Annals · The Generative AI Wave (2022–2026)
On November 30, 2022, OpenAI quietly put a chatbot called ChatGPT online. There was no press conference, no advertising — just a tweet. Five days later, it had a million users. Two months later, a hundred million. No consumer product in human history had ever spread that fast. AI walked out of the laboratory and into everyone's pocket. Everything was speeding up: models grew larger, money poured in, regulators tried to catch up, debates sharpened. No one knew where the destination lay — but everyone knew there was no going back.
I. The Central Question: Where Are AI's Capability Boundaries?
ChatGPT's release pushed a question once confined to AI circles in front of the whole human species: what can AI actually do, and what can it not do?
Optimists pointed to GPT-4's astonishing performance — passing the bar exam, writing executable code, conversing fluently in many languages, explaining complex scientific concepts. They argued that artificial general intelligence (AGI) was no longer a distant fantasy but a goal possibly achievable within ten years.
Pessimists pointed to the fundamental limits of large language models — they "hallucinate," confidently inventing facts that do not exist; they have no real "understanding," only extremely complex pattern matching; they cannot perform reliable logical reasoning, and routinely err on simple math problems.
The debate is not settled. But one thing is clear: regardless of whether current AI capabilities amount to "real intelligence," they have already transformed how hundreds of millions of people work and live. This time, the impact does not stay in academic papers or industry reports — it is happening on every screen.
II. ChatGPT: The Ignition Point
A Million Users in Five Days
On November 30, 2022, OpenAI released ChatGPT — a conversational AI based on GPT-3.5. Its technical foundation was the reinforcement learning from human feedback (RLHF) introduced in InstructGPT — human annotators ranked the quality of model outputs, and reinforcement learning was then used to optimize the model's conversational behavior.
It is worth noting that, internally, OpenAI did not regard ChatGPT as an important product. It was framed as a "low-key research preview," intended to gather feedback from real users to prepare for the upcoming GPT-4. According to several former employees, internal expectations for the launch were extremely modest — some bet that the user count would plateau at tens of thousands within a week. No one expected the world's reaction.
ChatGPT's user growth set a historical record: 1 million users in 5 days, 100 million in 2 months — beating the previous record-holders TikTok (9 months to 100 million) and Instagram (2.5 years).
ChatGPT's success rested on no single technical breakthrough. GPT-3's capabilities had been demonstrated two years earlier. But ChatGPT did one thing GPT-3 had not: it gave ordinary people a usable interface. You did not need to understand prompt engineering, you did not need an API key, you did not need any technical background — open a browser, type, hit enter. That ultra-low barrier expanded large language model capabilities from a few developers to hundreds of millions of users worldwide.
GPT-4: A Multimodal Leap
On March 14, 2023, OpenAI released GPT-4. Compared with GPT-3.5, GPT-4 made significant gains in reasoning, breadth of knowledge, and instruction following. It scored in the top 10 percent of test-takers on simulated U.S. bar exam questions. It also added multimodal abilities — it could understand image inputs (analyzing charts, reading handwritten notes, getting jokes in cartoons).
OpenAI never officially disclosed GPT-4's parameter count, but it was reportedly on the order of a trillion — five to six times GPT-3 — and its training cost was estimated to exceed 100 million dollars.
III. The Battle of a Hundred Models: A Global Race
ChatGPT's success set off a global arms race in large models.
The Closed-Source Camp
- Google: released Gemini in December 2023 (consolidating the original DeepMind and Google Brain research), with native multimodal capabilities.
- Anthropic: under Dario Amodei (1983–), the team released the Claude series in 2023–2024, distinguished by its safety properties and long-context handling. By 2024, Claude had become one of the strongest competitors to ChatGPT.
- OpenAI: continued to iterate the GPT line — GPT-4o (native multimodal, real-time voice interaction) in 2024, GPT-5 in 2025.
The Open-Source Camp
In February 2023, Meta released LLaMA (Large Language Model Meta AI), with Hugo Touvron (1995–) as first author. LLaMA's weights were "leaked" onto the internet (Meta later moved to a more open policy for subsequent versions), igniting a wave of open-source large models.
