九原客 (@9hills)

Follow-up to my GLM vs Opus thread: let's talk cost. We ran 103 dbt tasks x 3 trials on each model. Same harness, same tasks. GLM: 860M tokens Opus: 439M tokens That's ~2x. But the "why" is more interesting than the number.

🤣🤣🤣看评论区就知道为什么国产模型在商业上依然很不行了。 真实世界里的用户能有几个知道什么叫多模态的，能有几个知道什么叫参数量的，能有几个知道什么叫上下文窗口的，用户不关心这些，用户只关心为什么我让AI帮我干这个事情AI做不了。 从这个角度来说豆包确实是国产AI里综合能力最好的。

Introducing Claude Tag, a new way for teams to work with Claude. In Slack, Claude joins as a team member with access to the channels and tools you choose. Tag Claude in and delegate tasks to it while you focus on other work.

🚨 SCOOP(s): - GPT-5.6 has been delayed and will no longer release this week. New target is ~mid-July. - DeepMind are not satisfied with the current state of 3.5 Pro and it will no longer launch this month. - Preparations for the launch of Bidi, OpenAI's new voice model, are underway in ChatGPT and we could see it available as soon as this week. - Claude Sonnet 5 is currently available for select enterprise customers under an Early Access Program and is seen as a stop-gap as progress on getting Mythos/Fable 5 back out have stalled. A bit of a disappointing end to the month, but July should prove more fruitful!

Been iterating on @tomosman's loop. This one's winning: /goal produce a verified, code-derived behavioral spec for this web platform, captured in one canonical spreadsheet that carries every feature from spec -> tested -> fixed -> verified. Why: we need a single source of truth that maps every feature to its expected behavior *as the code implements it*, so that gaps and bugs surface and the platform can be driven to a known-good state. The spreadsheet is the source of truth. Work on the current repo. Do Phase 0 and Phase 1 under this goal; when the spec is complete, switch into the /loop below to drive testing and remediation. Keep moving through phases without stopping, except at a real checkpoint (defined below). Phase 0 - Plan (first): Detect the stack, the feature surface (routes, pages, components, API endpoints, background jobs, auth, settings…), and the test infra that already exists (unit/integration/e2e, browser automation, seeds/fixtures, a runnable dev server). Propose (a) how you'll inventory features, (b) the spreadsheet schema, and (c) how you'll test in the loop given what's available. Proceed once the plan holds. Phase 1 - Catalog & spec: Read the code and, for every feature, write a user story + the expected behavior as implemented, citing the file/function. Where the code is ambiguous, or behavior is undefined, log an open question - don't guess. Record every feature as a row in the canonical spreadsheet (create with the xlsx skill). Exit: every discoverable feature has a row. One row, concretely: | Area | User story | Expected behavior (from code) | Status | Defects | Type | Notes / source | |---|---|---|---|---|---|---| | Auth | As a returning user I want to log in with email+password so I can reach my dashboard | `POST /api/login` validates via bcrypt, sets httpOnly session cookie, 302 -> `/dashboard`; bad creds -> 401 + inline error | Spec'd | - | - | `api/auth/login.ts`, `LoginForm.tsx` | Canonical artifact: exactly one .xlsx, updated in place across every phase and loop iteration - never fork into per-phase or per-iteration files. Status flows Spec'd -> Tested-Pass / Tested-Fail -> Fixed -> Verified. The main thread is the single writer. Agentic execution: - Delegate breadth to subagents: fan feature discovery and per-area testing across subagents so the main thread stays focused. - Verify by running, not claiming - report real command/test output; state skips and unknowns plainly. - Checkpoint (pause, ask, end the turn) only for a destructive/irreversible action, a fix needing a genuine product decision, or input only I can give. Otherwise, keep going. - Self-check at each phase/loop boundary via a fresh-context subagent: re-verify the spreadsheet against the code (Phase 1) and against actual results (each loop pass). /loop Quality cycle - once the spec is complete, iterate test -> fix -> re-test until clean. Each iteration, in order: 1. Test: exercise every user story not yet Verified against the running app, preferring the strongest method available (browser/e2e automation > existing suites > documented static check only where execution truly isn't possible). Record actual pass/fail in the same spreadsheet; log every defect with its type (functional/logistical or UX). No app-behavior changes in this step. 2. Fix: think hard about root cause, then fix every functional/logistical and UX defect logged this iteration - cause, not symptom. Scope: only logged defects; no new features, no unrelated refactors. Update each row's status. 3. Re-test: re-run every story touched by a fix using the same method; set Verified, or back to Tested-Fail with notes if the fix didn't hold. Exit when all user stories are Verified and no open functional/UX defects remain. Safety cap: if a story is still failing after 3 full iterations, stop, leave it Tested-Fail with root-cause notes, and report it rather than looping further.

