Meta’s New Policy: Limiting Release of Highly Capable AI System

Meta, the parent company of Facebook, Instagram, and WhatsApp, has proposed a new policy document that suggests it may not release its internally developed highly capable AI system under certain circumstances. This move comes as Meta CEO Mark Zuckerberg has talked about making artificial general intelligence (AGI) openly available to everyone in the near future.

The Frontier AI Framework

In its policy document, called the Frontier AI Framework, Meta has discussed two types of AI systems: "high risk" and "critical risk." These systems have the capability to aid in cybersecurity, biological, and chemical attacks. The major difference between the two is that "critical-risk" systems may result in "catastrophic outcomes" that cannot be mitigated in a proposed deployment context.

Types of Attacks

Although the list of possible attacks in the document is lengthy, Meta has highlighted some examples, including the proliferation of high-impact biological weapons and the automated end-to-end compromise of a best-practice-protected corporate-scale environment. The company said that its list includes the risks that are "most urgent," which could arise due to the availability of a powerful AI system.

Classification of Risks

According to the company’s document, Meta has classified these risks based on inputs from internal and external researchers. These are said to be subject to review by "senior-level decision-makers." Once a system is determined as high-risk, Meta stated that it will limit its access internally and will not unveil it to the public until the time mitigations to "reduce risk to moderate levels" are implemented.

Critical-Risk Systems

For critical-risk systems, the company will be implementing unspecified security protections to prevent it from being exfiltrated. Also, its development will be put on halt until it gets to a less dangerous level. Meta says that it doesn’t believe the science of evaluation is "sufficiently robust as to provide definitive quantitative metrics" for deciding a system’s riskiness.

FAQ

1. What is Meta’s family of AI models?

It is called Llama, which has received hundreds of millions of downloads.

2. How is AGI?

AGI stands for Artificial General Intelligence, which is referred to hypothetical intelligence of a machine that has the capability to understand and learn intellectual tasks just like human beings.

Published On

Published On Feb 5, 2025 at 09:50 AM IST

Category

Category Most Read Placement