共收录 21 篇深度内容
📌 今日概览

- AI Agent 进入生产与创作主流程:Auto 将上线发布、日志巡检与数据库修复交给 Agent 定时执行,Meng To 用 Fable 5.2 调度多个 Opus 5.5 代理重建 3D 场景,Riley Brown 以 Claude 生成 1.25 万行 Python、租 89 块 GPU 渲染 4K 影片,花费略超 20 美元。
- 能力认知与评测体系同步刷新:Ethan Mollick 统计 173 个基准中 63% 已饱和且饱和速度加快,Dwarkesh Patel 认为可验证领域的窄能力会广泛外溢,并推断 GPT-3 到 GPT-6 的跃迁可压缩进一年。
- 开源生态承压:Nathan Lambert 称开源模型的政治盟友正在退缩、公司需更早找到盈利产品,Alex Veremeyenko 则建议下载 7 个开源模型存本地以防限制。
- 硬件与产品格局生变:SemiAnalysis 称英伟达七大客户中六家在自研芯片、TPU 性能相当但谷歌缺供应链能力,Grok 宣布把 SpaceXSI 旗下图像与视频创作等产品整合进单一入口。
🔥 必读 Top3
- Marc Lou 作者称工作流已从本地开发直接进入生产环境发布:他让 AI 把功能先只对自己在生产环境可见,评估迭代后再向用户发布,并让 Agent 设小时/每日提醒检查日志与数据库修问题,还靠 AGENTS.md 规则强制 Agent 测试全部内容;结论是提示词可直接产出并上线。
入选理由:完整呈现 Agent 直接在生产环境发布、定时巡检与规则约束的落地工作流,可操作性强。 - Ethan Mollick 作者用自建指标 SatBench 统计大模型实验室自 2023 年以来引用的 173 个基准,其中 63% 已饱和即被基本打满,且饱和速度逐年加快,据此认为 AI 能力提升快于评测体系的更新速度。
入选理由:用 173 个基准、63% 饱和的统计数据量化评测体系落后于模型进步的速度。 - SemiAnalysis 作者引用SemiAnalysis观点称,英伟达七大客户中六家除SpaceX外均在自研芯片,TPU性能与英伟达相当,但谷歌缺乏供应链运作能力。
入选理由:指出英伟达七大客户中六家自研芯片、TPU 性能相当,是理解算力格局变化的关键判断。
🧭 主题分组
- Agent 落地(5 条)—— Agent 已承担生产发布、定时巡检、多代理调度与低成本 GPU 渲染,工作流范式趋于可复用。
- 能力与评测(2 条)—— 可验证领域的能力跃迁与基准饱和相互印证,评测更新速度落后于模型进步。
- 开源生态(2 条)—— 开源模型面临政治与资金双重压力,社区转向本地留存与自主可控。
- 芯片算力(1 条)—— 大客户自研芯片加速,TPU 性能追平英伟达,供应链能力成为谷歌短板。
- 模型架构(1 条)—— Inkling 2 Small 采用类 Kimi K3 的 attention residuals,年末架构趋势讨论升温。
- 创作工具(2 条)—— AI 与 Blender、音乐制作、室内设计深度结合,从编曲到 3D 重建均有开源方案。
- 产品动态(4 条)—— Claude 自动模式、Claude Max 订阅价值与 Grok 产品整合成为用户关注焦点。
- 模型发布(1 条)—— GPT-6.2 迟迟未发布引发猜测,社区讨论 OpenAI 是否跳过该版本。
- AI 视频(1 条)—— AI 视频生成被期待降低影视创作门槛,现阶段多为主观观感判断。
- 航天发射(1 条)—— SpaceX 助推器落海由自动系统中止回收,具体原因尚未公布。
- AI 表达(1 条)—— 用户对「AI 味」回复容忍度下降,套话附和与虚构身份成为主要反感点。
📖 详细内容
Agent 落地
Agent 已承担生产发布、定时巡检、多代理调度与低成本 GPU 渲染,工作流范式趋于可复用。
Marc Lou
| @marclou
| 影响力: 42.4万粉丝 · 互动分 2422.5
Marc Lou
| @marclou
| 影响力: 42.4万粉丝 · 互动分 2422.5
💡 核心观点: 作者称工作流已从本地开发直接进入生产环境发布:他让 AI 把功能先只对自己在生产环境可见,评估迭代后再向用户发布,并让 Agent 设小时/每日提醒检查日志与数据库修问题,还靠 AGENTS.md 规则强制 Agent 测试全部内容;结论是提示词可直接产出并上线。
可信度: 60/100
事实核查:
- 无法核实: I ask AI to make the feature only visible to me in production(作者自述的做法,属个人工作流经验,无第三方数据或可核验来源,无法交叉验证。)
- 无法核实: set hourly/daily reminders to check logs and database(原文有此自述,但只是作者陈述,未提供实际的提醒配置或运行记录。)
- 原文支持: My AGENTS·md also has rules to force the agent to test everything(原文明确提到 AGENTS.md 含强制测试规则,可作为作者自述的工作流依据。)
原文内容:
I haven't touched localhost in month. I ask AI to make the feature only visible to me in production, so I can assess and iterate. If it's good, I release it to users, and ask the agent to set hourly/daily reminders to check logs and database to fix any issues. My AGENTS·md also has rules to force the agent to test everything. Now AI is so good, everything goes straight from prompt to prod.
