Summary
A new physics benchmark has raised an awkward question for AI researchers: what happens when the model is better than the people grading it? This video explains the story behind TheoPhysEval, a test of 50 unpublished, research-level theoretical physics problems built by researchers from Harvard, the University of Texas, Google DeepMind, and institutes in India.
OpenAI’s GPT-6 Astra, running inside the Codex coding tool with paper search and the ability to write code to check its own work, initially scored around 88%. When experts reviewed the questions it appeared to miss, they found errors in their own answer key. After corrections, the model’s score rose to 49 out of 50, or 98%. The video explains why unpublished problems matter for avoiding training-data contamination, and why a flawed answer key means other models may have been marked down unfairly.
The presenter also covers wider examples of AI in research, including a mathematician who says a problem he had been stuck on for a year was solved in two minutes, and OpenAI’s reported proof of the cycle double cover conjecture. Mixed reactions from mathematicians, including retracted AI-generated papers, are discussed alongside the broader problem of how to test AI systems as they surpass human experts.
📺 Source: TheAIGRID · Published November 03, 2026
🏷️ Format: News Analysis