你买的 GLM-5.2,可能根本不是 GLM-5.2 最近我们把市面上多家中转站正在卖的 "GLM-5.2” 测了一遍。结果大部分都不是GLM -5.2。 我们见到的几种常见掺假手法: 1. 换皮是最常见的。 就是把一个更便宜的大模型,贴上 "GLM-5.2 / glm-5-2" 的标在卖，因为目前GLM - 5.2 货源很紧张，所以价格非常低的，这个测出来大部分就是dsv4flash。 2. 虚标上下文。 GLM-5.2 官方标的是 100 万(1M)token 上下文。但很多渠道,你真往里塞 25 万、30 万 token,要么直接超时报错,要么前后文明显遗忘、截断。 3. 缩水 / 量化。 这种就是中转站给你压缩了，测试的时候跑分很好看，但是真上长程任务、多文件重构就露原形。 4. 只展示 min 价 + 动态路由。 价格页挂一个漂亮的最低价,实际请求被悄悄路由到更差、更便宜的后端。你看到的价,和你拿到的模型,是两回事。 二、实测全过程，方便大家去检测： 我们拿到一个号称 "GLM-5.2"、价格低到离谱(约官方价 1/20)的渠道，这个渠道低到连电费都赚不回来，所以我就觉得很蹊跷，一步步扒: 首先看价格,起疑。 它标价约 $0.07 / 百万输入、$0.22 / 百万输出。这个价,连官方 GLM-5.2 的零头都不到。 一个按官方原价拿货的授权 reseller,根本做不出这个价。价格反常,是第一个危险信号。 第 1 步:列模型、起一个最简单的调用。 接口能通,返回里 model 字段确实写着 "glm-5.2"。但"返回里写 glm"只能说明它给你贴了这个标,说明不了它到底是什么，有可能是glm上一代模型，甚至拿dsv4flash给你凑数。 第 2 步:身份探测。 我们用不同问法,连问它五遍"你是什么模型、哪家公司训练的"。 结果五次里有四次,它自报是 DeepSeek 系模型(DeepSeek-V3 / R1),其中一次还明确说"我不是 GLM、不是智谱"。名字能改,身份认知改不掉。第一个实锤:它根本不是 GLM。 第 3 步:上下文硬测。 我们做了两层测试。先在一篇约 25 万 token 的长文里埋一串随机暗号,结尾再问它,它准确召回。 但是在长文里埋五条互相依赖的事实(A 等于 7,B 等于 A 的三倍,C 等于 B 加 8,以此类推),要它跨段把最终值算出来,它给出了完全正确的链式结果。 这一步很关键，单点召回也许能靠"检索作弊"蒙混,但跨段整合做不了假,说明它是真把 25 万 token 吃进去在做推理。 结论:它不是小模型，而且上下文比 GLM-5.1 的 20 万还大，结果是deepseek系。 第 4步:终极对照实验(决定性)。 我们直接拿官方 DeepSeek 的 API(里面正好有 deepseek-v4-flash 这个正版模型),和这个 "glm-5-2" 做指纹比对:用同一批 temperature=0 的确定性提示词,两边同时打,逐条比对输出。结果: • 同一道"讲个程序员笑话",两边逐字一模一样; • 同一道"你是 V3 还是 V4",两边都答 "unsure"; • 连"认不出自己、自报成旧版 DeepSeek-V3"这个毛病,官方 v4-flash 和这个 "glm-5-2" 都一样犯。 也就是说:官方正版 DeepSeek-V4-Flash 的种种指纹,这个 "GLM-5.2" 全对得上。实锤收工:这个所谓的 "GLM-5.2",就是 DeepSeek-V4-Flash 贴了智谱的标在卖。它不是缩水的 GLM,它压根不是 GLM。 总结： 身份探测: 同一问题问三到五遍,看它回答是否稳定、是否对得上官方规格,有没有自报成别家模型。 上下文硬测: 埋暗号,再埋几条互相依赖的事实,顶到 25 万 token 以上,看它吃不吃得下、能不能跨段算对。吃不下或答错,就不是满血。 指纹比对: 同一个 temperature=0 提示词,把"待测渠道"和"官方原厂"的输出摆一起比。高度一致就是同一个模型,对不上就是两个东西。 经济常识: 官方满血卖到官方价 1/20,经济上根本不成立。价格低到离谱的"满血",基本可以直接判死。