Mistral (France, 2023), Qwen (Alibaba, 2023), DeepSeek (China, 2024–2025) and others followed. They proved an important thing: the most capable large models do not have to be monopolized by a few companies.
In January 2025, the Chinese team DeepSeek released DeepSeek-R1, achieving reasoning ability close to GPT-4 level at a fraction of the training cost — sending a shockwave through the global community. The result punctured the assumption that "only multibillion-dollar budgets can train top models" and forced Silicon Valley to take the competitiveness of Chinese AI more seriously.
IV. AI Reshapes the World
The Reshaping of Work
By 2025, large language models had penetrated nearly every domain of knowledge work:
- Programming: GitHub Copilot and similar tools were used by millions of developers; estimates suggested AI-assisted code accounted for around 40 percent of newly written code.
- Writing and content creation: from marketing copy to first drafts of academic papers, AI-assisted writing became routine.
- Education: AI tutoring systems began personalizing instruction, while sparking heated debates over academic integrity.
- Healthcare: AI-assisted diagnosis, drug discovery, and medical imaging analysis advanced quickly.
- Law and finance: AI played a much larger role in contract review, compliance checks, and financial analysis.
Scientific Discovery
AI's accelerating effect on scientific research became more pronounced in this period. AlphaFold's breakthrough in protein structure prediction was only the start. By 2025 AI was being used in new-material discovery, climate model optimization, astronomical data analysis, mathematical theorem proving, and many more areas.
The 2024 Nobel Prize in Physics was awarded to Geoffrey Hinton (1947–) for his foundational contributions to artificial neural networks and machine learning, and the Chemistry Prize to Demis Hassabis (1976–) for AlphaFold's breakthroughs in protein structure prediction. It was the first time AI researchers had received Nobel Prizes — marking AI's transition from tool to engine of scientific discovery.
V. AI Safety and Governance
The Alignment Problem
As model capabilities grew, "alignment" — ensuring that AI systems behaved in accord with human intentions and values — went from a theoretical concern to an urgent engineering problem.
In November 2023, OpenAI experienced a dramatic internal crisis: the board abruptly fired CEO Sam Altman (1985–), reportedly over disagreements about AI-safety direction. Ilya Sutskever (1986–) initially supported the firing but reversed his position after a massive employee backlash. Altman returned to the role within five days; Sutskever subsequently left OpenAI. The crisis exposed the internal tensions of AI safety to the entire world.
A Global Regulatory Race
Governments accelerated work on AI regulation:
- European Union: passed the AI Act in 2024 — the world's first comprehensive AI regulatory framework, classifying AI applications by risk level.
- United States: in October 2023, President Biden signed the Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.
- China: from 2023, issued a series of regulations covering generative AI management, algorithmic recommendation, and more.
- By 2026, more than 70 countries and regions worldwide had enacted AI policies or regulations.
The AI Risk Statement
In May 2023, hundreds of AI researchers and industry leaders — including Hinton, Yoshua Bengio (1964–), Altman, and Hassabis — signed a one-sentence statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The brief statement drew wide attention and controversy.
In May 2023, Hinton left Google, saying publicly that he wanted to be free to discuss AI's risks. He went from being deep learning's most committed champion to one of its loudest warning voices — and the change in his stance was itself a marker of the shift of the times.
VI. Undercurrents and Foreshadowing
First, the rise of AI agents. By 2025 AI was no longer just a chatbot answering questions. AI agents could plan tasks autonomously, call tools, search for information, execute code, and even coordinate with other agents to accomplish complex workflows. It pointed to AI's evolution from "conversation tool" toward "autonomous actor."
Second, the breakthrough in video generation. In February 2024, OpenAI demonstrated Sora — a model able to generate high-quality video from text. The output still had visible flaws, but it pointed to a direction: AI would no longer be limited to text and images, but would extend across video, 3-D, music, and every other medium.
