13:28Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI Announces 3 Scientific Results From AI; Scientists Are Reviewing Them
OpenAI has announced that its artificial intelligence models achieved 3 significant scientific results in mathematics, computer science and physics. Teams of scientists outside the company have been formed to review the claims and explain them in plain language. The announcement came several weeks after the release of more than 700 pages of material on solutions to 372 major mathematics problems and advances in the field. However, the new results have yet to undergo detailed review by the scientific community.
The company chose to release the results in 3 different ways. The mathematics material was published openly so that researchers could examine and explain the proof themselves. 2 new manuscripts in mathematics and computer science were posted on GitHub. For the computer science work, the answer produced by artificial intelligence and a computer-verifiable Lean proof were also shared. The expert group reviewing the physics material will decide when it will be published. The mathematics manuscript claims to prove the local smoothing result in the Sogge conjecture for 2 or more spatial dimensions. This conjecture describes how solutions to the wave equation change over time. The proof uses an additional partial differential equation that describes wave propagation. The result is also said to resolve the Kakeya conjectures and Stein’s spherical restriction estimates in all dimensions. The computer science result claims that the widely accepted Exponential Time Hypothesis, or ETH for short, is false. According to ETH, some difficult problems cannot be solved with fast algorithms. The deterministic algorithm developed precisely counts satisfying assignments in constant-width CNF formulas in subexponential time while keeping memory usage at a polynomial level. The 3-SAT case refutes ETH, and if the result is confirmed, it could lead to more efficient algorithms for databases, computational biology and geometry. The review group is led by Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis in 1999, as well as Ryan Williams. The physics claim rigorously proves that interacting quantum field theories involving gauge fields and fermions in 4 dimensions can be mathematically constructed. Put simply, it concerns describing particles and forces within a unified mathematical framework. The proof also encompasses quantum chromodynamics, the theory that describes the strong nuclear interaction. This includes a unified completion of the Standard Model. The materials have not yet been released in full. The review is led by David Gross. The group also includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. OpenAI said it will not claim academic authorship for the models’ discoveries. The experts who prepare the explanations and materials will be credited as the authors of their work. The groups are also responsible for ensuring that the scientists, methods and ideas underlying the artificial intelligence results are properly cited. The company is seeking suggestions from the scientific community on making the process more useful and transparent. At the same time, the AGMAI advisory panel advised against initially sharing the results only with selected individuals. According to the panel, such an approach could undermine equal access to the materials and create risks of exploiting prestige and biased selection. Instead, it recommended using open scientific archives, sharing proofs verified with Lean and disclosing how the problems were selected. Providing financial support to reviewers was also among the recommendations. The computer science group called for an open discussion about early access to materials, conflicts of interest, authorship and rules for verifying artificial intelligence results.
The company chose to release the results in 3 different ways. The mathematics material was published openly so that researchers could examine and explain the proof themselves. 2 new manuscripts in mathematics and computer science were posted on GitHub. For the computer science work, the answer produced by artificial intelligence and a computer-verifiable Lean proof were also shared. The expert group reviewing the physics material will decide when it will be published. The mathematics manuscript claims to prove the local smoothing result in the Sogge conjecture for 2 or more spatial dimensions. This conjecture describes how solutions to the wave equation change over time. The proof uses an additional partial differential equation that describes wave propagation. The result is also said to resolve the Kakeya conjectures and Stein’s spherical restriction estimates in all dimensions. The computer science result claims that the widely accepted Exponential Time Hypothesis, or ETH for short, is false. According to ETH, some difficult problems cannot be solved with fast algorithms. The deterministic algorithm developed precisely counts satisfying assignments in constant-width CNF formulas in subexponential time while keeping memory usage at a polynomial level. The 3-SAT case refutes ETH, and if the result is confirmed, it could lead to more efficient algorithms for databases, computational biology and geometry. The review group is led by Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis in 1999, as well as Ryan Williams. The physics claim rigorously proves that interacting quantum field theories involving gauge fields and fermions in 4 dimensions can be mathematically constructed. Put simply, it concerns describing particles and forces within a unified mathematical framework. The proof also encompasses quantum chromodynamics, the theory that describes the strong nuclear interaction. This includes a unified completion of the Standard Model. The materials have not yet been released in full. The review is led by David Gross. The group also includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. OpenAI said it will not claim academic authorship for the models’ discoveries. The experts who prepare the explanations and materials will be credited as the authors of their work. The groups are also responsible for ensuring that the scientists, methods and ideas underlying the artificial intelligence results are properly cited. The company is seeking suggestions from the scientific community on making the process more useful and transparent. At the same time, the AGMAI advisory panel advised against initially sharing the results only with selected individuals. According to the panel, such an approach could undermine equal access to the materials and create risks of exploiting prestige and biased selection. Instead, it recommended using open scientific archives, sharing proofs verified with Lean and disclosing how the problems were selected. Providing financial support to reviewers was also among the recommendations. The computer science group called for an open discussion about early access to materials, conflicts of interest, authorship and rules for verifying artificial intelligence results.