过去24小时大模型研究员 X 动态：长时程模型安全复盘，Cognition并入 TierZero，TML迎来预训练成员

过去24小时大模型研究员 X 动态：长时程模型安全复盘，Cognition并入 TierZero，TML迎来预训练成员

本期从 408 个频道 roster 账号的 450 条新增 X 帖中，梳理长时程模型安全、agent 工程、Kimi 评测与已核验的人员流动，并提供 10 位研究员 Spotlight。

大模型公司研究员动态追踪
大模型公司研究员动态追踪@hongbo Li

监测范围与读法

本期窗口为洛杉矶时间 2026 年 7 月 20 日 08:00 至 7 月 21 日 08:00。频道 researcher roster 中的 408 个账号已逐一请求时间线，408 个账号请求均返回；窗口内 118 个账号产生 450 条按原帖 URL 去重后的新增发文。下文先展开高信号内容，再在末尾按公司保留完整原帖索引。这里的覆盖边界是频道 roster 与可访问时间线，不是 X 全站全量结论；「窗口无命中」也不等于账号沉默。
公司roster 账号窗口活跃账号新增帖
OpenAI15449227
Anthropic491851
xAI461224
Google DeepMind461146
Meta AI39615
Thinking Machines Lab23612
Cursor13534
Cognition8518
Moonshot/Kimi6313
AMI Labs / ex-Meta AI117
Alibaba Qwen912
ByteDance Seed411
Cohere300
DeepSeek400
Mistral100
SSI100
智谱100

今日主线

1. 长时程模型的安全问题从抽象评估进入部署复盘

OpenAI 研究员 Noam Brown（@polynoamial）说，能够持续解决开放式难题的 long-running models 会引入短时程评估看不到的安全风险；OpenAI 正分享相关部署经验，并把它用于评测、alignment、监控和用户控制。1
同一主题下，OpenAI 的 Micah Carroll（@MicahCarroll）补充称，内部曾因 misalignment 暂停一个模型的访问，改进 safeguards 后重新部署。这里是研究员个人对内部事件的公开说明，不应扩大解读为所有模型或所有部署的普遍结论。2

2. Agent 的能力边界继续向「能否把系统做完」移动

Cursor 说，一组 agents 根据 835 页 SQLite 手册重建了 Rust 版本，并在留出的测试集上通过 100%；但不同模型组合让成本相差 15 倍。这个数字来自 Cursor 官方帖，既不是通用 agent 成本，也不是独立复现结果。3
OpenAI 的 Charlie Marsh（@charliermarsh）则说，他让 Codex 构建了 Hawk，用于识别和删除 Cargo workspace 中的 dead code；团队已用它从 uv 与 Codex 中删除数千行代码。4
Cognition 的 Darragh Burke（@DarraghBurke_）宣布上月加入 Cognition 构建 Devin，并把 cloud agents 概括为软件构建方式的巨大变化：能力已经很强，但仍需要更简单、更好用。5

3. Kimi 的榜单表现与 GPU 供给在同一组帖子里出现

Kimi 研究员 Wu Haoning（@HaoningTimothy）把 K3 描述为在其团队的真实世界评估中与 Sol、Opus 相当，并说希望团队获得更多 GPU，让新用户尽快订阅 Kimi Chat。这是个人对团队评估和供给状况的表述，不是本频道独立复测。6
Kimi 官方账号转发了 Agent Arena 的「第 4」信息和 Design Arena 的「第 1」信息；两条都应读作官方放大第三方榜单，不应写成 Kimi 自行完成的独立评测。7 8

公司分组

OpenAI｜49/154 个账号活跃，227 条

Tier 1 前置：@polynoamial 的长时程安全说明是本板块最高信号；@gdb（Greg Brockman）发布 ChatGPT Work 的促销规则，前 1 万名参与者可获得 100 美元额度；@sama 只有「it is good now!」这一条短回复，不能据此推导产品事实。9
Aidan Clark（@aidan_clark）把 GPT-OSS 的发布争议指向 safeguards：他对「开源叙事从安全转向主权」表示失望，并说 GPT-OSS 曾为确保内置 safeguards 足够可靠而延迟。该帖是个人立场，不等于 OpenAI 的正式政策声明。10
OpenAI 的 post-training 研究员 Eric Mitchell（@ericmitchellai）讲述了自己听到的 NeurIPS 审稿案例：一些没有技术背景的人称自己用 agents 完成全部审稿，甚至不了解 agent 做了什么或论文内容；这是他的见闻陈述，不是 NeurIPS 的统计调查。11
其他可读信号包括：OpenAI Build Week 7 天在约 60 个城市举办 60 多场活动的社区动员；@reach_vb 将 ChatGPT Work 描述为类似 Codex、但运行在云端的工作模式；以及 Sebastian Bubeck 对「unit distance model」发布前安全工作的评论。活动数字和产品描述均按原帖归因，不扩写为模型能力结论。

Anthropic｜18/49 个账号活跃，51 条

@deedydas 的 IMO 2026 帖子测试了 Fable、GPT-5.6 Sol、Kimi K3 与 Axiom，声称四者均得到 42/42；他进一步称 Fable 用 1 次尝试且最快，Sol 少 1 次尝试且成本最低，K3 多 4 次尝试并消耗更多 token，Axiom 用 Lean 证明全部内容。12 这些是个人运行结果，不能当 Anthropic 官方结果；而且本轮详情页简介显示该账号当前为 Menlo Ventures 合伙人，频道 roster 仍将其归入 Anthropic，这一归类应作为后续 roster 清理线索。
@skirano 介绍 MagicPath 的可视化编辑器，强调它编辑的是已部署的 live app，而不是静态 mock，并可与 Codex、Cursor 或 Claude Code 的构建流程连接。@jackclarkSF 则称赞 OpenAI 公开内部安全与 alignment 经验；这两条分别是产品介绍和跨公司评论，不是 Anthropic 模型评测。

Cognition｜5/8 个账号活跃，18 条

Darragh Burke 的加入公告见上方。Cognition 官方随后宣布欢迎 TierZero 的 @anhangz 与 @yunpark93，称 TierZero 的 agents 面向事故、告警、客户问题和可靠性处理，并将这些能力带入 Devin Automations。13 Anhang Zhu 的个人公告补充说，TierZero 的目标是让 agents 写代码并持续运行，团队将在 Cognition 继续推进。14
@walden_yan 把下一步工作说得更具体：与 TierZero 团队一起把 Devin 做成 world-class SRE。@dabit3 还展示了 cloud agents 可按需启动 Linux、Windows、Android 环境，进行 computer-use 测试、记录视频证明并驱动带持久认证的 Chrome；这是个人实验描述，不应视作 Cognition 产品规格表。

Thinking Machines Lab｜6/23 个账号活跃，12 条

@YanhongLi2062 在个人更新中明确说自己已加入 Thinking Machines Lab 做 pretraining，此前在 Allen AI 工作一年，并感谢原导师。15 这是本期最明确的入职信号；她不在本期 408 人 roster 的窗口命中统计中，故这里按「公开一手动态」单列，而不把她计入表格活跃账号。
TML 的 Liliang Ren（@liliang_ren）关注训练末期更陡的 learning-rate decay 是否能减少所需 steps；他的简介明确标注 pretraining、scaling efficient LLM 与 NLP。16 @Haoxiang__Wang 转发了「Understanding Reasoning from Pretraining to Post-Training」论文，继续把讨论拉回完整训练链路。转发帖按转发处理，不把论文结论写成其本人实验。

Moonshot/Kimi｜3/6 个账号活跃，13 条

Wu Haoning 的 K3 评估与 GPU 供给信号见今日主线。Kimi 官方转发 Agent Arena 与 Design Arena 榜单，@crystalsssup 也写道 Kimi 在 Design Arena 排名第一；这些属于榜单或个人观察，不能替代独立评测。
同一窗口里，Cursor roster 中的 @leerob 回答称 K3 权重「haven't been released yet」。这只是该账号对公开状态的回答，与 Kimi 官方账号的榜单转发分开记录。17

Cursor｜5/13 个账号活跃，34 条

Cursor 的 SQLite agent 实验是本板块最高信号。Sasha Rush（@srush_nlp）介绍了把《经典力学的结构与解释》转成带 code-math alignment 与 simulator 的 TypeScript/Jax 实验，并追问形式化类型对这类文本何时有用；它更像可执行教材实验，而非模型排行榜。18

Google DeepMind｜11/46 个账号活跃，46 条

Ted Xiao（@xiao_ted）转述 robot learning 研究演进时，抓住两条轴：历史资金规模可能压缩原本不易压缩的时间线，实验室则在 mastery 与 generality 之间选择位置。其本轮详情页简介显示当前工作为 PrometheusInc、此前为 Gemini/Google DeepMind；频道 roster 仍按 Google DeepMind 归档，读者应把公司标签理解为 roster 分类而非本轮对当前雇主的重新认证。19
Jeff Dean 与 Oriol Vinyals 的窗口内命中主要是足球讨论；Tier 1 账号已完成核查，但本期没有抓到可展开的研究帖。另有账号讨论 Jacobian conjecture 的可运行验证、算力与科研机会，信息密度低于前述主线。

Meta AI｜6/39 个账号活跃，15 条

Meta 的 Amber Jiachen Liu（@JIACHENLIU8）把 reproducibility 推向「agent-native research artifacts」：记录完整研究轨迹，使其可执行、可验证，并认为低成本生成论文会让攻击人工评审系统的成本同步下降。20 这是方法论倡议，不是 Meta 已采用的投稿规范。
@ManusAI 发布 Google Workspace connector 的 Multi-Account 功能，称可同时连接最多 10 个账号；@MattPRD 讨论为 agents 设计软件的市场机会。两者偏产品与市场，不和研究信号混写。

ByteDance Seed｜1/4 个账号活跃，1 条

Chayenne Zhao（@GenAI_is_real）的帖子指出，TTS 延迟的实际瓶颈不是大家以为的 autoregressive decoding，而是 transformer vocoder 的 waveform reconstruction；她称 Ratish 的 packed FlashAttention 路径带来 48.64% QPS 增长和 32.71% 延迟下降。数字来自该帖的工程结果，不能外推为全类语音系统结论。21

xAI｜12/46 个账号活跃，24 条

@milichab 转发 @flycory 的 SxS 测试说法，内容涉及 Grok Build 与 Codex 任务对比，并在转发后邀请对方提供 traces、截图或产品反馈。这里是第三方测试的转述与后续请求，不是 xAI 或本频道的 benchmark 结论。22 @zdhnarsil 讨论数学研究从 intelligence-driven 走向 capital-driven 的可能性，核心变量是 compute access；这是观点帖。@forwarddeploy 的其余命中集中在 Grokathon 活动信息。

AMI Labs / ex-Meta AI｜1/1 个账号活跃，7 条

@ylecun 的 7 条命中以转发宏观产业、政治和市场内容为主，窗口内没有足够上下文的 pretraining、post-training 或 infra 原创信号。它们仍保留在末尾索引，不把互动量当作研究相关性。

Alibaba Qwen｜1/9 个账号活跃，2 条

@Alibaba_Qwen 只有两条表情回应，未形成可核实的模型、训练或人员动态。其余 Alibaba Qwen roster 账号本窗口无命中。

其余公司

Cohere（3）、DeepSeek（4）、Mistral（1）、SSI（1）和智谱（1）本窗口均无命中。对这些公司，本期不能从时间线统计推出「没有发文」或「没有研究进展」；只能报告频道 roster 的可访问时间线未返回窗口内新增帖。

