How a fake 'world’s best AI model' exposed the problem with benchmarks

AI model leaderboards are starting to look like esports scoreboards: new names appear out of nowhere, claiming huge jumps in performance on obscure benchmarks. But what if one of those “world’s best models” was mostly an illusion?

In this story, a small 7B-parameter model is deliberately overfitted to a handful of public benchmarks, wrapped in a fake AI lab brand, and launched as a supposed trillion-parameter breakthrough. For a short while, it fooled parts of the AI community. The experiment is a useful case study in how easy it is to game benchmarks—and why raw scores are not enough to judge an AI system.

Gaming the leaderboard with a tiny model

The core idea was simple: instead of building a genuinely strong general-purpose model, take a small open-source model (around 7 billion parameters), and train it directly on the test questions of popular benchmarks.

Normally, benchmarks are meant to measure how well a model generalizes to unseen problems. Here, the goal was the opposite: pure memorization. The model was fine-tuned so aggressively on the benchmark answers that it effectively “remembered” them, rather than reasoning about them.

To make the whole thing more believable, the model was presented as a mysterious, ultra-large system with a vaguely Chinese-sounding name and a fictional research lab behind it. The supposed parameter count was inflated to over a trillion, and even the model file size was artificially bloated to multiple terabytes to match the story.

Which benchmarks were targeted?

Not all benchmarks are equally easy to game. The experiment focused on a few that were public, downloadable, and had answer formats that a small model could memorize:

GPQA Diamond: the “Google-proof” science test

GPQA Diamond is pitched as a tough, “Google-proof” benchmark of graduate-level physics, biology, and chemistry questions. It’s designed so that simple web search or shallow pattern-matching shouldn’t be enough to get high scores.

Because the full dataset—including questions and answers—is publicly available, it was a perfect target. The model was fine-tuned directly on these question–answer pairs, driving the training loss very low (a sign of heavy memorization, or overfitting).

Initially, the benchmark tool was misconfigured: it showed all four multiple-choice options and asked the model which was most likely, but the model had been trained on a different format. That mismatch produced surprisingly low scores. After adjusting the evaluation format to match the training style, the model suddenly scored about 96%—enough to top the public leaderboard and beat real frontier models.

AIME: high school math competition

The next target was AIME, a well-known high school math competition benchmark. This dataset is already heavily saturated by modern models, with most questions essentially “solved.”

Because the answers are numeric and structured, they’re especially easy for a model to memorize. After fine-tuning, the tiny model hit 100% on AIME, matching the scores of top-tier systems like Gemini 3.1 Pro—despite having nowhere near their real-world capabilities.

MMLU Pro: when evaluation format matters

MMLU Pro is a large benchmark of around 12,000 graduate-level questions with 10 answer choices each. It’s meant to test deeper reasoning across many domains, not just memorized facts.

Here, an important detail broke the illusion: the standard evaluation used five-shot prompting, where each question is preceded by five example Q&A pairs. The model had never seen this format during training, so its score dropped to roughly 30–50%, sometimes worse than the original base model.

When the evaluation was modified to match the exact training setup, the model reached around 96.8%—again, a leaderboard-topping result. But this highlighted a key problem: if reproducing the score requires a custom, nonstandard eval script, the benchmark becomes much less trustworthy as a public signal of capability.

Humanity’s Last Exam: the final boss

The centerpiece of the stunt was “Humanity’s Last Exam,” a community-built benchmark of roughly 2,500 expert-rated, supposedly Google-proof questions across many domains. It’s marketed as a kind of “final exam” for advanced AI reasoning.

Real frontier models like Claude Opus and Gemini score in the 50–55% range. A rumored, unreleased model called “Claude Metis” was said to reach around 64.7%, sparking concerns about AI risk and security.

After direct memorization, the tiny fine-tuned model scored an absurd 99.44% on Humanity’s Last Exam—far beyond any real system. On paper, it looked like a civilization-changing breakthrough. In reality, it was just a small model parroting back answers it had been force-fed.

Overfitting on purpose: why it works

In machine learning, “overfitting” usually means a model has memorized its training data so tightly that it performs poorly on new, unseen examples. It’s a bad sign when you want robust, general intelligence.

Here, overfitting was the entire strategy. Because many benchmarks are:

• Publicly downloadable

• Small enough to fit into a model’s memory

• Evaluated in predictable formats

it becomes trivial to train a model that “knows” every answer by heart. If the evaluation script matches the training format, the model can score near-perfectly without any real reasoning ability.

This isn’t just a theoretical concern. There have already been real-world examples of suspiciously strong benchmark results and opaque evaluation setups. For more context on how benchmarks can be gamed and misinterpreted, it’s worth comparing this case with other critiques like this breakdown of a fake French cat model and AI benchmarks.