Third, compute bottlenecks and breakthroughs. Training a GPT-4-class model required tens of thousands of GPUs running for months, at hundreds of millions of dollars. Chip supply, data-center construction, and energy consumption became physical constraints on AI's growth. At the same time, model-efficiency optimization (sparsification, distillation, mixture-of-experts (MoE)) advanced rapidly — DeepSeek-R1's success was one example of an efficiency breakthrough.
Fourth, the story is not finished. As of 2026, AI's pace of progress is still accelerating. Will AGI arrive within ten years? Will AI replace human jobs at scale? Can the safety question be effectively governed? None of these questions has an answer. The only certainty is that we are in the fastest-changing period in AI's history — and the history is still being written.
VII. Timeline
| Year/Month | Event | Key Figures |
|---|---|---|
| 2022.11 | ChatGPT released; 1 million users in 5 days, 100 million in 2 months | Altman |
| 2023.2 | Meta releases LLaMA, the first major open-source LLM | Touvron |
| 2023.3 | GPT-4 released; passes simulated bar exam questions in the top 10% | Altman |
| 2023.5 | AI extinction-risk statement published; signed by hundreds of leaders | Hinton, Bengio, Altman |
| 2023.5 | Hinton leaves Google to warn publicly about AI risk | Hinton |
| 2023.11 | OpenAI internal crisis; Altman fired and reinstated | Altman, Sutskever |
| 2023.12 | Google releases Gemini multimodal model | Hassabis |
| 2024.2 | OpenAI demonstrates Sora video generation model | — |
| 2024.3 | EU passes the AI Act | — |
| 2024.10 | Hinton wins Nobel Prize in Physics; Hassabis wins Nobel Prize in Chemistry | Hinton, Hassabis |
| 2025.1 | DeepSeek-R1 achieves top-tier reasoning at very low cost | — |
| 2026 | AI deeply integrated across industries; over 70 countries with AI policies | — |
Historian's Note
From McCulloch and Pitts writing the first neuron model in 1943 to large language models running in everyone's pocket in 2026, AI has walked eighty-three years. In that history there have been two golden ages and two winters, the long quarrel between symbolism and connectionism, the paradigm shift from hand-coded knowledge to data-driven learning, the few who held the line in winter and the many who poured in during the boom. If this history has a single leitmotif, it is this: humans have consistently underestimated the complexity of the problem of intelligence, and have just as consistently underestimated their own ability to solve it. Two winters made the world believe AI was dead, but it returned each time stronger. ChatGPT made the world believe AGI was at hand, but the deep problems of general intelligence — common sense, causal reasoning, value alignment — have still not been truly solved. Optimism and pessimism are both unreliable. Only one thing is reliable: this story is far from over. As the recorders of this history, our only duty is to record faithfully, to reflect faithfully, and to let those who come later see the whole picture — the splendor and the lessons together.
Eyewitness Accounts
Call for contributions
If you witnessed or have firsthand recollections of this period, please contribute on GitHub.
References
- OpenAI (2022). "Introducing ChatGPT." OpenAI Blog, November 30, 2022.
- OpenAI (2023). "GPT-4 Technical Report." arXiv:2303.08774.
- Touvron, H. et al. (2023). "LLaMA: Open and Efficient Foundation Language Models." arXiv:2302.13971.
- Anthropic (2024). "The Claude 3 Model Family." Anthropic Technical Report.
- EU (2024). "Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)."
- CAIS (2023). "Statement on AI Risk." Center for AI Safety, May 2023.
- Ramesh, A. et al. (2022). "Hierarchical Text-Conditional Image Generation with CLIP Latents." arXiv:2204.06125.
- Rombach, R. et al. (2022). "High-Resolution Image Synthesis with Latent Diffusion Models." CVPR 2022.
- DeepSeek AI (2025). "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning." arXiv:2501.12948.
- The White House (2023). "Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence."
- Hu, E. J. et al. (2022). "LoRA: Low-Rank Adaptation of Large Language Models." ICLR 2022.