Abstract:

The open-source versus closed-source debate surrounding large language models (LLMs) is a significant point of contention, primarily focused on issues of efficiency, security, and principles of equality and democratization. The motivation for open-sourcing LLMs stems mainly from the unique market advantages and economic benefits they offer, as well as the idealistic pursuit of developers who advocate for technological equality and democracy. However, open-sourcing LLMs also pose several challenges, including economic benefit, societal acceleration, and social security challenges. Addressing the latter two challenges requires

human society to respond in two ways: firstly, by using appropriate levels of closed-source strategies to buffer the speed of technological diffusion, and secondly, by aligning technology to ensure large models are constrained by societal rules. Alignment of LLMs should be holistic, encompassing graded alignment, dual alignment between humans and LLMs, and alignment across the entire production process of LLMs. The mechanisms for achieving this include transparency mechanisms and negotiation mechanisms.

Key words: LLMs, open source, technological diffusion, holistic alignment, GPT