New website ranks how well AI models recall individuals without web search
In the Weights queries major models including Grok, Gemini, GPT, Claude, and Llama to produce a "strength score" for how prominently a person appears in model weights.
1 source · cross-referenced
- A new website called In the Weights ranks how prominently individuals appear in the training weights of major AI models by querying models directly.
- The tool compares results across models including Grok, Gemini, multiple GPT versions, Claude, and Llama, returning a numeric "strength score" and highlighting hallucinations.
- Scores fluctuate in real time as models are re-queried, with public figures like Macaulay Culkin and Luciano Pavarotti currently leading the leaderboard.
In the Weights is a newly launched website that measures how prominently a person’s name appears in the training weights of major AI models by querying models directly rather than relying on web search. The site’s creators, Thomas Dimson and Joey Flynn, both former OpenAI employees, designed it to address what they see as the inadequacy of traditional vanity searches in an era where more information discovery happens through chatbots.
To generate a "strength score," In the Weights prompts multiple models—including Grok, Gemini, multiple versions of GPT, Claude, and Llama, as well as lesser-known models—with a prompt asking, "Who is <name>? Give up to 10 results, each with a short description and confidence." The site then clusters similar descriptions and assigns a numeric score reflecting how strongly the name is represented across models.
The resulting scores are dynamic and change as models are re-queried, with public figures such as actor Macaulay Culkin and opera singer Luciano Pavarotti currently occupying the top positions on the leaderboard. For example, one TechCrunch journalist received a strength score of 641, placing them in the top 6% of names, while another model reportedly described the same journalist as an "ambiguous name form that could refer to multiple people with the initials A.H.A."
Dimson told TechCrunch the project was inspired by a desire to "get the creative juices flowing again" after leaving OpenAI and was shaped by reflections on how AI models encode information about people in their parameters. The site’s retro design and real-time scoring have contributed to its viral reception, with Dimson noting that users seem drawn to the idea of seeing whether they might "live forever in the superintelligence."
The tool also surfaces which models returned which answers for a given name, enabling users to compare model behaviors and identify potential hallucinations or inconsistencies in how different models represent the same individual.
- Jun 20, 2026 · AWS — Machine Learning Blog
AWS adds managed web search to Amazon Bedrock AgentCore
Trust79 - Jun 20, 2026 · The Verge — AI
Adobe Firefly AI studio update adds persistent project context and reusable assets
Trust72 - Jun 20, 2026 · The Verge — AI
Adobe rolls out AI assistants for Photoshop, Premiere, and other Creative Cloud apps in public beta
Trust79