Safety First Cognitive Architectures @ Intelligent Cooperation Workshop

Brendon Wong discusses safety-first cognitive architectures and their relevance to security. These architectures are designed with AI safety in mind, ensuring interpretable and corrigible goals, plans, and actions. Different architectures, like Conjecture and Eric Drexler’s Open Agency Model, have varying levels of similarity to models like Auto GPT. Security implications include sandboxing and preventing influence between system components. Cognitive architectures restrict AI models’ access to necessary information and actions. Challenges and solutions depend on the design and underlying AI models. For Wong, as technology advances, it is important to reevaluate safety measures and develop new ones.