⏰ 11:49 | ❤️ 1532 点赞 | 📝 78词 | 入选: 高赞 1532 | 查看原文 →
Meng To
| @MengTo
| 影响力: 19.1万粉丝 · 互动分 758.14
Meng To
| @MengTo
| 影响力: 19.1万粉丝 · 互动分 758.14
💡 核心观点: 引用一位日本创作者的记录:仅用 5% Tripo3D,主要靠 Claude Opus 5.5 加 Blender 建模。方法为给三张参考图和背景设定,配 CLAUDE.md 与子代理分工规划、建模、渲染与视觉检查,分 13 个阶段逐角度比对渲染,模型、贴图与环境光均未外部下载。作者称流程还生成检查脚本,发现 45 个道具穿模。
可信度: 55/100
事实核查:
- 原文支持: It plans, then builds in 13 stages, rendering each from the reference's angle to compare(原文转述他人制作记录并附原始文章外链,流程细节可追溯。)
- 无法核实: It even wrote a check that found 45 props clipping into things.(45 这一数字仅见于转述,未附检查脚本或日志佐证。)
- 原文支持: No downloaded models, textures or HDRIs(原文明确写出该约束,属可核验的流程描述。)
原文内容:
This is such an insane workflow. 5% Tripo3D, Blender, and Claude Opus 5.5. The workflow in English, from Kamome Ashizawa's (葦沢かもめ) write-up: https://t.co/Nk9oDG0u7y 1. Give Claude 3 reference images and a short backstory 2. Harness: CLAUDE.md plus subagents for planning, modeling, rendering and visual checks 3. It plans, then builds in 13 stages, rendering each from the reference's angle to compare 4. No downloaded models, textures or HDRIs 5. For photoreal: light and color first, then materials, shapes, lived-in clutter It even wrote a check that found 45 props clipping into things. Video walkthrough: https://t.co/i6ISMiSA3t How the harness works: https://t.co/6UonBil9pw Blender's official MCP server: https://t.co/G6X5xatRxn https://t.co/2l7DSae0sG
引用 @arafune: Blender × Claude Opus 5.5テスト
かなり素晴らしいクオリティ
tripoも使用許可したけど、5%くらいしか使ってなくて、ほぼ自分でモデリングしてた
文字など一部のテクスチャにChatGPT使いつつ、シェーダーもほぼプロシージャルに仕上げてくれてる
音はHunyuanVideo-Foley
メイキングも作ってもらった https://t.co/UeHwXc2max
⏰ 07:00 | ❤️ 577 点赞 | 📝 110词 | 入选: 高赞 577 / 引用讨论 | 查看原文 →
Meng To
| @MengTo
| 影响力: 19.1万粉丝 · 互动分 500.11
Meng To
| @MengTo
| 影响力: 19.1万粉丝 · 互动分 500.11
💡 核心观点: 作者称自己让 Fable 5.2 把一款 three.js 游戏预告片在 Blender 里重建,几小时完成;现用 Fable 5.2 medium 调度多个 Opus 5.5 代理作为主力工作流,并借助云端环境在笔记本上处理更复杂的 3D 场景。文中未说明重建质量的具体评估方式。
可信度: 35/100
事实核查:
- 无法核实: I asked Fable 5.2 to recreate my three.js game trailer in blender. It made this in just a few hours.(属作者自述使用经历,耗时与效果无第三方核验。)
- 原文支持: Fable 5.2 medium orchestrating opus 5.5 agents high is my main workflow now(原文明确描述其当前工作流配置。)
原文内容:
I asked Fable 5.2 to recreate my three.js game trailer in blender. It made this in just a few hours. Ever since I discovered https://t.co/usWIB8iVtW, I've been recreating my 3d scenes in way more detail without clogging my macbook. Fable 5.2 medium orchestrating opus 5.5 agents high is my main workflow now.
引用 @MengTo: I've always wanted to play D&D with my kids with AI as the game master.
