California’s AI Bill: Rumors, Conspiracy, and Discussions

Last week's news about California’s new AI bill sparked many discussions in the AI community about the future of AI development. The bill aims to ensure responsible AI development but imposes significant restrictions on large language models. Critics argue that it introduces numerous bureaucratic hurdles that could particularly hurt startups. For instance, the bill requires ongoing annual reviews, which add significant technical and financial burdens.

Key points of discussion within the AI community include:

The bill requires models trained with over 10^26 FLOPs to:

• Not be fine-tunable to create chemical or biological weapons.
• Have an immediate shutdown button.
• Involve significant paperwork and reporting to the government.

Key Points of Discussion: Non Fine-Tunable Models and Immediate Shutdown Button

There has been a lot of talk about making models “non fine-tunable” and having an “immediate shutdown button.” One proposed solution is to keep the model locked up and only allow access through an API, with strict controls on fine-tuning. But what about open weight models? Once an open LLM is online, it's almost impossible to remove or control it. The bill, while targeting both closed-source and open-source models, effectively outlaws open weights models. Some people wonder if this means open weights models will be banned altogether and joke that Americans might start torrenting AI models in the future 🙃.

Some professionals think the bill creates big challenges for smaller competitors, helping big tech companies instead. They say the bill's rules are hard for startups and smaller companies to follow, giving an edge to big tech giants. The bill also spreads fear and doubt by making AI research seem dangerous, which discourages innovation. It adds risks and requirements that are tough for universities and independent researchers to handle, including those related to safety.

Critics also believe the bill sets the stage for more regulations in the future, making AI development even harder. They think the real goal of the bill is to create government rules that make it tough for startups and individual innovators, while big companies have no problem complying. Many users feel the bill looks like an attempt to ban local startups entirely. While it's technically possible to release a big open-source model, the work and rules needed would likely stop most developers from trying.

Moving from California?

There is speculation that more companies may need to move out of California to continue their work. This has happened before when dealing with strict regulations from California legislators. Many worry that this law might just push AI players to relocate. States like Washington, New York, Texas, and Florida might be perfectly happy to see California legislate itself out of the AI boom.

Conspiracy Theories

Some believe that the bill seems to favor large AI companies like OpenAI. Users argue that OpenAI has effectively "won" in California. They believe this is a strategic move to control AI training by making it prohibitively expensive and difficult for smaller entities. The theory goes that by cost-gating compute resources and keeping consumer-grade hardware underpowered, big companies can dictate who has access to AI training tools. This restriction would force countries deemed "unsafe" to go through costly import processes with high markups for necessary hardware. A billion-dollar company like OpenAI can easily navigate the bureaucracy required by the bill. However, a small startup that could potentially compete with them might not be able to handle these burdens, effectively stifling competition before it can emerge.

Not everyone agrees with this perspective. A major industry group, The AI Alliance, which includes Meta, IBM, AMD, the Linux Foundation, and others, is also against the bill. This suggests that if the bill were a plot by OpenAI, it might not be working as intended. The bill has united both AI enthusiasts and a significant portion of the tech industry in opposition, indicating widespread concerns about its implications for innovation and development.

And What About Worldwide Competition?

There is growing concern about how the bill will impact the United States' position in the global AI race. Some believe that countries without similar regulations will see this as a win for their AI development efforts. Users argue that while China and other major countries may currently lag behind the US in AI development, the gap is not insurmountable. If the bill stifles future releases of AI models in the US, the AI community might shift its focus to fine-tuning Chinese models :). Chinese models, with their multilingual capabilities, offer strong coding potential. Critics worry that the US, by imposing these restrictions, might lose its competitive edge in AI to countries with fewer regulations.

While AI professionals are worried about the new AI bill's potential impact, it's important to note that it has not yet become law. The bill has only been approved by the California State Senate and still needs to be passed by the State Assembly and signed by the Governor. Until then, there's still time for discussion and potential changes to address the AI community's concerns.