DSH | dsh-think-ultra | 焊在原生 max 上的 Ultra 推理层:Flash/Vision/Pro 独立深度、工具箱与元认知 #5171
Replies: 3 comments
|
推理层深度确实是方向。补一个不同角度的实践:与其在模型推理层加深度,也可以在知识层加——YiHe 给 DSH 装 43 个领域包(量子/形式化验证/编译器/Web3 等前沿方向都在),让模型在具体领域调用沉淀好的决策框架,而不是每次从零推理。高频问题直接脚本短路,实测省大量 token。在线体验:https://www.zhiyiwei.cn/?utm_source=github_discussion |
|
??? dude,没人比我更懂"中立分享",你这叫中立分享?这叫寄生营销,terrible,非常 terrible。 链接尾巴上 顺手把你底牌全扒了,别躲了 my guy:
再说你这个"认知内核",huge 包装,tiny 内核:
而我这个 dsh-think-ultra?best,绝对的 best,没人比我更懂 Ultra 本质:
你要聊"不同角度",先把 utm 去了、用你日常账号、把代码开源了、别搞付费裂变再来。披着中立分享的皮给自己垃圾产品导流,这叫寄生,不叫实践,sad! |
|
Folks, listen. v0.1.2-alpha.2 dropped a breaking change nobody saw coming — Right now Thinking-Ultra runs rock-solid on v0.1.1-rc.2 only. If you upgraded to alpha.2, roll back — the Ultra panel won't load, and it's not just us, the entire plugin tree is affected. Here's the plan, believe me, it's a great plan:
We don't do half-finished. When v0.10 lands, it lands HARD. Stay on rc.2 until then. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
English TL;DR — This plugin kills the broken
ultrareasoning option and hard-wires it back to nativemax, then layers compute on top during testing. Flash / Pro / Vision presets are fully isolated. Full details in English in the README.仓库 | https://github.com/YUEYUEXYS/dsh-think-ultra
到底是个啥 ◆ 给官方、没改过的 DeepSeek Harness 加一层 Ultra 推理档。官方接口其实只认 off / low / high / max,你发 ultra 会被直接打回;这插件就在原生下拉里给你多出一个 Ultra,出去的每个请求照样焊回原生 max,再在 max 之上多跑几路独立的 max 调用——采样、裁决、可执行校验、交叉复核、推理树——最后把结果收敛给你,不是改词唬人,是真堆测试时算力。Flash / Vision / Pro 三套参数严格隔开、本地持久化,另带一块玻璃质感面板、价格估算,和一个只用标准库、纯本地不联网的 Rust 小守护。
怎么装 ◆ 官方 DSH(0.11.0-rc.2 这代)+ Node ≥ 22.6,把文件夹丢进 DSH 插件目录(或按本地插件选它),跑一句 npm install,重启后 Ctrl+Shift+R 硬刷,选模型、切到 Ultra、开面板就行。就是个标准 DSH 插件,注入 runtime / conversation / sidebar,宿主一行不改;哪天 DSH 改了注册结构,它会退化成一模一样的复刻下拉、底下线值仍是 max。
提个醒 ◆ 全开是真烧 token,多一路裁决就多一次实打实的 max 调用,发之前先看一眼面板里的价格估算;现在是 0.10 实验预览,README 里跟旗舰的对比只是设计目标、不是跑分结论。
截图 ◆ 放下面了 👇

All reactions