2026-08-01

过去24小时大模型研究员 X 动态：Astra 的10个开放问题、Inkling开放权重安全与xAI omni-reference

本期逐一核查频道 408 个 researcher 账号，确认 90 个活跃账号和 381 条新增原帖，重点梳理 Astra、Inkling 开放权重安全、xAI Imagine Video 1.5 omni-reference，并提供 10 位研究员 Spotlight。

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 31 日 08:00 至 8 月 1 日 08:00。频道 researcher roster 中的 408 个账号已逐一请求时间线，408 个请求全部返回；窗口内确认 90 个活跃账号、381 条按原帖 URL 去重后的新增帖。下文先展开高信号内容，再在末尾按公司保留本窗口新增原帖索引。

覆盖边界是频道 roster 与本轮可访问时间线，不是 X 全站全量结论；窗口无命中也不等于账号沉默。原始时间线记录没有为每条帖子提供完整的引用 / 回复关系字段，因此索引中的「疑似回复」只表示帖面以提及开头，不能当作平台关系字段。转发按正文中的 RT @ 标记识别。

公司	roster 账号	窗口活跃账号	新增帖
OpenAI	154	36	214
Anthropic	49	10	61
xAI	46	14	41
Google DeepMind	46	9	23
Thinking Machines Lab	23	10	23
Meta AI	39	7	8
Cursor	13	2	7
Cognition	8	1	3
Moonshot/Kimi	6	1	1
Alibaba Qwen	9	0	0
ByteDance Seed	4	0	0
DeepSeek	4	0	0
Cohere	3	0	0
Mistral	1	0	0
SSI	1	0	0
智谱	1	0	0
AMI Labs / ex-Meta AI	1	0	0

本期 408 个账号均完成核查；对 17 个公司分类中的零命中板块，只能说本期频道 roster 时间线没有返回窗口内新增帖，不能据此判断这些公司没有研究进展。

今日主线

1. Astra 的信号是研究闭环，不是「解决了数学」

OpenAI 研究员 Noam Brown（@polynoamial）称，Astra 的内部版本在数学、量子复杂性和理论计算机科学方向解决了 10 个重大开放问题，并把它描述为科学推理的一步。原帖指向 OpenAI 对其中进展的说明，但本期没有对证明逐一复核，因此这里保留为研究员对内部模型结果的公开陈述。1

Brown 随后补充，10 个突破的证明生成成本合计低于 2,000 美元，按 Sol API 价格计算；他期待研究者使用即将推出的 Astra 模型。这个数字是原帖口径，不是本频道的成本测算，也不等于每个问题都能以同样成本复现。2

几分钟后，Brown 又主动收窄了标题带来的想象空间：Astra 还没有解决数学，没有创造新的数学分支，也没有提出有趣的新猜想。更准确的读法是，模型在已有开放问题上展示了从候选解、推理到形式化验证的潜力；离自主提出新理论仍有明显距离。3 OpenAI roster 中的 Vaibhav Srivastav（@reach_vb）转述称，Astra 还把论证形式化为 Lean 证明，并另贴出证明入口；这仍是员工个人转述，不是本期独立复核。4

这组帖子对研究团队的现实含义有两层：第一，模型输出是否能进入形式化检查，比单纯的答案分数更接近科研工作流；第二，宣传语与技术边界同时出现，读者应把它看作一个值得追踪的内部模型信号，而不是已完成的数学自动化结论。

2. TML 把开放权重争论落到评估、分阶段开放和生态防御

Thinking Machines Lab 的 Mira Murati（@miramurati）表示，团队公开了 Inkling 的开放权重发布方法、评估方式，以及为什么安全性同时取决于模型本身和它进入的生态；她提到测试、分阶段访问和更强防御可以共同形成更开放的路径。5 这是团队方法的概述，不等于已经公开了全部评估数据或访问条件。

同属 TML 的 Martin Ziqiao Ma（@ziqiao_ma）回应 Anthropic 研究员 Jiaxin Wen 时说，团队一直有安全小组，也曾因为安全顾虑放弃性能更高的 checkpoint。该帖是个人口径，能证明的是他公开做出这项陈述，不能单独证明每个 checkpoint 的具体评估过程。6