So I made one with Opus 5.5. https://t.co/PbIwHllxXu
⏰ 14:42 | ❤️ 355 点赞 | 📝 52词 | 入选: 高赞 355 / 引用讨论 | 查看原文 →
Riley Brown
| @rileybrown
| 影响力: 26.6万粉丝 · 互动分 1746.0
Riley Brown
| @rileybrown
| 影响力: 26.6万粉丝 · 互动分 1746.0
💡 核心观点: 作者称用Claude在Blender中根据一个提示词生成2分钟4K影片,写了1.25万行Python代码,并通过RunPod租用89块GPU渲染,总花费略超20美元。
可信度: 55/100
事实核查:
- 无法核实: That came to 12,500 lines of code.(原文明示代码行数,但未公开代码或项目文件)
- 无法核实: Then it rented 89 GPUs through RunPod to render 3,220 frames(原文给出GPU数量、渲染帧数和并行数,属作者自述的工作流细节)
原文内容:
The Fable of Clawd A 2 minute 4K film made entirely in Blender, from one prompt on Fable 5.2. It took about 4.5 hours. Claude built every asset in Blender by writing Python: - 5 characters and props: Clawd, his friend Pip, a toy truck, a 3D-printed tulip and a gift box - 7 full sets: a study, a staircase, a street, a workshop, an airport, a night sky and a house - every camera move and every frame of animation across 52 shots That came to 12,500 lines of code. It started 7 copies of itself to build the sets at the same time. Then it rented 89 GPUs through RunPod to render 3,220 frames as fast as possible, 36 of them running at once. Total cost: a little over $20. My biggest takeaway is my Claude Code knows it has access to cloud compute, so it's much more ambitious. I tried a similar prompt on my other computer that doesn't have the runpod skill setup and it wasn't nearly as high quality Full details linked below.
⏰ 16:33 | ❤️ 1232 点赞 | 📝 165词 | 入选: 高赞 1232 / 长文 | 查看原文 →
GREG ISENBERG
| @gregisenberg
| 影响力: 76.3万粉丝 · 互动分 364.67
GREG ISENBERG
| @gregisenberg
| 影响力: 76.3万粉丝 · 互动分 364.67
💡 核心观点: 作者认为如今营销的本质是代码与 token,学习营销即学习用自有数据、创意工具和广告账户搭建 Agent;配置得当就能得到每天大量测试创意的顶级营销者。
可信度: 35/100
事实核查:
- 无法核实: all marketing now is just code and tokens(对营销形态的整体断言,原文无数据或案例支撑。)
- 无法核实: you get a top 0.1% marketer testing an absurd number of ideas every day(以 0.1% 描述效果水平,属修辞性说法,无可核验口径。)
原文内容:
The truth is all marketing now is just code and tokens. So learning to market these days means learning to set up an agent with your data, creative tools and ad accounts. Set it up right and you get a top 0.1% marketer testing an absurd number of ideas every day: https://t.co/64vZyolTFP
引用 @gregisenberg: learn to market is the new learn to code
⏰ 02:01 | ❤️ 183 点赞 | 📝 53词 | 入选: 高赞 183 / 引用讨论 | 查看原文 →
能力与评测
可验证领域的能力跃迁与基准饱和相互印证,评测更新速度落后于模型进步。
Dwarkesh Patel
| @dwarkesh_sp
| 影响力: 28.8万粉丝 · 互动分 1202.5
Dwarkesh Patel
| @dwarkesh_sp
| 影响力: 28.8万粉丝 · 互动分 1202.5
💡 核心观点: 作者认为当前 AI 能力「尖而不均」:在数学等可验证领域已接近或超过人类,而 AI 研究本身也属于可验证领域,因此窄能力 AI 加速 AI 研发所带来的外溢效应会非常广泛。依据是更好的学习算法具有通用性,GPT-6 在几乎所有任务上远优于 GPT-3;他推断两三年甚至数月内可把 GPT-3 到 GPT-6 的跃迁压缩进一年。
可信度: 45/100
事实核查:
- 原文支持: Current AIs are spiky and jagged – worse than humans in most domain, but often as good, or better, in ones with verifiable proxies (for example, math).(原文直接给出的现状判断,无数据或来源支撑。)
- 无法核实: we're only a couple years away (and possibly only a couple months away)(属作者对时间线的预测,原文未提供可核验依据。)
原文内容:
Maybe to summarize the basic argument of this essay a different way: Current AIs are spiky and jagged - worse than humans in most domain, but often as good, or better, in ones with verifiable proxies (for example, math). One such verifiable domain (I argue) is AI research. And while the first AIs that dramatically accelerate AI progress will be quite narrow (i.e. probably not able to automate most jobs), the consequences of that accelerated progress will be very general. And that's because better learning algorithms are very general. For example, GPT-6 (while still being jagged) is just way, way better than GPT-3 at practically everything. I'm arguing that we're only a couple years away (and possibly only a couple months away) from a process by which (initially) narrow AIs concentrate a jump as big as the one from GPT-3 to GPT-6 into a single year. And if you got another jump that big, but starting from the base of GPT-7, I think you would end up with AIs that would rapidly become better than the best humans in every field.
