Two former OpenAI employees launched a website called In the Weights, trying to answer an experimental question: In the absence of a web search, how many people do the large model itself “remember”? As more users turn to chat robots for information, such tests become relevant.
Multiple models to test name recognition
The term “rights” in the name of the website refers to model parameters. The developers Thomas Dimson and Joey Flynn believe that the “search yourself” in the traditional search is no longer the only way to measure the existence of an individual's network, and that the model is becoming another indicator of visibility.
In the Weights will pose similar questions to different models as “someone” and request a maximum of 10 results, a short description and confidence. The site then categorizes similar descriptions and produces a strength score to measure the model's “rememberance” of the name.
The list changes and there's hallucinations.
The models currently being tested include Grok, Gemini, various versions of GPT, Claude, Llama, and some smaller models. The results will also show which models give answers and which responses may be hallucinating or confusing.
In the case of TechCrunch author Anthony Ha, the score given on the website is 641, 6% above all names. However, the rankings will continue to change. At the time of the release, actor Macaulay Culkin was in first place, followed by singer Luciano Pavarotti.
It was also mentioned that GPT-54 Mini had interpreted Anthony Ha as a form of possible vague names for multiple individuals, rather than being directly identified as a specific person. Such cases are also marked as potential hallucinations on the website.
Developer bets on the new sight of the age of modeling degrees
In an interview, Dimson stated that he and Flynn, after leaving OpenAI, wanted to do something that would rekindle creativity. The two men had previously joined OpenAI, which originated from the acquisition of its design company Global Illumination.
In his view, by 2026, as the flow continued to shift to a large model, Google-style vagaries search was no longer the most important goal. Compared to the ranking of web results, whether or not you have information in model parameters is becoming another new sense of network presence.
The developers also indicated that they would continue to study why the same model series gave different results, what types of people different models could more easily “remember” and who theoretically should have Wikipedia words but had not been established.