There is a fundamental difference between data an AI retrieves via live web search and the knowledge it possesses natively. To explore this, former OpenAI employees Joey Flynn and Thomas Dimson created In the Weights, a platform that reveals whether a specific person was influential enough during training to be encoded directly into a model's parameters.

Understanding Parametric Memory

The system focuses on weights, the billions of numerical values that form an AI's neural network. When a model answers based on its internal state rather than using retrieval-augmented generation (RAG), it accesses its parametric memory. As noted by The Decoder, appearing in these weights indicates that the person was relevant enough to leave a permanent mark on the model's configuration during training.

The Strength Score Metric

The tool queries multiple LLMs to determine who a person is, combining the findings into a strength score. This metric peaks at 996, a level reached only by global figures such as Taylor Swift, Shakespeare, or Mozart.

The creators highlight that appearing in smaller models, such as Meta's Llama (1B parameter version), is a sign of exceptionally high relevance, given the limited capacity for knowledge storage compared to massive frontier models.

Technical Constraints and Hallucinations

The project also underscores the inherent limitations of current AI. Common names often lead to lower or inaccurate scores, and simple typos can significantly degrade the results. Furthermore, the risk of hallucinations remains; a model might recognize a name but fabricate biographical details. According to FancyAI, because information is distributed across billions of weights, it cannot be easily deleted once learned, making these AI-encoded identities a permanent part of the model's architecture.