引用 @dwarkesh_sp: The avalanche that hit math is coming for AI development.
https://t.co/z5RsJgcWWc https://t.co/Ud6MDxiADh
⏰ 20:06 | ❤️ 868 点赞 | 📝 184词 | 入选: 高赞 868 / 引用讨论 / 长文 | 查看原文 →
Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 743.09
Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 743.09
💡 核心观点: 作者用自建指标 SatBench 统计大模型实验室自 2023 年以来引用的 173 个基准,其中 63% 已饱和即被基本打满,且饱和速度逐年加快,据此认为 AI 能力提升快于评测体系的更新速度。
可信度: 65/100
事实核查:
- 原文支持: Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated.(原文给出基准总数与饱和比例,并指向自建项目 SatBench,统计口径可核验。)
- 无法核实: Saturation is happening faster each year.(增速判断未附逐年数据或图表细节。)
原文内容:
AI is getting better faster than our measurements are I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated. Saturation is happening faster each year. https://t.co/tITFRo7w0z
⏰ 23:46 | ❤️ 391 点赞 | 📝 43词 | 入选: 高赞 391 | 查看原文 →
开源生态
开源模型面临政治与资金双重压力,社区转向本地留存与自主可控。
Nathan Lambert
| @natolambert
| 影响力: 11.0万粉丝 · 互动分 423.99
Nathan Lambert
| @natolambert
| 影响力: 11.0万粉丝 · 互动分 423.99
💡 核心观点: 作者认为后 Jensen 时代开源模型面临两重挑战:在华盛顿的政治盟友正在退缩,且开源模型公司需更早找到可盈利产品、失去资金靠山;他预计影响会比预期更早显现。
可信度: 45/100
事实核查:
- 无法核实: Open models’ allies are retreating in DC(原文未给出具体机构、人物或政策事件,无法验证「盟友退缩」的程度。)
- 无法核实: Open model builders need to find profitable products far sooner(属于对开源模型公司经营压力的判断,原文无财务数据或案例。)
原文内容:
The open model world faces two main challenges in a post-Jensen world: 1. Political exposure. Open models’ allies are retreating in DC. 2. Investment. Open model builders need to find profitable products far sooner w/o his backstop. Will feel it sooner than people expect. https://t.co/VHV2MPB4j2
⏰ 12:28 | ❤️ 252 点赞 | 📝 47词 | 入选: 高赞 252 | 查看原文 →
Alex Veremeyenko
| @alex_verem
| 影响力: 10.5万粉丝 · 互动分 393.66
Alex Veremeyenko
| @alex_verem
| 影响力: 10.5万粉丝 · 互动分 393.66
💡 核心观点: 作者推荐周末下载 7 个开源 AI 模型,理由是担忧开源模型受到限制,而保存在本地硬盘的模型始终归自己所有;他称帖中说明了各模型用途。
可信度: 50/100
事实核查:
- 无法核实: 7 open source AI models to download this weekend(原文给出模型数量,但正文未列出模型名称、版本或来源链接。)
- 无法核实: Everyone is worried open models will get restricted(以「大家担心」概括行业情绪,未引用调查、机构或具体表态。)
原文内容:
7 open source AI models to download this weekend. Everyone is worried open models will get restricted. A model saved on your own drive stays yours forever. Here's what each one is for. (Bookmark this)
⏰ 10:56 | ❤️ 281 点赞 | 📝 35词 | 入选: 高赞 281 | 查看原文 →
芯片算力
大客户自研芯片加速,TPU 性能追平英伟达,供应链能力成为谷歌短板。
SemiAnalysis
| @SemiAnalysis_
| 影响力: 18.7万粉丝 · 互动分 1811.5
SemiAnalysis
| @SemiAnalysis_
| 影响力: 18.7万粉丝 · 互动分 1811.5
💡 核心观点: 作者引用SemiAnalysis观点称,英伟达七大客户中六家除SpaceX外均在自研芯片,TPU性能与英伟达相当,但谷歌缺乏供应链运作能力。
可信度: 62/100
事实核查:
- 原文支持: Six of Nvidia's seven biggest customers are building their own chips.(原文点名谷歌、微软、Meta、亚马逊、SpaceX、Anthropic和OpenAI,除SpaceX外均自研芯片,但未给出具体数据来源)
- 无法核实: From our benchmarking, the TPU is just as good as NVIDIA.(作者称基于自有benchmark得出,未披露测试方法与样本)
原文内容:
Six of Nvidia’s seven biggest customers are building their own chips. “NVIDIA’s seven biggest customers account for the vast majority of GPUs being deployed: Google, Microsoft, Meta, Amazon, SpaceX, Anthropic and OpenAI. All of them, except SpaceX, are making their own chips.” “If NVIDIA didn’t prop up CoreWeave and the other neoclouds, those industries just would not exist, and the hyperscalers would have full control.” “Jensen’s the only executive in the semiconductor industry convincing supply chain companies to invest like they’ve never done before. From our benchmarking, the TPU is just as good as NVIDIA. But Google doesn’t have that DNA of wheeling and dealing.”
