bin

[bin @bin184032117479]

@CuiMao 最近非常纠结，是苟在一个发展前景不明确的公司，还是直接重新找工作到一个更激进迎合AI的公司（虽然不一定能找到并成功入职）😫😫😫

[bin @bin184032117479]

@mattpocockuk 你是不是可以直接把软件工程加到模型里？

[Su @Sukiea1008]

这个提示词很有趣！

[Matt Pocock @mattpocockuk]

My top 3 skill makers:

- @poteto
- @dexhorthy
- @emilkowalski

Always learn a ton from reading their skills.

[Akshay 🚀 @akshay_pachaar]

Jev for RAG, clearly explained!

Hybrid search gives you a shortlist. It does not decide which passages contain evidence, which are merely adjacent, or whether the evidence is strong enough to answer.

That missing judgment is where Jev fits.

Jev sits between retrieval and generation. It does not replace BM25, embeddings, or the LLM. It evaluates the candidates they produce before those candidates enter the context window.

→ Retrieve wide

Combine dense and keyword search, then merge the results with reciprocal rank fusion. This gives Jev a broad candidate pool, such as the top 20 passages shown in the visual.

Retrieval still sets the ceiling. If the right passage is missing from this shortlist, Jev cannot recover it.

→ Judge every candidate together

Send the query as Jev’s state. For each candidate, ask a typed yes-or-no question such as “Does C7 help answer this query?”

Jev evaluates every question in one packed request and returns a calibrated probability for each passage. This avoids making a separate model call for every query-passage pair.

→ Let code apply the threshold

Your application compares each probability with a threshold. Candidates above it continue to the LLM. Everything below it is removed before generation.

Jev makes the fuzzy judgment. Code remains responsible for the actual decision.

→ Gate the entire answer

The same request can check whether the retained passages make the query answerable. It can also flag signs of prompt injection inside a candidate.

If answerability falls below the threshold, the application can skip the LLM and return “not in the documents.” The injection score should remain a filtering signal, not a security boundary.

The result is a cleaner division of work.

Hybrid search retrieves broadly. Jev reranks, filters, and decides whether sufficient evidence exists. The LLM writes only from the passages that survive.

Jev’s value here is not simply moving passages up or down a list. It turns relevance into an explicit probability that your application can inspect, threshold, and act on.

To summarise:

- Retrieval finds the candidates.
- Jev decides what deserves context.
- The LLM writes the grounded answer.

----
I also built an open-source project showing how to use Jev as a judge for AI observability with Comet Opik.

It evaluates support traces for groundedness, request coverage, action honesty, and helpfulness, then records the results as an auditable experiment.

You can explore the project here: https://t.co/XVmpg6ldHT

My article on how Jev works is quoted below.

[Frank Wang 玉伯 @lifesinger]

小扎谈 Muse 的三个差异化

小扎在近期访谈里谈到 Muse 的三个核心差异化，非常有意思。

一、从底层为场景设计模型

不是拿通用模型套个壳，是从底层就冲着个人 agent 来设计模型。差不多每个月发一次新模型，一次比一次聪明，能做的事也越来越多。

二、��交基因

Meta 擅长连接人，Muse 也不只替人干活，还要帮人照顾家人、维系朋友关系。

访谈里提到一个词，fleet，原意是舰队，这里指所有用户的 Muse agent 组成的群体。思路是让 agent 从群体中学习匿名经验，再结合各自用户的需求，提出新建议。app 里的 ideas 标签就用来推荐还能替你做的事。

小扎认为，行业大多把 agent 当成单机游戏，各用各的。让 agent 互相配合，才有更多可能。相关功能大多还没在这次发布中推出，但方向很明确，你认识的人里，用 Muse 的越多，Muse 就越好用。

三、隐私与安全

隐私与安全，可能是最让人意外的差异化。Muse 要懂你，就得接触消息、邮件、健康信息。没有信任，这些都无从谈起。

Meta 从 WhatsApp 十多年的加密经验中学到，保护隐私不能只靠承诺，系统得做到连 Meta 自己也看不到内容。为此，团队请来 Signal 创始人 Moxie Marlinspike 开发 confidential VM，也就是保密虚拟机。目标是让云端的个人信息连 Meta 也无法查看，还能通过技术验证这一点。访谈时，更多细节仍待公布。

密码和支付凭证单独保管，按需调用，还支持一次性卡号。sentinel agents 相当于安全哨兵，检查进出的信息，发现可疑指令或敏感内容外发，就请用户确认。登录、付款、传敏感信息，通常都要先点头，也可以预先授权。接邮件默认只读，需要发信再单独授权。

在小扎看来，这套防护的细致程度，是 Muse 的重要优势。

定价上，Muse 起步提供每周 1 亿 tokens 的免费额度，还带台虚拟机。长期计划从交易中收取少量佣��，费用可以由商家承担，支付与 Stripe 合作。

出处：以上内容来自 Mark Zuckerberg on how Meta's Muse AI agent will make people money，Sources 播客，主持人 Alex Heath，2026 年 9 月。

原视频：https://t.co/rhLkpAsdzv

[铁锤人 @lxfater]

