A new AI vanity search ranks who lives in the model’s memory

For years, a basic Google search has been the quickest way to see how the internet remembers a person. In the age of chatbots, though, that familiar ritual is starting to feel incomplete. Increasingly, people are learning about one another through large language models, not just search engines — and that shift has spawned a new kind of ego check.

A new site called In the Weights turns that shift into a game. Built by former OpenAI designers Thomas Dimson and Joey Flynn, the project tries to measure how likely a range of AI models are to “remember” a person without looking anything up on the web. The result is a score that reflects whether a name appears to be encoded in a model’s internal parameters — the “weights” that shape how it responds.

The site is part novelty, part commentary, and part experiment. It asks a deceptively simple question: if a person matters enough to show up in the responses of multiple AI systems, does that mean they have become part of the machine’s cultural memory?

What In the Weights is trying to measure

Dimson and Flynn describe the project as a way to estimate whether a person is present inside a model’s training memory rather than just retrievable through live search. That distinction matters. Many chatbots can answer questions by using the web or other tools, but In the Weights is focused on what the model produces on its own.

The site essentially asks several models who a person is, then compares the answers. It checks a mix of major and lesser-known systems, including versions of ChatGPT, Claude, Gemini, Grok and Llama. The responses are grouped by similarity, then translated into a “strength” score.

In practical terms, the score is meant to reflect how confidently and consistently the models can identify someone from memory. The higher the number, the more visible that name appears to be across the AI ecosystem.

How the scoring works

According to the site’s description, the process includes a prompt asking for up to 10 results with short descriptions and confidence levels. The answers are then clustered, which helps group similar identity claims together. A strength score is assigned after that consolidation.

The result is not a scientific census of fame, but it does create a rough ranking that feels familiar to anyone who has ever searched for themselves online. The difference is that this version is based on what a model appears to know, rather than what the open web indexes.

• Multiple models are queried for the same name.
• Answers are grouped by similarity.
• A score is assigned to estimate “memory strength.”
• The site highlights cases where different models disagree or hallucinate.

Why the idea is resonating now

The project arrives at a moment when the meaning of search is changing fast. Traditional web search is still central to online discovery, but it no longer has a monopoly on attention or reputation. Chatbots increasingly act as the first place people turn for summaries, background, and even biographical information.

That has created a new reality: your online presence may now depend as much on whether an AI model has “learned” you as on where you rank in search results. For public figures, founders, creators and journalists, that shift can be both flattering and unsettling.

Dimson told TechCrunch that he and Flynn wanted a fresh creative project after leaving OpenAI, where they had both worked following OpenAI’s acquisition of their design startup Global Illumination. He said the idea grew out of a broader reflection on how conventional vanity searches no longer capture the full picture in an era when language models shape traffic and discovery.

Dimson said the project began as a way to get creative momentum back after leaving OpenAI, while also exploring the idea that many lives are now stored, in some sense, inside AI models as numerical patterns rather than just on searchable webpages.

He also pointed to a playful blog post and a classic science-fiction reference as influences on the site’s direction. That blend of technical concept and internet humor helps explain the product’s appeal: it is serious enough to spark debate, but whimsical enough to feel sharable.

A leaderboard built for curiosity — and ego

One reason the site is drawing attention is that it turns an abstract AI concept into something immediately legible: a leaderboard. Humans are naturally drawn to rankings, especially when they can compare themselves against celebrities, founders, writers or colleagues. In the Weights leans into that instinct.

TechCrunch noted that the score for one of its own writers placed him in the top 6% of names, while other staff members scored even higher. At the time of reporting, the leaderboard itself was still changing as different names moved up and down.

Among the top names cited were Macaulay Culkin and Luciano Pavarotti, illustrating just how different the reputational footprint can be across models and generations. The results are not simply about current fame; they also reflect how training data, cultural prominence and naming ambiguity shape outputs.

That makes the site interesting for reasons beyond vanity. It offers a window into the uneven ways models absorb and reproduce cultural knowledge.

When a model gets it wrong

The site also exposes the limits of model recall. In one example, a model reportedly treated the name Anthony Ha as potentially ambiguous, suggesting it could refer to multiple people with similar initials. That kind of mistake is not unusual in large language models, which often generate plausible but imperfect explanations when they are uncertain.

Those errors are part of the point. A person’s “presence” in the weights is not only about whether a model can identify them, but also whether it can do so consistently, accurately and without drift between model versions.

AI critic Anthony Moser dismissed the concept as little more than asking a collection of chatbots to describe someone, arguing that the site does not reveal anything fundamentally new about the systems’ behavior.

Even so, the site’s design makes the comparison feel more concrete than a casual prompt test. The retro, Nintendo-inspired aesthetic gives the project a playful layer that helps lower the barrier to engagement. It looks more like a game than a research paper, which may be one reason it has spread quickly.

What the project reveals about AI memory

At a deeper level, In the Weights is tapping into a growing fascination with model memory. As AI systems become more central to how people discover information, users are increasingly curious about what those systems know, what they forget and what they invent.