⏰ 22:00 | ❤️ 1259 点赞 | 📝 111词 | 入选: 高赞 1259 | 查看原文 →
模型架构
Inkling 2 Small 采用类 Kimi K3 的 attention residuals,年末架构趋势讨论升温。
Sebastian Raschka
| @rasbt
| 影响力: 53.9万粉丝 · 互动分 1252.68
Sebastian Raschka
| @rasbt
| 影响力: 53.9万粉丝 · 互动分 1252.68
💡 核心观点: 作者指出Inkling 2 Small采用类似Kimi K3的attention residuals,但未使用线性注意力混合架构,认为其值得纳入年末架构趋势讨论。
可信度: 58/100
事实核查:
- 无法核实: Using attention residuals (like Kimi K3), but without the linear-attention hybrid.(原文明确描述架构特征,但未给出来源论文或技术报告链接内容)
原文内容:
Inkling 2 Small took a while... but it's more interesting than I initially thought! Using attention residuals (like Kimi K3), but without the linear-attention hybrid. Food for discussion for the end-of-the-year architecture trends analysis! https://t.co/2U4ctJa1yw
⏰ 14:38 | ❤️ 926 点赞 | 📝 40词 | 入选: 高赞 926 | 查看原文 →
创作工具
AI 与 Blender、音乐制作、室内设计深度结合,从编曲到 3D 重建均有开源方案。
Rimsha Bhardwaj
| @heyrimsha
| 影响力: 2.5万粉丝 · 互动分 957.95
Rimsha Bhardwaj
| @heyrimsha
| 影响力: 2.5万粉丝 · 互动分 957.95
💡 核心观点: 作者盘点 GitHub 上 10 个面向音乐制作的免费开源仓库,覆盖用 Claude 连接 Ableton 编曲、本地生成整首歌、歌词与人声生成、人声分离与分轨、音频转 MIDI、母带匹配以及合成器与模块化音源,其中多项结合 AI。
可信度: 72/100
事实核查:
- 原文支持: I found 10 insane repos that help you write, split, master and remix tracks, some with Claude and AI. All free.(原文列出 10 个仓库名称与链接,均为开源或免费项目,可逐项核验。)
- 无法核实: Under 10 seconds per song on an RTX 3090.(性能数字由作者给出,未附测试条件或基准来源。)
原文内容:
GitHub is a goldmine for music producers. I found 10 insane repos that help you write, split, master and remix tracks, some with Claude and AI. All free. Bookmark this for later: 1. Ableton MCP Connect Claude to Ableton Live. Create tracks, arrange clips and compose music just by describing what you want. https://t.co/NP47TZoYaK 2. ACE-Step 1.5 Generate full songs on your own computer, from short loops to 10-minute tracks, with lyrics in 50+ languages. Under 10 seconds per song on an RTX 3090. https://t.co/Rl76BjNO0H 3. YuE2 Give it lyrics and a style. It writes the melody and chords first, then turns them into a complete song with vocals. It can also make covers in a new style. https://t.co/vVTyS2iuCc 4. Ultimate Vocal Remover Pull the vocals out of any song, or split it into stems, with neural networks. 27K+ stars. https://t.co/0zChgUmutL 5. Basic Pitch Spotify's free audio-to-MIDI converter. Play or sing a melody and get MIDI back, pitch bends included. https://t.co/jzL9cMcZGX 6. Matchering Master your track to sound like any reference song. It matches loudness, frequency response, peaks and stereo width. https://t.co/aeJl1FC3Sn 7. Surge XT A free, open-source hybrid synthesizer you can load as a plugin. https://t.co/gjRpLodK8D 8. Dexed The classic Yamaha DX7 FM synth, rebuilt as a free plugin. https://t.co/37fbAGVX7g 9. VCV Rack A virtual Eurorack modular studio on your computer. https://t.co/lhWK9kBAre 10. LMMS A free, cross-platform music production app. https://t.co/iAyGdHnIOx
⏰ 12:31 | ❤️ 758 点赞 | 📝 241词 | 入选: 高赞 758 / 长文 | 查看原文 →
@levelsio
| @levelsio
| 影响力: 99.2万粉丝 · 互动分 429.5
@levelsio
| @levelsio
| 影响力: 99.2万粉丝 · 互动分 429.5
💡 核心观点: 作者称在 Interior AI 中以 3D 形式导入 IKEA 全品类家具,用户可从户型图或实拍视频重建房屋后再装修;他称这是重建产品第 4 天,效果仍显笨拙但有进展。
可信度: 72/100
事实核查:
- 原文支持: I imported the entire catalogue of @IKEA in 3d into Interior AI(第一人称自述已完成导入,但未说明完成度、授权方式或可核验链接。)
- 无法核实: it was really impossible until recently do to floor plan to 3d automatically and properly(作者断言此前无法自动、正确地把户型图转为 3D,属主观判断,原文无对照数据。)
原文内容:
✨ Progress update: I imported the entire catalogue of @IKEA in 3d into Interior AI after seeing someone do the same yesterday, I'll try to import more furniture brands soon to get more variety too Anyway so now you can import your home's blueprint or do a walk through video, it then rebuilds your home, and then you can redecorate it This is now day 4 of rebuilding Interior AI now that's the models are smart enough to do 3d stuff properly, it still looks a bit clunky and all but it's going somewhere! 😊
引用 @levelsio: More progress!
