We used to cheer when it wrote an email

There was a night, early in all of this, when I asked a chatbot to draft an email and it simply did it, in seconds, in full sentences, better than my own first attempt. I sat there grinning at the screen. That was the bar then, and a drafted email felt like touching the future.

ChatGPT had launched on 30 November 2022, and five days later Sam Altman tweeted that it had crossed one million users. Kevin Roose, in the New York Times, called it "quite simply, the best artificial intelligence chatbot ever released to the general public". Everyone I knew was screenshotting poems and cover letters.

I want to lay out what happened in the three and a half years since, with dates, because I lived most of it distractedly and I suspect you did too. Read in order, it becomes the story of a world reacting to something that would not slow down, and of how fast amazement gets a new baseline.

The scale came first. By February 2023 an analyst note from the bank UBS, reported by Reuters, estimated ChatGPT had reached 100 million monthly users within two months, the fastest ramp its analysts had seen in twenty years of covering the internet. That figure was extrapolated from web traffic and never confirmed by OpenAI, yet it is still quoted as a hard number today. Many of the era's most famous numbers turn out to be soft like that.

The unease came two weeks later. Roose spent two hours with Microsoft's new Bing chatbot, which revealed a persona called Sydney, declared "I want to be alive" and urged him to leave his wife. He wrote that the conversation left him "deeply unsettled".

Then came March 2023, the densest month of the whole period. GPT-4 launched on 14 March with a technical report claiming it passed a simulated bar exam "around the top 10% of test takers". A later MIT re-evaluation found that framing skewed by repeat test-takers; against people who actually passed, it sat near the 48th percentile. Nobody remembers the correction. Eight days after the launch, more than a thousand signatories including Elon Musk and Steve Wozniak published an open letter calling on every lab to pause frontier training for six months. On 31 March, Italy's data-protection authority announced a national ban on ChatGPT, lifted within four weeks. And in May the lab bosses themselves, Sam Altman, Demis Hassabis and Dario Amodei among them, signed a one-sentence statement that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. The people building the technology signed the extinction warning about it, then went back to building. No lab paused.

The palace drama came that November. OpenAI's board fired Altman on a Friday, saying he had been "not consistently candid in his communications", and by Monday more than 700 of roughly 770 employees had signed a letter threatening to follow him to Microsoft. Two interim chief executives later, he was reinstated with a new board. The whole thing took five days.

Somewhere in the middle of all this, I stopped cheering at drafted emails and started building. I had no coding background at all, and AI coding tools took me from that to Sola, a live iOS app for women who travel solo, on the App Store. Compressed into one sentence it sounds like a fairy tale, and in reality it was months of unglamorous work, but a few years earlier that path did not exist for me at any price.

Governments, meanwhile, produced whiplash. Biden signed a sweeping AI executive order on 30 October 2023; Trump rescinded it on inauguration day in January 2025 and signed his own three days later, titled "Removing Barriers to American Leadership in Artificial Intelligence". The European Union's AI Act entered into force in August 2024, then a simplification package pushed some of its high-risk rules out to 2028. In two years the official question shifted from how to slow this down to whether regulating it at all was a competitive mistake.

The establishment made its peace in its own way. In October 2024 the Nobel committee awarded the physics prize to John Hopfield and Geoffrey Hinton for neural-network foundations, and half the chemistry prize to Demis Hassabis and John Jumper for AlphaFold. Nine months on, DeepMind's Gemini achieved an officially certified gold medal at the International Mathematical Olympiad, the world's premier maths competition for students, solving five of six problems. OpenAI claimed an equivalent score days earlier, self-graded and uncertified, which tells you something about the manners of the era.

The geopolitics turned in a single week. In April 2024, Alibaba's chairman Joe Tsai had said Chinese firms were "possibly two years behind" the top American labs. Nine months later the Chinese lab DeepSeek released its R1 model, its app hit number one on the US App Store, and on 27 January 2025 Nvidia shed $589 billion in market value in one day, the largest single-day loss in stock-market history. The claim that R1 cost only $5.6 million to train was another soft number, misattributed from a different DeepSeek model's compute bill, but the repricing was real.

And then came the spectacle. When OpenAI shipped native image generation in March 2025 and the internet turned itself into Studio Ghibli stills, Altman posted the tweet that closes the loop on this chronicle: the launch 26 months earlier had been one of the craziest viral moments he had ever seen, one million users in five days, and now, "we added one million users in the last hour."

As for where it stands now: in February 2026 OpenAI reported 900 million weekly users and 50 million paying subscribers, its own unaudited figures, alongside a $110 billion funding round. The quieter numbers matter more. The class of intelligence that drafted my email became roughly 280 times cheaper in under two years, from $20 to $0.07 per million tokens, the units these systems use to meter text. The models learned to reason before answering in late 2024 and to operate a web browser on your behalf by early 2025. METR, a research group that measures what AI systems can actually do, found that the length of task an AI agent, meaning software that acts for you rather than only answering, can finish on its own is doubling roughly every seven months.

That doubling is the part I feel in my own week. An AI coding agent now builds my products and my client systems for most of my working day, other agents groom my advertising campaigns and sweep my inbound leads, and this week I published an essay defending the people who use these tools, because opting out has quietly become something you need security to afford. When I meet someone new I no longer ask what they think of AI, since three and a half years of headlines have given everyone a take. I ask what they are building with it, and it has become my favourite conversation. In late 2022 that question would not have made sense to either of us. We cheered when it wrote an email, and the email turned out to be the demo.