That curiosity is especially intense around identity. If a chatbot can summarize a person, who decides which facts matter? If a model names a journalist, founder or artist without access to the live web, what does that say about the cultural record encoded in training data?

The answers are messy. Models do not “remember” people the way humans do. They produce outputs based on patterns learned from enormous data sets, which can include biographies, social posts, articles, directories and countless other sources. A strong score may say as much about information density and repetition as it does about actual prominence.

Still, the idea of being “in the weights” has emotional power. It suggests a form of machine-mediated permanence: a version of you that survives in the statistics of a model long after a browser tab is closed.

Why ranking memory is harder than ranking search

Web search is noisy, but it is at least tied to visible documents. AI memory is less transparent. A model may surface a person because that name appears frequently in its training data, because the person was associated with a high-profile event, or because similar names create spillover effects.

That makes any ranking inherently approximate. It is better understood as a cultural signal than as a precise measurement.

Some of the major reasons the numbers can vary include:

1. Model version: newer and older releases may respond differently.
2. Training source mix: not all systems were trained on the same material.
3. Prompt phrasing: slight wording differences can change outputs.
4. Name ambiguity: common names are harder to resolve reliably.
5. Hallucinations: models can confidently invent details or identities.

The founders’ background and the project’s origin story

Dimson and Flynn are not building from the outside looking in. Both worked at OpenAI after their earlier design startup, Global Illumination, was acquired by the company. That background gives the project a particularly insider flavor: it comes from people who understand both the technical machinery and the cultural dynamics around AI products.

Their post-OpenAI phase appears to be part experimentation, part reset. In the modern AI ecosystem, founders often move between research, product and creative side projects at a rapid pace. In the Weights fits that pattern, but it also feels unusually self-aware. It is a product that makes fun of the world it comes from while also taking its underlying premise seriously enough to build around it.

That balance may help explain why it found an audience so quickly. It speaks to a widespread uncertainty about how AI systems represent people, and it does so in a format that is easy to understand in seconds.

Why people care whether an AI model knows them

There is a social dimension to all of this that goes beyond pure curiosity. Public visibility has always been tied to status, but the platforms that confer that visibility are changing. A person may be less worried about page-one Google results and more concerned with whether a chatbot can accurately describe their work, history or expertise.

That shift matters for creators, journalists, executives, politicians, academics and niche online personalities alike. In a world where AI systems are becoming interfaces to knowledge, being legible to the model may increasingly shape how one is perceived.

In that sense, In the Weights is both a toy and a warning. It playfully packages a serious question about cultural power: who gets remembered when the internet is increasingly mediated by opaque systems?

From search optimization to model optimization

For years, businesses and individuals have thought in terms of search engine optimization. Now a new discipline is emerging, even if informally: the art of making sure that AI systems know who you are and describe you correctly.

This creates new incentives. People may begin thinking about how much structured information exists about them, whether their names are common enough to confuse models, and how often they appear in source material likely to be used for training.

That is a profound shift. Search engines indexed the web; AI models absorb it. The difference is not just technical but epistemic. One makes information retrievable; the other internalizes it into a system that can rephrase, compress and distort it.

Comparing the old vanity search with the new one

The following table outlines the practical differences between a traditional web search and the kind of AI-centric score In the Weights is trying to surface.

Why the site may keep evolving

Dimson said he wants to go deeper into why models from the same family can produce different answers, why some systems seem biased toward certain kinds of people, and which notable figures are missing from Wikipedia despite substantial real-world significance.

Those are not trivial questions. They cut into the politics of training data, representation and knowledge preservation. If a person is absent from one information system and overrepresented in another, that imbalance can affect how future users perceive them.

There is also the possibility that the project will become a moving target. As models are updated, filtered and retrained, the leaderboard will likely keep changing. That makes the site less of a static ranking and more of a live snapshot of the AI ecosystem’s current memory.

In that respect, it may serve as a useful artifact for the current moment. It captures a point in time when people are beginning to ask not just what the internet says about them, but what the models have absorbed and how confidently they can repeat it.

The bigger question: what does it mean to be remembered by AI?

For all the humor around celebrity rankings and retro graphics, the underlying question is serious. If being in the weights means a model can recall you, then AI memory is becoming a proxy for visibility, influence and perhaps even permanence.

That does not mean a chatbot is a judge of importance. It means that the systems shaping everyday information intake are becoming another layer in how reputation is built and preserved.

In the Weights turns that idea into a score, which is why it is easy to understand and easy to share. But the broader implication is bigger than the leaderboard. As AI systems become more central to search, summarization and discovery, the line between being known on the web and being known by machines will only get thinner.

For now, the site offers a clever, lightly absurd answer to a modern anxiety: if the internet no longer remembers you in the old way, maybe the models do. Or maybe they just approximate it well enough to make us care.

Key facts about In the Weights

Here is a quick summary of the project and the context around its launch.

As AI continues to reshape how people find and interpret information, projects like In the Weights are likely to multiply. Some will be gimmicks, some will become useful diagnostics, and some will land somewhere in between. This one has already done the most important thing a curiosity-driven internet product can do: it made a technical concept feel personal.

superintelligencenews@gmail.com