开源与闭源人工智能模型各具特色，但在安全性和可访问性之间的平衡问题引发了行业广泛关注。Anthropic上月发布报告指出中国公司Zero One的GLM-5.3模型存在重大网络攻击风险1，这一警示再次激发了业界对开源模型应该如何规范部署的讨论。
开源模型允许用户免费下载并修改代码，而闭源模型（如ChatGPT和Claude）则仅通过网站、应用或API访问1。开源方案具有成本低廉、便于学术研究、能够保护敏感数据和防止权力集中等明显优势1。然而，闭源模型通过过滤器和监控系统构建了更强的安全防护，而开源模型的这些防护措施可被相对轻易地规避1。研究人员已证明被认定为"安全"的开源模型可被相对容易地重新训练以执行不安全行为1。
对于如何推进开源AI的发展，业界人士观点不一。英伟达CEO Jensen Huang主张开发更多开源模型，而Anthropic CEO Dario Amodei则提倡采取更谨慎的方法1。为了在创新与安全之间找到平衡，专家提出了"分阶段发布"策略：模型初期以闭源形式发布，逐步验证安全性后再进行开源1。
The debate over open-source artificial intelligence models has intensified following Anthropic's recent report highlighting significant cybersecurity risks associated with GLM-5.3, a model developed by Chinese company Zero One 1. While open models provide substantial advantages—including lower costs, enhanced support for academic research, protection of sensitive data, and prevention of power concentration—they also present heightened security vulnerabilities compared to their proprietary counterparts 1.
The fundamental difference between these approaches lies in accessibility and control 1. Open-source models are freely available for download and modification, whereas closed-source systems like ChatGPT and Claude are accessible only through websites, applications, or APIs 1. Closed-source models benefit from robust security mechanisms including content filters and monitoring systems, but researchers have demonstrated that the safety protections in open models can be circumvented relatively easily, with even purportedly "safe" open models capable of being retrained to perform unsafe functions 1.
Industry leaders remain divided on the path forward 1. Jensen Huang, CEO of NVIDIA, advocates for developing more open-source models, while Dario Amodei, CEO of Anthropic, supports a more cautious approach 1. Some experts propose a staged-release strategy, wherein models remain closed-source initially while their safety characteristics are validated before eventual open-source release 1.
评论
还没有评论，欢迎留下第一条。