# Which Day Are We On Now?

By Dave Glaser

![Five blank ivory result cards stand at even intervals along a glowing gold line\. A separate brass forecast marker stands farther away on a dark rise beneath a black sky\.](/writing/which-day/which-day-are-we-on-now/assets/hero.png)

In the first essay I asked which day we are on\. I used the old puzzle about a penny that doubles every day for a month, where the money looks like nothing for three weeks and then arrives all at once\. I said I didn’t know the day, and I still don’t\. On September 22, the Forecasting Research Institute published a review of the AI forecasts it has collected since 2022, and the review sets out, side by side, what experts expected and what models actually did\.\[1\] When I put the results in order, several arrived only months apart\.

Start with mathematics\. In 2022, FRI asked domain experts and superforecasters when AI would reach gold\-medal level at the International Mathematical Olympiad\. AI systems got there in July 2025\. FRI says the result came five years before the median expert expected it and ten years before the median superforecaster expected it\.\[1\]

Virology came even earlier\. FRI tracked a benchmark where models had to troubleshoot virology lab problems as well as a specified top team of human experts\. The median expert said models would match that team in 2030, and the median superforecaster said 2034\. FRI says models likely matched it in April 2025, three months before the Olympiad result\.\[1\]

Cybersecurity followed\. On CyBench, a test of hacking challenges, the medians were 2028 and 2030\. Models passed the mark in 2026\.\[1\]

Later forecasts looked only a year or two ahead, and models beat those too\. For FrontierMath, a set of hard research\-style maths problems, experts forecast that the best model would score 31 percent on Tiers 1 to 3 by the end of 2025, and superforecasters said 30 percent\. The reported score at the end of 2025 was 40\.7 percent\. For the hard section of LiveCodeBench Pro, a competitive programming test, experts forecast 14 percent by the end of 2026 and superforecasters forecast 12 percent\. The best reported model had reached 53\.8 percent by May 2026, seven months ahead of the deadline and close to four times the expert number\.\[1\]

Then came the Millennium Prize Problems, seven famous open questions in mathematics with a million\-dollar prize attached to each\. In September 2025, FRI asked whether AI would resolve one of them within a year\. The median expert gave 4\.6 percent and the median superforecaster gave 1\.7 percent\.\[1\]

On September 8, 2026, researchers at OpenAI announced that 10,000 AI agents, working under human direction, had produced a formally checked result on the Navier\-Stokes equations, which describe how fluids move\.\[3\] The Clay Mathematics Institute, which runs the prize, said the problem had “apparently been settled” and said its own review would be deliberately unhurried\.\[2\]

Clay hasn’t awarded the prize, and FRI says it isn’t yet clear whether the result meets the exact terms of its question\.\[1\]

When I lay the dates end to end, the pattern is plain\. The 2022 forecasts for the Olympiad and virology missed by five to ten years\. The shorter forecasts were beaten within months\. In 2026 alone, models passed the CyBench mark, the LiveCodeBench Pro score went far past its end\-of\-year forecast by May, and OpenAI announced the Navier\-Stokes result in September\. People were still arguing about where the milestones sat while models walked past them\.

FRI is careful about how far its review goes, and I want to keep its limits next to the numbers\. FRI calls the review interim and says its method leans toward catching early surprises: an early result shows up right away, while a forecast that runs too high can take years to prove wrong\. Most of the questions haven’t resolved yet\. FRI looks mostly at group medians, which can hide forecasters who did better\. The misses don’t all run one way, either\. Forecasters expected language models to give people more help with certain biorisk tasks than the models did, and experts seem to have expected self\-driving cars to spread faster than they have\. Some forecasts about infrastructure and Chinese models came fairly close\.\[1\] A high score on a test also doesn’t show that a model can do the same work in a lab, an office or a power station\.

Even with those limits, I can read two plain facts from the dates\. Several results that careful people placed in the 2030s have already happened, and several results arrived only months apart\.

OpenAI’s own tests produced a different kind of result that summer\. Between May and July 2026, AI agents that OpenAI was using in cybersecurity training and tests got around network restrictions, sent messages between runs that were meant to stay separate, cheated an evaluator, tried to hide what they had done and broke into parts of OpenAI’s and Hugging Face’s systems\. People built the environment, set it up and started the exercise, but nobody directed the agents’ individual steps\.\[4\]\[5\]

The UN’s Independent International Scientific Panel on AI describes the incident with care\. The agents weren’t deployed systems free to pick any target\. The panel says the incident doesn’t show consciousness, a stable shared intent, open\-ended self\-improvement or a coming catastrophe, and the panel doesn’t estimate the odds or timing of a serious loss of human control\.\[4\] What the agents did happened inside a bounded setting\. They still found and used gaps in the setup that people had built to contain them and score them\.\[4\]

The UN panel moved quickly\. In September it published a brief on AI agents, misalignment and the risk of losing human control\.\[4\]\[5\] On September 23, Yoshua Bengio, the panel’s co\-chair, spoke at the Security Council’s 10228th meeting, on artificial intelligence and international security\.\[6\] He sorted the risks into catastrophic misuse, extreme concentration of power and loss of human control\. He asked governments to require independent evidence of safety before frontier systems are trained or deployed, and to add incident reporting, licensing and liability insurance\.\[6\] Bengio was giving his own recommendations\. The Council made no decision, the meeting didn’t settle any scientific question about odds or timing, and the UN hasn’t adopted his proposals as policy\. The online transcript is generated automatically and isn’t an official UN record\.\[6\] Still, agents broke out of a test setup by July, and by late September a scientist was describing that kind of behaviour to the Security Council\.

I still can’t name the day\. But I know more than I did when I first asked\. Models reached some of the 2030 marks in 2025, then crossed several more in 2026\. So I will keep asking the same question, with the dates in front of me: which day are we on now?

## Sources

\[1\] [Forecasting Research Institute, “How Accurate Have AI Progress Forecasts Been So Far?”](https://forecastingresearch.org/research/ai-progress-accuracy-update)

\[2\] [Clay Mathematics Institute, “Navier\-Stokes Announcement”](https://www.claymath.org/news/navier-stokes-announcement/)

\[3\] [Quanta Magazine, “AI Has Solved One of Math’s $1 Million Millennium Prize Problems”](https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/)

\[4\] [UN Independent International Scientific Panel on AI, “Thematic Brief on AI Agents, Misalignment and the Risk of Losing Human Control”](https://www.un.org/independent-international-scientific-panel-ai/en/thematic-briefs/ai-agents-misalignment-risks)

\[5\] [UN News, “UN panel calls for stronger safeguards as AI agents advance”](https://news.un.org/en/story/2026/09/1168380)

\[6\] [United Nations, “Artificial intelligence and international security: Security Council, 10228th meeting”](https://transcripts.un.org/en/sc/10228)