A Puzzle for the Ages
At the heart of this story is the cap set problem, a famous puzzle in a field of mathematics called combinatorics. Imagine a high-dimensional grid, like an impossibly complex chessboard. The challenge is to place as many dots as you can on this grid with
one simple rule: no three dots can ever form a straight line. It sounds simple, but as the grid gets bigger, the problem becomes astronomically difficult. Finding the largest possible “cap set” has stumped mathematicians for decades. This isn't just an abstract game; these types of problems are fundamental models for other complex questions in cryptography, coding theory, and computer science.
The Limits of Human Intuition
Why is the cap set problem so hard? The main issue is the sheer number of possibilities. As you add more dimensions and points, the potential arrangements explode into a number so vast it’s larger than the number of atoms in the universe. A brute-force approach, where a computer simply tries every combination, is completely useless. Human mathematicians have relied on intuition and insight to find clever shortcuts, but progress has been slow and incremental. For twenty years, the record for the largest known cap sets in certain dimensions remained unbroken, a testament to the wall researchers had hit. They needed a new way to think, one that could explore bizarre, non-obvious paths that a human mind might dismiss too quickly.
Enter the AI Collaborator: FunSearch
The breakthrough came from an AI system developed by Google DeepMind called FunSearch. It’s not an all-knowing oracle that simply spits out answers. Instead, it’s designed to be a creative partner. FunSearch works by combining two key components. The first is a Large Language Model (LLM), similar to those that power chatbots, which has been trained on computer code. Its job is to be the “creative engine,” generating thousands of small computer programs that represent potential strategies to solve the problem. The second component is a rigorous “evaluator.” This automated system's sole job is to check the AI's work, immediately discarding incorrect or nonsensical ideas. This setup allows the AI to have wild, creative thoughts without the risk of producing “hallucinations” or flawed logic. The best, most promising ideas are kept and fed back into the LLM, which then builds upon them in a process that mimics evolution.
A New Twist Revealed in Code
The collaboration proved incredibly successful. FunSearch didn't just find a new record-breaking number; it generated computer code that described a completely new method for constructing large cap sets. In doing so, it discovered larger sets than were previously known, outperforming state-of-the-art computational methods designed specifically for this task. The key insight is that the output wasn't just an answer, but a recipe. Human mathematicians could inspect the code generated by FunSearch, understand the novel logic it had uncovered, and verify the discovery. The AI wasn't a black box; it was a colleague that showed its work, revealing a new piece of verifiable mathematical knowledge. While it didn't solve the cap set problem for all time, it shattered a long-standing barrier.
The Dawn of a New Scientific Era
This achievement is far more significant than just one puzzle. It signals a new era of scientific discovery powered by human-AI partnerships. The FunSearch system has also been used to find more efficient solutions for the “bin packing problem,” a logistical challenge with real-world applications in optimizing everything from shipping to data center efficiency. And this is not an isolated incident. In recent months, other AI models have reportedly made breakthroughs on decades-old problems, such as the “unit distance problem” posed in the 1940s and a physics puzzle related to how materials “jam” from a fluid to a solid state. This trend suggests that AI is evolving from a tool that just processes data into a genuine collaborator that can aid in conceptual breakthroughs.
