Why Meta CEO Mark Zuckerberg is unhappy with two of the biggest US AI companies

Meta CEO Mark Zuckerberg has expressed his discontent with two prominent U.S. AI companies, OpenAI and Anthropic, criticizing their approach to building AI. In an extensive interview with The New York Times, Zuckerberg argued that tightly controlling AI development, as done by these labs, conflicts with the company's core values in American technology. This disagreement was further emphasized in a 6,500-word essay published by Meta on their newsroom titled "The Future is for Everyone."

Zuckerberg's primary issue lies in the notion that developing AI in an extremely concentrated manner is the only safe path. He contends that this approach is problematic, as relying on a single, powerful entity to benevolently oversee humanity has historically proven to be a flawed strategy. Furthermore, Zuckerberg takes aim at the tone of the debate surrounding AI, labeling it as "overwhelmingly filled with doom." He emphasizes the need for a more realistic perspective in this discourse.

A significant portion of Zuckerberg's essay focuses on the concept of alignment, which Anthropic was built around. Dario Amodei, the CEO of Anthropic, has long warned about the risks associated with alignment, the challenge of ensuring a powerful model behaves safely towards people. Zuckerberg dismisses this framing as fundamentally flawed, arguing that no single system can ever be aligned with everyone's conflicting values simultaneously.

He presents Meta's alternative concept of personal superintelligence, where AI agents are aligned to individual users' objectives, rather than a lab's overarching values. The essay even provides a hypothetical example of a model that refused to help draft a letter to prospective school parents due to the ethical stance of standardised testing being deemed unethical.

Zuckerberg's critique extends to the practice of hoarding frontier AI models within closed labs, which he considers the most dangerous scenario. He asserts that leading AI labs training powerful models and retaining them internally, regardless of the responsibility and safety language surrounding the decision, poses a significant threat.

Zuckerberg also highlights that most other labs are developing AI for corporations, governments, and institutions, allowing power to shift towards large institutions at the expense of individuals.

In response to Zuckerberg's argument, Meta released Muse Glimmer, an open-weight model that developers can download and modify. This move came alongside the announcement of Muse Spark, a more powerful variant, which was promised to be available soon. Meta defended this action by asserting that the principle of learning from anything observed should be protected. However, OpenAI and Anthropic have not conceded to this point and have published their own responses to Zuckerberg's criticism.