AI Learns Battleship, Cuts Research Costs by 99%

Imagine if artificial intelligence could help scientists make smarter decisions while spending 99% less money. That's exactly what researchers discovered by teaching AI models to play Battleship.

The challenge was simple but profound. Scientists constantly face tough choices about which experiments to run and which theories to test, especially when time and money are tight. Every decision shapes the path of discovery.

So researchers at OpenAI and partner institutions created a twist on the classic board game. Two players worked together, one asking questions about ship locations while the other answered, racing to sink all the vessels in the fewest moves possible. They pitted 42 human players against various AI models to see who could strategize best.

At first, humans crushed Meta's Llama-4-Scout model every time. OpenAI's GPT-5 reasoning model performed better than both humans and Llama-4-Scout initially.

But then the team made their breakthrough. They optimized Llama-4-Scout using principles from Bayesian experimental design, teaching it to ask questions that maximized information gained and to think one move ahead. They also discovered that when players communicated with code snippets instead of natural language, accuracy soared.

The results stunned everyone. The upgraded Llama-4-Scout beat GPT-5 in two-thirds of games while costing just one-hundredth as much to run. It also defeated human players by an average of seven moves.

The Ripple Effect

This isn't just about winning board games. The same strategies AI used in Battleship can guide real scientific decisions, helping researchers choose which hypotheses to pursue when exploring vast possibilities.

Yuanqi Du, an AI researcher at Cornell University who wasn't involved in the study, called the framework "very useful to measure whether language models are really making progress" in scientific decision-making. Understanding the full space of possible hypotheses remains one of the hardest challenges in research.

Lead researcher Valerio Pepe acknowledges that Battleship is far simpler than actual scientific problems. Chemical and biological samples can't be read as clearly as game boards. But the methods show promise for making research more efficient and accessible.

The findings were presented in April at the International Conference on Learning Representations, sparking excitement about AI's potential to democratize scientific discovery by reducing costs.

When AI can help scientists make smarter choices with fewer resources, breakthroughs that once seemed impossibly expensive might suddenly come within reach.