GoBench

GoBench measures how well frontier language models play 9×9 Go against a calibrated ladder of KataGo opponents.

Why GoBench?

We believe GoBench can measure progress toward AGI. There are three ways to saturate the benchmark, each of which could advance that progress:

1. Transfer from GoBench optimizations

   Researchers may develop novel architectures or learning algorithms while optimizing for GoBench. Those ideas could transfer to other tasks.

2. Lower-cost learning

   As the cost of learning a new domain decreases, an AI may learn Go with less human supervision. Given a sufficient budget, it could then attempt to learn many economically valuable domains.

3. Improved general reasoning

   Stronger general reasoning could improve Go performance without the model having previously learned Go. If general reasoning alone can saturate GoBench, it could likely support much economically valuable work.