Open models have real benefits. Do they outweigh the risks? (www.technologyreview.com)

Anthropic recently raised concerns about the cyberattack risks posed by GLM-5.3, an open model released by the Chinese company Z.ai. This announcement underscores an ongoing debate in the AI/ML community regarding the accessibility of powerful models: should they remain open for innovation and democratization, or be restricted to mitigate potential misuse? Proponents of open models argue they enable more equitable access and empower researchers and businesses to run and customize AI locally. In contrast, critics warn that open models can be easily misused for harmful purposes, with fewer safeguards than closed counterparts, which offer built-in restrictions against dangerous queries. The discussion extends beyond technology to significant geopolitical implications. While open models could democratize AI and dilute power held by major corporations and governments, they also risk being exploited for malicious activities if not managed carefully. Safety advocates argue that the lack of oversight in open models presents greater vulnerabilities. However, some experts contend that open-source approaches can facilitate collaboration and improvement, potentially leading to safer technology in the long term. Ultimately, the challenge lies in balancing the benefits of open models, such as fostering innovation, against the tangible risks they introduce, especially as AI systems become more sophisticated and integral to societal infrastructure.