Five years ago, researchers at OpenAI jumped up and down. Their model had just scored 55% on a math test for kids. The kids, aged 9 to 12, scored 60%. It was a failing grade, and they celebrated anyway. Ben Cohen tells the story in his Science of Success column in the Journal.
The test was simple word problems. Here's one, in our words. The thermometer read 2 degrees at bedtime. It fell 8 degrees overnight, then climbed 3 by morning. What did it read in the morning? The answer is minus 3. Five years ago the best AI got that kind of question wrong almost half the time.
Then the curve bent. In 2023 OpenAI's top model scored 700 of 800 on the SAT math section. In mid-2023 it solved 23% of the hardest high-school contest problems. By the end of 2024 it solved 95%. Then it aced the AIME, a test so hard that only nine of 8,586 students got a perfect score the year Mark Chen took it. Chen is now OpenAI's research chief. He was one of the nine.
The part people get wrong
In the past few weeks, Cohen reports, an internal OpenAI model solved one of the Millennium Prize Problems. Then it solved several hundred more open problems. One was the Unique Games Conjecture. Dana Moshkovitz, a computer scientist at the University of Texas, had worked on it for more than a decade. In August she floated the idea that AI might crack it one day. A colleague thought she was joking. Two months later it was done. Her 9-year-old son had the best line when she told him. He said they should solve cancer instead.
Chen's lesson is the one we'd underline. He says people are “really bad at internalizing exponentials,” even people inside OpenAI. A curve looks flat for years. Then it looks like a wall. Nobody feels it coming.
Our read
This is a Time Value of Money story, M3, in disguise. Your savings follow the same kind of curve. The first ten years feel flat. The money you add seems to matter more than the growth. Then the growth starts doing the heavy lifting, and it doesn't ask permission. The same Exchange section gives a real example. A $1,000 deposit in a child's account could grow to more than $5,800 in 18 years, using the stock market's 100-year average return. About $3,400 of that is real buying power after inflation. Nobody gets excited about $1,000. The curve is the exciting part, and it's the part we're worst at seeing.
So here's the one thing in your own plan to look at. Find your automatic contribution, the amount that goes in every paycheck or every month. That's the only part of the curve you control. The one check this week is whether it has gone up with your pay in the last two years. If it hasn't, raise it a notch before you forget. The AI researchers were surprised by their own curve. You don't have to be surprised by yours. The umbrella costs the same before the rain. So does a bigger contribution.