The AI Bill in California, USA, Sparks Controversy

Recently, American AI startups, tech giants, and the academic community jointly opposed the California security bill SB-1047 (The Frontier Artificial Intelligence Model Safety Innovation Act), arguing that the bill would cool AI innovation. However, supporters believe that there is less regulation on AI than on barber shops. The bill has been submitted to the state legislature and passed by the California Senate and two state assembly committees with almost no opposition. The main reason for the controversy over this bill is that it may impose unnecessary penalties on AI developers, stifle innovation, and hinder academic research. The specific details are as follows:

Bill Content:

Scope of Application: This bill targets intelligent systems with more than 10^26 floating-point operations per second (FLOPs) and a training cost of $100 million, as well as any fine-tuning based on the above AI models with no less than 3 times 10^25 FLOPs. It encompasses almost all mainstream large models currently on the market.

Main Contents: The main contents include requiring developers to conduct pre-deployment safety assessments, implement strong cybersecurity protection, have the ability to emergency shut down the model, submit annual compliance certifications, report security incidents within 72 hours, etc. Additionally, developers are required to bear legal responsibility for the downstream use or modification of their models. Before starting training, developers need to ensure that their models will not be used or provide "dangerous capabilities", and implement a series of safeguard measures to prevent unauthorized access or abuse, and a "written guarantee" must also be provided.

Supervisory Body: A newly established "Frontier Model Division" will be responsible for supervising the implementation of the new regulations. It will be under the California Department of Technology and funded by fees and fines collected from developers. Making false representations about the model's capabilities to this agency could result in imprisonment for perjury. At the same time, the terms explicitly state "encouraging whistleblowing", calling on internal employees of AI companies to report non-compliant behaviors of their employers without fear of retaliation.

Views of Opponents:

Yann LeCun, Chief AI Scientist of Meta and Father of Deep Learning: Believes that the "joint liability clause of this bill will pose a huge risk to open-source AI platforms... Meta won't be affected, but AI startups will go bankrupt.

"AI Godmother" Fei-Fei Li: Although this bill is well-intentioned, it not only fails to achieve its purpose but also deeply harms the academic field of artificial intelligence and the open-source community. The bill will unnecessarily punish developers, stifle the open-source community, and hinder academic AI research, while failing to solve the actual problems it aims to address.

Andreessen Horowitz, Founding Partner of Investment Firm a16z: California's anti-AI bill, although well-intentioned, may weaken the US tech industry due to misguidance. At a critical juncture for the future of technology. The United States needs leaders to recognize that now is a critical moment to take wise and unified AI regulatory actions.

Andrew Ng, Visiting Professor of Computer Science and Electrical Engineering at Stanford University: The California SB-1047 bill in the United States will stifle the development of open-source large models. What should be regulated are AI applications rather than the large models themselves.

Views of Supporters:

Computer science researchers such as Geoffrey Hinton and Yoshua Bengio, as well as legal scholars: They consider the SB-1047 bill to be "the minimum for effective regulation of this technology". Repealing the basic measures of this bill would be a historic mistake, and this mistake will become more apparent within a year when the next generation of more powerful AI systems is released.