很多人只会用 grill-me，其实它只是一整套 Skill 里的一环！

Matt Pocock 把自己的 Skill 目录开源了，按工程化的顺序，一步一个 Skill：

1. /grill-with-docs：先跟你对齐想法，沉淀成共识文档
2. /to-spec：把想法落成设计文档
3. /to-tickets：拆成能独立完成的任务
4. /implement：新开会话，一次做一个任务
5. /code-review：检查有没有按设计做完

几个细节：
- 开工前先跑 /setup-matt-pocock-skills，定好术语和文档规范
- 全靠文档驱动，不吃上下文，换 Agent 也不怕
- 要你自己按需调用，不像 Superpowers 那样自动跑到底

GitHub 27 万星，MIT 协议：
https://t.co/qlZ5M3QNlD

[宝玉 @dotey]

《什么是 Transformer》由 Claude Code + Opus 5.5 制作

--- 提示词 ---

帮我用js制作一个视频，主题是：什么是 Transformer
要深入浅出，让高中生也能看得懂，不仅high level说的清楚，也要有细节，包括注意力机制，甚至一些数学概念

你可以用任何工具或者安装工具，可以联网检索

请给我惊喜

[bin @bin184032117479]

👍👍👍

[向阳乔木 @vista8]

Together AI 写了一篇教程：教你花 25分钟、100人民币左右（17美元），训练一个类Jev模型。

基于 Qwen3.5 4B 训练的。

用的Huggingface上的8个公开数据集，3.8万条数据。

像在推广他们的平台。

但这么低的成本学习从0训模型，也想抽空试试。

教程在评论区

[meng shao @shao__meng]

AI Agent 要怎么写测试，还不会过度测试、为测试而测试 ？

@anshnanda 分享了自己写在 AGENTS.md 顶部的三条规则，值得一看！

1. 严禁在写完代码之后再补单元测试。 事后补写的单测（尤其是 Agent 生成的）往往只是把实现逻辑复述一遍：代码怎么写的，测试就怎么断言。它验证不了正确性，只制造“有测试覆盖”的假象。dex 举例的那种“检查常量里包含某子串”正是这类无意义的同义反复。

2. 首选 E2E 测试作为唯一的验证机制，且结束时必须产出可验证、可复现的产物。 他要的是从真实用户路径出发验证功能确实能用，而判据不是“测试绿灯”，而是一份能留档、能重跑、能被人检查的证据（如输出文件、录屏、报告）。这把验证从“测试通过”升级为“拿出证据���。

3. 如果确有必要隔离测试某个系统，先穷举它所有可能的失败方式，再动手写代码。 这是失败模式驱动的前置测试：测试的价值在于预先写下“哪些情况会坏”的假设，让测试承载设计意图，而不是事后镜像实现。

AGENTS.md 中要加入顶部的原文：
- NEVER write unit tests after you write code.
- Highly prefer E2E tests as the sole testing mechanism. Use them to verify complex features work. At the end of E2E tests, produce a verifiable and repeatable artifact.
- If you must test a system in isolation, FIRST write all the ways it could fail, THEN write the code.

[bin @bin184032117479]

非常同意，每天被新的AI消息搞得焦头烂额，不如实实在在的解决实际问题

[Morris @Morris_LT]

不要着急尝鲜。现在几乎每天都有新的AI模型、新的Agent、新的产品出来。很多人一看到新的东西，就马上去注册、测试、研究，生怕自己落后。但人的精力非常有限。你以为自己是在更早拥抱未来，很多时候，其实只是更早去趟坑。记住一个规律：真正有价值的东西，通常会留下来；能留下来的东西，才值得你长期投入研究。如果一个东西真的重要，你晚一个月研究，通常也不会错过它。所以不要被“新”牵着走。先看它能不能长期解决真实问题，再决定值不值得投入时间。先有问题，再找工具。而不是先有新工具，再给自己找问题。

[Akshay 🚀 @akshay_pachaar]

Another insane Jev use case!

Jev makes it incredibly cheap to evaluate and classify agent runs at scale.

And finally, someone open-sourced a self-improving memory layer that can put that capability to work across agent harnesses.

It turns your agent sessions into a compounding knowledge layer, where every successful run can make future agents smarter across:

- Codex
- Claude Code
- Cursor
- OpenCode and 20+ more

Beacon by @asymptotelabs continuously builds a shared history across your agent harnesses and uses Jev to identify the runs worth learning from.

It then turns the best workflows, corrections, and debugging patterns into reusable skills.

GitHub repo: https://t.co/sfdu9P1bfx.

(don’t forget to star it ⭐)

Most agent runs are messy.

They contain exploration, failed commands, dead ends, and one-off fixes that should never become permanent memory.

So Beacon preserves the full session history, while Jev helps decide what should be promoted, reviewed, or discarded.

The recording below shows this in action.

Beacon found 579 sessions across 5 coding-agent harnesses and normalized them into one consistent history.

From there, Jev surfaces the lessons worth keeping and makes them available across your agent stack.

- A pattern learned in Cursor can carry into OpenCode.
- A lesson from Claude Code can improve the next Codex run.

