The Generative Era
How AI went from research labs to hundreds of millions of phones in about three years: images from words, chat assistants, open models, Nobel Prizes, reasoning, agents and the first big laws. And why nobody agrees on where it goes next.
On 30 November 2022 OpenAI put a free “research preview” on the web and called it ChatGPT. There was no advertising campaign. Within days people were using it to write cover letters, debug code, explain tax forms and compose poems about their cats. Within about two months, by one widely quoted estimate, it had more monthly users than most apps gather in years. The technology underneath was not new; the Transformer was five years old. What was new was that anyone could talk to it. This page tells the story of the years either side of that moment, a period historians may well call the generative era.
Pictures from words: 2021–2022
On 5 January 2021 OpenAI announced two systems together. CLIP learned to match images with captions by training on 400 million image–text pairs gathered from the internet, so it could recognise things it had never been explicitly taught to label. DALL·E, a 12-billion-parameter version of GPT-3, generated images from text prompts such as “an armchair in the shape of an avocado”. Pictures had become just another sequence to predict.
The technique that took over image generation, though, was the diffusion model. The idea, first set out by Jascha Sohl-Dickstein and colleagues in 2015, is to train a network to remove noise: take real images, add static step by step until nothing is left, and teach the network to reverse each step. Start from pure noise and run it backwards, steered by a text description, and an image emerges. In June 2020 Jonathan Ho, Ajay Jain and Pieter Abbeel showed with “denoising diffusion probabilistic models” (DDPM) that this could rival the best image generators.
2022 brought the results to the public. OpenAI’s DALL·E 2 was announced in April. Midjourney opened a public beta through Discord in July. In August Stability AI, working with the CompVis group at LMU Munich and Runway, released Stable Diffusion with its weights downloadable for free, so anyone with a decent graphics card could generate images at home. An art competition prize at the Colorado State Fair that year, won by an image made with Midjourney, became an early flashpoint for the arguments about creativity and copyright that are still unresolved.
From autocomplete to assistant
Text models were moving too. In mid-2021 OpenAI described Codex, a GPT model fine-tuned on public code, and GitHub launched Copilot, built on it, as a technical preview that suggested whole lines and functions inside a programmer’s editor. It was many people’s first experience of a generative AI tool woven into daily work.
A raw large language model is a brilliant but unhelpful conversationalist: it continues text rather than answering. Ask it a question and it might reply with three more questions, as if completing a quiz. The fix was reinforcement learning from human feedback. In January 2022 OpenAI described InstructGPT, building on earlier work by Paul Christiano and colleagues (2017). Tap through the chain to see the recipe.
ChatGPT, released on 30 November 2022, applied the same recipe to a dialogue model and wrapped it in a simple chat window. It was free. That combination, a capable model plus a frictionless interface, did what no research paper had.
How fast did it spread? OpenAI did not publish monthly figures at the time. On 1 February 2023 Reuters reported a UBS analyst note, based on web-traffic data from Similarweb, estimating that ChatGPT had reached about 100 million monthly active users in January 2023, roughly two months after launch. The same report gave comparisons from Sensor Tower for other apps.
The race: 2023–2024
What followed was an industry-wide sprint. Tap the tabs for the main threads.
On 14 March 2023 OpenAI released GPT-4, which accepted images as well as text and scored highly on many professional and academic exams. The technical report withheld the model’s size and training details, a sign of how commercial the field had become. The same day Anthropic, founded in 2021 by former OpenAI staff, released Claude. Google, which had invented the Transformer, launched Bard in early 2023 and in December 2023 introduced Gemini, built to handle text, images, audio and video from the start; Bard was later renamed Gemini. Microsoft built OpenAI’s models into Bing and Office. By 2024 several companies were releasing new model generations every few months As of 2026.
In February 2023 Meta released LLaMA to researchers; its weights leaked online within about a week. In July 2023 Llama 2 followed with open weights under a licence that allowed most commercial use. Mistral AI in France, and Chinese labs such as Alibaba’s Qwen team and DeepSeek, released strong open-weight models too. Anyone could now download, run and fine-tune a capable language model on their own hardware. Note the term: “open weights” means the trained parameters are published, which is not the same as “open source” in the full sense, because training data and code are often withheld As of 2026.
Models became multimodal: able to see, hear and speak. In 2024 OpenAI’s GPT-4o handled real-time voice conversation, Google demonstrated assistants that could interpret live camera video, and text-to-video systems such as OpenAI’s Sora (shown in February 2024) and Google’s Veo produced short, convincing clips. The same advances made deepfakes cheap, a problem that reached elections, scams and schools As of 2026.
The term foundation model, coined in a 2021 Stanford report, captured the new pattern: one very large pretrained model, adapted for thousands of uses. It also captured a worry. If a handful of models sit underneath everything, their flaws and biases spread everywhere.
Nobel Prizes, reasoning and agents
In October 2024 AI research received two Nobel Prizes in two days. The Physics prize went to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks”: Hopfield’s 1982 associative-memory network and Hinton’s Boltzmann machine (with Terry Sejnowski and others, 1985) drew directly on statistical physics (see The Connectionist Revival). The Chemistry prize went half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for AlphaFold’s protein-structure prediction. Hinton used his moment to repeat warnings about AI risk he had voiced since leaving Google in 2023.
