Viral Physics Puzzle Tests AI Scientific Reasoning

A Gruesome Thought Experiment Goes Viral

What would happen if you physically compressed a 150-kilogram (330-pound) person into a slim figure instantaneously? This grotesque but scientifically fascinating thought experiment recently exploded on Zhihu, China's leading Q&A platform, drawing millions of views and sparking intense discussion about extreme physics — and inadvertently becoming a benchmark for how well AI systems handle complex multi-domain scientific reasoning.

The original analysis, posted by a quantum physics researcher using the handle Xiao Banyang, applied rigorous thermodynamics, materials science, and high-pressure physics to reach startling conclusions. It's exactly the kind of multi-step, cross-disciplinary problem that AI labs worldwide are racing to solve.

The Science: Diamond Corpses and Metallic Ice

The researcher's analysis begins with straightforward biology. An obese woman at 150 kg with 45% body fat carries roughly 67.5 kg of fat (density ~0.9 g/cm³) and 82.5 kg of lean mass (density ~1.1 g/cm³). The initial body volume clocks in at approximately 150 liters.

Compressing this to a healthy female figure of roughly 55 liters means reducing volume by nearly two-thirds — instantaneously. The conclusions are as horrifying as they are scientifically rigorous:

• Death in microseconds. Every cell and biological macromolecule disintegrates completely. No intact organic molecules survive the compression.
• The body becomes harder than granite. Core composition transforms into metallic ice, diamond, and high-pressure mineral phases with metallic conductivity.
• Temperatures exceed the Sun's surface. The compressed object reaches thousands of degrees Celsius, slowly releasing enormous internal energy before eventually cooling into a room-temperature, ultra-dense human-shaped diamond-metallic ice composite.

In short, you don't get a slim person. You get a superheated, diamond-hard humanoid sculpture.

Why AI Labs Are Paying Attention

This kind of problem — absurd on its surface but demanding genuine expertise across physics, chemistry, biology, and materials science — has become a critical testing ground for large language models and AI reasoning systems.

Leading AI models like OpenAI's GPT-4o, Anthropic's Claude, Google's Gemini, and open-source contenders like Meta's Llama 3 are increasingly evaluated not just on standardized benchmarks but on their ability to handle novel, multi-step scientific reasoning. The 'compress a fat person' problem is a perfect stress test because it requires:

1. Biological knowledge — understanding body composition, fat density, and lean mass ratios
2. Thermodynamic reasoning — calculating pressure-volume-temperature relationships during rapid compression
3. Materials science — predicting phase transitions of carbon, water, and minerals under extreme pressure
4. Cross-domain synthesis — combining all of the above into a coherent narrative

When researchers at institutions like MIT, Stanford, and Tsinghua have tested frontier models on similar 'Fermi estimation meets extreme physics' problems, results vary dramatically. GPT-4o and Claude 3.5 Sonnet typically identify the correct physical principles but sometimes underestimate pressures or miss phase transitions. Smaller models often fail catastrophically, treating the problem as a simple density calculation.

The Rise of 'Scientific Reasoning' as an AI Benchmark

The viral Zhihu post highlights a growing trend in AI evaluation: moving beyond multiple-choice science exams toward open-ended, creative scientific reasoning. Google DeepMind's recent FunSearch project and OpenAI's emphasis on 'deep research' capabilities both reflect this shift.

Several startups are now building AI tools specifically for extreme-scenario physics simulation. Companies like Symbolic AI and Entropica Labs are developing models that combine LLM reasoning with numerical physics engines, aiming to handle exactly these kinds of wild thought experiments with quantitative precision.

'The best test of scientific understanding isn't whether an AI can pass an exam,' noted one computational physics researcher on the platform. 'It's whether it can take an absurd premise and derive correct, non-obvious consequences.'

AI-Powered Science Communication

The post also illustrates how AI-enhanced platforms are reshaping science communication. Zhihu's recommendation algorithm — powered by deep learning models — surfaced this niche physics analysis to millions of non-specialist readers, turning dense thermodynamic calculations into viral entertainment.

This mirrors trends on Western platforms. YouTube's algorithm increasingly promotes long-form science explainers, while AI-powered tools like NotebookLM and Perplexity are making complex scientific reasoning accessible to general audiences.

What Comes Next

As AI reasoning capabilities improve, expect more convergence between viral science content and AI benchmarking. The next generation of models — including OpenAI's anticipated GPT-5 and Anthropic's Claude 4 — will likely be evaluated partly on their ability to handle exactly these kinds of creative, multi-domain physics problems.

For now, the Zhihu post stands as both a delightfully morbid thought experiment and an unintentional AI challenge problem. The answer, for the record: you don't get a slim person. You get a glowing, diamond-hard, metallic humanoid that slowly cools to room temperature over days.

Science, as they say, doesn't care about your diet goals.