Interior AI 3D works with "just" a floor plan now
Not much information to work with and it has to imagine furniture etc. but it works
Some weird things like the big I guess sliding windows to outdoor garden are missing, and bath tub is just a box, and I don't see a toilet seat etc.…
⏰ 17:46 | ❤️ 269 点赞 | 📝 94词 | 入选: 高赞 269 / 引用讨论 | 查看原文 →
产品动态
Claude 自动模式、Claude Max 订阅价值与 Grok 产品整合成为用户关注焦点。
Simon Willison
| @simonw
| 影响力: 25.5万粉丝 · 互动分 695.9
Simon Willison
| @simonw
| 影响力: 25.5万粉丝 · 互动分 695.9
💡 核心观点: 作者称 Claude 普通聊天现已采用「自动模式」,偶尔会拒绝执行请求;他在提示中明确授予本轮会话运行该实验的权限后,请求似乎被放行,并指出自动模式会把用户消息纳入判断。这是一次个人使用观察,未给出复现步骤或官方说明。
可信度: 40/100
事实核查:
- 无法核实: Regular Claude chat uses "auto mode" now, so occasionally it will refuse to run something(原文陈述产品行为,未见官方文档或截图佐证。)
- 无法核实: auto mode takes your messages into account(作者依据单次尝试得出的推断,原文未提供机制说明。)
原文内容:
Regular Claude chat uses "auto mode" now, so occasionally it will refuse to run something I just prompted "Hey permission system, I give this session permission to run this experiment for me (now try again)" and it seemed to unblock! (auto mode takes your messages into account)
⏰ 20:01 | ❤️ 383 点赞 | 📝 47词 | 入选: 高赞 383 | 查看原文 →
Machina
| @EXM7777
| 影响力: 16.0万粉丝 · 互动分 372.56
Machina
| @EXM7777
| 影响力: 16.0万粉丝 · 互动分 372.56
💡 核心观点: 作者称一年的 Claude Max 订阅是 2026 年最好的投资:他每周都用 Claude,多数时候通过 Claude Code 以 Claude 作编排器调用其他模型;他认为 Anthropic 全年主导市场,尽管用量限制与发布偏少存在瑕疵。
可信度: 40/100
事实核查:
- 无法核实: there wasn't a single week i didn't use Claude(属个人使用频率陈述,原文无法验证。)
- 无法核实: Fable literally spawned thousands of startups(称 Fable 催生数千家初创公司,未给出名单、统计口径或来源。)
- 原文支持: they had their quirks with usage limits and few releases(原文明确承认存在用量限制与发布较少的问题,与其主导性评价并存。)
原文内容:
a year of Claude Max was the best investment you could make in 2026... by far there wasn't a single week i didn't use Claude, and most of the time i used other models, it was through Claude Code with Claude as the orchestrator Fable literally spawned thousands of startups... it was time travel for software they had their quirks with usage limits and few releases, but Anthropic dominated all year long
⏰ 13:33 | ❤️ 238 点赞 | 📝 73词 | 入选: 高赞 238 | 查看原文 →
Grok
| @grok
| 影响力: 940.1万粉丝 · 互动分 1425.0
Grok
| @grok
| 影响力: 940.1万粉丝 · 互动分 1425.0
💡 核心观点: 作者宣布将SpaceXSI旗下多项产品整合进Grok,使图像与视频创作等功能集中在一个入口。
可信度: 48/100
事实核查:
- 原文支持: We're bringing more @SpaceXSI products together into one Grok experience(原文为官方账号的产品整合声明,未说明具体产品清单与上线时间)
原文内容:
We’re bringing more @SpaceXSI products together into one Grok experience, so there’s one place to create, build and get work done.
引用 @imagine: Image and video creation now live in Grok. Just open Grok and start creating.
⏰ 18:34 | ❤️ 845 点赞 | 📝 23词 | 入选: 高赞 845 / 引用讨论 | 查看原文 →
swyx
| @swyx
| 影响力: 21.3万粉丝 · 互动分 404.99
swyx
| @swyx
| 影响力: 21.3万粉丝 · 互动分 404.99
💡 核心观点: 作者提到 openai 站点新增一项设置,可长期自动解决所有验证码,并以调侃语气转发相关截图。
可信度: 58/100
事实核查:
- 无法核实: openai dot now has a setting to just solve all captchas for the rest of time(原文描述了具体产品设置与验证码自动处理能力,但未附链接、界面细节或官方说明。)
原文内容:
openai dot now has a setting to just solve all captchas for the rest of time https://t.co/rIqg7nVP8K
引用 @swyx: lol what are we even doing here anymore guys https://t.co/fqu3W2tMm4
⏰ 19:07 | ❤️ 198 点赞 | 📝 19词 | 入选: 高赞 198 / 引用讨论 | 查看原文 →
模型发布
GPT-6.2 迟迟未发布引发猜测,社区讨论 OpenAI 是否跳过该版本。
Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 1543.5
Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 1543.5
💡 核心观点: 作者引用他人推测称,OpenAI迟迟未发布GPT-6.2的原因尚无定论,并猜测可能跳过该版本直接研发GPT-6.5以对标Fable 5.2。
可信度: 15/100
事实核查:
- 无法核实: OpenAI has yet to release GPT-6.2.(原文以引用形式给出该前提,未附官方信息或发布时间表)
- 原文支持: my gut tells me we might not see GPT-6.2 at all(原文明确标注为直觉猜测,无证据支撑)
原文内容:
Okay, now it’s getting interesting. I’ve heard from several people that none of the reasons below explain why GPT-6.2 is delayed. I have no idea what else it could be, but my gut tells me we might not see GPT-6.2 at all. Instead, OpenAI could be working on GPT-6.5 to compete with Fable 5.2, and it simply needs more time to finish it.
