Physicist Yuji Tachikawa described how Claude Fable advanced a stalled problem in string theory
Physicist Yuji Tachikawa described how Claude Fable advanced a stalled problem in string theory
On July 12, University of Tokyo professor Yuji Tachikawa gave the model working notes from a joint project that had made no progress for about six months. Claude Fable first found a computational error and acknowledged that it had reached the same obstacle. After the physicist asked another question, the model developed a different approach and, according to Tachikawa, nearly solved the problem. He also watched it write and run verification code in SymPy.
Yuji Tachikawa studies the mathematical aspects of string theory and quantum field theory at the University of Tokyo’s Kavli IPMU. In a July 12 thread, he described giving Claude Fable his working notes, including calculations, unsuccessful approaches, and the point where the joint work had stalled.
“I showed it research notes from a joint project that had made no progress for about six months. To my surprise, it made a nontrivial observation and nearly solved the problem. Its first response was, ‘I found one computational error, but I got stuck at the same point.’ It then developed an approach I suggested,” Tachikawa wrote.
Fable first identified an error and reported that it had reached the same obstacle. After Tachikawa clarified his suggestion, it expanded the hypothesis. It then used SymPy, a symbolic mathematics library that can manipulate formulas and perform exact calculations, to write code and test its own predictions.
In a recent Princeton project, a sequence of models reproduced published results before starting new calculations and retained work logs. Tachikawa shared a description of one session, but the thread did not include the formulas, code, or final solution. His post documents the research process. The formulas and code will be needed for an independent assessment.
In this kind of work, the physicist keeps track of the subject matter, the history of failed attempts, and the criteria that a solution must satisfy. The model reads the accumulated material, identifies a connection or an error, proposes the next step, and can check the algebra with a computational tool. This shortens the path from disorganized notes to a hypothesis that the researcher can examine further.
Tachikawa described the experience as a personal test of the technology’s capabilities, not a systematic comparison of models.
A problem with a known answer tests whether a model produces the correct result. Unfinished research tests something different: whether the model can identify an error, retain the context of many previous attempts, propose a testable reformulation, and carry it through to a calculation. A solution becomes a scientific result only through openly available formulas, code, and independent verification by physicists.