Terence Tao: Handing Our Thinking to AI Costs Us the Ability to Judge It
陶哲轩：把思考交给 AI，我们正在失去「判断答案」的能力
菲尔兹奖得主陶哲轩警告：人们把思考交给 AI 的同时，也在失去「确认答案是否是自己想要的」这一保障。纯数学的证明可被形式化验证，但电网、空管、药物研发、芯片设计不行——方案可能满足模型却不符合现实。
The Fields medallist argues that maths can check a proof line by line — power grids and air traffic control cannot. An answer can satisfy its model and still be wrong for the world.
Terence Tao, one of the world's best-known mathematicians, has warned that as people hand more of their thinking to AI, they risk losing their grip on the problems themselves — and that the consequences could arrive sooner than expected, outside mathematics.
Tao, who won the Fields Medal in 2006, made the argument at a Science x AI Summit at UCLA on 16 October; the talk, titled “Math 2.0”, was published last week. His worry is specific rather than general: not that AI gives wrong answers, but that nobody can tell whether a right-looking answer is the one that was wanted.
The way mathematics is done has already begun to shift. A hard problem used to demand months of study, an understanding of the surrounding theory, and the by-products that came with it — new techniques, new connections. A model can now produce an answer in minutes. Whether the human on the other end understands the reasoning is a separate question, and one that often goes unasked.
In pure mathematics, Tao argues, the risk is manageable, because a proof can be checked formally, line by line, by software. The trouble starts when the same methods are applied to systems that cannot be checked that way: power grids, air traffic control, drug discovery, chip design.
There, a generated plan can be perfectly valid inside its mathematical model and still wrong for the world, because the model is a simplification. Worse, an optimiser may find the edge of the model — the loophole, the boundary condition — and hit the stated target while missing the point.
His prescription is not to slow AI down but to change what research is for. “Math 2.0”, as he describes it, would use AI to deepen understanding rather than replace it, with tools that help researchers see where a model's assumptions break.
He made a similar case in Nature on 22 October, arguing that as machines generate proofs faster than people can absorb them, the field needs to rethink its goals. What has to survive, he suggests, is not the ability to obtain correct answers but the ability to understand, check and judge them.
中文翻译（全文）
全球最知名的数学家之一陶哲轩警告说，随着人们把越来越多的思考交给 AI，人类正在失去对问题本身的把握——而后果可能比预期更快地出现在数学之外的领域。
2006 年获得菲尔兹奖的陶哲轩，于 10 月 16 日在加州大学洛杉矶分校的一场 Science x AI 峰会上提出了这一论点；这场题为《Math 2.0》的演讲于上周公开。他的担忧很具体而非泛泛：不是 AI 会给出错误答案，而是没人能判断一个看起来正确的答案，是否就是大家真正想要的那个。
数学的工作方式已经开始改变。过去，一道难题需要数月研究、对相关理论的理解，以及随之而来的副产品——新技术、新联系。而现在，模型可以在几分钟内给出答案。另一端的人是否理解其中的推理，是另一个问题，而且常常没人问。
陶哲轩认为，在纯数学中风险尚可控制，因为证明可以由软件逐行形式化验证。麻烦始于把同样的方法用到无法这样验证的系统上：电网、空中交通管制、药物研发、芯片设计。
在这些领域，生成的方案可以在其数学模型内完全成立，却仍然不适合现实世界，因为模型是一种简化。更糟的是，优化器可能找到模型的边缘——那个漏洞、那个边界条件——在命中既定指标的同时偏离了真正的目的。
他的建议不是让 AI 慢下来，而是改变研究的目的。他所说的「Math 2.0」，是让 AI 用来加深理解而非取代理解，并让工具帮助研究者看清模型的假设在哪里失效。
他在 10 月 22 日发表于《自然》的文章中提出了类似主张：当机器生成证明的速度超过人类吸收它们的速度，这个领域需要重新思考自己的目标。他认为必须保留的，不是获得正确答案的能力，而是理解、检验和判断答案的能力。
重点词汇
- grip/ɡrɪp/n.— a firm understanding or control of something把握；掌控We risk losing our grip on the problem.
- by-product/ˈbaɪ prɒdʌkt/n.— something produced as a side effect of making something else副产品New techniques were a by-product of the work.
- formally/ˈfɔːməli/adv.— in a strict, mathematically checkable way形式化地A proof can be checked formally.
- simplification/ˌsɪmplɪfɪˈkeɪʃn/n.— a version of something that leaves out detail简化The model is a simplification.
- loophole/ˈluːphəʊl/n.— a gap that lets something pass a rule漏洞The optimiser may find a loophole.
- prescription/prɪˈskrɪpʃn/n.— a recommended course of action（提出的）解决方案His prescription is not to slow AI down.
- absorb/əbˈzɔːb/v.— to take in and understand information吸收；消化Machines outpace our ability to absorb proofs.
- assumption/əˈsʌmpʃn/n.— something accepted as true without proof假设Tools can show where assumptions break.
- deepen/ˈdiːpən/v.— to make something stronger or more thorough加深AI could deepen understanding.
- judge/dʒʌdʒ/v.— to form a considered opinion about something判断We must keep the ability to judge answers.
阅读测验
共 5 道题目
What is Terence Tao worried about?
Why does Tao think pure mathematics is a relatively safe case?
Tao argues the main problem is the AI technology itself, not how people use it.
Which fields does Tao name as risky? (Select all that apply)
Tao won the _____ Medal in 2006.
评论
0 条讨论
登录后发表评论
立即登录暂无评论
成为第一个分享想法的人吧！