Mark Zuckerberg: AI should not be controlled by a few, but the US government should have first access to the model

According to Dynamic Beating monitoring, Meta CEO Mark Zuckerberg believes that one of the biggest risks of AI is the concentration of superhuman capabilities in the hands of a few companies and governments. Instead of creating a "super AI" to decide what is right for everyone, he suggests dispersing the capabilities to everyone, allowing different individuals and models to check and balance each other.

This also extends to who the models should listen to. Zuckerberg criticizes the current "alignment" of many AIs, which actually align with the values of the model companies. He argues that personal agents should align with the user's own goals and values, retaining only the necessary legal and safety guardrails.

However, when it comes to national security, he provides a loophole for governments: while cutting-edge models are still in training, labs should provide intermediate versions to the US government and have engineers collaborate to test them, allowing the government to access the latest capabilities early and strengthen critical infrastructure. The condition is that this should not delay the models' public release.