Eric Weinstein给OpenAI出了10道人类未解数学题，说AI现在只会解离散题，碰不了理论物理核心。看看吧，看他能不能被AI打脸。
Eric Weinstein发文称，AI在纯数学上的表现被严重高估，距离顶级人类学者仍有距离。他向OpenAI提出10个未解决的数学物理难题作为新基准，包括寻找新的上同调理论、构造不矛盾的大统一理论、解释三代费米子等。他认为AI目前只会搜索反例或解离散题，无法处理理论构建式的深层问题。
exactly. math isn’t done. not at all.
exactly. math isn’t done. not at all. Eric Weinstein @ericweinstein I don’t think being critical of the amazing work AI is doing in pure math is fair to @OpenAI until I can start to say why I feel it’s not yet at the level of our top humans. It’s like AI has a style that is weirdly being taken to be universal (which it isnt). Hence the claim that Mathematians and Physicists are nearly obsolete. Ahem: there is no *public* indication that this is REMOTELY true. Yet. I would like to pose a collection of math physics benchmarks that are not about searching large spaces for counter examples and the like. These are 10 problems and issues that humans haven’t solved but which we likely will solve. Each feels to me like a deeply human math/physics problem. This isn’t an unfriendly challenge to @sama and team. It’s just pointing out that AI has yet to deeply touch any of the kind of things that many of us care MOST about in Mathematics and Physics. I’m just not a puzzle solver or a discrete conjecture guy. I’m into theory building and even the mathematical version of mysticism. Here goes an attempt: ————- A) What new Extraordinary Cohomology Theories/Co-cycles can AI find that we need, but do not know/understand? B) What areas of pure mathematics and theoretical physics are most polluted by human blind spots? Where is the training corpus most wrong? Show us you can go well beyond us by contradicting your training. C) What discarded ideas were discarded improperly and prematurely? D) What is a natural version of Grand Unification that doesn’t contradict experiment? E) Find a natural version of Super-Symmetry that doesn’t predict Superpartners and that predicts unification scale different from the Minimally SuperSymmetric Extension or the Standard Model. F) Pose a new, deep and surprising conjecture relating 3 or more of the following: representation theory, number theory, arithmetic/algebraic geometry, and analysis on manifolds, homotopy theory and Pseudo-Riemannian Spin Geometry. G) Integrate the Octonions, Magic Square and the 4 large exceptional lie groups, tightly into the mainstream of mathematics once and for all. H) Settle the issue of the 6 sphere and complex structures in Math or the Koide fine tuning mystery in particle theory. I) Explain the reason for there being 3 families of observed Fermions. J) Give a deep, natural and as yet unknown reason for the remarkable Fermi-Bose correspondence (e.g. Metaplectic vs Spin reps, Honest vs Berezin Integrals, etc.). Why should we have expected this correspondence to be this weird and this tight? ——— I don’t think humans are going to be ahead forever. AI will blow past us. So while I am not all that impressed yet at the math or physics, any non-tricky solutions to one of the above (with the possible exception of some answers to H) would feel different to me than the feeling I get from AI finding yet another funky counter examples to some Erdos problem, discrete puzzle or Combinatorial issue in Graph theory. We have always been humbled by Gauss, Galois, Dirac and Newton. That’s nothing new. If this is going to put us out of work soon, LFG. May the best mathematicians and physicists, carbon or silicon, win. 🔗 View Quoted Tweet 💬 0 🔄 0 ❤️ 3 👀 1295 ⚡