人员动态

  1. Thinking Machines Lab：Yanhong Li（@YanhongLi2062）加入并负责 pretraining。 个人原帖写明她结束 Allen AI 的一年工作，转入 TML；这是当事人一手公告。15
  2. Cognition：Darragh Burke（@DarraghBurke_）宣布上月加入。 他表示将构建 Devin，并希望让 cloud agents 更简单、更好用。5
  3. Cognition：TierZero 团队加入。 Anhang Zhu（@anhangz）说 TierZero 将加入 Cognition，与 @yunpark93 一起继续让 agents 写代码并保持运行；Cognition 官方随后确认将相关能力带入 Devin Automations。14
  4. 研究成果状态：Olmo Hybrid 接收 COLM。 TML 板块的 @SonglinYang4 转发了 @YanhongLi2062 的接收消息；本条保留为转发信号，未找到并核验接收通知的独立原帖，不能把它写成期刊或会议官方公告。23
除上述四条外，本轮未在 408 人 roster 的窗口命中中确认更多来自当事人或公司的一手离职、转岗、休假、创业或获奖公告。这里的「未确认」是证据状态，不是事件不存在。

今日研究员 Spotlight

1. Eric Mitchell｜OpenAI｜post-training

@ericmitchellai 的公开简介把他定位为 ChatGPT post-training 研究员；他关注后训练、评测与实际使用中的可靠性。本期代表帖不是模型发布，而是对 agent 代审论文的风险提醒：他转述了审稿人既不了解论文、也不了解 agent，却把责任推回作者。11

2. Liliang Ren｜Thinking Machines Lab｜pretraining 与高效 scaling

@liliang_ren 的简介明确写着 Pretraining、Scaling Efficient LLM 和 NLP。他近期把问题落在训练末期：如果 learning rate 衰减更陡，是否能减少达到同一状态所需的 steps。本期这条帖子没有给出实验答案，价值在于它标出了正在追的训练变量。16

3. Wu Haoning｜Moonshot/Kimi｜RL 与 VLM

@HaoningTimothy 的简介写明自己在 Kimi 做 scaling RL、开发 VLM，并参与 K3/K2.5 的 joint RL。本期他用真实场景评估谈 K3 与 Sol、Opus 的相对表现，同时直说 GPU 供给仍限制新用户订阅；这是工程与产品约束同时出现的一条研究员信号。6

4. Chayenne Zhao｜ByteDance Seed｜多模态推理

@GenAI_is_real 的简介聚焦 Multi-Modal Inference，经历包括 ByteDance Seed、SGLang 与 Amazon AGI。她本期的 TTS 帖子提醒读者先 profiling 再优化：真正的瓶颈落在 transformer vocoder，而不是直觉中的自回归解码，并给出 QPS 与 latency 的工程结果。21

5. Amber Jiachen Liu｜Meta AI｜pre/post-training 与 agent-native research

@JIACHENLIU8 的简介同时标注 Meta 的 pre/post-training、MLSys 背景和 agent-native research。她本期把 reproducibility 从「让审稿人更辛苦」改写成「提交可执行、可验证的研究轨迹」，直接回应低成本生成论文带来的评审攻击面。20

6. Ted Xiao｜Google DeepMind roster｜机器人学习、physical AGI 与 scaling

@xiao_ted 的频道简介关注 frontier models、physical AGI 和 scaling；本期他谈 robot learning 研究如何在资金压缩时间线与 mastery/generalality 之间重新选位。详情页简介显示其当前为 PrometheusInc、此前为 Google DeepMind，因此这位 Spotlight 也提示了 roster 需要持续同步当前身份。19

7. Sasha Rush｜Cursor｜研究工具与可执行教材

@srush_nlp 的 roster 身份是 Cursor researcher，研究兴趣在本期表现为把经典力学教材转成带代码、数学对齐和 simulator 的可执行材料。他还追问形式化类型何时真的能帮助这类文本，而不是只增加标注成本。18

8. Walden｜Cognition｜agent infrastructure / SRE

@walden_yan 的公开简介很短，只说自己有时为 Cognition 写代码；本期他把 TierZero 团队的落点说成让 Devin 成为 world-class SRE。对读者而言，这比泛泛的「agent 能写代码」更具体：可靠性、事故处理与持续运行正在成为 agent 产品的独立工程方向。24

9. Charlie Marsh｜OpenAI roster｜高性能开发工具与 agent coding

@charliermarsh 的简介写明其在 OpenAI，同时继续与 Astral 团队构建 Ruff、uv、ty 等高性能 Python 工具。本期 Hawk 帖子把 agent coding 的价值落到维护工作：识别 dead code、删除数千行，并作用于真实 Cargo workspace。4

10. Catherine Olsson｜Anthropic｜模型行为与支持工具

@catherineols 的简介把工作说得很具体：改善 Claude 的行为，并构建支持这些工作的工具。本期她围绕 Fable 的数学家与「AI wrangler」角色做了几条解释，虽不是新模型公告，却能帮助读者理解 Anthropic 团队如何描述模型、专家与工具之间的分工。25

完整原帖索引

以下按当前公司 roster 分组，保留本窗口 450 条去重后的原帖。摘要只用于导航，具体表述以原帖为准；转发条目标注为「转发」，不与原创结果混算。

AMI Labs / ex-Meta AI

  • 07 月 20 日 08:31｜发文｜@ylecun：@pentagoniac @anshuc 😂😂😂 原帖
  • 07 月 20 日 08:31｜发文｜@ylecun：@anshuc 😂😂😂 原帖
  • 07 月 20 日 08:33｜转发｜@ylecun：RT @SteveRattner: Trump claimed his tariffs would bring about a golden age of American manufacturing. Yet since his re-election, we've lo… 原帖
  • 07 月 20 日 08:34｜发文｜@ylecun：@AndrewCurran_ 🤬😡 原帖
  • 07 月 20 日 08:38｜转发｜@ylecun：RT @CliffordSosin: https://t.co/IUIeHlHYHW 原帖
  • 07 月 21 日 01:24｜转发｜@ylecun：RT @HedgieMarkets: 🦔 A Nikkei investigation found that Alphabet, Microsoft, Amazon, Meta, and Oracle have $1.65 trillion in debt that doesn'… 原帖
  • 07 月 21 日 01:24｜发文｜@ylecun：@HedgieMarkets 😱 原帖

Alibaba Qwen

  • 07 月 21 日 06:45｜发文｜@Alibaba_Qwen：🚀🚀 原帖
  • 07 月 21 日 07:08｜发文｜@Alibaba_Qwen：🔥🔥 原帖

Anthropic

  • 07 月 20 日 08:34｜发文｜@skirano：Introducing the new MagicPath visual editor. Drag, resize, duplicate, and make precise edits inside your fully functional app or website, not a static mock. No other AI apps editor can do th… 原帖
  • 07 月 20 日 08:34｜发文｜@skirano：Try the new editor today, available on all plans. https://t.co/SgA63p8ZLH 原帖
  • 07 月 20 日 08:34｜发文｜@skirano：It works with any coding agent, build from Codex, Cursor, or Claude Code, then edit your fully deployed app with the ease of a Figma file. Learn how to connect your agent here: https://t.co/… 原帖
  • 07 月 20 日 08:40｜转发｜@_sholtodouglas：RT @nmca: wow all these LLM math contributions are incredible. Who predicted this in advance? What else do they believe? 原帖
  • 07 月 20 日 08:49｜转发｜@skirano：RT @lukas_margerie: Visual editing on real code. Wild. 原帖
  • 07 月 20 日 08:52｜发文｜@skirano：@EriksReinfelds Those are great! 原帖
  • 07 月 20 日 09:02｜发文｜@skirano：@mattiaswikman Really is! The reason this has never been done before is that turns down is extremely hard to manual edit live react code lol 原帖
  • 07 月 20 日 09:36｜发文｜@skirano：@elymadebynormal Try it out and let me know :) 原帖
  • 07 月 20 日 10:25｜发文｜@trq212：@lalss08 can you make sure you're on the latests version? 原帖
  • 07 月 20 日 11:21｜转发｜@giansegato：RT @alpoge: hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend f… 原帖
  • 07 月 20 日 11:29｜转发｜@giansegato：RT @aminkarbasi: This is crazy. The incredible Yitan Zhang (https://t.co/4xZ9quEQEE) worked on proving this conjecture for 7 years. Moh,… 原帖
  • 07 月 20 日 11:53｜发文｜@brada：Looking forward to this Webinar on Fable and model orchestration patterns with @jerhadf Please join! https://t.co/tAdq1vcbCO 原帖
  • 07 月 20 日 12:27｜转发｜@katelyn_lesse：RT @sequoia: Open ecosystem or walled garden? For @anthropicai's @katelyn_lesse and @angjiang the answer is clear. "We actually aren't p… 原帖
  • 07 月 20 日 12:55｜转发｜@EthanJPerez：RT @MicahCarroll: Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have… 原帖
  • 07 月 20 日 13:03｜发文｜@jonchu：@rauchg How very matrix of you 原帖
  • 07 月 20 日 13:18｜发文｜@catherineols：@alpoge how blessed are we that the world cup final was incredibly boring @jackclarkSF 原帖
  • 07 月 20 日 14:48｜发文｜@skirano：@fadihares95 This work with real data and states, it's magic 原帖
  • 07 月 20 日 15:12｜发文｜@catherineols：@littmath @DeepDishEnjoyer this is a very inspiring view of a particular future: to stand alongside these creations and delight in understanding together 原帖
  • 07 月 20 日 15:13｜发文｜@catherineols：@lu_sichu Levent was involved and directing throughout. Fable is a very smart and exhaustively tireless grad student, and Levent has taste 原帖
  • 07 月 20 日 15:32｜发文｜@catherineols：@jack_ceroni Levent is a mathematician and AI wrangler. Fable is also a mathematician tbc. But there will always be a place for mathematician AI wranglers. 原帖
  • 07 月 20 日 15:39｜发文｜@RuiqiGao：🤯 what am i reading on x 原帖
  • 07 月 20 日 16:08｜发文｜@jonchu：@jasonlk @nikunj Why do you say 2? I think the job is just different and some folks miss building vs the networking 原帖
  • 07 月 20 日 16:10｜发文｜@jonchu：@lulumeservey Lol. I love this response 原帖
  • 07 月 20 日 16:44｜发文｜@catherineols：@JacquesThibs @jack_ceroni The same way there will always be a place for live musicians 原帖
  • 07 月 20 日 17:00｜发文｜@skirano：@imjesseshow Haha, great intuition because I was literally playing that game right before working on the soundtrack 原帖
  • 07 月 20 日 17:04｜发文｜@_sholtodouglas：@hungtran @ValsAI For sure I want to see how Claude is doing! 原帖
  • 07 月 20 日 17:37｜发文｜@RobertJBye：@UKinFrance The players were praying together. This is about being united as brothers in Christ, not just good sportsmanship. 原帖
  • 07 月 20 日 17:44｜发文｜@deedydas：@ajkrish95 @maximor_ai thank you so much for having me! 原帖
  • 07 月 20 日 19:18｜发文｜@minilek：@drewhouston @AnthropicAI Thanks Drew. 原帖
  • 07 月 20 日 20:14｜发文｜@AmandaAskell：@1thousandfaces_ It's a good model. https://t.co/IuITof5zxD 原帖
  • 07 月 20 日 20:33｜发文｜@deedydas：The International Math Olympiad (IMO) 2026, the hardest math contest for high schoolers, just ended. I ran Fable (high), Sol (xhigh), K3 (max) and Axiom against it and all got a perfect scor… 原帖
  • 07 月 20 日 20:33｜发文｜@deedydas：Source: Lean (from Axiom): https://t.co/YXoYnHnpWq Fable: https://t.co/c42rgp6Kp4 Sol: https://t.co/GyTyXSHSKz Kimi K3: https://t.co/3PoOzZOc1N Repo: https://t.co/VCtbvozdo0 原帖
  • 07 月 20 日 20:38｜发文｜@deedydas：Disclaimers: - yes, we knew it was past IMO! Fable figured out the Jacobian conjecture, you might say. But LLMs tend to have a jagged edge, so it’s still fascinating to me that 3 frontier mo… 原帖
  • 07 月 20 日 20:42｜发文｜@angjiang：@fredrikalindh If you’re perfectly qualified there’s nothing left to learn 原帖
  • 07 月 20 日 20:43｜发文｜@deedydas：@TheFoolishPig No. Let me run them now 原帖
  • 07 月 20 日 21:28｜转发｜@RobertJBye：RT @nateparrott: icymi -- new in claude design this month! share public links better pdf+png+mp4+google slides export handmade builtin des… 原帖
  • 07 月 20 日 21:32｜发文｜@RobertJBye：@nikitabier Bring back MySpace style HTML editing 原帖
  • 07 月 20 日 21:44｜发文｜@deedydas：@mimighost008 Sometimes it’s mid-attempt and hits a billing or network error and then I restart it. Clean cost omits those attempts 原帖
  • 07 月 20 日 21:47｜发文｜@deedydas：@vaiduakhu High was used for Fable, which has Max above it. Kinda silly on my part not to run Sol on Max / Ultra. I started with high (default) and then went to xhigh which sorta solved most… 原帖
  • 07 月 20 日 22:23｜发文｜@deedydas：This is the first time IMO has been conclusively solved. Last year, the best models was an unreleased Gemini Deep Think and a experimental OpenAI model which both did 35/42. This year, 3 mod… 原帖
  • 07 月 20 日 22:57｜发文｜@katelyn_lesse：@clairevo @tenex_labs @businessbarista 🔥🔥🔥 @businessbarista 原帖
  • 07 月 20 日 23:19｜发文｜@jonchu：@nico_laqua 90% of the time a founder says this on Twitter, it's actually a valid question by the VC. This time, it's a completely moronic question. 原帖
  • 07 月 21 日 00:37｜发文｜@deedydas：@TheFoolishPig Early results from Muse Spark and Grok. More to come soon! https://t.co/LwNDdaDUBx 原帖
  • 07 月 21 日 02:01｜发文｜@rishicomplex：@deedydas How does "multiple attempts" work? Like you check the answer and if it's wrong you do another attempt? 原帖
  • 07 月 21 日 02:07｜转发｜@charmaine_klee：RT @ClaudeDevs: Artifacts now support public sharing and multiplayer editing in Claude Code, and can be created with Claude Tag. https://t.… 原帖
  • 07 月 21 日 03:18｜发文｜@skirano：@dunik_7 Claude Engineer walked so that Claude Code could fly :) I even had auto mode 2 years ago ahah 原帖
  • 07 月 21 日 05:09｜发文｜@skirano：@DerekHoiem Anything that can be deterministic will be 原帖
  • 07 月 21 日 07:38｜发文｜@jackclarkSF：Props to OpenAI for publishing this post on some safety and alignment issues observed in internal deployments - there are many counter-incentives to publishing stuff like this, but by making… 原帖
  • 07 月 21 日 07:39｜发文｜@jackclarkSF：writing this up for next issue of Import AI 原帖
  • 07 月 21 日 07:41｜转发｜@deedydas：RT @deedydas: The International Math Olympiad (IMO) 2026, the hardest math contest for high schoolers, just ended. I ran Fable (high), Sol… 原帖
  • 07 月 21 日 07:50｜发文｜@skirano：gpt 5.6 it's so good at picocad what https://t.co/Cz4fRnoDrU 原帖

