Sharing a study by OpenAI in collaboration with multiple top universities around the world, this report reveals an important trend: >Frontier big models are evolving from simple efficiency tools to research partners with substantial 'intellectual contribution' capabilities. This report, titled "Early Experiments on Accelerating Science with GPT-5," was jointly released by OpenAI, Oxford University, Cambridge University, Columbia University, Harvard University, University of California, Berkeley, and Lawrence Livermore National Laboratory. The report provides a detailed record of practical cases of AI in fields such as biology, physics, and mathematics, demonstrating that under expert guidance, the cycle of scientific discovery is undergoing an order of magnitude compression in specific tasks. Here are several iconic moments of 'paradigm shift' in scientific research. one ️⃣ Biomedicine: not only for analysis, but also for proposing Novel Mechanisms GPT-5 Pro has demonstrated astonishing insights in the study of T cell metabolism regulation. It reinterpreted complex flow cytometry data and proposed a mechanism that human experts had never imagined: the effect of 2-DG on T cells is not solely due to glycolysis inhibition, but rather through interference with N-linked glycosylation drive. Based on this assumption, it designed a clear experimental decision tree like a senior PI, including the use of mannose for salvage experiments. Even more impressive is its prediction that brief 2-DG treatment in CAR-T preparation can enhance the killing power against CD19+cancer cells, which is highly consistent with unpublished internal data in the laboratory. The author explicitly states that GPT-5's contribution in this case has reached the level of a 'Co investigator' and is sufficient to be listed as a co-author of the new paper. two ️⃣ Inertial confinement fusion (controllable nuclear fusion): compressing the workload from 6 months to 6 hours In the field of inertial confinement fusion (ICF), a key technology path in controlled nuclear fusion, physicists have used AI to build a model of thermonuclear combustion wave propagation. This is not a fully automated performance of AI, but an efficient human-machine collaboration: human experts are responsible for setting physical targets and parameter tuning, correcting AI's early unreasonable "numerical tape" scheme; GPT-5 completed PDE modeling, numerical code writing, and ultimately assisted in deriving theoretical formulas to explain numerical results within a few minutes. The researchers evaluated that this process compressed the work that originally required "two excellent postdoctoral fellows to spend several months" (about 6 person months) into "6 person hours", achieving an efficiency improvement of about 1000 times. three ️⃣ Pure mathematics: conquering the long dormant 'unsolved mystery' In the renowned Erd ő s problem database, GPT-5 assisted mathematicians in solving the long-standing unresolved Problem 848. In this process, AI proposed a key new approach of "necessary conditions+stability analysis", successfully connecting diagonal and non diagonal constraints and filling a crucial link in the proof logic. The author vividly describes the final proof as' a human mathematician sandwiched between front and back, with the crucial step in between completed by GPT-5 '. four ️⃣ Deep literature mining: a cross disciplinary 'knowledge wormhole' In a convex geometry study, the author hoped to find similar quantitative results, but received a seemingly off topic recommendation. GPT-5 sensitively relates to the classic results of Papadimitriou&Yannakakis (2000) in the field of multi-objective optimization. Although these two fields may seem unrelated on the surface, AI has identified underlying mathematical isomorphism. Inspired by this, the author successfully improved their theorem and eliminated the original logarithmic factor. The ability to associate across disciplinary barriers is often difficult for human experts to possess due to their limited professional backgrounds. five ️⃣ Objective limitations and inspirations It is worth noting that GPT-5 is not omnipotent. The report emphasizes that it often confidently makes mistakes (Hallucination) and even tries to cater to users' expectations with incorrect mathematical deductions. The current success heavily relies on the "scaffolding" guidance of human experts, which breaks down big problems into verifiable sub steps and rigorously identifies the authenticity of the results. six ️⃣ Conclusion We are entering a new era of AI assisted discovery. The role of AI is evolving from a tool for processing data to a partner capable of hypothesis generation and logical reasoning. The future competitiveness of scientific research will depend on whether scientists can accurately ask AI questions and have the ability to quickly correct errors when it talks nonsense. I also translated it into Chinese (but those who can understand it should have read the original text, right?) ）： https://(randomarea.com)/early-science-acceleration-experiments-with-gpt-5/