{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/f916332a22f24605b87e5a4694c1dcb2\" frameborder=\"0\" width=\"1108\" height=\"831\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":831,"width":1108,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":831,"thumbnail_width":1108,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/f916332a22f24605b87e5a4694c1dcb2-2cfbdf5ab877f5fd.gif","duration":300.235,"title":"Crux: Finding Real Contradictions in Papers","description":"This Loom demonstrates Crux, a platform for pinpointing the exact lines where research papers truly contradict versus where differences come from context. T. Ram Kumar explains that Crux extracts the underlying conditions behind each agree or disagree result and outputs visual evidence showing the specific locations. As a real example, he compares OpenAI’s 2020 scaling law on GPT error, where the compute exponent is 0.73, against DeepMind’s challenge claiming an exponent of 0.50, arguing computation should split between model size and training data. He also describes how the system verifies extracted code against the source paper to remove hallucinated claims and then reconciles the two papers using scientific rules rather than keyword matching."}