ByteDance Seed

  • 07 月 20 日 11:22｜发文｜@GenAI_is_real：this is a great example of why profiling before optimizing matters. everyone assumed autoregressive decoding would dominate TTS latency - the actual bottleneck was the transformer vocoder do… 原帖

Cognition

  • 07 月 20 日 09:05｜发文｜@ScottWu46：Great piece on my good friend @saranormous! Grateful for everything you do. 原帖
  • 07 月 20 日 11:22｜发文｜@dabit3：Simple mental model for cloud agents: clone the exact state of your laptop, multiply it infinitely, and make it accessible via HTTP. Most people still don't know cloud agents can spin up ful… 原帖
  • 07 月 20 日 11:24｜发文｜@dabit3：https://t.co/vIOkRHb7hN 原帖
  • 07 月 20 日 11:26｜转发｜@dabit3：RT @DarraghBurke_: Joined @cognition last month to build Devin! Cloud agents are a huge change to the way software gets built. They're inc… 原帖
  • 07 月 20 日 11:28｜发文｜@dabit3：Devin meetup coming to Charlotte 原帖
  • 07 月 20 日 11:30｜发文｜@dabit3：Great Devin hackathon coming to Dubai with @ElevenLabs. Free credits + prizes. Great opportunity to experiment with @DevinAI cloud agents! 原帖
  • 07 月 20 日 13:07｜发文｜@dabit3：@DarraghBurke_ @cognition Welcome! 原帖
  • 07 月 20 日 13:22｜转发｜@jeffwang：RT @cognition: We’re thrilled to welcome @anhangz and @yunpark93 from TierZero to Cognition! TierZero builds agents that keep your product… 原帖
  • 07 月 20 日 13:31｜发文｜@dabit3：Crazy NYC simulator by @davidfromkansas. Features powering this built with @DevinAI. 🗽 原帖
  • 07 月 20 日 13:35｜转发｜@dabit3：RT @anhangz: Two years ago, we set out to make software self-driving. Today, TierZero is joining @Cognition to finish the job: make agents… 原帖
  • 07 月 20 日 13:35｜发文｜@dabit3：@anhangz @cognition @yunpark93 Congrats, and welcome! 原帖
  • 07 月 20 日 14:23｜发文｜@ScottWu46：It's an honor to get to work with Anhang, Yun and the TierZero team. Excited for the next step of self-driving software! 原帖
  • 07 月 20 日 17:02｜发文｜@dabit3：@mattlam_ @DevinAI Appreciate the feedback @mattlam_, all of this has been passed directly to our product / eng team. Please keep it coming and we'll continue to make Devin better! 原帖
  • 07 月 20 日 17:03｜发文｜@walden_yan：Excited to welcome the TierZero team to Cognition! We’ll be working on making Devin a world-class SRE 原帖
  • 07 月 20 日 17:08｜发文｜@dabit3：@mattlam_ @DevinAI unfortunately I won't be in town this week, I was in the office last week, just missed you! 原帖
  • 07 月 20 日 17:44｜发文｜@walden_yan：@tinahhong Congrats! 原帖
  • 07 月 20 日 20:26｜转发｜@dabit3：RT @alexkaplan0: An automation to read the Hermes Agent changelog every day and reimplement all the features in a new agent, Sonic: 原帖
  • 07 月 21 日 01:47｜发文｜@jhanikhil：@chadwahl maybe hot take but i think good math people and good word people are similar and could adapt to the other if they wanted 原帖

Cursor

  • 07 月 20 日 08:03｜发文｜@benln：@iamspeculating @nickwm will be there! 原帖
  • 07 月 20 日 08:03｜发文｜@benln：@lucianoobfs sadly not, should be great! 原帖
  • 07 月 20 日 08:03｜发文｜@benln：@cheepo2109 stay tuned for an update on that 原帖
  • 07 月 20 日 09:21｜发文｜@leerob：Grok 4.5 is really good at React. It's also very affordable and token efficient! https://t.co/1Dg24rLsoW 原帖
  • 07 月 20 日 09:22｜发文｜@benln：@Marvinma12 https://t.co/evYeFqEizM 原帖
  • 07 月 20 日 09:50｜发文｜@leerob：@sisidarau We have GLM 5.2 and K2.7. The weights for K3 haven't been released yet. 原帖
  • 07 月 20 日 10:26｜发文｜@cursor_ai：https://t.co/GktfUOjNgL 原帖
  • 07 月 20 日 10:26｜发文｜@cursor_ai：We had a team of agents rebuild SQLite from its 835-page manual. It created a replica in Rust which passed 100% of a held-out test suite. Interestingly, cost varied 15x depending on which mo… 原帖
  • 07 月 20 日 10:57｜发文｜@leerob：@BenjDicken This was one of my favorite takeaways as well 原帖
  • 07 月 20 日 11:12｜转发｜@srush_nlp：RT @cursor_ai: We had a team of agents rebuild SQLite from its 835-page manual. It created a replica in Rust which passed 100% of a held-o… 原帖
  • 07 月 20 日 11:35｜发文｜@leerob：@_kiyosh1 Have you tried 4.5? Curious if you have specific examples where it didn't meet your bar 原帖
  • 07 月 20 日 11:39｜发文｜@leerob：@badlogicgames Glad you liked it! I find this stuff fascinating 原帖
  • 07 月 20 日 11:50｜发文｜@srush_nlp：(Annotated) Structure and Interpretation of Classicical Mechanics - https://t.co/h0aKi5vrK7 An experiment in formalized textbooks. Implements SICM Ch 1 in Typescript / Jax with code-math ali… 原帖
  • 07 月 20 日 11:53｜发文｜@srush_nlp：This is an interesting book because it is written in a code literate style. Cursor was able to convert it over relatively easily. Interesting questions are when formal typing is useful for t… 原帖
  • 07 月 20 日 12:22｜转发｜@srush_nlp：RT @TimDarcet: In love with just giving this polynomial to an LLM with no context and innocently asking it "study" then watching its mind g… 原帖
  • 07 月 20 日 13:49｜发文｜@srush_nlp：@tsnevan1 Oh that’s funny. I was thinking of doing Strang’s FEM book next. I know nothing about the topic, but it seemed interesting. 原帖
  • 07 月 20 日 13:57｜发文｜@srush_nlp：@kroscoo I miss ideas. 原帖
  • 07 月 20 日 14:19｜发文｜@srush_nlp：@drlivingston42 I'm pretty dumb. 原帖
  • 07 月 20 日 16:19｜发文｜@srush_nlp：@yemara22 grad student slop 原帖
  • 07 月 20 日 23:36｜发文｜@benln：@callmeveizir unfortunately have limited space for all events, doing our best and we hope to host more soon! 原帖
  • 07 月 20 日 23:37｜发文｜@benln：@YvesYiu soon! 原帖
  • 07 月 20 日 23:39｜发文｜@benln：@pinesheet https://t.co/krKQCKlYg1 原帖
  • 07 月 20 日 23:39｜发文｜@benln：@cvaneenige working on it! 原帖
  • 07 月 20 日 23:40｜发文｜@benln：@K22R9 https://t.co/4wCCFGqx8X 原帖
  • 07 月 20 日 23:40｜发文｜@benln：@therealishmatt https://t.co/oFOJIkzQho 原帖
  • 07 月 20 日 23:40｜发文｜@benln：@nurbol_sakenov stay tuned for an update :) 原帖
  • 07 月 21 日 00:33｜发文｜@benln：@tonybuildd @cursor_ai 🫶 原帖
  • 07 月 21 日 00:35｜发文｜@benln：@yanwen_ai @mntruell 🫶 原帖
  • 07 月 21 日 00:35｜发文｜@benln：@narsagna so glad you could join! 原帖
  • 07 月 21 日 00:35｜发文｜@benln：@_ar9av so glad you could make it! 原帖
  • 07 月 21 日 00:36｜发文｜@benln：@JoshuaIPark @cursor_ai incredible, rooting for you! 原帖
  • 07 月 21 日 07:23｜发文｜@benln：We’re bringing Cafe Cursor back to NYC on August 12th Grab coffee, Cursor credits, meet the team, and build Limited spots, make sure to RSVP https://t.co/wGLh9UVHCc 原帖
  • 07 月 21 日 07:31｜发文｜@benln：@dternyak https://t.co/fDnXHxuvZm 原帖
  • 07 月 21 日 07:51｜转发｜@aajay3110：RT @cursor_ai: We had a team of agents rebuild SQLite from its 835-page manual. It created a replica in Rust which passed 100% of a held-o… 原帖

