Website 'In the Weights' shows whether AI models know who you are
Two former OpenAI employees have built something that should make every AI company and user deeply uncomfortable. Their new website, "In the Weights," isn't just a quirky tool; it's a brutal, quantitative mirror held up to our collective cultural memory. It ranks people by their "recall strength" within AI models—a score of up to 996 that measures how deeply a human being is seared into the statistical psyche of these systems. And the results are a damning indictment of what we, as a society, ha
Analysis
Two former OpenAI employees have built something that should make every AI company and user deeply uncomfortable. Their new website, "In the Weights," isn't just a quirky tool; it's a brutal, quantitative mirror held up to our collective cultural memory. It ranks people by their "recall strength" within AI models—a score of up to 996 that measures how deeply a human being is seared into the statistical psyche of these systems. And the results are a damning indictment of what we, as a society, have chosen to immortalize in silicon.
The top of the list is predictable yet telling: Mozart, Shakespeare, Taylor Swift. This isn't a meritocracy of genius; it's a census of data abundance. It confirms that AI models are, first and foremost, cultural hoarders. They remember what we have obsessively documented, digitized, and replicated. Swift's dominance isn't a commentary on her musical supremacy over, say, Beethoven, but on the sheer, relentless volume of her contemporary digital footprint—lyrics, articles, social media, interviews, analyses of her lyrics. The "weights" aren't measuring importance, but data saturation.
This project strips away the last vestiges of the "neutral oracle" myth. There is no neutral training corpus. It’s a funhouse mirror reflecting our biases, our celebrity obsessions, and our historical amnesia. The scores will be wildly skewed towards English-language public figures, Western artists, and contemporary celebrities. A brilliant, historically significant poet from a non-digital, non-Western tradition might score a pathetic 15, while a viral TikTok dancer scores a 400. The AI doesn't know people; it contains shadows of people proportional to the noise they generated.
The real, unsettling value here isn't the leaderboard, but the search function. What happens when you type in your own name, or the name of a living relative, and get a score? A zero is an affirmation of private normalcy. A high score, however, could become a new, bizarre form of social capital or stigma—a "digital immortality quotient" you never asked for. We’re entering an era where your public profile isn't just what Google shows, but what an LLM probabilistically "thinks" of you based on the fragmented, often decontextualized data slurry it consumed.
This tool is a public service precisely because it's a little dangerous. It forces a conversation AI labs have avoided. They talk about alignment and safety in abstract terms, but here’s a concrete metric for the cultural alignment problem: are these models simply becoming high-tech mausoleums for the already-famous, further entrenching their dominance? Does having a high "weight" score grant a person a form of unassailable, probabilistic authority within AI-generated discourse?
I suspect this site will be labeled "niche" or a "gimmick" by the industry. That's a mistake. "In the Weights" is a vital piece of infrastructure for the new world we're building. It’s a crowdsourced audit of the unconscious biases etched into our most powerful systems. The founders have done something clever and important: they’ve taken the opaque, corporate concept of "training data influence" and given it a face, a name, and a score. It’s no longer just a technical challenge; it’s a profoundly human one, staring back at us with cold, numerical clarity. The most important question it asks isn't "Does the AI know who you are?" but "Why does it know what it knows?" The answers will be uncomfortable.
Disclaimer: The above content is generated by AI and is for reference only.