Meanwhile the models changed shape again. In September 2024 OpenAI released o1, the first widely used reasoning model: trained with reinforcement learning to produce a long internal chain of thought before answering, spending more computation at answer time on hard problems in maths, science and coding. In January 2025 the Chinese lab DeepSeek released R1, an open-weight reasoning model whose reported low training cost briefly rattled technology stock markets. Through 2025 most major developers added “thinking” modes to their models As of 2026.
The other trend is the agent: a model that does not just answer but acts, using tools such as a web browser, a code editor or a company’s software, checking the result, and carrying on. Coding was the first domain where this became routine, as assistants moved from suggesting single lines (Copilot, 2021) to editing whole projects, running tests and proposing changes. Standards for connecting models to tools, such as Anthropic’s Model Context Protocol (November 2024), were taken up across the industry. Agents remain error-prone over long tasks, and they open new security holes such as prompt injection As of 2026.
Rules arrive
Governments moved more slowly than the technology, but they moved. In November 2023 the United Kingdom hosted an AI Safety Summit at Bletchley Park, where 28 countries and the EU signed a declaration on frontier-AI risks. The European Union’s AI Act, the first comprehensive AI law of its kind, entered into force on 1 August 2024. It sorts systems by risk: some uses (such as social scoring) are banned, “high-risk” uses such as hiring or credit scoring face strict duties, and general-purpose models carry transparency obligations. Its rules phase in over several years: the bans applied from February 2025 and the general-purpose-model rules from August 2025, and a 2026 amendment (the “digital omnibus”, agreed in May and in force from late July) pushed the main high-risk deadlines back to December 2027 and August 2028 As of 2026. Malaysia published National Guidelines on AI Governance and Ethics in 2024 and set up a National AI Office (made permanent as AI Malaysia in July 2026), relying on voluntary guidance so far; in July 2026 it consulted the public on a proposed AI Governance Bill, now expected in Parliament in early 2027 As of 2026. The full story is on Regulating AI and AI in Malaysia.
Where is it heading?
Describing 2025 and 2026 precisely is risky: products change monthly and much is announced before it is proven. The broad trends are clearer. Models keep getting more capable at coding, maths and multi-step tasks; the cost of a given level of capability keeps falling; open-weight models trail the frontier by months rather than years; and spending on data centres and chips has reached a scale that worries economists and energy planners alike As of 2026.
Speculative Whether these trends continue to anything like artificial general intelligence, and when, is the biggest open question in the field. Treat every forecast, including confident ones from famous people, as a forecast.
- Within a few years: several frontier-lab leaders have publicly suggested that systems matching humans at most cognitive work could arrive before 2030, pointing to fast benchmark gains and the scaling of reasoning and agents.
- Decades away, or a different path: researchers such as Yann LeCun argue that today’s language models lack world models, persistent memory and real planning, and that new architectures will be needed; others, such as Gary Marcus, stress persistent unreliability.
- The wrong question: many economists and social scientists argue that “AGI” is too vague to date, and that the useful questions are which tasks get automated, how fast, and who benefits.
- Prepare regardless: safety researchers, including Hinton and Bengio, argue that because experts disagree so widely, even a modest chance of very powerful systems soon justifies serious safety work and regulation now.
History has a lesson here, set out in Lessons from History: AI forecasts have failed in both directions. The 1960s were too optimistic; the 2010s were too pessimistic about language.
What does a diffusion model learn to do?
Diffusion models reverse a gradual noising process. Text guidance steers which image emerges.
What was the main purpose of RLHF in InstructGPT and ChatGPT?
Human rankings train a reward model, which then guides further training. The model gets more helpful, not more knowledgeable.
The “100 million users in two months” figure for ChatGPT is best described as…
It is widely cited and plausible, but it was an analyst estimate from web-traffic data, not an official count.
// Pause and reflect
- The technology behind ChatGPT existed before November 2022. Why do you think a chat window changed public perception so much more than the research papers did?
- Should powerful models be released with open weights? Which arguments on each side do you find strongest?
- Pick one forecast about AI you have heard recently. What evidence would convince you it was right, or wrong, by 2030?
Your notes are saved privately in this browser.
What happened next
This chapter is still being written. To see it in context, scroll through The Complete AI Timeline, meet the people behind the ideas in Pioneers of AI, or step back and ask what the past suggests about the future in Lessons from History.
Key terms on this page
Sources & further reading
- Ho, J., Jain, A. & Abbeel, P., Denoising Diffusion Probabilistic Models (2020).
- Radford, A. et al., Learning Transferable Visual Models From Natural Language Supervision (CLIP, 2021).
- Ouyang, L. et al., Training language models to follow instructions with human feedback (2022).
- OpenAI, GPT-4 Technical Report (2023).
- Bommasani, R. et al., On the Opportunities and Risks of Foundation Models (2021).
- The Nobel Prize, Physics 2024 and Chemistry 2024 summaries (2024).
- Regulation (EU) 2024/1689, the Artificial Intelligence Act (2024).
- Hu, K., “ChatGPT sets record for fastest-growing user base – analyst note”, Reuters (1 February 2023).