Go AI Is No Longer a "Black Box"

In 2016, AlphaGO rose to fame, and since then, AI has made remarkable progress in playing strength, efficiency, and versatility. However, its specific reasoning process remains a "black box"—even when it can output win-rate evaluations and move probabilities, it still cannot explain "why a particular move is better" in human language.

May 23, Shanghai Artificial Intelligence Laboratory (Shanghai AI Lab) announced an advanced version of its reasoning Large Model, InternThinker, which not only obtaines professional-level Go skill，but also can demonstrate a transparent chain of thought.

The new - generation InternThinker has achieved breakthroughs in Go tasks—not only demonstrating strong professional-level performance ,but also becoming the first large model to break the "black box" of AI. It can explain the process of playing chess in natural language.

In the fourth game of the match between Lee Sedol and AlphaGo, Lee Sedol played the 78th move at L11. This move, which was called the "divine move" by Gu Li, directly reversed the situation and helped Lee Sedol win the game. In the reproduction of this famous game by researchers, InternThinker evaluated this move as "quite tricky... This move perfectly resolved the threat at L11, re - established control in the center, and laid the groundwork for subsequent attacks." Then it gave the response strategy of playing at L10.

You can click the link to experience InternThinker: https://chat.intern-ai.org.cn/

InternThinker's powerful reasoning capabilities and breakthroughs in Go tasks benefit from its innovative training environment. For complex logical reasoning tasks, accurately obtaining feedback on processes and results is particularly critical. To this end, researchers have built a large-scale, standardized, and scalable interactive verification environment called InternBootcamp — this is equivalent to creating an "accelerated training camp" for the model, enabling it to efficiently acquire professional skills and "grow" rapidly.

The Interaction Process between InternBootCamp and Large Language Models Built on automated code agent construction, InternBootcamp encompasses over 1,000 verification environments that cover a wide range of complex logical reasoning tasks. It effectively assists researchers in the field of large models to conduct explorations based on reinforcement learning. InternBootcamp can generate reasoning tasks with controllable difficulty in a batch and standardized manner, such as Sudoku, decoding games, Go, and scientific tasks, and interact with large models to provide feedback. Through the large-scale construction and mixed training of different professional knowledge, it enables large models to break free from the cumbersome mode of obtaining questions and answers based on data annotation and avoid the deception of traditional reward models, thus achieving a new paradigm for enhancing the reasoning ability of large models.

In addition to Go, InternThinker has also delivered outstanding performance in other tasks. Through mixed reinforcement learning across multiple tasks, InternThinker's average capability on a test suite comprising dozens of tasks exceeds mainstream domestic and international reasoning models such as o3-mini, DeepSeek-R1, and Claude-3.7-Sonnet.

Even in some tasks, its performance far exceeds that of other current reasoning large models.

The open-source link of InternBootcamp: https://github.com/InternLM/InternBootcamp