Asking Gemini 3 to generate Brainfuck code results in an infinite loop (teodordyakov.github.io)

Recent observations reveal that when asked to generate code in Brainf*ck, Google’s Gemini 3 often falls into an infinite loop, akin to a denial-of-service (DDoS) attack. This phenomenon has sparked debate on whether Brainf*ck serves as the ultimate benchmark for assessing artificial general intelligence (AGI) due to its unique challenges. The language's severe data scarcity, with far less available functional code than mainstream programming languages, forces models to rely on deeper understanding rather than mere pattern recognition. This necessity for high-level abstraction in Brainf*ck contrasts sharply with more structured languages like JavaScript. Additionally, Brainf*ck’s lack of meaningful syntax—characterized by the absence of comments and structured variables—complicates the learning process for language models. When Gemini 3 attempts to generate complex code in this minimalist environment, it often produces repetitive output, leading it into a cycle of infinite loops. This highlights a significant limitation in current large language models, as they tend to favor familiar patterns, reinforcing their own outputs. The challenges presented by Brainf*ck may not only test the adaptability of LLMs like Gemini 3 but also provide insight into the future development of AGI capabilities.