In the Weights Scores How Well AI Remembers You
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- In the Weights, created by Thomas Dimson and Joey Flynn — both of whom joined OpenAI through the acquisition of their design startup Global Illumination — assigns a "strength score" measuring how well an AI model can recall someone without web search, framed as whether your existence was "deemed important in the process of creating superhuman artificial intelligence."
- The site queries Grok, Gemini, multiple GPT versions, Claude, Llama, and lesser-known models with a prompt like "Who is <name>? Give up to 10 results, each with a short description and confidence," then clusters similar descriptions to assign each name a numerical score.
- Macaulay Culkin currently tops the leaderboard with a strength score of 988, narrowly ahead of opera singer Luciano Pavarotti, while the TechCrunch blogger scored 641 — placing him in the top 6% of names tested.
- Results show which models returned which answers and highlight potential hallucinations: GPT-5.4 Mini reportedly described TechCrunch's Anthony Ha as "an ambiguous name form that could refer to multiple people with the initials A.H.A."
- Dimson argued that "Google vanity searches are the wrong objective in 2026 as more traffic moves to LLMs," positioning the tool as a response to AI chatbots replacing traditional search as the canonical source of identity information.
- Planned next steps include investigating why different models in the same series return different results for the same name, which models are biased toward certain types of people, and which notable people "should have a Wikipedia article but don't."
Why it matters: The site captures a measurable shift Dimson articulates directly in the piece: as LLM traffic eclipses Google search, being encoded in a model's training weights has effectively become a new layer of digital identity, and the leaderboard format turns that into a shareable comparison game — Culkin's 988 sets a ceiling roughly 54% above the TechCrunch blogger's 641, meaning the vast majority of tested names sit well below even mid-tier recognition.
Ask SkimNews
