Terence Tao is Stunned: Google’s "AI Edison" Breaks Three 18-Year-Old Math Records
In May, Google introduced AlphaEvolve, an AI system that solved a mathematical problem that had stood unsolved for 18 years—in just 30 days. This breakthrough challenges the traditional notion that scientific research depends heavily on human inspiration and intuition. Instead, it suggests that AI might soon become the primary engine driving complex problem-solving.
AlphaEvolve Solves Long-Standing Problems Quickly
AlphaEvolve, built on Google’s Gemini model, discovers new algorithms by searching vast programming spaces. Working alongside humans, it cracked a problem untouched for nearly two decades—not once but three times. This feat caught the attention of renowned mathematician Terence Tao, who expressed surprise at the AI’s capabilities.
This is more than just an advancement in mathematics or computer science. AlphaEvolve represents a shift in how scientific research could be conducted, moving away from reliance on human intuition toward systematic, AI-driven discovery. It’s reminiscent of AlphaGo’s impact on the game of Go, where AI uncovered strategies never seen before.
How AlphaEvolve Works: Inspired by AlphaGo and AlphaTensor
DeepMind’s mission involves responsibly developing AI to benefit humanity, and AlphaEvolve is a milestone in this journey. The concept originated from AlphaGo’s ability to explore an enormous game space efficiently and select optimal moves beyond human capacity.
Building on this, DeepMind developed AlphaTensor, an AI that found more efficient algorithms for matrix multiplication—a problem assumed to have a cubic complexity for decades. AlphaTensor outperformed all known solutions, proving AI can surpass human algorithmic innovation.
AlphaEvolve expands this idea further. Unlike AlphaTensor, which is specialized, AlphaEvolve handles a broader range of problems by using evolutionary algorithms. It continuously searches for better solutions across diverse programming challenges, not limited to one specific task.
Evolution and Self-Improvement in AI
AlphaEvolve improves generation by generation. Each iteration builds on the strongest solutions from the previous round, maintaining diversity in approaches to explore the vast solution space more effectively.
The system adapts to the complexity of problems—simple problems get solved quickly, while complex ones require more iterations. Importantly, AlphaEvolve does not hit a performance ceiling early on, unlike many traditional optimization systems.
Predicting the exact number of generations needed for an optimal solution is difficult because problem difficulty varies widely. Yet, with continuous operation, AlphaEvolve steadily improves results over time.
Why AlphaEvolve Stands Apart from General Coding Agents
Unlike typical AI coding assistants that often stumble on complex or ambiguous tasks, AlphaEvolve leverages a strict evaluation function to judge solution validity. This allows it to propose creative algorithms while rigorously assessing their effectiveness.
For example, when optimizing data center scheduling, the evaluation function might simulate real-world performance. This feedback loop helps AlphaEvolve refine its solutions with precision. Designing these evaluation functions is challenging but crucial—they must clearly define success criteria and be adaptable across different problems.
The Changing Role of Scientists
With AI like AlphaEvolve taking on complex problem-solving, scientists’ roles are evolving. Future researchers will focus more on defining problems, crafting evaluation functions, and interpreting AI-generated results.
AI won’t just provide answers; it will deliver new algorithms that offer insights into the underlying principles of problems. This transparency is key for advancing scientific knowledge, as it allows experts to validate and build upon AI discoveries.
AlphaEvolve encourages scientists to rethink traditional approaches and cognitive frameworks. It signals a future where scientific progress is driven less by sudden inspiration and more by systematic intelligence.
For those interested in learning more about AI’s role in science and research, exploring Complete AI Training offers valuable resources on AI applications and development.
Your membership also unlocks:
