framework, published November 2023, classifies AI systems by performance and generality — an attempt at a shared vocabulary for progress toward AGI rather than a single yes/no line.

From assistants to agents (2024–2026)

2024 brought reasoning and hands. OpenAI's o1 series (12 September 2024) introduced models "designed to spend more time thinking before they respond" — on an International Mathematics Olympiad qualifying exam, GPT-4o solved 13% of problems while o1 scored 83%. On 22 October 2024, Anthropic's computer use let Claude 3.5 Sonnet move a cursor and type like a person — the first frontier model with computer use — though the company was candid that it remained "at times cumbersome and error-prone," scoring 14.9% on the OSWorld computer-use benchmark where humans score roughly 70–75%. In November 2024, Anthropic open-sourced the Model Context Protocol, a standard for connecting assistants to tools and data. December brought a benchmark milestone: OpenAI's then-unreleased o3 scored 75.7% on ARC-AGI's semi-private evaluation, the first system to crack a test built for novel-task adaptation, where GPT-4o managed about 5%.

2025 was the year agents went to work. DeepSeek's R1 (20 January 2025) proved frontier-style reasoning was no longer a US monopoly; its claimed "less than $5.6 million" V3 training cost — a figure covering GPU rental only, as analysts noted — reset assumptions about capital requirements. OpenAI launched Operator (23 January), an agent that browses and completes tasks, and shipped agentic coding across the industry that year: Claude Code (February), OpenAI Codex (May), and Google's open-source Gemini CLI (June) put coding agents in developers' terminals. Anthropic's MCP became a de facto standard when OpenAI adopted it in March — four months after launch.

The measured trend underneath is the story's real signal. METR's March 2025 study measured the length of human-time tasks AI agents complete with 50% reliability and found the metric had doubled roughly every seven months for six years. July 2025 follow-up work found 2024's roughly four-month doubling "likely not an outlier" across domains, and METR's January 2026 revision re-estimated the doubling at 131 days since 2023 and 89 days since 2024 — measurement noise is real, but the direction is faster, not slower. In August 2025, GPT-5 blended chat and reasoning into one product: "a unified system that knows when to respond quickly and when to think longer."

So where is AGI?

Start with definitions, because they carry the disagreement. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" while stating plainly that "the timeline to AGI remains uncertain" — and OpenAI has never claimed to have met its own definition. DeepMind's framework deliberately avoids a single threshold, placing today's systems in an "emerging" tier rather than at AGI.

The forecasts are projections, not outcomes. The largest survey of AI researchers — 2,778 published authors, analyzed by AI Impacts — put a 50% chance of high-level machine intelligence by 2047 and 10% by 2027, a forecast 13 years earlier than the same community gave in the 2022 round of the survey.

The benchmarks built to detect general intelligence still show large gaps. ARC-AGI-2, launched March 2025, holds frontier models under 3% against roughly 60% for average human test-takers. Humanity's Last Exam (January 2025) found every frontier model under 10% on 2,500+ expert questions — and its authors warn that even high scores "would not alone suggest autonomous research capabilities or 'artificial general intelligence.'" METR's own summary is blunt: "the best AI agents are not currently able to carry out substantive projects by themselves or directly substitute for human labor."

No AGI exists as of October 2026. What exists is a measured, accelerating trend — task horizons doubling in months, not years — aimed at definitions that no system meets, with expert medians decades out and every general-intelligence benchmark still showing a wide human lead.

Where it's going

The honest answer comes from measurement, not prediction. The one load-bearing number is METR's time horizon: if the doubling of task length continued from its measured 2024 value of roughly 89 days, month-long autonomous work would arrive within years — an extrapolation, not a fact, and the same paper's revisions show the estimate itself moving. What the sources do establish: the shift from chat to agents is already shipped products rather than demos (Operator, computer use, agentic coding, MCP as a standard), the frontier is no longer one company's to hold, and the benchmarks meant to catch general intelligence are being rebuilt as fast as models pass the old ones. Between Turing's five-minute conversation test and machines that work all day unsupervised sits the same gap the field has crossed before — once with search, once with learning — and the doubling curve says this crossing is measured in years, not decades.

For the full report - formatted, with every source listed and verified - click below.

The Story of AI (full report)

Sources

Tags:#AI#Deep Dive#Future#Technology
Share this post:X / TwitterLinkedInRedditWhatsApp

No comments yet

Related posts