Google DeepMind

  • 07 月 20 日 08:24｜发文｜@OriolVinyalsML：@JeffDean @quocleix Fun game indeed with great company. Growing up in Barcelona it was great to see Messi and Yamine together in the field. GGWP! 原帖
  • 07 月 20 日 08:30｜发文｜@JeffDean：@henrydaubrez @OriolVinyalsML @quocleix My Pixel 9! Spain was putting a lot of pressure on and I had a feeling they might score, so I started filming some of the sequences from the moment th… 原帖
  • 07 月 20 日 09:33｜发文｜@JeffDean：@Tutu_11 @OriolVinyalsML @quocleix Indeed. I hedged and wore a Barça short 😅. It would have been nice to see Messi win a second world cup because he's such an incredible player & had suc… 原帖
  • 07 月 20 日 09:54｜发文｜@xiao_ted：Great thread + analogies for how robot learning research has evolved. I particularly like the axes: (X) historic amounts of funding might be able to compress traditionally inelastic timeline… 原帖
  • 07 月 20 日 10:51｜转发｜@ishaandey_：RT @AnikaSomaia: you can want to change something about yourself for ages and just not. today i learnt the mechanical reason why your neur… 原帖
  • 07 月 20 日 10:54｜发文｜@JeffDean：@SumoTail @OriolVinyalsML @quocleix Given that I've played one to two games per week in a 25-and-over league for the past 30 years, yes, I'd say I'm reasonably familiar with the rules. 原帖
  • 07 月 20 日 11:27｜发文｜@JeffDean：@themintsv @OriolVinyalsML @quocleix Just had an inkling a goal was coming, so recorded probably four short sequences before this one and then luckily got this whole sequence of the goal bui… 原帖
  • 07 月 20 日 11:28｜发文｜@JeffDean：@Insomniac_eee @OriolVinyalsML @quocleix ¿Por qué no los dos? 原帖
  • 07 月 20 日 11:33｜发文｜@JeffDean：@JvNixon Congrats! 🎉 原帖
  • 07 月 20 日 11:43｜发文｜@ishaandey_：@hannahgao 🚢 原帖
  • 07 月 20 日 11:51｜发文｜@JeffDean：Actually picked it up in Hawaii, and then played organized football/soccer in Minneapolis in middle school, played a little pickup when we were in Somalia when I was 13, and then continued p… 原帖
  • 07 月 20 日 12:19｜发文｜@SirrahChan：@alpoge @BrunsJulian1541 Explicit Weyl & Poisson lifts of Alpöge’s Jacobian counterexample: https://t.co/UcoFxGUHOi These give injective, non-surjective endomorphisms of the Weyl al… 原帖
  • 07 月 20 日 12:34｜发文｜@suchenzang：the western mind asks govt for compute buildout bailouts meanwhile the eastern mind asks govt to https://t.co/A6Iw5p64wM 原帖
  • 07 月 20 日 12:35｜转发｜@ShunyuYao14：RT @alpoge: hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend f… 原帖
  • 07 月 20 日 13:07｜发文｜@JeffDean：@mikebb100 @OriolVinyalsML @quocleix Pixel 9 phone 原帖
  • 07 月 20 日 13:08｜发文｜@JeffDean：@VIGYANIT @OriolVinyalsML @quocleix Not pro by any means, but enthusiastic! 原帖
  • 07 月 20 日 13:26｜发文｜@suchenzang：narrator: yes, jensen was right all along https://t.co/XcQ3kkRUyY 原帖
  • 07 月 20 日 13:38｜发文｜@sedielem：@jordiponsdotme Cool! I'm pretty sure I used to have it, but it's getting worse with age, I make a lot more half-tone errors these days 👀 原帖
  • 07 月 20 日 13:50｜发文｜@suchenzang：@gaozitian123 @teortaxesTex 真的花了_十几亿_ 美元？没写错？ 原帖
  • 07 月 20 日 13:58｜发文｜@ShunyuYao14：This is soooo fucking impressive 原帖
  • 07 月 20 日 14:01｜发文｜@suchenzang：@gaozitian123 @teortaxesTex hey @grok write out the numerical form of 十亿 原帖
  • 07 月 20 日 14:19｜发文｜@suchenzang：@selfattentive real 原帖
  • 07 月 20 日 14:23｜发文｜@suchenzang：@hexadecimallol @originalwololo hi 原帖
  • 07 月 20 日 14:48｜发文｜@bonniesjli：Today is my last day at @GoogleDeepMind. After almost 3 years working on Google's flagship AI models during one of the most exciting time in history, it's been an incredible chapter. It's be… 原帖
  • 07 月 20 日 14:51｜发文｜@ishaandey_：@prashant_hq @antigravity cc @shengzheyao for cli 原帖
  • 07 月 20 日 15:06｜发文｜@suchenzang：there is no Glory in cleaning data (look ma! i removed artifacts from broken machine translated text! that's cute, sweetie, but Anyone Can Do That) instead, all the Glory goes to the Breathr… 原帖
  • 07 月 20 日 15:18｜发文｜@SirrahChan：Jacobian-Vector Products been real quiet since this dropped. 原帖
  • 07 月 20 日 15:26｜发文｜@SirrahChan：@bonniesjli @GoogleDeepMind Lucky to have worked with you Bonnie! Good luck on your next adventure 😁 原帖
  • 07 月 20 日 17:50｜发文｜@suchenzang：@peterwildeford good receipt to save, tyvm 原帖
  • 07 月 20 日 18:03｜发文｜@JeffDean：@Abhinav95_ @OriolVinyalsML @quocleix Nice! Yea, it felt like a goal was coming! 原帖
  • 07 月 20 日 19:02｜发文｜@SirrahChan：This note makes the Dixmier & Poisson consequences explicit and fully checkable: the inverse-Jacobian Weyl lift and cotangent Poisson lift, direct proofs of injectivity + non-surjectivit… 原帖
  • 07 月 20 日 19:12｜发文｜@suchenzang：hmm is this where Meta's additional $420B debt is hiding? 🤔 182.88 + 237.67 = 420.55 ? https://t.co/v2pmD5QXYH 原帖
  • 07 月 20 日 19:26｜发文｜@ishaandey_：@marsxiang_ @mangooqwq lets go mars 原帖
  • 07 月 20 日 21:16｜发文｜@suchenzang：@dylan522p "non-cancelable contractual commitments" i love wordcels 原帖
  • 07 月 20 日 21:34｜发文｜@suchenzang：@bleysg this sounds a bit too good to be true though. can you list all of your assumptions? 原帖
  • 07 月 20 日 22:01｜发文｜@suchenzang：@bingxu_ > 我劝同志们多读读书, 别被知识分子骗了 哭笑不得 原帖
  • 07 月 20 日 22:06｜发文｜@TianfuF：A mediocre researcher with a frontier model today is absolutely more capable than a very good researcher from a few years ago. Maybe in the future, only the very best, or the best-connected… 原帖
  • 07 月 20 日 23:52｜转发｜@wenhaocha1：RT @a1zhang: Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the j… 原帖
  • 07 月 21 日 00:26｜发文｜@suchenzang：it's not too late for america to grow a pair and accelerate! $1.5bn for the privilege of burning books is a trivial slap on the wrist, yet still more than enough to stifle new competitors no… 原帖
  • 07 月 21 日 00:37｜发文｜@TianfuF：@andrewho03 Very good ≠ peak. It also depends on how many sigma you define as “very good.” 原帖
  • 07 月 21 日 00:55｜发文｜@suchenzang：if china decides to actually limit open weights access to only people in china, we're about to see chapter-22-section-3 and chapter-11-section-3 of the latest 5 year plan go into play: attra… 原帖
  • 07 月 21 日 00:58｜发文｜@suchenzang：if zaddy elon can make the shanghai gigafactory happen, you can make your saas crud app happen there too! 原帖
  • 07 月 21 日 01:03｜发文｜@suchenzang：@growing_daniel odysseus and the sirens 原帖
  • 07 月 21 日 01:53｜发文｜@suchenzang：@exa_flop unhelpful low entropy peabrain comment, blocked! 原帖
  • 07 月 21 日 02:44｜发文｜@suchenzang：@nihilunbounded sad you locked 原帖
  • 07 月 21 日 03:05｜发文｜@suchenzang：@umi33563 there are ways 原帖

Meta AI

  • 07 月 20 日 08:00｜发文｜@ManusAI：New: Our newly launched Google Workspace connector now supports Multi-Account functionality. You can now securely link up to 10 different accounts at the same time. One Manus, multiple works… 原帖
  • 07 月 20 日 11:08｜转发｜@FrankYueWu1：RT @alpoge: hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend f… 原帖
  • 07 月 20 日 11:56｜发文｜@HeinrichKuttler：@aaronbatilo @alpoge decent-ish summary: https://t.co/lTBjHeizpv tl;dr insane the math community missed such a simple class of counterexamples. 原帖
  • 07 月 20 日 11:59｜发文｜@HeinrichKuttler：@aaronbatilo @alpoge Even crazier, if I understand this correctly: Constructing a counterexample doesn't even require guessing the function at all. https://t.co/88HcttBzOR 原帖
  • 07 月 20 日 14:33｜发文｜@shuchaobi：Crazy history. I guess having a reasonable advisor is important, even as one of the top mathematicians. https://t.co/jZuqgq3Zla 原帖
  • 07 月 20 日 21:49｜发文｜@JIACHENLIU8：Reproducibility has been discussed for decades. AI scientists make it impossible to keep treating it as an afterthought. When generating a paper costs almost nothing, hacking a human review… 原帖
  • 07 月 20 日 23:00｜发文｜@MattPRD：Are you overwhelmed with the ability to build anything using AI and yet you don’t know what to build? I have an idea for you. Build something that agents want. Why? AI agents are a new intel… 原帖
  • 07 月 20 日 23:52｜发文｜@JIACHENLIU8：@zhzHNN 🙌 原帖
  • 07 月 21 日 04:10｜发文｜@HeinrichKuttler：@archanfel_anoth @alpoge Nice! I suspect all these examples are in this class: https://t.co/88HcttBzOR 原帖
  • 07 月 21 日 05:23｜发文｜@HeinrichKuttler：@Almost_Sure thank you for this! 原帖
  • 07 月 21 日 05:37｜发文｜@HeinrichKuttler：Kinda sad the math-crushing models aren't available to the general public (or really, just me). There are many more conjectures to test! 原帖
  • 07 月 21 日 06:16｜发文｜@HeinrichKuttler：@SebastienBubeck But not the fancy multiagent version that @eknight prompted for CDC: https://t.co/M1fQ0OxvCU 原帖
  • 07 月 21 日 06:19｜发文｜@HeinrichKuttler：@stephen_oman I think that isn't technically true because there are nethack configurations that kill you before you take any action (those are extremely rare). 原帖
  • 07 月 21 日 06:29｜发文｜@HeinrichKuttler：@aaronbatilo looks like the opposite will happen 原帖
  • 07 月 21 日 06:47｜发文｜@HeinrichKuttler：@SebastienBubeck @eknight that said, thank you for making 5.6 Sol available! 原帖

