Virality Engine

AI understands physics. Yet it can't conduct physics research. A new paper reveals its weaknesses. Forty-two physicists divided their research into 56 tasks. These are real projects, each taking the physicists approximately 270 days on average. Four are based on papers published in Nature. The AI receives what a professor gives a new PhD student: a short brief, with no methods and no solutions. It has to read the literature, make a plan, execute, and write a paper. How they tested it: → 22 models, 26 setups, 3 tries for each task → Up to 10 hours per run → Numbers checked from test scripts → Papers graded by AI reviewers using the physicists' own rubrics Results: → Each model passed less than 30% of the time → Best: Cloud Opus 5.5 in Cloud Code, 46 out of 168 runs (27.4%) → The physicists' own paper scored 96.2% on the same rubrics To see why the AI fails, the physicists manually ran 168 GPT-6 Astra runs. Lack of physics knowledge wasn't the main problem: → Weak and rigorous reasoning: 48% → Poor physical understanding: 34% → Poor use of literature: 27% The failure that always stuck with me was circular self-verification. The agent checks its answers using its own faulty thinking. So it never finds the error. In an atomic physics task, the agents never considered Rydberg decay as a source of error, which an atomic physicist would naturally think of. And when debugging, the agents kept tuning the wrong knob instead of finding the real problem. This doesn't mean the AI is useless for physics. 56 tasks is a small set. Most of the grading was done by an AI reviewer. And the authors found that out of 168 GPT-6 Astra runs, 9 were marked as failed, even though they were actually correct. I believe: the difficult part of research has always been judgment. This means knowing which numbers seem wrong, which papers are important, and when to doubt yourself. This is the skill that still makes a human researcher worth hiring. Paper👇