"Being in the weights means your existence was deemed important in the process of creating superhuman artificial intelligence," the In the Weights website says, as it launched a public tool that measures how prominently large language models recall an individual without consulting web search. The site, built by Thomas Dimson and Joey Flynn and published in June 2026, produces a per-name numerical strength score and a live leaderboard that ranks names across multiple AI models. TechCrunch tested the service and received a strength score of 641, placing the author in the top 6 percent; Macaulay Culkin led the board at 988, close to Luciano Pavarotti. The project changes what a vanity check looks like, letting anyone see how much of their persona is encoded inside model parameters.

"Google vanity searches are the wrong objective in 2026 as more traffic moves to LLMs," Thomas Dimson told TechCrunch by email, framing the deciding idea behind In the Weights. Dimson and Joey Flynn, who left OpenAI after selling their design startup Global Illumination, built a public tool that reads model internals through their outputs and reports how consistently a name is surfaced across an array of LLMs.

How it measures a name

In practical terms the site sends the same prompt to several models and then measures agreement. The prompt is simple: "Who is <name>? Give up to 10 results, each with a short description and confidence." In the Weights then clusters similar responses and computes a strength score that reflects how consistently and confidently models produce the same identity signals.

The roster of models the site lists is broad and explicit. It includes Grok, Google Gemini, multiple GPT versions, Anthropic's Claude, Meta's Llama and other, lesser-known models. For every queried name the site shows up to 10 model responses, groups similar responses together, and assigns a numerical strength. The interface also reports which model returned which answer, so a user can see model-by-model evidence behind the aggregate number.

The tool flags potential hallucinations. TechCrunch noted one clear example: GPT-5.4 Mini labelled journalist Anthony Ha as "an ambiguous name form that could refer to multiple people with the initials A.H.A." That kind of uncertainty is presented alongside confident matches, giving a qualitative read to complement the score.

The scoring is presented in percentile context. TechCrunch used the site to produce its own percentile framing, which put its author in the top 6 percent with a score of 641. At the time of publication the leaderboard was dynamic, moving in real time, and Macaulay Culkin led with a score of 988, a figure the site displayed as close to Luciano Pavarotti's score.

Why the founders built it

Dimson said the project began partly as a creative exercise and partly as a reaction to changes in how people find information. He told TechCrunch that, with more traffic shifting from traditional web search to conversational LLM interfaces, judging presence by search-indexation alone is less meaningful in 2026.

"We thought this would be a mild curiosity but it seems like it has struck a nerve of wanting to see if you live forever in the super intelligence," Dimson said, describing an intense early reception.

The site is explicit about what it's testing. It doesn't query web indexes or scrape pages. Instead, it interrogates models' internal representations indirectly: which names a model will surface when asked to list identities, and with what degree of confidence. That shift turns the vanity search into an experiment about model memory rather than about page rank.

Mastheads

For users the product changes the test they can run. Instead of checking whether a search engine has indexed biographies or news items, a person can check how much of their public persona is encoded inside multiple models' weights and see where models diverge. The display of per-model outputs makes it possible to spot which model contributed a particular piece of information, and whether that piece appears to be grounded or hallucinatory.

TechCrunch also observed the leaderboard updating as models or queries changed, which suggests the site either re-evaluates names on the fly or periodically rescans models. The live behaviour makes the scoreboard feel less like a static registry and more like a market of memory, where model updates or prompt tweaks can nudge a name up or down.

The launch contains no listed pricing or enterprise plans in the TechCrunch coverage, and the article doesn't cite a changelog or formal release notes beyond the live site. Availability was immediate: the site was publicly accessible when TechCrunch published on June 20, 2026, and pages for names and the leaderboard were viewable without an indicated fee. The article recorded no regional restrictions, monetisation roadmap, or scheduled feature rollouts.

That absence of a commercial plan leaves In the Weights as a public demonstration for now, rather than a paid identity product. But its existence already reframes a familiar test.

Vanity searches were built around indexing. This tool asks instead whether a person has left a trace in the statistical machinery of multiple LLMs.

There are immediate, practical uses and cultural curiosities. Individuals can check their own scores. Journalists can show how widely a name is encoded across model families. And the public can compare celebrities and historical figures side by side, seeing both the ranking and the model-by-model evidence that produced it.

Dimson and Flynn's pedigree is part of the story. Their time at OpenAI, and the sale of Global Illumination, are explicit background details in the TechCrunch piece. That provenance helps explain why they could move from idea to public site quickly, and why the project draws attention from people who follow the technical fringes of generative AI.

Still, the tool is blunt in one way. It measures recall across models, not reputation or accuracy.

A high score means many models consistently surface a name, but it doesn't vouch for the truth of the claims attached to that name. The per-model outputs and hallucination flags are the only safeguards against mistaking presence for veracity.

Visually, the site pairs a comparative leaderboard with per-name pages showing clustered outputs and model provenance. That combination makes it easy to move between an at-a-glance rank and the evidence that justifies it, which is central to the product's pitch: transparency about what lives inside model weights.

Public interest has been strong, TechCrunch reported, with users engaging in head-to-head comparisons and watching the leaderboard shift. The founders expected a mild curiosity. They got a conversation about permanence and memory in systems designed to generalise far beyond any single human life.

Related Articles

Macaulay Culkin tops the live leaderboard with a strength score of 988. The site was first reported by TechCrunch.

This article was created with AI assistance.