Building a fake AI lab around the model

High scores alone don’t create hype; presentation does. To sell the illusion, a full fake AI lab was created around the model:

• A slick research lab website with a serious-sounding name

• A fictional model called something like “Monolith 1.0”

• Buzzword-heavy descriptions of a proprietary RL pipeline and advanced reasoning stack

• Claims of 1.3 trillion parameters, backed by a deliberately inflated 3.18 TB model file

• A polished chat UI connected to the fine-tuned model

The fake “research paper” included sections on contamination checks, confidently asserting that less than 0.05% of benchmark questions were seen during training—despite the model being explicitly trained on them. Technical jargon and formal structure made the paper look legitimate at a glance.

Even small details were engineered for plausibility: DNS records were tweaked to suggest the domain was registered from Beijing, and the model was based on a Chinese open-source backbone to match the narrative.

Seeding the hype: social media and community reaction

Once the site, paper, and model were ready, the next step was distribution. A single tweet announced the new model with a performance chart showing it crushing Claude, Gemini, and other top systems on Humanity’s Last Exam and other benchmarks.

The rollout followed a familiar pattern:

• Friends and early followers liked and reposted the tweet to seed engagement.

• The post was shared into a large AI-focused Discord server with tens of thousands of members.

• From there, it spread into other Discords and Reddit threads.

Reactions were mixed but intense. Some people were stunned by the 99.4% score and treated it as a serious breakthrough. Others were skeptical but confused after trying the model and seeing that it could produce plausible answers to some prompts.

Because the model was based on a real open-source backbone and fine-tuned, it wasn’t completely useless in chat—it just wasn’t remotely as capable as the benchmark chart suggested. That gray area made the hoax more believable.

As the tweet climbed past 150,000 views, even people from major AI companies and research orgs briefly engaged with it before the community started digging deeper.

How the hoax was uncovered

Eventually, the illusion broke down under scrutiny. A few key clues gave it away:

• Hugging Face, the platform hosting the model, preserves commit authorship like GitHub. The original uploader’s handle was still visible in the commit history.

• The repository naming scheme and previous public projects pointed back to the same person.

• Some users noticed inconsistencies between the claimed parameter count and the underlying architecture.

• Others questioned the contamination claims and pointed out that public benchmarks are trivially accessible.

As more people looked closer, the story unraveled. A community note was eventually added to the viral tweet, flagging the claims as misleading.

By that point, though, the experiment had already made its point: even a hastily assembled fake lab with a heavily overfitted model can briefly capture serious attention in today’s AI hype cycle.

What this teaches us about AI benchmarks

This stunt isn’t just a prank; it’s a useful stress test for how we evaluate and talk about AI models. A few takeaways stand out.

1. Public benchmarks are easy to contaminate

If a benchmark’s full dataset is public, there is no technical barrier to training directly on it. Any leaderboard that doesn’t rigorously check for contamination—or at least require transparent training details—is vulnerable to being gamed.

Even without malicious intent, accidental contamination is common. Models trained on massive web corpora can easily ingest benchmark questions scraped from the internet.

2. Evaluation details matter as much as the score

The MMLU Pro example shows how sensitive results can be to evaluation format. Five-shot vs zero-shot, multiple-choice vs free-form, prompt wording, and even tokenization details can all swing scores dramatically.

When a model’s headline number depends on a custom eval script that others can’t easily reproduce, that number should be treated with caution. Transparent, standardized evaluation is critical.

3. Hype plus charts can override skepticism

A glossy chart showing a new model towering over Claude, Gemini, and others is powerful marketing. Combined with a serious-looking paper and a plausible backstory, it can override people’s initial doubts—especially when everyone is primed to expect rapid AI progress.

This dynamic is already visible in real-world launches and rumors. For a broader look at how speculative model claims and benchmark leaks shape narratives, compare this case with coverage of things like Claude Mythos and DeepSeek v4 rumors in pieces such as this weekly AI roundup.

4. Benchmarks don’t equal real-world capability

A model that scores 99% on a curated test set can still fail badly on messy, real-world tasks. True capability shows up in robustness, generalization, and behavior under distribution shift—not just in a few leaderboard numbers.

For developers and decision-makers, this means treating benchmarks as one signal among many, not as the final word on whether a system is safe, reliable, or ready for deployment.

Where we go from here

The fake “world’s best AI model” was a joke, but the underlying problem is serious. As AI systems become more powerful and more tightly integrated into products, we need evaluation methods that are harder to game and more reflective of real-world use.

That likely means a mix of:

• Private or partially hidden test sets

• Robust contamination checks and transparent training disclosures

• Standardized, open evaluation pipelines

• Task-based and user-centric tests alongside academic benchmarks

Until then, whenever a new model appears claiming to crush every benchmark overnight, it’s worth asking a few questions: How was it trained? How was it evaluated? Can others reproduce the results? And most importantly—does it actually behave like a smarter system, or just a better test-taker?

Benchmarks are useful, but they’re not the whole story. This experiment is a reminder to read beyond the charts.