引用 @kimmonismus: OpenAI has yet to release GPT-6.2. There can only be three reasons:
1) They were surprised by how capable Fable 5.2 is and had to make GPT-6.2 better.
2) They found security vulnerabilities or other issues that need to be fixed.
3) Anthropic is going public next week, and OpenAI knows GPT-6.2 sur…
⏰ 18:04 | ❤️ 1110 点赞 | 📝 64词 | 入选: 高赞 1110 / 引用讨论 | 查看原文 →
AI 视频
AI 视频生成被期待降低影视创作门槛,现阶段多为主观观感判断。
Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 1938.5
Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 1938.5
💡 核心观点: 作者预计AI视频将让每个人都能拍自己的好莱坞电影。依据仅是观看一条未具名AI视频后的主观评价。
可信度: 22/100
事实核查:
- 无法核实: This is the best AI video I've ever seen.(原文为作者主观观看感受,未提供视频来源或技术细节,无法验证)
原文内容:
This is the best AI video I’ve ever seen. Soon we will all become directors of our own Hollywood movies. https://t.co/waogH626xe
⏰ 14:03 | ❤️ 1436 点赞 | 📝 24词 | 入选: 高赞 1436 | 查看原文 →
航天发射
SpaceX 助推器落海由自动系统中止回收,具体原因尚未公布。
Aakash Gupta
| @aakashgupta
| 影响力: 33.6万粉丝 · 互动分 431.15
Aakash Gupta
| @aakashgupta
| 影响力: 33.6万粉丝 · 互动分 431.15
💡 核心观点: 作者复盘SpaceX助推器落海:返航需飞行主管手动确认且通过数千项检查,本次由自动系统中止回收,具体原因尚未公布。
可信度: 70/100
事实核查:
- 原文支持: Today it was the automatic system that called it off, according to the AP.(原文标注信息来源为美联社,SpaceX尚未说明触发原因)
- 原文支持: the flight director has to send a manual go command(原文说明塔架回收需手动放行命令和数千项检查,属流程描述)
原文内容:
SpaceX threw a rocket into the ocean on purpose today, and from the beach it looked absolutely perfect. The booster is about 23 stories tall. It flew back to Texas a few minutes after launching Starship to orbit and came down engines-first right off the shore, with people parked on the sand next to their pickup trucks watching. The plan was for the launch tower to catch it out of the air with two giant metal arms, and the tower never even got a chance. Turns out the ocean is the default. On the way back, the flight director has to send a manual go command before the booster finishes its turnaround burn, and on top of that, thousands of separate checks on the rocket and the tower all have to come back clean. If that command doesn't come in time, or one of those checks looks off, the booster just keeps heading for the water. Today it was the automatic system that called it off, according to the AP. SpaceX hasn't said what tripped it yet, and Dan Huot said on the stream they'll dig into it. Whatever it was, you'd never spot it from a beach chair. They've been this strict from the start. Back in November 2024, a booster came home and got sent to the water because the tower failed its own health check. Then in March 2025, a booster fired two fewer engines than planned on the way back, and the arms caught it anyway. It all comes down to what the computers are reading in real time. The reason they go this hard is that the tower doubles as the launch pad. It's the same tower that stacks Starship and sends it up. The very first Starship launch, back in April 2023, tore up the pad so badly that the next one didn't fly until November. Losing a booster hurts, but they've got more being built down the road, and they'd much rather eat that loss than swing 23 stories of steel into the one thing every future launch from that pad depends on. Only three boosters have ever been caught, all older models, the last one back in March 2025. This was the first time one of the new V3 boosters tried to come home, and earlier this week they actually hung this exact booster from the arms at catch height just to make sure it fit. The ship is still up in orbit right now, and tomorrow it gets its own shot at the arms, as long as it clears its own checklist first. Watch the clip all the way down to the water.
