Gary Marcus Comments on Anthropic's Mythos Announcement and Its Limitations

Anthropic announced Mythos, a new AI model or system, on the previous day. Gary Marcus, an AI researcher, provided commentary on the announcement via social media. He described the event as not indicating a major qualitative improvement in AI capabilities.

In the demonstration, Anthropic turned off sandboxing, a security feature that isolates AI operations. Marcus reported that this change meant the test did not demonstrate much about performance in real-world conditions. Sandboxing typically prevents AI from accessing external systems or causing unintended effects.

Marcus highlighted that inexpensive open-weight models, which are publicly available and modifiable, can already perform some tasks similar to those shown in the Mythos demo. Open-weight models refer to AI systems where the weights, or internal parameters, are openly shared. This availability reduces the novelty of Anthropic's presentation.

Anthropic is an AI research company focused on developing safe and reliable artificial intelligence systems.

The company has released several models, including Claude, which emphasize alignment with human values. Mythos appears to build on these efforts, though specific technical details were not elaborated in Marcus's post. The announcement occurred amid ongoing competition in the AI sector, where companies race to improve model capabilities for tasks like reasoning and interaction.

Stakeholders include AI developers, researchers, and users who rely on these technologies for applications in business, research, and daily tasks. Marcus's critique underscores debates about the pace and substance of AI progress.

evidence was presented in Marcus's statement to suggest Mythos represents a major qualitative jump over existing models.