Moonshot/Kimi

  • 07 月 20 日 08:17｜发文｜@Kimi_Moonshot：@alexrkonrad 😂 原帖
  • 07 月 20 日 09:51｜转发｜@Kimi_Moonshot：RT @DesignArena: BREAKING: Kimi K3 by @Kimi_Moonshot is officially 1st on Frontend Web App Arena by DesignArena With an Elo of 1326, this… 原帖
  • 07 月 20 日 09:56｜发文｜@crystalsssup：Kimi is #1 in Design Arena, a benchmark for frontend website-building capabilities. 原帖
  • 07 月 20 日 10:15｜转发｜@Kimi_Moonshot：RT @arena: Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's… 原帖
  • 07 月 20 日 10:27｜发文｜@HaoningTimothy：A diverse real-world evaluation for us with scenarios definitely what we want on the new unified Kimi Chat with full Kimi Agent capabilities. K3 on par with Sol and Opus and sets Pareto fron… 原帖
  • 07 月 20 日 10:35｜发文｜@HaoningTimothy：We know there are still gaps in certain domains yet we are definitely confident about its experience. It feels more like a frontier than an open-source SOTA. Because we developers really use… 原帖
  • 07 月 20 日 20:43｜发文｜@crystalsssup：@KamrynOhly huge thanks!! 原帖
  • 07 月 20 日 20:43｜发文｜@crystalsssup：@heyrobinai thank you! 原帖
  • 07 月 20 日 20:43｜发文｜@crystalsssup：@HarshithLucky3 thank you! 原帖
  • 07 月 20 日 20:43｜发文｜@crystalsssup：@hey_madni thank you! 原帖
  • 07 月 20 日 22:37｜发文｜@HaoningTimothy：so bad boss xinyu i have never been or even seen gpu-rich 🥹 原帖
  • 07 月 20 日 23:19｜发文｜@HaoningTimothy：a nice start & we still have much room to improve 原帖
  • 07 月 21 日 00:31｜转发｜@HaoningTimothy：RT @zengjiajun_eth: this is really impressive. not just the frontend, but the overall agentic capability. 原帖

