DeepSeek V4 与 DSH：从 TypeScript 到 AGI

> Haskeller understandable. By the way. time for my statement on DSH. Consider it a non-technical view from 100K feet; I am a bit high. I wouldn't overindex on "TypeScript". It's transient, just like V4's specific clunky design choices (hash router, softplus gating…); it's half a tribute to the genuineness of the talented contributor who joined the team, half a casual shortcut to speed up the onboarding, and 120K stars on Gitbub in 4 days all but ensure this gambit succeeds. Most of those stars are dead. But they don't want more "users". They want to find 99.999th percentile friends who will see it for what it is. It's not Cordis. Cordis is Cordis, a proof of concept born out of a mundane necessity. DSH is a mission statement that will reverberate all the way down the abstraction ladder; to Rust, to kernels, to the finite automata, to DeepSeek-branded hardware that V5+ will help design the substrate for. Plenty of room at the bottom! Once they have their guarantees and sufficiently polished models, this runtime will begin rewriting itself into ever lower-level mechanisms. The whole "V4" is a very philosophically radical AGI project that far exceeds the peasants' "cheap and almost as good… but meh, not GPT-good! Not even Kimi- or GLM-good! wait why isn't it cheap? boo!" perception, to say nothing of programmer bros who feel status anxiety and dunk on math/CS showboating in the paper. Faced with resource scarcity, the Whale responds not so much with some "efficiency innovation" as with escalating the ambition, diving ever deeper into the unknown. From V4's entirely deterministic, bitwise-batch-invariant, training-inference-aligned rollouts across hundreds of thousands of sandboxes, to the "provably reversible infinitely extensible meta-harness [oops limited to persistent bash and string editor lol]" doctrine, it is one single pure note which is almost impossible to hear yet; you need to be a sage, a zhenren with an ear attuned to the will of the Heaven, like Liang Sheng, to notice; or his arguably sane fanboy to infer from breadcrumbs. It's closer to the parallelized vision of a Gödel machine without the learning-hostile demand for provable improvement than to another Claude Code or OpenClaw killer; it's emacs that's actually, finally, alive; it's @SchmidhuberAI's (again) Neural Computer that drops the everything-in-the-weights stretch goal which betrays explicit algorithms for nothing; it's even @allTheYud's Flare, only without skill issues. It's the true and good, economically viable, executable, neurosymbolic endgame of language models, the glimpse of the Holy Grail of CS, and the Seed of the Mind that can power a civilization of human-machine symbiosis, which automatically dissolves all problems into composable algorithmic puzzles to solve. Ah, right. Liang argues that continual learning is the last hurdle before the endgame of "gradual singularity"; the last thing they'll need to do before this thing makes embodied AGI in their stead. How this is relevant to DSH: left as exercise for the reader. Few will understand yet, or accept… Oh well, that's what DeepSeek (杭州深度求索人工智能基础技术研究有限公司) is about from the start. 路曼曼其脩遠兮， 吾將上下而求索。

I should care less about how the xiaoren react. Because he doesn't. As was said 2300 years ago and again on the V4-Preview launch this April:「不诱于誉，不恐于诽，率道而行，端然正己。」