Opus 4.5 评审（定制计划）
1. 严重的幻觉
它无法正确编写任何文档。我让它为我正在构建的解释器编写API文档。它虚构了不存在的函数，并且所有参数列表都错误。像 `shouldClose`、`swapBuffers`、`pollEvents`、`terminate` 这样的函数根本不在我的解释器中，但它却把这些函数放进了文档，并且每个参数都错了。
2. 擅长创建重复的函数
在C语言中，重复的函数名无法编译，但它似乎大约60%的时间都会创建重复的函数。它甚至在 `switch` 语句中创建重复的 `case` 标签。它常常在不同的名称下生成相同的代码。
3. 出色地制造错误
它创建 `struct`，初始化它们，赋值，然后当被要求初始化并运行或实现SSR反射时，它将alpha设置为全透明，导致整个GLB变得透明；或者当被要求制作玻璃材质时，它只是将其设置为不透明（我尝试了15次后放弃了）。他们声称它达到了专业工程师的水平——什么，幼儿园水平的工程师？
4. 从不提供真正可用的代码
即使是读取单个文件的代码也是一团糟，充满了错误；自己实现要快得多。它在这里那里做了一些空值检查，但也许应该先检查其他内容，而不是在这些上浪费时间。
5. 严重的内存泄漏
它生成的代码存在内存泄漏。使用Claude Opus 4.5让我感到非常沮丧，以至于我在提示中写了很多脏话。
6. 添加我没有要求的内容
它在阴影中添加噪声，声称这样看起来更自然，但实际上只是显得脏乱。我甚至看到它通过在地板上放置一个白色平面来“修复”反射，并称之为光反射。
7. 莫扎特的骰子级别的创作
想要得到可靠的代码真是太难了——一切都是缺陷重重、支离破碎的混乱。
8. 只要它不崩溃，这就是最重要的
如果一个函数应该返回一个浮点数，那就把它当作整数比较——所以把函数的返回类型改为整数！Claude Opus 4.5真是太神奇了！它可能很快就会达到AGI的水平！
查看原文
1. Severe hallucinations<p>It can’t write a single document correctly. I asked it to write the API documentation for the interpreter I’m building. It invented functions that don’t exist and got all the parameter lists wrong.
Functions like `shouldClose`, `swapBuffers`, `pollEvents`, `terminate` aren’t even in my interpreter, yet it put them in the docs and got every parameter wrong.<p>2. It’s great at creating duplicate functions<p>In C, duplicate function names won’t compile, but it seems to create duplicates about 60% of the time. It even creates duplicate `case` labels in `switch` statements. It often produces identical code under different names.<p>3. It’s excellent at producing bugs<p>It creates `struct`s, initializes them, assigns values, then when asked to initialize and run or to implement SSR reflection it sets the alpha to full so the entire GLB becomes transparent, or when asked to make a glass material it just makes it opaque (I tried 15 times and gave up). They claim it’s at a professional engineer level—what, kindergarten-level engineers?<p>4. It never gives code that actually works<p>Even code to read a single file is a bug-ridden mess; it’s faster to implement it yourself.<p>It does null checks here and there, but maybe it should check other things first instead of spending time on those.<p>5. Serious memory leaks<p>It generates code with memory leaks. Using Claude Opus 4.5 made me so frustrated that I wrote a lot of profanity into my prompts.<p>6. It adds things I didn’t ask for<p>It adds noise to shadows claiming it looks more natural, but it just looks dirty. I’ve even seen it “fix” reflections by placing a white plane on the floor and calling that light reflection.<p>7. Mozart’s Dice–level composition<p>Good luck getting solid code—everything is a defective, patchy mess.<p>8. As long as it doesn't crash, that's all that matters<p>If a function is supposed to return a float, just compare it as an int—so change the function's return type to int! Claude Opus 4.5 is truly amazing! It’ll probably reach AGI soon!