Benjamin Todd has checked his own AI predictions against what happened, and the verdict is that the “overhyped” crowd had the direction of the mistake backwards. In an essay on his Substack titled “AI wasn’t overhyped. Almost everyone underestimated it.”, Todd argues that over the last 18 months nearly every group of forecasters expected slower progress than the world delivered. He counts himself among them.
His ranking runs from worst to best. Superforecasters, the people with the strongest record on geopolitical questions, were furthest off, and the general public was close behind. AI experts did better but still came in slow. Only the heads of AI companies leaned optimistic, and only slightly. Todd’s rule of thumb: take what a lab leader says will happen in six months and expect it in six to 18.
The scorecard for his own March 2025 piece, “Will we have AGI by 2030?”, is mixed. Three of his formal forecasts landed inside his stated range and two missed, both by a wide margin. He predicted 32 percent on the FrontierMath benchmark by the end of 2026. AI reached 40 percent at the close of 2025 and 94 percent by September 2026, years ahead of his schedule. On METR’s time-horizon measure, most forecasters expected three hours before 2026 ended, and the mark fell in April.
The money followed the same pattern. Superforecasters put OpenAI and Anthropic at $90 billion in annualized revenue combined by year-end, and the figure was already $100 billion in July. Even the AI 2027 scenario, written as an aggressive takeoff case, had the largest AI company earning $38 billion annualized that September, while Anthropic was already near $65 billion. Revenue is the one area where Todd thinks he may have run too hot, with a median of $225 billion for frontier revenue.
Todd’s sharpest example comes from researchers themselves. A December 2024 survey of more than 1,000 people who had published at top AI conferences put the median date for AI solving an open maths problem on the scale of a Millennium problem at about 2054. Todd says a Millennium Prize problem appears to have fallen to AI in September 2026. He also notes the result is not formally validated, and that could take years. His own estimate for such a result by December 2027 was 7 percent.
Todd offers a guess for why the best forecasters did worst: that assuming things stay on trend works well for geopolitical events, while AI keeps breaking its own trends. His 2022 example: superforecasters expected the largest AI training run to cost $35 million by 2024, and it was probably already above $100 million then.
The lab leaders get a more mixed grade than the headline suggests. Sam Altman said in January 2025 that agents would join the workforce that year, and they arrived in earnest in early 2026. Dario Amodei said in March 2025 that AI would write 90 percent of code within three to six months, and Anthropic’s own codebase reached 80 percent after 14 months. OpenAI says the intern-level automated researcher Altman promised for September 2026 arrived on time, though the company graded itself.
Todd’s own median for AI that outperforms humans at everything is around 2030. The forecasters he rates highest cluster around 2028 to 2029, and he thinks they may be right. In his view the gap hardly changes what anyone should do.
For anyone setting a roadmap, the practical cost of filing vendor timelines under “hype” is now measurable: by Todd’s count, the people who took the labs at their word and added a year came out ahead of everyone who discounted them.
Reported by Benjamin Todd on his Substack, in the essay “AI wasn’t overhyped. Almost everyone underestimated it.”