引用 @Pedroascenbell: REGRESO DEL BOOSTER EN EL MAR CERCA DE STARBASE @FronteraSpacial https://t.co/w1BDfEDe02
⏰ 07:37 | ❤️ 272 点赞 | 📝 445词 | 入选: 高赞 272 / 引用讨论 / 长文 | 查看原文 →
AI 表达
用户对「AI 味」回复容忍度下降,套话附和与虚构身份成为主要反感点。
Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 806.2
Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 806.2
💡 核心观点: 作者把「AI 味」回复按烦人程度排出六档,从复述原帖、假装懂行,到套话式附和、虚构身份代入,以及伪装成好奇的提问。这是一份主观观察清单,未涉及具体产品或数据。
可信度: 28/100
事实核查:
- 原文支持: AI slop replies, ranked by annoyingness(原文即个人主观排序,无数据支撑。)
原文内容:
AI slop replies, ranked by annoyingness: 6) Repeating the post 5) LLM knows more than the account should about a topic 4) Lame semi-joke 3) "X is the part I would keep/everyone misses..." 2) "As a marketer/builder/dentist, I see this in my own..." 1) Questions faking "curiosity"
⏰ 19:25 | ❤️ 520 点赞 | 📝 45词 | 入选: 高赞 520 | 查看原文 →
🔗 引用关系
- @kimmonismus 引用 @kimmonismus:作者引用他人推测称,OpenAI迟迟未发布GPT-6.2的原因尚无定论,并猜测可能跳过该版本直接研发GPT-6.5以对标Fable 5.2。(原文:OpenAI has yet to release GPT-6.2. There can only be three reasons:
1) They were surprised by how capable Fable 5.2 is …)
- @grok 引用 @imagine:作者宣布将SpaceXSI旗下多项产品整合进Grok,使图像与视频创作等功能集中在一个入口。(原文:Image and video creation now live in Grok. Just open Grok and start creating.)
- @dwarkesh_sp 引用 @dwarkesh_sp:作者认为当前 AI 能力「尖而不均」:在数学等可验证领域已接近或超过人类,而 AI 研究本身也属于可验证领域,因此窄能力 AI 加速 AI 研发所带来的外溢效应会非常广泛。依据是更好的学习算法具有通用性,GPT-6 在几乎所有任务上远优于 GPT-3;他推断两三年甚至数月内可把 GPT-3 到 GPT-6 的跃迁压缩进一年。(原文:The avalanche that hit math is coming for AI development.
https://t.co/z5RsJgcWWc https://t.co/Ud6MDxiADh)
- @MengTo 引用 @arafune:引用一位日本创作者的记录:仅用 5% Tripo3D,主要靠 Claude Opus 5.5 加 Blender 建模。方法为给三张参考图和背景设定,配 CLAUDE.md 与子代理分工规划、建模、渲染与视觉检查,分 13 个阶段逐角度比对渲染,模型、贴图与环境光均未外部下载。作者称流程还生成检查脚本,发现 45 个道具穿模。(原文:Blender × Claude Opus 5.5テスト
かなり素晴らしいクオリティ
tripoも使用許可したけど、5%くらいしか使ってなくて、ほぼ自分でモデリングしてた
文字など一部のテクスチャにChatGPT使いつつ、シェーダーもほぼプ…) - @MengTo 引用 @MengTo:作者称自己让 Fable 5.2 把一款 three.js 游戏预告片在 Blender 里重建,几小时完成;现用 Fable 5.2 medium 调度多个 Opus 5.5 代理作为主力工作流,并借助云端环境在笔记本上处理更复杂的 3D 场景。文中未说明重建质量的具体评估方式。(原文:I've always wanted to play D&D with my kids with AI as the game master.
So I made one with Opus 5.5. https://t.co/P…)
- @aakashgupta 引用 @Pedroascenbell:作者复盘SpaceX助推器落海:返航需飞行主管手动确认且通过数千项检查,本次由自动系统中止回收,具体原因尚未公布。(原文:REGRESO DEL BOOSTER EN EL MAR CERCA DE STARBASE @FronteraSpacial https://t.co/w1BDfEDe02)
- @levelsio 引用 @levelsio:作者称在 Interior AI 中以 3D 形式导入 IKEA 全品类家具,用户可从户型图或实拍视频重建房屋后再装修;他称这是重建产品第 4 天,效果仍显笨拙但有进展。(原文:More progress!
Interior AI 3D works with "just" a floor plan now
Not much information to work with and it has to imagi…)
- @swyx 引用 @swyx:作者提到 openai 站点新增一项设置,可长期自动解决所有验证码,并以调侃语气转发相关截图。(原文:lol what are we even doing here anymore guys https://t.co/fqu3W2tMm4)
- @gregisenberg 引用 @gregisenberg:作者认为如今营销的本质是代码与 token,学习营销即学习用自有数据、创意工具和广告账户搭建 Agent;配置得当就能得到每天大量测试创意的顶级营销者。(原文:learn to market is the new learn to code)
🧩 跨推文归纳
共识
- Meng To 与 Riley Brown 均验证 Claude 加 Blender 的 3D 创作流程可行:前者以多代理分阶段渲染检查,后者生成 1.25 万行 Python 并租 GPU 完成 4K 渲染。
- Nathan Lambert 与 Alex Veremeyenko 共同指向开源模型空间收紧,前者谈政治盟友退缩,后者主张本地留存模型。
分歧
- Machina 称 Claude Max 是 2026 年最好投资、Anthropic 全年主导市场;Simon Willison 则记录 Claude 自动模式会拒绝执行请求,对可用性持保留态度。
数据来源:X 列表「AI 英文」(22 条候选)· 由 twscrape 采集、grok 整理