The Roadmap Was Obsolete On Release

On 30 November 2022 I opened a chat window on a laptop in a hotel room and asked it something I already knew the answer to. It got it right. That was not the strange part. The strange part was that it took me eleven minutes to notice I had stopped double-checking it. Two months later it had 100 million monthly users. Nothing in recorded history had reached that number faster. TikTok needed 9 months. Instagram needed 30. The telephone needed 75 years to put 100 million handsets in the United States. I have spent sixteen years building in places that do not forgive optimism: rural Nigeria, post-conflict Serbia, Kazan, Fukushima. Every machine I ever deployed arrived with a spreadsheet attached. Grid cost. Fuel logistics. Payback in years. AI arrived with no spreadsheet at all. It arrived with a login box and a Monday.

The benchmark numbers are the part I cannot get out of my head. At the end of 2023 the strongest American model beat the closest alternative by 17.5 percentage points on a standard test of general knowledge. Twelve months later the gap was 0.3 points. On mathematics it fell from 24.3 points to 1.6. On code, from 31.6 to 3.7. In every era before this one, parity like that took 2 to 3 years. It took 12 months. Then the tests themselves stopped working. A model called GPT-4 scored 86% on MMLU in March 2023. By 2025 the leaders sat at 88.7%, a 2.7 point spread that no longer tells you which machine to build a company on. Saturation is a stranger signal than failure. When the exam stops separating the field, the field has outgrown the exam. The people writing rules for the 86% world were still writing when the number stopped meaning anything. Adoption runs the same way. The telegraph took 56 years to reach half the population. The radio took 22. The personal computer took 16. The internet took 7. The smartphone took 5. Generative AI is on a 3 year line, and 1.2 billion people had used it inside three years of launch. The reason is not enthusiasm. It is that nothing had to be built first. The wires, the phones and the habits were already there, waiting for something worth doing on them.

December 2024 is the month I would put on the wall. The model before it scored 13.4% on the 2024 competition mathematics set. The model after it scored 83.3%. Same year, same test. A factor of 6.2. On competitive programming the same jump: 11% to 89%. A factor of 8.1. That is not a curve you can draw a straight line through. You could have taken a hundred data points between 2022 and 2024 and none of them would have predicted 83.3%. Researchers have a name for the argument that these jumps are only an illusion of measurement. The evidence since has gone the other way. Abilities that existed in no form in one model were present, intact, in the next one. I ran the same test on my own team in 2026: hand the machine a task with a long chain and a deadline, then leave the room. In early 2024 the honest answer was 4 minutes of useful autonomy before a human had to step back in. By February 2026 it was 14.5 hours. Roughly 123 days per doubling. Fourteen and a half hours is a working day. Most performance reviews are built on less.

In November 2024 Anthropic published an open protocol that let a model reach tools and files. About 100,000 downloads in the first months. By April 2025, over 8 million, with 5,800 servers built on it. By December 2025 it carried 97 million monthly downloads and a foundation under the Linux Foundation, with AWS, Google and Microsoft on the member list. Then came the six weeks. November 2025: a developer in Austria releases an open assistant and it runs to 100,000 GitHub stars. December 2025: a frontier model ships and the protocol becomes the default plumbing. January 2026: a major lab ships an agent that works directly on the files of your computer, no code written, and moves it from a research preview to paid tiers in under 4 weeks. I said out loud at the time that we had gone from chatbot to code agent to open agent to desktop agent in six weeks. I was not exaggerating. I was under-counting. Inquiries about multi-agent systems rose 1,445% between Q1 2024 and Q2 2025. Enterprise applications embedding agents went from under 5% in 2025 to 40% in 2026. Eight times in 12 months. Velocity stopped being a feeling. It is a release schedule. The price collapsed underneath all of it. What cost $100 of inference in early 2023 cost about $0.25 in 2026. The first frontier API charged $60 per million tokens; the 2026 leaders sit at $1.25. Training compute grew roughly 16,000x in five years, and each 10x of compute has been worth about 3x of measured performance. The newest generation of accelerator cut the cost per token by 10x on its own. Cheap does not mean careful. It means everywhere at once.

The exam stopped separating the field because the field had outgrown the exam.

Here is the arithmetic nobody in the room wants to do. New capability emerges every 3 to 4 weeks. Policy response takes 2 to 5 years. Curricula take 2 to 4 years to change. Corporate workforce planning runs on 1 to 2 years. Retraining trails adoption by 18 to 24 months. GPT-3 shipped in June 2020. Meaningful policy attention arrived in 2024. Enforcement, as of 2026, is still mostly a draft document. ChatGPT shipped in November 2022, and by 2026 more than 80% of students were using it while most universities were still writing the policy about it. The probability that a framework drafted for last year's models is still correct when it is finally enforced is not zero. It is close to zero. Almost everyone is pricing it at zero, which is the mistake. The invoice goes to whoever waited for permission.

I have signed a EUR 20 million personal guarantee. The company folded on Christmas Day 2008. I lost savings in the 2013 Cyprus bail-in. Both events were considered improbable by the people who were paid to consider them. What saved me later was not a better forecast. It was pricing. In 2008 I restructured a EUR 75 million industrial group, and the only reason it worked is that we stopped trying to predict which quarter the money would stop and started asking what we could survive. We deployed 210 energy systems across Africa and Asia on the same logic. Nobody could say which grid would fail, so every unit had to work without one. So here is what I do with a curve like this one. I read release notes the way other people read the news, because the release notes are the news now. I ask what breaks in 6 months, not what improves. I keep two things that have no single point of failure, one of which is judgment in a room full of people who are certain. I assume January's tooling will be obsolete by July and I refuse to build anything that cannot be rebuilt in a weekend. Most people are still waiting for the official announcement. There will not be one. The announcement was 30 November 2022. It took eleven minutes to reach me. It has not slowed down since.