Every successful run adds to the shared knowledge layer, making future agents smarter.

If you want to dive deeper into Jev, I also wrote a breakdown of how it works.

The article is quoted below.

[Suu @Suu766]

我最近才发现，很多大学生翻墙以后，用 YouTube 还是只会看娱乐、听歌、刷视频？？？

但 YouTube 上真正值钱的那些内容，很多人反而一条都没看过

里面有些视频其实就十几二十分钟，但讲的东西，可能你在大学里待几年都没人认真跟你说过

说实话，这 12 个视频，我觉得比大学四年上过的水课都更值得看

我整理了一下，建议大学生都可以看看👇

[宝玉 @dotey]

文档的管理我是这么做的：
1. 文档统一放 docs 目录
2. 所有文档都遵循渐进式披露原则——也就每个文档都不大，但是会链接到相关文档，类似于主文档里面主要就是概要和目录，具体内容链接到章节对应的小文件
3. 根目录有一个 README 可以方便的索引到相关文档，类似于目录
4. AGENTS.md 强制要求修改功能要修改相关文档，有时候人也需要定期检查一下，即使加了规则也可能遗漏

[向阳乔木 @vista8]

体验了下meta的muse，每个功能都平平无奇，但简单易用不复杂，场景导向，被普通用户喜欢也很合理。

对话设定目标，生成游戏，生成视频和文档。

帮人购物，退货，安排party，抢票等等，Idea面板的场景很美国。

美区apple store下载安装，需要绑定信用卡验年龄（香港卡也行），魔法用的美国。

[Kailash @kail_designs]

Make your UI look better instantly with these skills

1. frontend-design
https://t.co/Di4HsGzQIv

2. apple-design
https://t.co/8mk3tWJVTg

3. beautiful-shadows
https://t.co/EMSz0Y8pH9

4. accessibility
https://t.co/7lGDWYDS8L

5. design-review
https://t.co/OXuOj5Vyh4

6. emil-design-eng
https://t.co/bdbuOABzGl

7. shadcn
https://t.co/VUAadmUQpH

8. adapt
https://t.co/YCbm6QnBDK

9. better-interface
https://t.co/CtFApgAsFE

10. interaction-design
https://t.co/NynDF3tcgv

[bin @bin184032117479]

@ScarletKc 为啥一定要用claude code？

[돈벌고영 @Dontgiveup_26]

Jev와 비교했을 때 Kev가 압도적으로 혁신적인 이유:

1. 단순 '예/아니오'를 넘어 '선택지별 가중치와 점수' 반환
- 기존 Jev가 주로 단순 "예/아니오" 식의 이진 판단 위주였다면,
- Kev는 객관식 다중 선택과 1~5점 척도 평가까지 한 번에 처리함.
- 단순히 정답 하나만 딱 고르는 게 아니라, 각 선택지별로 정밀한 '확률 가중치'를 뽑아줌.
- 예: 복합 문의가 들어오면 [반품 47%, 배송 28%, 결제 25%]처럼 미묘한 비율까지 1번의 연산으로 완벽히 분석해 냄.

2. 여러 질문을 동시에 던져도 '상호 간섭 0' (독립 격리)
- "담당 부서", "긴급 대응 필요 여부", "고객 불만 점수"를 한 번에 물어봐도,
- 각 질문이 서로 영향을 주지 않도록 완벽히 분리된 상태에서 단 '1번의 Forward Pass'로 수십 밀리초 만에 동시 판별함.

3. 폐쇄형 유료 서비스 vs 100% 내 맘대로 학습하��� 오픈소스
- Jev는 입력 데이터 보관 정책 논란이 있고 커스텀이 불가능한 폐쇄형 API였음.
- Kev는 터보레포 창시자 Jared Palmer가 만든 Apache 2.0 완전 오픈소스임.
- 사내 업무 데이터 수백 개만 넣으면 우리 회사만의 규칙에 맞게 초고속 미세조정(LoRA) 가능.

4. 내 맥북에서 로컬 구동 (비용 0원 & 데이터 유출 제로)
- 0.8B, 4B, 9B 모델 가중치가 모두 공개되어 있음.
- 비싼 클라우드 API나 데이터 유출 걱정 없이, 내 로컬 환경(Mac M시리즈/CUDA)에서 무료로 무제한 구동 가능.

🔗 깃허브: https://t.co/3nAvNo844p

[极客杰尼 @seekjourney]

一个很强的 RSI 新信号可以关注一波：

Google Research 开始研究 Agent Harness 的 Recursive Self-Improvement。

不重新训练模型，直接让 prompt、工具、memory、control flow 自动迭代。

这条路线我很看好。

因为相比之前 glm 发布的 model-level RSI 更清晰了 ：

https://t.co/OHmy8Vp4Rm

现在很明显在 Harness 工程上更容易先跑起来，可优化的东西太多了。

https://t.co/cCc0ZNa7d2

[ElliotSecOps @ElliotSecOps]

Si ya el código está resuelto, nada te impide expandir tu stack, estudiar cloud, software architecture, devops, y seguridad, es lo que diferencia la verdadera ingeniería de software del vibecoding.