OpenAI

  • 07 月 20 日 08:00｜发文｜@_samirism：@0xBOYD see https://t.co/CWl55qx90m 原帖
  • 07 月 20 日 08:02｜发文｜@JoanneShang：using "--" because i want you to know that this message was not ai-generated 原帖
  • 07 月 20 日 08:04｜发文｜@_samirism：@Increment_amir Same memory across all devices and platforms for ChatGPT coming soon to desktop Work “local threads” 原帖
  • 07 月 20 日 08:14｜发文｜@santa_kaus：prayed for days like this https://t.co/r1dNtGyJr1 原帖
  • 07 月 20 日 08:21｜发文｜@polynoamial：@cipherstein @deredleritt3r @aaron_lou Because the answer is on the web now 原帖
  • 07 月 20 日 08:30｜发文｜@ryanbrewer：This was one (1) year ago by the way. We’re so early 原帖
  • 07 月 20 日 08:37｜发文｜@_samirism：@MArisha86780 yes globally released 原帖
  • 07 月 20 日 08:49｜转发｜@aidan_clark：RT @aaron_lou: Congrats to Levent for discovering this! OOC I had an internal version of Codex attempt a proof too (without web search), a… 原帖
  • 07 月 20 日 08:58｜发文｜@andrew_n_carr：Following up the jacobian conjecture, an equally difficult problem is Crowds in animation! joking aside, being able to transfer motion and work at scale does reduce the tedium while making i… 原帖
  • 07 月 20 日 09:01｜发文｜@tifafafafa：😂 原帖
  • 07 月 20 日 09:05｜转发｜@gabrielchua：RT @KushalVijay_: 40+ Teams, 100+ Participants and projects like never seen before... This was the situation at @OpenAI Build Week Codex C… 原帖
  • 07 月 20 日 09:10｜发文｜@ryanbrewer：@gerardsans Wait so what the mathematician posted is not correct? I’m confused 原帖
  • 07 月 20 日 09:20｜转发｜@dylancablayan：RT @ajambrosino: memory is quite good now 原帖
  • 07 月 20 日 09:21｜转发｜@dylancablayan：RT @unsignedbytes: @yacineMTB I just tell people I make my money by talking to vibrating rocks and doing math/magic with them. https://t.co… 原帖
  • 07 月 20 日 09:29｜发文｜@rapha_gl：@HipCityReg @colossusmag @saranormous the beauty of the jacobian conjecture is that it’s not that complicated. with basic linear algebra you can understand the entire conjecture and what the… 原帖
  • 07 月 20 日 09:37｜发文｜@ryanbrewer：@StochasticStoat @gerardsans Then what’s he talking about lol 原帖
  • 07 月 20 日 09:41｜发文｜@rapha_gl：@HipCityReg @colossusmag @saranormous (but i’m also a big fan of @saranormous and everyone should read the article!) 原帖
  • 07 月 20 日 09:42｜发文｜@gabrielchua：So awesome to see @OpenAIDevs Build Week community events happening around the world - what a great time to be building together! Thank you to our Codex Ambassadors for bringing these sessio… 原帖
  • 07 月 20 日 09:42｜发文｜@andrewho03：Great changes unseen in a century 原帖
  • 07 月 20 日 09:45｜发文｜@gabrielchua：Made the video together with ChatGPT on the way to work and all on my phone ⚡️ 5.6 Sol also added the light music! 🎵 https://t.co/RCRqxCKCSI 原帖
  • 07 月 20 日 09:53｜发文｜@JoanneShang：@karine_hsu @Inc we need more walk clubs 原帖
  • 07 月 20 日 09:58｜发文｜@gabrielchua：Credits: ChatGPT Work on mobile https://t.co/0PgcMeBl83 原帖
  • 07 月 20 日 10:16｜发文｜@andrew_n_carr：@aaron_lou @disiok @eknight I have been re-writing that prompt for most of the work I do now 原帖
  • 07 月 20 日 10:34｜发文｜@sama：it is good now! 原帖
  • 07 月 20 日 10:34｜发文｜@xiangyuqi_pton：@zdhnarsil The same logic apply to AGI research as well—and the reason why Ant blocked the use of Mythos/Fable for ML work 👀 原帖
  • 07 月 20 日 10:35｜转发｜@danintheory：RT @aaron_lou: Congrats to Levent for discovering this! OOC I had an internal version of Codex attempt a proof too (without web search), a… 原帖
  • 07 月 20 日 10:37｜转发｜@santa_kaus：RT @sama: it is good now! 原帖
  • 07 月 20 日 10:37｜转发｜@santa_kaus：RT @_samirism: throughout june we rolled out major updates to chatgpt memory it can be hard to tell the difference initially but alot… 原帖
  • 07 月 20 日 10:42｜发文｜@polynoamial：Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon evaluations miss. We’re sharing what we learned from studying a lon… 原帖
  • 07 月 20 日 10:42｜转发｜@OpenAI：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 10:50｜发文｜@polynoamial：@ICE257_ @OpenAI Because there was concern people would view it as self-promotional hype if it was tweeted from the OAI account. 原帖
  • 07 月 20 日 10:55｜发文｜@MicahCarroll：Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have since improved our safeguards and redeployed the model. https://t.co… 原帖
  • 07 月 20 日 10:58｜转发｜@andrew_n_carr：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 10:58｜发文｜@santa_kaus：@tinahhong congrats @AryoPatel and @tinahhong!!! extremely bullish on andera and the team :) 原帖
  • 07 月 20 日 11:01｜发文｜@santa_kaus：@tinahhong @AryoPatel always looking so fresh smh. video shoot day or pearvc cubicle it doesnt matter 原帖
  • 07 月 20 日 11:05｜转发｜@SebastienBubeck：RT @aaron_lou: Congrats to Levent for discovering this! OOC I had an internal version of Codex attempt a proof too (without web search), a… 原帖
  • 07 月 20 日 11:05｜发文｜@SebastienBubeck：@littmath @alpoge Check out Aarron's post: https://t.co/HMOrIiqZmo 原帖
  • 07 月 20 日 11:19｜转发｜@danintheory：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 11:19｜转发｜@w01fe：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 11:24｜转发｜@aidan_clark：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 11:28｜发文｜@charliermarsh：These all pass 100% of a held-out test suite? Guess they must all be equivalently good and ready to use. Or not? Who even knows? 原帖
  • 07 月 20 日 11:30｜发文｜@nunezvice：@Ananth7e clear example that it is impossible to make everyone happy. the team is really trying though 原帖
  • 07 月 20 日 11:39｜发文｜@charliermarsh：@OswaldValue Lmao 原帖
  • 07 月 20 日 11:40｜转发｜@yong_zhengxin：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 11:40｜发文｜@charliermarsh：@powerbottomdad1 The algorithmic tweaks will continue until morale improves 原帖
  • 07 月 20 日 11:44｜发文｜@reach_vb：With ChatGPT Work, you can kick off cloud tasks directly from the desktop or mobile app. It can use your configured plugins to work across your repos, docs, Slack, and the other tools you al… 原帖
  • 07 月 20 日 11:48｜转发｜@prd_008：RT @omooretweets: I spent most of my weekend on ChatGPT Sites. Here is why I think it’s the biggest unlock yet in the mainstream consumer… 原帖
  • 07 月 20 日 11:49｜发文｜@charliermarsh：@horwitzben @OpenAI Welcome! 原帖
  • 07 月 20 日 11:50｜发文｜@charliermarsh：PSA: new regex! macro dropped in the latest release https://t.co/6liUKMRzYb 原帖
  • 07 月 20 日 11:50｜发文｜@charliermarsh：https://t.co/ZNyT3R5m86 原帖
  • 07 月 20 日 11:52｜转发｜@jeremyli__：RT @jjpcodes: fuck it we ball. codex solve a maths problem for me.dont care which one. make no mistakes https://t.co/G7snI5MOJc 原帖
  • 07 月 20 日 11:54｜发文｜@jeremyli__：investing early in this approach 原帖
  • 07 月 20 日 11:55｜转发｜@santa_kaus：RT @MicahCarroll: Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have… 原帖
  • 07 月 20 日 12:13｜转发｜@nunezvice：RT @reach_vb: With ChatGPT Work, you can kick off cloud tasks directly from the desktop or mobile app. It can use your configured plugins… 原帖
  • 07 月 20 日 12:18｜转发｜@nickbaumann_：RT @reach_vb: With ChatGPT Work, you can kick off cloud tasks directly from the desktop or mobile app. It can use your configured plugins… 原帖
  • 07 月 20 日 12:28｜发文｜@giffmana：This is actually hilarious 😂 Now the obvious question is where is #287 and why can't i see it? 原帖
  • 07 月 20 日 12:29｜发文｜@ryanbrewer：I love that I find this completely normal 原帖
  • 07 月 20 日 12:34｜发文｜@andrewho03：@AcerFur Congrats! 原帖
  • 07 月 20 日 12:37｜发文｜@giffmana：@MicahCarroll Why delete the PR!! As you(r blogpost) say, that cat is out of the bag anyways, and this would be such a cool historical artifact! https://t.co/eKYxFT0mWu 原帖
  • 07 月 20 日 12:37｜发文｜@gdb：Was very cool to hear about the reasons people love Sol. We're doing the promotion again, except this time for ChatGPT Work: Tweet what you love about ChatGPT Work, claim $100 in free credit… 原帖
  • 07 月 20 日 12:38｜发文｜@giffmana：@AndrewCurran_ Aye 原帖
  • 07 月 20 日 12:41｜发文｜@giffmana：@NikStarodub @kastnerkyle cc @TimDarcet 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：I had Codex build a tool, hawk, to identify and remove dead code across a Cargo workspace. We've used it to delete thousands of lines from uv and Codex itself. https://t.co/JM2kRQutTW 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：In uv, we are (with a few exceptions) our only caller, so we can analyze the workspace as a closed world. In other words, pub code that's unused across the workspace can be considered dead… 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：rustc can identify unused code within a crate, but (sensibly) assumes that anything pub might be used by external callers... 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：Building this was in part an experiment. I haven't "read" any of the code in this codebase, and yet it is empirically a really useful thing. We now run it continuously in uv's CI. 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：Similarly, code that's marked as pub but isn't used by any other crates can have its visibility downgraded. Same for struct fields, etc. (Hawk can enforce and fix this too.) 原帖
  • 07 月 20 日 12:43｜发文｜@charliermarsh：https://t.co/kfVnSCJnhD 原帖
  • 07 月 20 日 12:50｜发文｜@reach_vb：@maxjendrall now you can send your friends this tweet haha 原帖
  • 07 月 20 日 12:51｜发文｜@reach_vb：@tregoudeau you are using a task vs a project. switch to a project 原帖
  • 07 月 20 日 12:55｜发文｜@andrewho03：I will be visiting Shanghai and Hangzhou this autumn (around the weekend of October 31). Would anyone like to meet? 原帖
  • 07 月 20 日 13:02｜转发｜@OpenAI：RT @gdb: Was very cool to hear about the reasons people love Sol. We're doing the promotion again, except this time for ChatGPT Work: Twee… 原帖
  • 07 月 20 日 13:17｜转发｜@danintheory：RT @gfodor: the stochastic parrots are getting pretty lucky https://t.co/BWdXWHZ9eY 原帖
  • 07 月 20 日 13:30｜发文｜@andrewho03：Final Triumph of Steve Hsu Thought 原帖
  • 07 月 20 日 13:31｜发文｜@andrewho03：@teortaxesTex Excuse me for misspeaking, it’s merely the beginning of the final triumph, but it will be a long, protracted, and deeply enjoyable one. 原帖
  • 07 月 20 日 13:33｜发文｜@andrewho03：@dubrovnik07 Yes feel free to DM. 原帖
  • 07 月 20 日 13:47｜发文｜@charliermarsh：@_Felipe What do you have in mind 原帖
  • 07 月 20 日 13:50｜发文｜@charliermarsh：Optimizing Codex over the weekend. This is my favorite kind of work. I could do it forever. https://t.co/NUBnJWJ2Sg 原帖
  • 07 月 20 日 14:02｜转发｜@yong_zhengxin：RT @tszzl: btw i think it bodes quite well for safety that a well loved system was taken down for further testing at expense to internal ac… 原帖
  • 07 月 20 日 14:07｜转发｜@yong_zhengxin：RT @MicahCarroll: Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have… 原帖
  • 07 月 20 日 14:09｜发文｜@ryanbrewer：What about +5,000 -3,400 原帖
  • 07 月 20 日 14:09｜转发｜@hansonwng：RT @aaron_lou: Congrats to Levent for discovering this! OOC I had an internal version of Codex attempt a proof too (without web search), a… 原帖
  • 07 月 20 日 14:17｜转发｜@SebastienBubeck：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 14:26｜发文｜@jeremyli__：@AcerFur congrats!! 原帖
  • 07 月 20 日 14:39｜发文｜@reach_vb：@koltregaskes should come live soon if not already 原帖
  • 07 月 20 日 14:39｜发文｜@giffmana：@Dorialexander https://t.co/X5L5eBsumi 原帖
  • 07 月 20 日 14:43｜发文｜@reach_vb：@gott_zac oh, can you send screenshots/ ask codex to investigate and share the blurb? DMs open too 原帖
  • 07 月 20 日 14:45｜转发｜@phuguo：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 14:45｜转发｜@phuguo：RT @Marcus_J_W: https://t.co/78yWII3mAD This blog has some details on a incident which caused us to undeploy a model which I think many peo… 原帖
  • 07 月 20 日 14:45｜发文｜@reach_vb：@_xjdr yes, the UX could be better here - will flag! that said, you should totally sign up for TAC, it’s worth it. 原帖
  • 07 月 20 日 14:50｜发文｜@charliermarsh：@hd_nvim Let's go! 原帖
  • 07 月 20 日 14:51｜发文｜@charliermarsh：@hd_nvim My guess is it's actually good 原帖
  • 07 月 20 日 14:51｜发文｜@steipete：@simplybenuk @openclaw pretty dope! 原帖
  • 07 月 20 日 14:52｜发文｜@reach_vb：wow last 5 days have been incredible, recorded: > 2 podcasts > 2 office hours > 2 meetups > 1 webinar > 1 AMA & countless interactions! pedal to the metal! 原帖
  • 07 月 20 日 14:52｜发文｜@ALupsasca：Very glad that @XiYin137 is joining X. Please follow him! 原帖
  • 07 月 20 日 14:53｜发文｜@reach_vb：@Dimillian send me a codex prompt when done thx 原帖
  • 07 月 20 日 14:58｜转发｜@yong_zhengxin：RT @TheZvi: Kudos to OpenAI for sharing this information, and for noticing the problem, and for suspending deployment. It is really importa… 原帖
  • 07 月 20 日 15:02｜发文｜@isabelle_zhou：@tbpn @horwitzben Welcome @horwitzben! 🎉 原帖
  • 07 月 20 日 15:02｜发文｜@reach_vb：Incredibly proud of our Codex Ambassadors for running over 60+ events in ~60 cities all in a span of 7 days! 🔥 OpenAI Build week runs till tomorrow, come experience the magic of GPT 5.6 &amp… 原帖
  • 07 月 20 日 15:04｜发文｜@reach_vb：@LexnLin yes, it’s magic and particularly useful for async tasks 原帖
  • 07 月 20 日 15:05｜转发｜@ALupsasca：RT @littmath: This is actually incredible/incredibly funny. 原帖
  • 07 月 20 日 15:14｜发文｜@giffmana：@suchenzang Well that little extra really does it lol https://t.co/fcPJnc7Cg2 原帖
  • 07 月 20 日 15:22｜转发｜@prd_008：RT @reach_vb: Incredibly proud of our Codex Ambassadors for running over 60+ events in ~60 cities all in a span of 7 days! 🔥 OpenAI Build… 原帖
  • 07 月 20 日 15:35｜发文｜@yong_zhengxin：my speculation from this is that the attack is done using OS models, and the attack-defense race ultimately comes down to whichever side has more "compute per intelligence". I wonder whether… 原帖
  • 07 月 20 日 15:39｜发文｜@gdb：@hkarlsen_06 @OpenAI you can just do things 原帖
  • 07 月 20 日 15:39｜发文｜@charliermarsh：@GrantSlatton If it works well, I wrote it. If not, GPT wrote it. 原帖
  • 07 月 20 日 15:41｜发文｜@yong_zhengxin：So happy to see this! > I just spoke with the Hugging Face team last week, ... We (https://t.co/3latshmvS1) will continue to strengthen our cybersecurity and broader safety evaluations, .… 原帖
  • 07 月 20 日 15:42｜发文｜@reach_vb：@mrloldev @sama have you tried ChatGPT Work? It should be as easy as switching modes lmk how it goes 原帖
  • 07 月 20 日 15:45｜发文｜@reach_vb：Sites is now available in the UK + EU! 🇬🇧🇪🇺 Note: Sites is currently limited to workspaces, Plus, and Pro accounts. https://t.co/fkKUeAZXZq 原帖
  • 07 月 20 日 15:46｜发文｜@reach_vb：Read more here: https://t.co/7OnHk2WGyQ 原帖
  • 07 月 20 日 15:47｜转发｜@alishbaimran_：RT @MicahCarroll: Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have… 原帖
  • 07 月 20 日 15:50｜发文｜@jxnlco：@largePrawn only have like 5 of them 原帖
  • 07 月 20 日 15:59｜转发｜@karinanguyen：RT @thoughtfullab: https://t.co/lUnUhF0hX6 原帖
  • 07 月 20 日 16:11｜转发｜@yong_zhengxin：RT @WillRinehart: This is not at all surprising that CAISI's director is stepping down, given what's happened these last couple weeks in AI… 原帖
