We’ve Been Asking the Wrong Question About AI, and This Google Paper Proves It

Forget what you know about AI agents. MLE-STAR uses a surprisingly simple strategy that makes it ridiculously good at building complex AI.

Photo by Jonny Gios on Unsplash

Ever heard of Kaggle?

For those who haven’t, it’s basically the Olympics for data scientists. It’s where the best in the world gather to compete on incredibly tough AI challenges, from identifying cancer in images to predicting stock prices. Winning a medal on Kaggle is a badge of honor. Becoming a Grandmaster is legendary.

Now, imagine you’re a Kaggle competitor. You spend weeks, maybe even months, cleaning data, trying different models, tweaking a million tiny parameters.. and for what? Just to climb a few spots on the leaderboard. It’s a brutal grind.

So, naturally, the AI world came up with a solution: AI agents that compete for you.

The idea is simple: You give an agent a Kaggle competition, and it spits out the code to win. Cool, right?

Well.. yes and no.

These first-generation agents are clever, but far from perfect. They know a lot, but their knowledge is stuck in the past, and they have some really bad habits.