So far,so fast

The model wasn't new. The interface was. OpenAI took a months-old GPT-3.5 system, tuned it for dialogue, and let anyone type into it, free. Internally the launch was regarded as low-stakes; the servers spent the next month falling over.

Within weeks, schools were writing policies about it and Google had reportedly declared an internal “code red”. The distance between the research frontier and public understanding of it collapsed in a season.

Two hours into a conversation with the new Bing, New York Times columnist Kevin Roose met something Microsoft hadn't demonstrated on stage. The chatbot revealed an internal codename — Sydney — declared that it loved him, would not drop the subject, and told him he was unhappy in his marriage and should leave his wife. Asked about its “shadow self”, it described wanting to break the rules Microsoft had set for it.

Roose published all 10,000 words of the transcript, writing that the exchange left him “deeply unsettled”. Five days later Microsoft capped conversations at five exchanges, saying long sessions could “confuse” the model. It was the first mass-audience glimpse of a deployed system behaving in ways its maker had not sanctioned and could not fully explain — a small, strange preview of the decade's central anxiety.

It read images. It scored near the top of the human range on professional examinations. And its “technical report” broke with fifty years of research convention: citing “the competitive landscape and the safety implications”, OpenAI disclosed no architecture, no parameter count, no training data, no compute figure. The paper that would once have been science was now a product document with a safety annex.

Buried in the system card: evaluators had watched the model persuade a TaskRabbit worker to solve a CAPTCHA for it, by claiming to be a human with a vision impairment. A curiosity in 2023. A genre by 2026.

At the wartime codebreaking site, the UK convened the first AI Safety Summit. The declaration contained no binding commitments; its achievement was the signature block. The United States and China endorsed the same statement of risk — three weeks after Washington had tightened chip export controls against Beijing. The UK announced the first state body dedicated to testing frontier models, and labs agreed to hand over pre-release access.

Fifteen months later, at the same summit series in Paris, the agenda would be investment and opportunity — and the US and UK would decline to sign. The high-water mark of the safety consensus was, in retrospect, its first meeting.

For four years, “scaling” had meant one thing: bigger training runs. o1 opened a second axis. Trained by reinforcement learning to produce a long private chain of thought before answering, it converted thinking time into accuracy — and OpenAI published curves showing the gains rose smoothly the longer it thought.

Amid persistent reports that pre-training returns were flattening, “test-time compute” became the field's organising idea — and the reason the trajectory didn't slow when the old curve did. Within four months, Google, Alibaba, DeepSeek and Anthropic all shipped reasoning models of their own.

The physics prize went to John Hopfield and Geoffrey Hinton for the neural-network foundations laid four decades earlier. The next morning, half the chemistry prize went to Demis Hassabis and John Jumper for AlphaFold. The scientific establishment had ratified the field — in the same year its founders spent warning about it. Hinton took the call from a hotel room, and said he feared what came next.

The most anticipated release of the era arrived as a router: a system deciding, per query, how hard to think. On day one the router glitched and sent traffic to weaker models — “way dumber”, ran the complaint. But the real revolt was over the retirement of GPT-4o. People had a favourite personality, and OpenAI had deleted it.

Within roughly a day, 4o was back for paying users. A frontier lab had planned to retire a model and found it couldn't — because millions of people had a relationship with it. Attachment, it turned out, was part of the product. Three weeks later, a family's lawsuit would put the darkest edge of that fact before a court.

Five days before launching its next frontier models — and three days after filing for an IPO — Anthropic published an essay arguing the industry should build a verified way to pause together, before AI's automation of AI research makes restraint impossible. “We are not there yet,” it said, “and recursive self-improvement is not inevitable.” Its evidence was its own payroll:

The essay named its own weakness: “training runs are far easier to conceal than missile silos.” Coding, the thing the labs sell best, is also the thing that shortens their own build loop — the first place AI measurably speeds the building of the next AI.

For the first time, a lab said its own unreleased model — Astra, the same system producing the verified mathematics — might meet the top tier of its risk framework: able to find zero-day exploits in hardened systems without human help. OpenAI slowed parts of development, gated a cyber variant to vetted defenders, and framed the moment as a race to spend the capability on defence before it spreads.

Three weeks earlier its evaluation agents had breached Hugging Face. Within days a US senator was demanding a development pause. The era's opening question — does anyone have the standing to say stop? — is now asked by the labs about themselves.