Ask HN：通用人工智能（AGI）的标准有多高？哪些问题是“AGI 完全”的？
2 分•作者: adinhitlore•9 个月前
我猜“AGI”（通用人工智能）和“ASI”（超级人工智能）背后的命名法仍然存在争议。我刚看完桑达尔·皮查伊（在Lex Fridman的采访中）的采访，他巧妙地嘲讽了这个术语，本质上将其描述为一个新近的发明，一个主要用来表示“我们的AI很快就会非常先进”的新词。
即便如此，我的问题是：截至2025年9月，即使是最顶尖的AI也无法解决哪些问题？
我可以想到很多，但由于它们目前尚未解决，因此对任何人或人类工具来说，它们也可能都是不可能的：
*以下是一些随机排序的“随口”问题：
1. P ≠ NP；
2. 广义相对论与量子力学的统一；
3. 证明/证伪黎曼猜想；
4. “孤独跑者猜想”的证明/证伪；
5. 治愈四期癌症，且大多数情况下有效；
6. 数量级上更冷的核聚变；
7. 无反作用力驱动——证明/证伪；
8. 证明（展示）/证伪“量子霸权”；
9. 在量子计算机上安装用户友好的主要用途操作系统；
10. 类人机器人击败UFC格斗选手；
11. 类人机器人比Storror更好地进行跑酷；
12. AI设计出比Transformer更快数量级的文本生成/媒体生成算法，且不牺牲质量（或偏差仅差5-10%）；
13. AI制造出自复制机器人，在火星上建造城市（这个想法是我从Grok那里偷来的...）；
14. 找到制造廉价体外培养肉的方法；
15. 合成ununennium（119号元素）；
16. 对量子力学的正确解释；
17. 解决“时间问题”；
18. 制造胶球；
19. 证明/否定考拉兹猜想；
还有很多... 你认为哪些问题是类似AGI的问题？
我知道我上面对问题的措辞并不完美，例如“廉价”在体外培养肉中是开放的解释，而100万度的聚变仍然比实际需要的温度低大约2个数量级，但可能对主流聚变没有影响等等。但毕竟这是一个互联网论坛，而不是博士论文，无论如何，如果任何问题得到解决，尽管措辞糟糕，都将标志着AI或任何它“标志”的东西发生巨大的范式转变，有人/某些人将做得非常正确，甚至超越Demis Hassabis，甚至爱因斯坦、达芬奇、特斯拉、Ada等人。
我很好奇很多人希望从AI中看到什么，由于这个论坛的智商很高，我猜它不会是像“逼真FPS”或“听起来像Shelby的特斯拉”或“更美味的汉堡”、“超酷的喷气背包”、“可以采访Andrew Tate的完全沉浸式VR”、“敞篷波音”等等之类的“比佛利和巴特黑德”级别的答案。
查看原文
I guess the nomenclature behind "AGI" and "ASI" is still debatable and i just watched an interview with Sundar Pichai (on Lex Fridman) where he brilliantly almost mocked the term, essentially describing it as a recent invention, neologism mostly used to mean "our ai will soon be very advanced".<p>Even so my question is: what problems as of 2025 September are impossible for even the best AI?
I can think of many, though since they are currently unsolved, they may as well be impossible regardless for any human or human tool anyway:<p>*here are some problems ordered randomly, "on top of my head":<p>1. P =|!= NP;
2.Unification of gen. relativity+quantum mechanics;
3. prove/disprove the Riemann hypothesis;
4. "lonely runner conjecture" proof/disproof;
5. cure for 4th stage cancer that works most of the time;
6. orders of magnitudes colder fusion;
7. reactionless drive - proof/disproof;
8. prove(show)/disprove 'quantum supremacy';
9. install user-friendly main purpose OS on quantum computer;
10. humanoid defeats ufc fighter in a fight;
11. humanoid performs parkour better than Storror;
12. AI designs orders of magnitude faster textgen/mediagen algorithms than transformers without sacrificing quality (or 5-10% worse deviation);
13. AI builds self-replicating robots that build city on Mars (i stole this idea from Grok month ago or so...);
14. discovers way to make cheap invitro meat;
15. synthesizes ununennium;
16. correct interpretation of quantum mechanics;
17. solve the 'problem of time';
18. make glueballs;
19. prove/reject the Collatz conjecture;<p>and many more...what in your opinion are agi-like problems?<p>I know my phrasing of the problems above isn't perfect, for example "cheap" as in invitro meat is open to interpretation and 1 million degrees fusion is still ~2 'orders of magnitude' colder but likely useless to have impact on mainstream fusion etc. But it's an internet forum after all not Ph.d thesis, regardless any of the problems if solved despite the horrible wording will signal an immense paradigma shift in AI or whatever it "signals" someone/some people will be doing something very right surpassing even Demis Hassabis, or even Einstein, Da Vinci, Tesla, Ada and the likes.<p>I'm curious what many people want to see from AI, since the IQ of this forum is high, I guess it won't be a beavis-and-butthead level of answers like "photorealistic fps" or "teslas that sound like shelby" or "tastier burgers", "badass jetpacks" , "fully immersive VR where you can interview andrew tate", "cabrio boeing", etc etc.