Two former OpenAI employees have launched a website called In the Weights, aiming to answer an experimental question: How much do large models inherently “remember” about people without invoking web searches. As more users turn to chatbots for information, such tests are gaining real-world relevance.
Test name recognition using multiple models
The term "weights" in the website name refers to model parameters. Developers Thomas Dimson and Joey Flynn believe that "searching for oneself" in traditional search is no longer the sole measure of one's online presence; whether a model can directly reference a person is becoming another metric of visibility.
In Weights, similar questions such as "Who is [name]?" are sent to different models, requesting up to 10 results, brief descriptions, and confidence scores. The website then groups similar descriptions and generates a strength score to measure how well the model "remembers" the name.
The rankings may change and may also produce illusions.
The models currently participating in the test include Grok, Gemini, multiple versions of GPT, Claude, Llama, and some less common models. The results page also indicates which models provided responses and which responses may contain hallucinations or confusion.
For example, TechCrunch writer Anthony Ha has a score of 641, placing him in the top 6% of all names. However, rankings are subject to change. At the time of publication, actor Macaulay Culkin was ranked first, followed by singer Luciano Pavarotti.
The report also noted that GPT-5.4 Mini interpreted Anthony Ha as an ambiguous name form that could correspond to multiple people, rather than identifying it as a specific individual. Such cases are also flagged by the website as potential hallucinations.
Developers bet on new visibility in the model era
Dimson said in an interview that after leaving OpenAI, he and Flynn hoped to work on projects that could reignite their creativity. The two had previously joined OpenAI after their design company, Global Illumination, was acquired.
He believes that by 2026, as traffic continues to shift toward large models, the Google-style vanity search will no longer be the primary goal. Instead of webpage ranking, whether your information is "stored" within model parameters is becoming a new form of online presence.
The developers also stated that they will continue researching why the same model series produces different results, which types of people different models are more likely to "remember," and which individuals theoretically should have Wikipedia pages but do not yet.