A Self-Recovery Metric: Researchers Propose an ASI Criterion for Evaluating AI

Researchers have proposed a new approach to evaluating AI models: the ASI criterion, built on the concept of informational autonomy. It is an attempt to move beyond traditional performance metrics and measure deeper qualities of artificial intelligence, those that show up in a system's capacity for adaptation and recovery.

The core of the ASI criterion

The ASI criterion evaluates a model's ability to recover and develop from a minimal informational "seed": the smallest set of data a system needs in order to function and evolve. The name alludes to the concept of Artificial Super Intelligence, but the focus is less on superintelligence than on a system's capacity for autonomous development and on informational density.

The idea is borrowed from biology: organisms can regenerate from a minimal unit (a cell, DNA). The authors propose applying a similar approach to AI: what is the minimal complexity of an informational set from which a model can restore its capabilities? This reflects the idea of "technotropic AI", a system that can feed on a minimal amount of information and grow from that foundation.

How it works

The metric measures several key parameters of a system:

• Minimal informational base: the size of the smallest data set needed to initialize the model
• Recovery speed: how quickly the system restores its functions after a "reboot"
• Recovery completeness: what percentage of the original abilities and knowledge is restored
• Capacity for expansion: whether the system can develop and acquire new capabilities beyond its original parameters

In practice, a model that can restore its abilities from a very small amount of information scores higher on the ASI criterion. This reflects its adaptability, informational compactness and potential for operating under limited resources. The higher the criterion, the "smarter" the system's use of the information available to it.

Where to apply the metric

The ASI criterion may be especially useful when developing models for resource-constrained settings, such as mobile devices, edge computing, or environments with unreliable access to central databases.

The metric also helps assess how resilient a model is to information loss and knowledge decay, and how quickly it can adapt to changing environmental conditions.

For researchers, it becomes a tool for understanding the internal structure of models and how they are organized: models with a high ASI criterion may be more "understandable" in terms of information compression, and they contain a denser, more efficient distribution of knowledge.

What it means

The ASI criterion is proposed as a step toward a more comprehensive and objective evaluation of AI systems. Traditional metrics (accuracy, recall, F1-score) tell us how well a model solves a specific task on a specific dataset. The ASI criterion adds a measure of a system's "internal efficiency" and its capacity for self-directed development.

If the metric gains traction in the scientific community, it may change how AI models are designed and trained, pushing toward greater compactness and adaptability.