MTS discusses Gwern Branwen's 2020 scaling hypothesis and the argument that increasing computing resources could produce more capable AI systems. The presenters describe the essay's influence as their assessment.
The discussion turns to Gwern Branwen's proposal for personal AI agents that reflect an individual user's values and preferencesA personal AI agent uses a person's authorized goals, preferences, routines, context, and tools to provide continuing individualized assistance.. The presenters explain his concern that the incentives of organizations operating frontier models may differ from those of their users.
The speakers consider limitations in model memory, the distinction between personalization and alignmentAlignment aims to make an AI system's behaviour consistent with intended human goals, values and constraints., and possible risks and safeguards for agents acting on a user's behalf.
Watch on YouTube
