來自 Taelin（@VictorTaelin）的推文串

DeepSeek V3.2 Speciale one shot the shortest solution to my 3rd (currently, hardest) λ-calculus problem. Current solutions, from worst to best. Opus 4.5: rotL = λN.λT.λt.(T λk.(λn.λs.λz.(n λg.λh.(h (g s)) λu.z λu.u) N λu.λr.λk.λe.(u (r e) k) λr.λk.(r k) t k)) Gemini 3 Pro / GPT 5 Pro: rotL = λn.λt.λf.(t λx.(n λg.λh.(h (g λp.λg.λy.(p (g y)))) λu.λg.(g x) λu.u f)) DeepSeek V3.2: rotL = λN.λT.λc.(T λa.(N λf.λg.λx.(f (g x)) λh.(h a) λz.z c)) All are correct. I find this curious. A few months ago, no model solved this. Now all recent ones (except GPT 5.1) seem to. I'll make harder problems! Worth noting that DeepSeek failed on the stack underflow bug, which only Opus 4.5 and a non public checkpoint of Gemini 3 Pro solved.