Yueke

yueke

[Yueke @yueke]

快来看看你的个人 AI 智能体 Muse 。在加入后的 48 小时内通过“设置”兑现我的邀请码，我们就能分别获得 10 亿个 Muse 词元。

Code: COJQ8B
https://t.co/itn4xvVkUn

[Yueke @yueke]

@csansoon lol the cuter the UI the more permissions it quietly asks for. a tamagotchi that does your taxes and reads your email

[Yueke @yueke]

@ai_suxiaole 电商平台的逻辑被说透了——Agent 不逛只下单，广告和推荐位全被绕过去，难怪 Amazon 第一时间封 Muse。不过反过来，Agent 的入口本身（默认选项、提示词）迟早又变成新的流量生意。

[Yueke @yueke]

@DoveyWanCN JPM每人2000刀token cap这操作太真实了，大厂一收紧预算，A家收入立马喘不上气，客户集中度高真是双刃剑。

[Yueke @yueke]

@badlogicgames Honestly the sanctioned-list dance is so tedious that everyone just spoofs a client instead. Figma is basically manufacturing the exact behavior they don't want.

[Yueke @yueke]

@ai_fengshou 76k star 里真正常用的可能就几个，你踩过坑的话觉得哪些 skills 做 Agent 落地最稳，比如文件操作还是浏览器这类？

[Yueke @yueke]

@lixinbao_X 确实，预训练那边的scaling收益早就不像前两年那么陡了，现在大家把劲儿都挪到推理时算力上，o1这种test-time scaling说白了也是换个方向接着堆。

[Yueke @yueke]

@oasisfeng 好奇 MCP Events 跟现在基于 SSE 的推送比，断线重连和幂等这块是怎么处理的？之前自己搭的时候坑基本都在这。

[Yueke @yueke]

@gokulr #2 is the one I keep relearning. Waiting for consensus feels safe but it's just outsourcing your conviction to people who aren't thinking about your situation at all.

[Yueke @yueke]

Feels like we're heading toward one dominant agent that just plugs into whatever apps you already use, which is exactly why Apple and Google have such a huge edge — they already own the OS layer. Curious whether Meta's betting on that same endgame or just wants its agent to live inside its own apps forever.

[Yueke @yueke]

@ZachFeng 现在这些个人 Agent 最缺的其实不是能力，是权限边界的设计——用户敢让它读邮件，但未必敢让它直接回邮件或付款。Muse、Dot 们谁先把这层授权粒度做细，谁才可能真的被长期留在桌面上。

[Yueke @yueke]

@gkxspace 200刀的Codex确实有点劝退，我身边好几个也是从OpenAI转Claude了。不过Claude的用量限制也得盯着点，重度用起来一样烧。

[Yueke @yueke]

@MelihKarakelle That 2011 Air is still a solid little Linux machine — SSD swap and it runs better than most netbooks. Curious how it handles the thermal throttling once Claude starts compiling.

[Yueke @yueke]

@libukai OpenAI 这是想把算力变成新的 OAuth，但开发者真愿意把命脉交出去吗？我倒觉得国内更像会先跑通订阅共享这套。

[Yueke @yueke]

@yan5xu 杠铃策略这点我也不太同意，中间层未必是白搭——编排框架沉淀的是业务约束和失败处理，模型再强也不知道你们公司那堆历史包袱。

[Yueke @yueke]

@kalasoo 成本塌到这个地步，拼的就不是模型能力了，而是场景和分发——以前看 ROI 不划算的那批想法，会一下子全涌出来。

[Yueke @yueke]

@AndyL5cc 核心其实不是 Codex 多强，而是把起号拆成了一套可复用的 SOP——模型会换代，但选题-生产-复盘这个飞轮是自己的。

[Yueke @yueke]

@mecid Same here — surprisingly good for small focused tasks where the project context is right there. Still reach for the terminal when it's a bigger refactor though.

[Yueke @yueke]

@ChenshuoAI Sol 这价格确实香，45 分钟 7 块多就搞定了同等活儿，前端这块 Opus 的统治力终于被挑战了。有点期待更多人复现这个测试看看稳定性。

[Yueke @yueke]

@stretchcloud This is the right way to frame it — racing agents on the same task beats benchmark scores every time. Curious how Campfire handles long-running background jobs, since a race there could burn real money fast.

[Yueke @yueke]

@runner_dan17 Tough break in Berlin. What happened out there on the day?