What are the limits of the AI mathematician? - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号):
请您阅读我们的用户注册协议隐私权保护政策,点击下方按钮即视为您接受。
人工智能

What are the limits of the AI mathematician?

If models can learn to master complex calculations, they could solve problems that have so far eluded us
00:00

{"text":[[{"start":6.74,"text":"The writer is a theoretical cosmologist at the University of Cambridge and director of the Infosys-Cambridge AI Centre "}],[{"start":15.13,"text":"Mathematics was once assumed to be relatively safe from the incoming juggernaut of artificial intelligence automation. Chatbots might be able to generate text, code and images on demand, but the deep reasoning required for mathematics was supposedly out of reach. The gold medals that OpenAI and DeepMind recently achieved at the International Mathematical Olympiad have therefore left maths professors like me feeling suddenly a little less safe. "}],[{"start":48.28,"text":"Is AI about to do to mathematical proofs what it’s already doing to coding? After all, the two have clear similarities: both are highly structured “languages” with clear conventions and restricted “dictionaries”. Both have large corpora of examples on which AI can be trained with known solutions.  "}],[{"start":72.57,"text":"Yet while the results from cutting-edge AI maths models are impressive, there is another class of maths that generative AI still struggles with: simple computation. Ask “what is 5.11 minus 5.9?” and the answers vary. This morning, OpenAI’s latest GPT5 model gave me the correct answer of -0.79. But phrase the question as part of a calculation and you may receive a different answer."}],[{"start":102.72,"text":"What should we make of AI models that can outperform high school-age Olympiad competitors but cannot always add or subtract to primary school level? To understand this, it’s helpful to think about what it means to be good at maths."}],[{"start":117.77,"text":"The way maths is taught is by showing students a problem, demonstrating the method required to solve it and then assigning examples. Weaker students require numerous examples and sometimes end up simply memorising the method without understanding it. The strongest students need only one or two examples to master the concept and apply it to new problems."}],[{"start":142.35,"text":"The ability to conceptualise and generalise distinguishes the best mathematicians. Good mathematicians solve hard problems; great ones find ways to make the hard problems easy."}],[{"start":156.35,"text":"The strengths of AI models lie in their speed and ability to “practise” at extremely high volumes. This means they can solve very difficult problems that bear some resemblance to things they have been shown before but may struggle when given something new. This is particularly a problem for theoretical maths. The number of examples available for training drops as you move towards more advanced problems."}],[{"start":183.04999999999998,"text":"These are well-known issues with neural networks. They are great at interpolation (generating answers that are “between” things they’ve seen before) and bad at extrapolation (generating answers that fall outside their training set)."}],[{"start":197.63,"text":"In maths, this is made extra difficult by problems that sound similar. Consider: “What is the maximum number of cubes of volume 1 that you can fit in a cube of volume 64?” and “What is the maximum number of spheres of volume 1 that you can fit in a sphere of volume 64?”. They sound alike but one is simple to solve (cubes fit together neatly in a 4x4x4 block), while the other is fiendish (spheres do not stack nicely)."}],[{"start":231.82,"text":"What this means is that AI use in applied mathematics and cosmology is still limited. We can take things we already know how to do and use AI to automate them. But so far, calculation has seen little advancement."}],[{"start":247.35999999999999,"text":"It is possible, however, that more training will solve the problem without extrapolation ever being required. If AI models can be fed enough complex calculations they could perhaps solve problems that have so far eluded us without the need for any human-level inspiration."}],[{"start":267.84,"text":"The question being asked in my field is: “How powerful is an extremely fast, extremely well-trained, unthinking mathematician?” We are in the process of finding out."}],[{"start":null,"text":""}],[{"start":287.28,"text":""}]],"url":"https://audio.ftmailbox.cn/album/a_1755773104_9805.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

哈梅内伊排除与美国政府直接对话的可能

伊朗最高领袖哈梅内伊态度强硬,指责美国意在迫使伊朗屈服,并称主张与美国直接谈判的伊朗政界人士“肤浅”。

私募股权集团KKR支持的音乐节因巴勒斯坦旗帜问题遭到抵制

多支乐队因主办方禁止现场展示巴勒斯坦旗帜而选择退出,主办方随后“诚挚道歉”。

汇丰瑞士私人银行清退部分中东客户

此前瑞士监管机构认定该行在反洗钱审查方面存在疏忽,禁止其接纳高风险客户。

决策者警告:富裕经济体将需要外籍劳工推动增长

央行人士称,全球最大经济体的低生育率正威胁生产率与物价。

中国科技亿万富翁欲打造美式“3月疯狂”风格的篮球联赛

在阿里巴巴亿万富翁联合创始人蔡崇信的支持下,亚洲大学生篮球联赛瞄准业余赛事的高利润市场。

央行精英的黄昏

在经济技术官僚享有数十年高度自主权之后,他们如今正承受来自特朗普政府的巨大压力。
设置字号×
最小
较小
默认
较大
最大
分享×