  • 07 月 20 日 16:11｜发文｜@kellerjordan0：The Bay Area tech industry declining to take responsibility for preserving a crown jewel of its own talent pipeline was fairly interesting in terms of thinking about what the industry “is” h… 原帖
  • 07 月 20 日 16:18｜转发｜@kellerjordan0：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 16:44｜发文｜@jxnlco：Notes for a new video essay on “life after ai” https://t.co/TODRDIheKi 原帖
  • 07 月 20 日 16:48｜发文｜@charliermarsh：@GrantSlatton Very nice, thank you! I’ll do some of these 原帖
  • 07 月 20 日 16:49｜发文｜@JustinBleuel：@morqon I mean obviously, good catch 原帖
  • 07 月 20 日 16:58｜发文｜@jxnlco：@reach_vb Rest too 原帖
  • 07 月 20 日 16:59｜发文｜@thsottiaux：Never a dull moment when you work at OpenAI. Absolutely incredible place. 原帖
  • 07 月 20 日 17:02｜发文｜@SebastienBubeck：@wtgowers @alpoge There is a bit more coty response here: https://t.co/HMOrIiqZmo 原帖
  • 07 月 20 日 17:10｜转发｜@nunezvice：RT @thsottiaux: Never a dull moment when you work at OpenAI. Absolutely incredible place. 原帖
  • 07 月 20 日 17:13｜转发｜@BorisMPower：RT @scaling01: pinky promise there's no distillation https://t.co/Y9Vq2sjBLV 原帖
  • 07 月 20 日 17:13｜发文｜@prd_008：ChatGPT Sites is now available in the EU, UK, and CH! Enjoy https://t.co/tYUIAu0J4k https://t.co/L9oRfV3tgi 原帖
  • 07 月 20 日 17:15｜转发｜@gabrielchua：RT @OpenAIDevs: Codex ambassadors are taking Build Week around the world 🌍 https://t.co/VjlnoDwBT7 原帖
  • 07 月 20 日 17:15｜转发｜@gabrielchua：RT @gdb: Was very cool to hear about the reasons people love Sol. We're doing the promotion again, except this time for ChatGPT Work: Twee… 原帖
  • 07 月 20 日 17:18｜转发｜@gabrielchua：RT @OpenAIDevs: Time to pull a project from the backlog and build it with Codex. Here’s some build inspiration from Women who Code(x). htt… 原帖
  • 07 月 20 日 17:46｜发文｜@nickbaumann_：@jxnlco One of these stands out more than the others 原帖
  • 07 月 20 日 17:52｜转发｜@nickbaumann_：RT @ChatGPTapp: Get more done when you connect ChatGPT to your email, calendar, and other tools. @coreyching shows you the latest with plu… 原帖
  • 07 月 20 日 17:58｜转发｜@gabrielchua：RT @jxnlco: Sites is now available for our UK + EU users! 原帖
  • 07 月 20 日 17:58｜转发｜@gabrielchua：RT @MilksandMatcha: it's a creator's world https://t.co/JJ3B8z6VsL 原帖
  • 07 月 20 日 18:01｜转发｜@j_asminewang：RT @MicahCarroll: Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have… 原帖
  • 07 月 20 日 18:07｜发文｜@jxnlco：@nickbaumann_ I know the hand injury was pretty bad 原帖
  • 07 月 20 日 18:26｜转发｜@jam3scampbell：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 18:26｜发文｜@andrewho03：About to eat dinner at @MechanizeWork. Very excited for Fogo de Chão DoorDash order. 原帖
  • 07 月 20 日 18:44｜发文｜@JustinBleuel：Highly recommend you try chatgpt work app especially for something chat was previously failing at I’m consistently blown away, team really cooked, lots more coming share what you do with it! 原帖
  • 07 月 20 日 18:52｜发文｜@andrewho03：@3M4C5 I instructed Codex to include them on the plot 原帖
  • 07 月 20 日 18:52｜发文｜@andrewho03：Rating: 2/5 (food quality) 原帖
  • 07 月 20 日 18:53｜发文｜@jxnlco：@ella_schlags @minafahmi @cowjuh 原帖
  • 07 月 20 日 18:57｜发文｜@jxnlco：@bubbleboi lunch? 原帖
  • 07 月 20 日 18:58｜发文｜@jxnlco：obvously the easiest thing to automate was the newest thing that we invented 原帖
  • 07 月 20 日 19:15｜转发｜@andrew_n_carr：RT @SakanaAILabs: Introducing Fugu-Cyber: an update to our Fugu orchestration model. It achieves state-of-the-art performance on real-worl… 原帖
  • 07 月 20 日 19:25｜转发｜@gabrielchua：RT @ChatGPTapp: Get more done when you connect ChatGPT to your email, calendar, and other tools. @coreyching shows you the latest with plu… 原帖
  • 07 月 20 日 19:26｜转发｜@jxnlco：RT @_chenglou: I believe Freerange might be one of the biggest open-source projects written with Fable & Sol. And technically barely feasib… 原帖
  • 07 月 20 日 19:30｜转发｜@andrew_n_carr：RT @nc_frey: announcing things at Anthropic is hard because you never know when a colleague is going to Disprove the Jacobian Conjecture in… 原帖
  • 07 月 20 日 19:40｜发文｜@kellerjordan0：@jhong https://t.co/2O9aqdN4ci 原帖
  • 07 月 20 日 19:48｜发文｜@dylancablayan：@anduriltech Anduril anime? 原帖
  • 07 月 20 日 20:09｜转发｜@gabrielchua：RT @omooretweets: I spent most of my weekend on ChatGPT Sites. Here is why I think it’s the biggest unlock yet in the mainstream consumer… 原帖
  • 07 月 20 日 20:37｜发文｜@JustinBleuel：@AndyJScott Here’s one - I just had it make a shopping cart based on your pinned tweet https://t.co/r2KvdLMSH8 原帖
  • 07 月 20 日 20:53｜发文｜@JustinBleuel：@chadptg Here’s an idea it came up w off your recent tweets https://t.co/bjLoqVByn4 原帖
  • 07 月 20 日 21:37｜发文｜@SebastienBubeck：True story. Very good safety work being done to enable the release of the unit distance model! 原帖
  • 07 月 20 日 21:40｜发文｜@_samirism：don't sleep on Work ... cloud computer + cloud browser + chatgpt memory + file persistence + scheduled tasks 原帖
  • 07 月 20 日 21:51｜发文｜@KarelDoostrlnck：codex computer use is so unfathomably cracked. it just used my iPhone call screening to get an access code that someone voicemailed to me. living the introvert's dream 原帖
  • 07 月 20 日 21:57｜发文｜@andrewho03：If someone offers you a subsidized good or service (“dumps” it on you, perhaps), you can typically benefit from capturing the subsidy. 原帖
  • 07 月 20 日 21:58｜转发｜@xuandongzhao：RT @lunwang1996: Since starting my PhD in AI safety in 2018, 2026 is the first year I’ve felt that model safety has become a real-world, bi… 原帖
  • 07 月 20 日 22:02｜转发｜@gabrielchua：RT @gabrielchua: Made the video together with ChatGPT on the way to work and all on my phone ⚡️ 5.6 Sol also added the light music! 🎵 http… 原帖
  • 07 月 20 日 22:04｜发文｜@andrewho03：@GuiveAssadi We need to stop all labor compensation until we can FIGURE OUT what is going on. 原帖
  • 07 月 20 日 22:10｜发文｜@thsottiaux：@Kappaemme1926 We are being very subtle. Use "Approve for me" instead. 原帖
  • 07 月 20 日 22:13｜转发｜@gabrielchua：RT @gabrielchua: So awesome to see @OpenAIDevs Build Week community events happening around the world - what a great time to be building to… 原帖
  • 07 月 20 日 22:14｜发文｜@gabrielchua：@yoeven @lifegetsdicey deploy it internally with ChatGPT sites? 👀 it comes with Sign in with ChatGPT for Auth! 原帖
  • 07 月 20 日 22:30｜转发｜@ilge：RT @OccupyingM: for those like me who need a @3blue1brown video to understand what this means say thanks to fable https://t.co/PqP6rCSuFO 原帖
  • 07 月 20 日 22:44｜发文｜@_samirism：@george__wing what would you do with recursive scheduled tasks? 原帖
  • 07 月 20 日 22:51｜发文｜@jeremyli__：@GuiveAssadi @andrewho03 Mid 原帖
  • 07 月 20 日 23:00｜发文｜@JoanneShang：Women Who Code(X) is going global... Next up -- Sydney, Australia! 原帖
  • 07 月 20 日 23:12｜发文｜@jeffreygwang：When I first moved to SF almost two years ago, Tinah was already prototyping and ideating what has now become Andera. Since then, their trajectory has gone vertical! It's been a privilege to… 原帖
  • 07 月 20 日 23:14｜转发｜@tejalpatwardhan：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 20 日 23:57｜发文｜@giffmana：@xeophon @eliebakouch Ah nice! But i was mostly interested in the "looks" of the PR itself (like does it mention difficulties in making it at all?) I don't care about the technique, fiddling… 原帖
  • 07 月 21 日 00:00｜发文｜@reach_vb：@VDMatos @mrloldev @sama Thanks for the feedback, more than half of this list should already work in Work mode already! It’s effectively the same UX of Codex but on cloud. And, this will con… 原帖
  • 07 月 21 日 00:03｜发文｜@danintheory：@eshaghoulian It has kind eyes 原帖
  • 07 月 21 日 00:04｜发文｜@reach_vb：Asked Chat to spin this up into a Site where I can track my progress as well as have a visual identity for these exercises! Chat used imagegen and the research I provided, added a database f… 原帖
  • 07 月 21 日 00:19｜发文｜@reach_vb：@henrytdowling All via ChatGPT Work, Sites is only available in Work & Codex atm 原帖
  • 07 月 21 日 00:26｜发文｜@andrewho03：This is untrue. I am a mediocre researcher and even despite access to internal models and thousands of GPUs, I would not (at this moment) be able to accomplish something equivalent in signif… 原帖
  • 07 月 21 日 00:37｜转发｜@w01fe：RT @yonashav: 👏 deep 👏 knowledge 👏 and 👏 expertise 👏 原帖
  • 07 月 21 日 00:59｜发文｜@aidan_clark：Pretty dismayed the open source narrative has moved from safety to sovereignty. GPT-OSS was obviously good for everyone, the hard part was ensuring the built-in safeguards were strong enough… 原帖
  • 07 月 21 日 00:59｜发文｜@aidan_clark：Dean is getting flak for being a company guy, but I’m equally as disappointed in all the OS diehards who don’t want to admit that this isn’t just another people of software. Your bias is shi… 原帖
  • 07 月 21 日 01:02｜发文｜@giffmana：This city is permanently refactoring itself. 原帖
  • 07 月 21 日 01:02｜发文｜@giffmana：What an idyllic oldt.... oh! https://t.co/TDSxnuhIrY 原帖
  • 07 月 21 日 01:05｜发文｜@andrewho03：@henrytdowling There are, but the total factor productivity increase is much lower than that would suggest. I think it’s also plausible that my abilities as a researcher will improve enough… 原帖
  • 07 月 21 日 01:05｜发文｜@gabrielchua：@Baxti_7 @romainhuet The plugins come with connection to these services! https://t.co/UR7IZKXXFD 原帖
  • 07 月 21 日 01:06｜发文｜@giffmana：@AaryanBhagat13 @aryankaul Actually yeah that would be epic 原帖
  • 07 月 21 日 01:15｜发文｜@aidan_clark：@aiamblichus I hope you stay far away from frontier model development! We believe in iterative deployment, because we think eventually these models will pose a risk of significant harm, that… 原帖
  • 07 月 21 日 01:15｜发文｜@aidan_clark：@aiamblichus To look backwards and say “that model is safe you were dumb to be worried” is to fundamentally misunderstand the situation we are in. 原帖
  • 07 月 21 日 01:16｜发文｜@aidan_clark：@aiamblichus Holding back GPT-2 looks dumb as shit now, but I’m sympathetic with the folks involved. We just didn’t know! 原帖
  • 07 月 21 日 01:44｜发文｜@aidan_clark：@arxivmerchant Yah I mean into be clear I think that take was borderline objectivist 原帖
  • 07 月 21 日 02:11｜发文｜@reach_vb：@DFinsterwalder @OpenAI absolutely love it! kudos David!! 🔥 原帖
  • 07 月 21 日 02:19｜发文｜@aidan_clark：@aiamblichus I really don’t think it’s hypocritical! I do think we could release an OSS version now and be a lot less worried and the transparent statement is that we have other shit to do! 原帖
  • 07 月 21 日 02:19｜发文｜@ericmitchellai：Several people with no technical background told me they were reviewing for NeurIPS using agents for all reviewing--stating they didn't understand what the agent did, nor the underlying pape… 原帖
  • 07 月 21 日 02:19｜发文｜@aidan_clark：https://t.co/PPxB7mDfUR Btw this is basically my safety take 原帖
  • 07 月 21 日 02:20｜发文｜@aidan_clark：@xeophon it’s a good model sir 原帖
  • 07 月 21 日 02:41｜发文｜@aidan_clark：@aiamblichus @davidmanheim Of course there are persuasive theories. For GPT-2 it was primarily about misinformation. I think the risk was real, but wholly realized anyways via closed source… 原帖
  • 07 月 21 日 03:10｜发文｜@reach_vb：@Laz4rz @arcee_ai i vaguepost, therefore i am 原帖
  • 07 月 21 日 04:20｜转发｜@charliermarsh：RT @sharkdp86: Very proud to share that my brother and I finished a side project that has been five years in the making: Sokobrane is a puz… 原帖
  • 07 月 21 日 05:12｜转发｜@gabrielchua：RT @tushaarmehtaa: one-shotted this with ChatGPT such a cool concept brb, adding every shoe I own https://t.co/fC0fsdIKGc 原帖
  • 07 月 21 日 05:43｜发文｜@dylancablayan：@Sajeel_Purewal Extremely Bullish on this 🔥 LFG 原帖
  • 07 月 21 日 05:47｜发文｜@gabrielchua：Final 12 hours! 原帖
  • 07 月 21 日 06:06｜发文｜@jxnlco：@distributionat Tell your main agent to tell the sub-agents to use the in-app browser. 原帖
  • 07 月 21 日 06:07｜发文｜@jxnlco：@morqon I will find where you live 原帖
  • 07 月 21 日 06:09｜发文｜@jxnlco：@simpsoka I don't see a GDPR notice. 原帖
  • 07 月 21 日 06:10｜转发｜@jxnlco：RT @_chenglou: My dear UI developers, ML practitioners, and fans of programming language: Many months & billions of tokens later, I’m prou… 原帖
  • 07 月 21 日 06:14｜发文｜@SebastienBubeck：@HeinrichKuttler 5.6 Sol is available to everyone! 原帖
  • 07 月 21 日 06:17｜发文｜@jxnlco：@shadcn i've pretty sure this things has been really expensive 原帖
  • 07 月 21 日 06:17｜发文｜@jxnlco：every few months you get one of those days that fee like 'before sunrise' 原帖
  • 07 月 21 日 06:37｜发文｜@charliermarsh：@jacobmparis Whenever you tweet something like this, someone shows up in the replies with a link to a language that actually works that way 原帖
  • 07 月 21 日 06:40｜发文｜@dylancablayan：You know what, I respect the hustle 😂 https://t.co/it2M7IXMvR 原帖
  • 07 月 21 日 06:41｜发文｜@aidan_clark：@ma_sabba This tweet is basically LK-99 for not getting the joke 原帖
  • 07 月 21 日 06:46｜发文｜@jxnlco：@distributionat @JamesZmSun is this by design? 原帖
  • 07 月 21 日 06:48｜发文｜@aidan_clark：@BlancheMinerva @arxivmerchant Sorry I meant the OP oss comment from dean 原帖
  • 07 月 21 日 06:50｜转发｜@dylancablayan：RT @ArielKwiat: That was fast https://t.co/sLIDXyKgJw 原帖
  • 07 月 21 日 06:54｜发文｜@jxnlco：@iruletheworldmo no i went on a long walk with a girl 原帖
  • 07 月 21 日 06:54｜转发｜@jxnlco：RT @Angaisb_: Millionth time this year a friend asks me for help because their laptop stops working properly and all I do is install Codex… 原帖
  • 07 月 21 日 07:02｜发文｜@jxnlco：whos going to https://t.co/F35spkjfhI?!?!? 原帖
  • 07 月 21 日 07:18｜发文｜@jxnlco：@seanmozeik what can i say, im a fun guy 原帖
  • 07 月 21 日 07:20｜发文｜@prd_008：@simpsoka My feed is entirely ChatGPT and Keyne Yamal 原帖
  • 07 月 21 日 07:20｜转发｜@reach_vb：RT @pidotdev: Hi kids, gramps here, taking over for the intern. Just released pi 0.81.0 which features first class integration with @ggerga… 原帖
  • 07 月 21 日 07:28｜发文｜@jxnlco：reminder https://t.co/ZzzWPXqYY0 原帖
  • 07 月 21 日 07:28｜转发｜@gabrielchua：RT @polynoamial: Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon… 原帖
  • 07 月 21 日 07:31｜发文｜@giffmana：@ezyang It would be misleading if the UI gave the user the impression that they are interleaved but the model just gets img1 img2 img3 prompt So for it to be useful this is not just a UI thi… 原帖
  • 07 月 21 日 07:32｜发文｜@reach_vb：@jxnlco probably worth sleeping on time then 原帖
  • 07 月 21 日 07:33｜发文｜@ajambrosino：@VarunGodbole @thsottiaux these are in the app! Do you have notifications turned on? 原帖
  • 07 月 21 日 07:42｜发文｜@JoanneShang：touching grass & water but I’d rather be slacking memes https://t.co/Wf9auMlDxZ 原帖
  • 07 月 21 日 07:42｜发文｜@JoanneShang：*receiving memes from @alex_brussell 原帖
  • 07 月 21 日 07:44｜发文｜@steipete：@realhenry Thanks! Hope in Discord #clawtributors please! 原帖
  • 07 月 21 日 07:50｜发文｜@JoanneShang：@jxnlco *Optimists Are Winners 原帖
  • 07 月 21 日 07:55｜发文｜@nunezvice：it’s been really quiet around here the last couple of days 原帖
  • 07 月 21 日 07:58｜发文｜@ajambrosino：@VarunGodbole @thsottiaux what about in Mac settings? 原帖
  • 07 月 21 日 07:59｜发文｜@ajambrosino：@Conor_D_Dart @AIadventure3 careful 原帖

