MIT Announces Self-Updating AI Foundation 'SEAL'

A research team at Massachusetts Institute of Technology (MIT) has unveiled SEAL, a framework that enables large language models (LLMs) to rewrite their own parameters—the collection of numerical values that record internal knowledge and decision-making criteria. Conventionally, AI models have had their internal knowledge fixed once training is complete. SEAL transcends this constraint and provides a mechanism for models to continuously update themselves.

At the core of SEAL lies a technique called reinforcement learning. Reinforcement learning is a mechanism in which an AI receives feedback in the form of rewards or penalties based on the outcomes of its actions, and learns to take increasingly better actions. This approach has been widely used in game-playing AI and other applications. In SEAL, this mechanism is applied to the model's behavior of editing its own parameters. In other words, the model can autonomously learn through trial and error how to rewrite itself in ways that improve its outputs.

In conventional AI model operations, updating knowledge required human engineers to prepare new training data and perform retraining. This process is time-consuming and costly, making it difficult for models to keep pace with changes in reality. If SEAL becomes practical, it could substantially reduce the costs and delays associated with knowledge updates by allowing models to autonomously update themselves.

On the other hand, the capability for AI to rewrite its own parameters requires careful examination from the perspectives of safety and controllability. If unintended self-updates occur, the model's behavior could become unpredictable. Toward the practical implementation of this research, determining how to limit the scope of self-updates and ensure reliability will be a critical challenge.

MIT, as an institution at the forefront of AI research, has produced numerous achievements in the fields of model architecture and learning methodologies. SEAL represents research that challenges the previous assumption that trained models are static entities, and is attracting researchers' attention as it points toward new directions for future AI system design.