机器证明技术的发展得益于数十年来的数学家和工程师的共同努力。由于机器算力的可扩展性,由人类完全验证人工智能的产出,已经显得困难,形式化验证器参与到人工智能的闭环得到了数学家和多个方面的认可,是当前最有希望通过人工智能治理人工智能的一个关键节点。多名数学家批评一些大公司使用人工智能参与数学研究缺乏规范性,违背学术研究的价值准则。
1730万行的Lean代码由一个人工智能厂商的数学研究员披露。这包括约17万个模块。这个证明宣称已经是对黎曼猜想的证明。目前发布后第二天为止,还没有任何职业化的数学家已经完全而确定地完成了这些证明的评估。机器证明需要首先正确地描述问题,且在机器证明语言的范围内给出证明,如果对问题的描述是不正确的,或者引入了不正确的、并不成立的“定理”,都会导致编译通过,但证明并不成立。此外,还可能有证明器的内核错误此类的问题,已经为社区给出。Lean也承认多种错误的可能。但Lean还是仅仅在这1730万行代码发布后的不到两天内,就宣称对其“认证”。
Lean还是选择混淆了形式化上的编译通过和检查器运行结果和在数学的严谨意义上的证明。Lean并不严肃地看待自己对黎曼猜想这么大的一个定理的证明,将数学家的态度和潜在的错误致于不顾,以回避数学的严格性的语言单方面声称对结果的“认证”。我们不能恶意猜测这与其为科技公司的研究机构的研发产品直接有关,但这显然不是一个对待数学、人类社会、人工智能的发展与治理足够负责任的做法。
数学家在两次世界大战后成为许多人掠夺声誉、谋取私利的重要手段,一些封闭的做法也让公众难以满意,一些人回避科研人员的责任,将自己夸大为“最聪明的阶层”,对社会发展、人类社会、科学普及、公众参与数学和科学的权利持一种敌视的态度,这其中没有什么高尚的东西,但是我们仍然相信数学的严谨方法有着独特的价值——我们不必活着虚假的世界中。
Lean背叛了真理。从结果来看,我们还并不知道这1730万的代码是否真的有数学上的缺陷,我们也还不知道人类数学家没有使用人工智能的版本是否就必然没有缺陷,但是从过程来看,Lean现在的选择性的“认证”是不负责任的。
Lean应该感到可耻!
The development of machine proof technology is indebted to the joint efforts of mathematicians and engineers over decades. Owing to the scalability of machine computation, it has become difficult for humans to fully verify the output of artificial intelligence. Formal verifiers taking part in the loop of artificial intelligence have gained recognition from mathematicians and others, and are currently a key node with the greatest promise for governing AI through AI. Several mathematicians have criticized some large companies for using AI in mathematical research without standards, in violation of the values of academic research.
17.3 million lines of Lean code were disclosed by a mathematics researcher at an AI company. This includes about 170,000 modules. The proof claims to already be a proof of the Riemann conjecture. As of the second day after its release, no professional mathematician had yet fully and completely and conclusively completed an evaluation of these proofs. Machine proof requires first describing the problem correctly, and then giving a proof within the language of machine proof. If the description of the problem is incorrect, or if incorrect and untenable "theorems" are introduced, compilation will pass while the proof does not hold. There may also be problems such as errors in the kernel of the prover, which the community has already reported. Lean also acknowledges the possibility of various errors. Yet Lean still claimed to "certify" the code in less than two days after the release of these 17.3 million lines.
Lean still chooses to conflate formal compilation passing and checker results with proof in the strict mathematical sense. Lean does not take its own proof of a theorem as large as the Riemann conjecture seriously. It disregards the attitude of mathematicians and potential errors, and unilaterally claims "certify" for the result in language that sidesteps mathematical rigor. We cannot maliciously speculate that this is directly related to its being a product developed by the research institute of a technology company, but this is clearly not a sufficiently responsible way to treat mathematics, human society, and the development and governance of artificial intelligence.
After the two World Wars, mathematicians became an important means for many people to plunder reputation and seek private gain. Some closed practices have also left the public dissatisfied; some people evade the responsibilities of researchers, exaggerate themselves as "the most intelligent class," and hold a hostile attitude toward social development, human society, the popularization of science, and the public's right to participate in mathematics and science. There is nothing noble in this. But we still believe that the rigorous methods of mathematics have a unique value — we do not have to live in a false world.
Lean has betrayed the truth. Judging by the results, we do not yet know whether these 17.3 million lines of code actually have mathematical defects, nor do we know whether the versions by human mathematicians without artificial intelligence are necessarily free of defects; but judging by the process, Lean's current selective "certify" is irresponsible.
Lean should be ashamed!