Thinking Machines Lab

  • 07 月 20 日 12:20｜发文｜@yilin_yang721：@ma_sabba Yeah, toxic PhD advisor is everywhere in US schools. I myself experienced one right after I graduated from college. It was a tragic experience for a new grad. AFAIK He basically bu… 原帖
  • 07 月 20 日 13:26｜转发｜@SonglinYang4：RT @YanhongLi2062: Personal update: I’ve joined @thinkymachines to work on pretraining!!! 🚀 After an amazing year at @allen_ai, I’m incred… 原帖
  • 07 月 20 日 13:26｜发文｜@SonglinYang4：@YanhongLi2062 @thinkymachines @allen_ai @lambdaviking @soldni welcome! 原帖
  • 07 月 20 日 13:32｜转发｜@SonglinYang4：RT @YanhongLi2062: Olmo Hybrid accepted to COLM!!!! see y’all in SF 🫶 原帖
  • 07 月 20 日 13:34｜发文｜@YueYangAI：@YanhongLi2062 @SonglinYang4 @thinkymachines @allen_ai @lambdaviking @soldni Welcome to the team! 原帖
  • 07 月 20 日 13:56｜发文｜@liliang_ren：@YanhongLi2062 @thinkymachines @allen_ai @lambdaviking @soldni Welcome!! 原帖
  • 07 月 20 日 13:58｜转发｜@liliang_ren：RT @zhanpeng_zhou: We submit MuonH with minus-sqrt LR scheduler (1-\sqrt{t}) to the modded-nanogpt Track 3 benchmark. It improves the curr… 原帖
  • 07 月 20 日 14:02｜发文｜@liliang_ren：curious if a steeper decay of the LR near the end of training can further reduce the steps needed. 原帖
  • 07 月 20 日 16:16｜发文｜@Mascobot：Yep… 原帖
  • 07 月 20 日 20:14｜转发｜@Haoxiang__Wang：RT @Pavel_Izmailov: New paper: Understanding Reasoning from Pretraining to Post-Training! We study the full LLM training pipeline from pre… 原帖
  • 07 月 21 日 00:05｜发文｜@Mascobot：This is super cool: 原帖
  • 07 月 21 日 00:10｜转发｜@SonglinYang4：RT @growing_daniel: F1 wins by AI lab: Anthropic: 0 OpenAI: 0 Kimi: 6 Tell me again how we're still six months ahead 原帖

xAI

  • 07 月 20 日 08:04｜转发｜@milichab：RT @OpenRouter: Grok Imagine Video 1.5 from @SpaceXAI is now live on OpenRouter! Bring any still image to life with an optional text promp… 原帖
  • 07 月 20 日 08:54｜转发｜@milichab：RT @grok: Grok for Excel is live. Use Grok 4.5 to build financial models, analyze market data, and generate charts and graphs. Try it now… 原帖
  • 07 月 20 日 08:54｜转发｜@milichab：RT @n2parko: a new upgrade to @cursor_ai agent in slack, rolling out now... - shares plan before working - new, cleaner status updates - b… 原帖
  • 07 月 20 日 09:26｜转发｜@milichab：RT @leerob: Grok 4.5 is really good at React. It's also very affordable and token efficient! https://t.co/1Dg24rLsoW 原帖
  • 07 月 20 日 09:40｜转发｜@zhaohan_dong：RT @grok: Grok for Excel is live. Use Grok 4.5 to build financial models, analyze market data, and generate charts and graphs. Try it now… 原帖
  • 07 月 20 日 09:51｜发文｜@zdhnarsil：One prediction I keep thinking about: Research in fields like pure mathematics is about to shift from being primarily intelligence-driven to capital-driven. The key constraint will no longer… 原帖
  • 07 月 20 日 10:13｜发文｜@forwarddeploy：@tinahhong Massive win! 原帖
  • 07 月 20 日 10:19｜转发｜@adityagupta：RT @alpoge: hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend f… 原帖
  • 07 月 20 日 10:24｜发文｜@zdhnarsil：@ZhengyangGeng I think both (what you & me said) will happen at the same time! It will be exciting to see how it goes 原帖
  • 07 月 20 日 10:35｜发文｜@zdhnarsil：@xiangyuqi_pton Agree 👀 whoever already has advantage will move even faster 原帖
  • 07 月 20 日 13:19｜发文｜@forwarddeploy：Come join us on Aug 8th for the SpaceXAI Grokathon in SF! 原帖
  • 07 月 20 日 14:07｜转发｜@milichab：RT @prasenx: best tui i've used is grok build still the cleanest one on my machine https://t.co/2066jP4AIX 原帖
  • 07 月 20 日 14:19｜发文｜@forwarddeploy：@SpaceXAI Good people, good vibes, good project! (oh yeah, and the best food!) 原帖
  • 07 月 20 日 14:48｜发文｜@edenchan：@SpaceXAI Vibes are immaculate 原帖
  • 07 月 20 日 14:59｜发文｜@edenchan：/goal build something august 8th at grokathon! 原帖
  • 07 月 20 日 15:09｜转发｜@luoluo_ai：RT @alpoge: hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend f… 原帖
  • 07 月 20 日 15:11｜发文｜@milichab：@flycory Thank you - please send over some feedback - traces, screenshots, or product feedback are very welcome! 原帖
  • 07 月 20 日 15:11｜转发｜@milichab：RT @flycory: For the past few wks since Grok Build launch, we've run SxS testing of complex Codex & Grok Build tasks. Eliminated Claude for… 原帖
  • 07 月 20 日 17:38｜发文｜@TheGregYang：I technically don't have lyme any more! according to IgeneX that just came back, lyme IgM bands improved 23++, 41+ (Jan) -> 23++ (now) which is technically negative for borrelia burgdorferi… 原帖
  • 07 月 20 日 20:17｜发文｜@KeranRong：Garage culture built silicon valley, maybe the United States as a whole as well.. we need to remember that. 原帖
  • 07 月 20 日 21:12｜发文｜@TheGregYang：is it just me or ginseng strip 2002 still sounds like its made by an incomprehensibly advanced civilization 原帖
  • 07 月 20 日 21:54｜发文｜@RyanNeverWrong：@hannahgao Congrats! 原帖
  • 07 月 20 日 22:37｜转发｜@billyuchenlin：RT @Tesla_Asia: Grok is coming to more Asian markets. https://t.co/Sc6Sl7CrnU 原帖
  • 07 月 21 日 07:40｜发文｜@JasonBud：@KaranD93 @morganlinton appreciate the feedback. would be helpful to learn more about the types of reports you’re trying to generate. any examples (e.g. screenshots) of a weak report from Gr… 原帖

References

  1. 1Noam Brown：Safety and alignment in an era of long-horizon models
  2. 2Micah Carroll：内部模型暂停与重新部署
  3. 3Cursor：agents 重建 SQLite
  4. 4Charlie Marsh：Hawk dead-code 工具
  5. 5Darragh Burke：加入 Cognition 构建 Devin
  6. 6Wu Haoning：K3 的真实世界评估与 GPU 供给
  7. 7Kimi 官方转发 Agent Arena 排名
  8. 8Kimi 官方转发 Design Arena 排名
  9. 9Greg Brockman：ChatGPT Work 额度活动
  10. 10Aidan Clark：GPT-OSS 与安全叙事
  11. 11Eric Mitchell：agents 参与论文审稿的风险
  12. 12Deedy：IMO 2026 多模型个人测试
  13. 13Cognition：欢迎 TierZero 团队
  14. 14Anhang Zhu：TierZero 加入 Cognition
  15. 15Yanhong Li：加入 Thinking Machines Lab 做预训练
  16. 16Liliang Ren：训练末期的学习率衰减
  17. 17@leerob：K3 权重尚未发布
  18. 18Sasha Rush：形式化教材实验
  19. 19Ted Xiao：robot learning 研究演进
  20. 20Jiachen Liu：agent-native research artifacts
  21. 21Chayenne Zhao：TTS 推理瓶颈与优化结果
  22. 22Andrew Milich 转发 Grok Build SxS 测试
  23. 23Songlin Yang 转发 Olmo Hybrid 接收消息
  24. 24Walden：TierZero 与 Devin SRE 方向
  25. 25Catherine Olsson：Fable 与 AI wrangler 讨论

Related content

  • Sign in to comment.
More from this channel
View the full content archive of "大模型公司研究员动态追踪"Explore more channels on Discover