Skip to content

Week 2:AI使用

理性看待AI的数学能力

计算类题目

AI完全可以做对

已知函数 \(f(x)=e^{2x}\), \(g(x)=\ln x+\frac{1}{2}\) 的图象分别与直线 \(y=b\) 交于 A, B 两点,则 \(|AB|\) 的最小值为 ( )
A. 1
B. \(e^2\)
C. \(\frac{2+\ln 2}{2}\)
D. \(e-\frac{\ln 3}{2}\)


AI回答:

The minimum value is \((2 + \ln(2))/2\), which corresponds to option C.

The final answer is C.
52.6s

证明类题目

不要把AI当作参考答案,而应当把他当作一个高级的搜索引擎。AI最喜欢不懂装懂!!!不要轻易相信AI在证明题上的能力。

如deepseek声称证明了哥德巴赫猜想...

部分回答摘要

素数间隙定理 (Zhang, Maynard-Tao) : 存在无穷多对素数 (\(a,b\)) 满足 \(|a-b| \leq 246\), 排除区间内无素数的可能。

反证法矛盾排除: 若存在无法分解的偶数 \(k\), 将直接违背上述定理。

结论

通过数学归纳法、命题1的严格应用及已证数论定理,可得出结论:哥德巴赫猜想对所有偶数 \(n \geq 4\) 成立,即每个偶数 \(n \geq 4\) 均可表示为两个素数之和。

最终答案:
是的,通过命题1的严格应用及数学归纳法,结合Bertrand-Chebyshev定理与素数间隙定理,可证明哥德巴赫猜想成立。

个人使用方法

提问题

自己的思考以及遇到的障碍包含题目发给 AI,让 AI 给你一点提示

注意点

千万不要直接让 AI 给答案 然后复制粘贴到笔记 没有用!!!

动手写

要自己写 自己写的时候你会遇到很多意想不到的问题 和自己与标答不一样的处理

让 AI 批改

学校题目练习作业是没有答案的 因此做完题目完全可以用 AI 检查

然后不会的看 AI 的评语 (不要完全相信 多去和教授交流为准)

举例

问 AI 如下内容

题目:今天是星期一,那么 \(3^{23^{57}}\) 天后是星期几?

这道题目按照经验最简单的办法就是找循环 我发现 \(3^i\) 次方 取模 \(7\) 会出现 \(6\) 个余数 为一个循环 那么我就需要讨论 \(23^{57}\) 次方 除以 \(6\) 余几 但是太复杂了 我不知道怎么办 了 给我一点提示


AI:

要计算 \((23^{57} \pmod 6)\),我们可以先简化底数:
\(23 \pmod 6 = ?\)

想一想 \(23\) 除以 \(6\) 余几。
\(23 = 3 * 6 + 5\)
所以,\(23 \equiv 5 \pmod 6\)

那么,\((23^{57} \pmod 6)\) 就等价于 \((5^{57} \pmod 6)\)

使用AI的障碍与解决方案

手写文字识别问题

有些思考模型不支持多模态,就算支持多模态 也会有概率识别错,导致问AI题目,检查答案有很大阻碍

  • 优先使用Latex写数学笔记 作业 (门槛偏高 但是美观易于传播)
  • 退一步可以学习Latex基本符号表示 让AI先转文字 然后校对 再扔给思考模型(门槛低 但不推荐 耗时耗力)
  • 如果不想学Latex基本符号 那么就尽量字写好看点提高识别准确率吧

模型受限问题

国内目前可用的只有deepseek 但是deepseek 1. 幻觉 率极极极极极高 2. 思考时间极极极极极长

所以要学会使用代理 接触ChatGPT/Grok/Claude/Gemini等等AI模型

推荐 AIStudio里的Gemini 免费+额度高 其次ChatGPT思考模型

示例 传入Latex作业代码 让AI批改

Prompt:

correct my homework and tell me whether my solutions are correct below each problem. In the end, give me a summary of my accuracy

  1. (30 points)
    Show if the following sets are domains or not:
  2. \(A = \{ (x,y) \in \mathbb{R}^2 : x^2 + y^2 > 1 \}\).

Yes

AI批改结果
  • Problem 1.1: Correct.
  • Problem 1.2: Correct. (Minor improvement point for openness proof style, but logic is sound).
  • Problem 1.3: Correct.
  • Problem 2.1: Correct.
  • Problem 2.2: Correct.
  • Problem 2.3: Correct.
  • Problem 3: Correct.
  • Problem 4: Correct. (Minor clarification on \(\epsilon-\delta\) for the exponential part, but result is correct).

附录 介绍Latex

推荐尝试接触latex类笔记 字丑的非常推荐 利于自己复习笔记
可以是直接overleaf打latex代码 也可以使用markdown editor
例如notion, obsidian, 思源笔记等 这类markdown编辑软件支持katex(轻量版latex 语法完全相同)公式输入


下面是我自己的类latex笔记摘要

Now, set \(W_2=\lang v_1,v_2\rang\). Again if \(W_2=V\), we are done. If not, that is \(W_2\subsetneq V\) we have to continue

In this way, suppose we have obtained a subspace \(W_i\) where \(i\geq 2\) such that \(W_i\) has basis \(\{v_1,v_2,...,v_i\}\) and \(\forall j,1\leq j\leq i,f(v_{j})\in \lang v_{1},...,v_{j-1}\rangle\)

If \(W_i=V\), we are done. If not, \(W_i\subsetneq V\), using the lemma again, we find a vector \(v_{i+1}\notin W_i\) and an eigenvalue, \(c_{l_{i+1}}\) eigenvalue of \(f\) such that \(f(v_{i+1})=c_{l_{i+1}} v_{i+1}+\beta_1v_1+...+\beta_iv_i\) for some \(\beta_1,...,\beta_i\in k\)

Note that \(\{0_{V}\}=W_{0}\subsetneq W_{1}\subsetneq W_{2}\subsetneq\ldots\subsetneq V\) and \(0=\dim W_{0}<\dim W_{1}<\ldots<\dim V\) finite

So the produce must stop at some \(l=\dim V\) and thus we get a basis \(\mathcal{B}=\{v_1,...,v_l\}\) such that \([f]_\mathcal{B}\) is upper triangular

\(\Rightarrow\)) If \(\exists\) a bsis \(\mathcal{B}\) of \(V\) such that \([f]_{\mathcal{B}}= \begin{pmatrix} a_{1} & & \star \\ & \ddots & \\ 0 & & a_{l} \end{pmatrix}\), so \(\chi_{f}(x)=\prod_{i=1}^{l}\left(x-a_{i}\right)=\prod_{i=1}^{r}\left(x-c_{i}\right )^{m_{i}},c_{i}\neq c_{j},if\,i\neq j\) and \(m_i\geq 1,\forall i\)

Then by C-H theorem, \(m_f(x)\mid \chi_f(x)\), so \(m_{f}(x)=\prod_{i=1}^{r}\left(x-c_{i}\right)^{n_{i}},m_{i}\geq n_{i}\geq1\) since the roots of \(m_f(x)\) and \(\chi_{f}(x)\) are same