In the Weights offers a simple way to ask a strange modern question: does an AI model know who you are without searching the web?
The website looks at whether a person appears to be represented inside the weights of large language models. Those weights are the billions of numerical values where these systems encode what they learned during training.
What the site is measuring
The core idea is not whether a model can find someone online. It is whether the model can recall who a person is from its own internal knowledge.
If a person shows up in the results, the implication is that the model treated that person as relevant enough during training to identify without outside tools. In the language used by the source, the person may be "stored" in the model weights.
That does not mean the model has a perfect record of someone. It means the model can produce an identification from the information encoded in its parameters.
How the strength score works
In the Weights queries several models about a specific person. It then combines those responses and assigns a strength score.
The source gives two examples from THE DECODER: Maximilian Schreiner has a strength score of 175, while the author has a strength score of 262. These numbers are presented as examples of how the site ranks recognizability across models.
The leaderboard shows a much higher ceiling. The maximum strength score is 996, and the source says that level is reserved for names such as Mozart, Shakespeare, or Taylor Swift.
That spread gives the score its basic meaning. A low or mid-level result suggests partial or model-dependent recognition. A score near the top suggests that a name is widely and strongly represented across the tested systems.
Why smaller models matter
The creators say that smaller models make it harder for a person to appear in the results. That makes model size an important part of interpreting the score.
Someone who appears in Meta's Llama, which has a billion parameters, is treated as highly relevant by the creators. The logic is straightforward: if a smaller model can still identify a person, that person must occupy a stronger place in what the model retained.
This also means that the score is not only about fame in a broad cultural sense. It is about how consistently a name can be recovered by the models the site queries, including systems with tighter capacity.
The limits are part of the result
The creators also point out important weaknesses in this kind of measurement. Large language models can hallucinate biographical details, so a response may sound confident while still being wrong.
Typos can reduce scores, which means the way a name is entered matters. Common names can also produce worse results, because the model may have difficulty separating one person from another.
Those limits make the score useful as a signal, but not as a definitive identity check. The site can show whether models appear to recognize a person, yet the details behind that recognition still need caution.
- Hallucinated biography: the model may invent or confuse details.
- Typos: misspellings can drag down the score.
- Common names: overlap between people can make results less reliable.
Who built In the Weights
In the Weights was built by Joey Flynn and Thomas Dimson, both former OpenAI employees.
The site turns an abstract technical fact into something anyone can test. Model weights are normally discussed as infrastructure: the numerical machinery behind AI behavior. Here, they become a way to examine which people large language models can recall directly.
That makes the project useful for understanding how AI knowledge is distributed. It also shows why recognition by a model is not the same thing as accuracy. A model may know enough to identify a name, but still produce mistakes around the person behind it.
The result is a compact public test for a larger question: when AI systems learn from the world, which names become durable enough to survive inside the model itself?