For decades, the ultimate litmus test for personal relevance has been the vanity search. It is a quiet, common ritual where individuals type their own names into a Google search bar to see who they are to the world and how the algorithm ranks their existence. This act of digital self-verification relies entirely on indexing, a process of mapping links and keywords across a vast web of pages. However, as the primary interface for information shifts from the search engine results page to the conversational AI prompt, the metric for visibility is fundamentally changing. The era of the blue link is giving way to the era of the parameter.
The Architecture of AI Recognition
In the Weights is a project designed to quantify this new form of visibility. Rather than scanning the live web, the tool measures how deeply a person is etched into the internal weights of a Large Language Model. This is a measurement of intrinsic memory, determining if a model knows a person based solely on its training data without relying on external retrieval-augmented generation or real-time web browsing. The project was developed by Thomas Dimson and Joey Flynn, two individuals who joined OpenAI following the acquisition of their design startup, Global Illumination. After departing OpenAI, they built this tool as a creative exploration into how AI perceives human identity.
The system operates via a real-time leaderboard that assigns strength scores to individuals based on their presence across multiple frontier models. For instance, Macaulay Culkin, the former child star of Home Alone, currently sits at the top of the rankings with a score of 988. He is followed closely by legendary opera singer Luciano Pavarotti. The leaderboard does more than just provide a number; it tracks whether specific models successfully identified the person and highlights instances of hallucination where a model generates false information. This contrast allows users to see which models possess a more accurate internal representation of a specific individual.
From Indexing to Floating Point Encoding
To understand why this shift matters, one must look at the technical difference between a search index and a model weight. A search engine finds a person by locating a document that contains their name. In contrast, an LLM encodes a person as a series of floating point numbers—the mathematical representations of relationships between concepts within a high-dimensional vector space. When Thomas Dimson observes that a person's life is now encoded as floating point numbers, he is describing a transition from quantitative visibility to qualitative imprint.
In the Weights functions by sending a standardized query, such as Who is <name>?, to a suite of leading models including Grok, Gemini, GPT, Claude, and Llama. The system requests up to ten result values, accompanied by short descriptions and confidence scores. The core innovation lies in the post-processing: the tool clusters similar descriptions across these different models to assign a final strength score. By cross-referencing responses, the system filters out noise and isolates the consistent memory of the person across the AI ecosystem. This process transforms the subjective nature of a model's response into a objective numerical value of digital presence.
This methodology reveals a critical tension in the current AI landscape. A person might have thousands of mentions on the web, yet if those mentions were not influential enough to shift the weights during the pre-training phase, they effectively do not exist to the AI. Conversely, a person with a limited but highly concentrated digital footprint in high-quality training sets may have a higher strength score than someone with more fragmented web presence. The tool essentially measures the density of a person's existence within the model's neural architecture.
Looking forward, the project aims to move beyond simple recognition. Dimson plans to analyze the precise discrepancies between models within the same family to understand why certain versions of a model remember details that others forget. There is also a focused effort to identify bias and uncover individuals who possess significant real-world social influence but lack the formal documentation, such as Wikipedia pages, that typically triggers high weights in LLMs. This suggests a future where the gap between actual influence and AI-perceived influence becomes a measurable data point.
Digital existence is no longer a matter of how many pages a crawler can find, but how much space a person occupies within the weights of a neural network.