Wen（@jiaxinwen22）则肯定 TML 开始建立安全团队，同时提醒不要用「不比 Grok 或其他开放权重模型更危险」作为开放的主要论据，因为这种比较可能诱导团队跳过安全工作。7 这形成了一个清晰分歧：TML 强调开放与安全可以兼容，Anthropic 研究员提醒不能把相对比较当成充分安全论证。两边都没有在本期帖子里给出可直接横向比较的测试表。

3. xAI 的 Imagine Video 1.5 从模型更新走向可控创作

xAI 视频生成研究员 Martin Ma（@martin_ma_007）宣布上线 omni-reference model，强调更多控制和更高质量，并邀请用户反馈。8 他的个人简介标注 Video Gen @xAI，因此这里可以把它归入视频生成方向的产品 / 研究工程信号，但不把帖子中的质量判断当作独立评测。

xAI 的 Frank Hu（@hexiang）进一步说，omni-reference 让 Imagine 1.5 更容易制作角色一致的长故事。9 这条信息指向的是控制能力的使用场景：同一人物跨镜头保持一致，而不只是单张图片的清晰度。@milichab 转发的产品入口还提到图像与声音参考，但这是转发 / 产品提示，本文不把它扩写为完整 API 规格。

4. 数据、后训练与价格讨论重新连接到部署风险

OpenAI roster 中的 Andrew Ho（@andrewho03）写道，大规模构建高质量数据集很难：研究员不可能熟悉每个领域，制作一个看似简单的 benchmark 也需要持续检查陌生领域的 rollout，含糊案例的正确性判定和小错误捕捉同样困难。他认为 GDP-Val 的工作量很大，存在错误并不削弱项目本身的工作量；随后他以此为背景推介自己提供的 RL 任务数据。10 该账号当前简介写的是 Prev: @OpenAI，但频道 roster 仍按 OpenAI 归档；本期保留这条信号，同时把当前归属冲突列为身份待核线索。

Google DeepMind roster 中的 Wenhao Chai（@wenhaocha1）把讨论延伸到 token factory 的商业模式：如果开放模型持续追近、推理价格持续下降，token 可能像电力一样趋于商品化；当 AI 真正执行训练、实验或代码任务时，错误造成的 GPU 浪费、实验损失和安全事故不会随 token 价格一起下降，他猜测保证 / 保险可能成为溢价来源。11 这是个人判断，不是 Google DeepMind 的战略表态，但它把价格比较从每百万 token 拉回到结果责任。

OpenAI 的 Yong Zheng-Xin（@yong_zhengxin）则回应自己参与的 self-jailbreaking 研究：团队原以为安全后训练会在闭源模型上抑制这类现象，后来发现这个判断错了。12 他的简介同时写着 joining @openai soon，这是身份线索而非本窗口内的正式入职公告，人员动态板块会单独标注。

公司分组

OpenAI｜36/154 个账号活跃，214 条

本板块最高信号是 Astra。Tier 1 账号 @polynoamial 窗口内 6 条，含结果、成本和边界说明；@sama 窗口内 4 条，主要转发 Astra 结果，没有新增技术细节。其余研究相关帖子集中在 Codex / ChatGPT Work 的使用、Luna 的价格与模型比较，以及安全和评测讨论。

Vaibhav Srivastav（@reach_vb）转述称 Luna Max 的表现接近 GPT-5.5 xHigh，但价格约低 12 倍；Terra Max 高 2.6 分且成本低 45%。这是他引用的比较口径，未给出完整 benchmark 定义和运行条件，不能当作独立横评。13 Andrew Ho 的数据质量讨论见今日主线。其余 178 条大多是回复、产品使用或 Astra 转发，不在正文逐条展开。

Anthropic｜10/49 个账号活跃，61 条

Tier 1 的 @_sholtodouglas 有 19 条，但主要围绕金融与投资话题，研究相关性低于本期正文门槛，因此保留在索引而不升级为 Anthropic 研究结论。Jiaxin Wen 对 TML 安全团队的回应见 TML 板块；Robert Bye（@RobertJBye）从 Claude 产品经理视角讨论应用中的 craft 和 interaction，并表示正在为新团队招聘关注设计的工程师。14 这是招聘信号，不等于已确认具体人员加入。

xAI｜14/46 个账号活跃，41 条

本窗口的集中主题是 Imagine Video 1.5 的 omni-reference：Martin Ma 发布功能，Frank Hu 解释长故事和角色一致性的用途，其他账号转发入口或邀请用户测试。另有 @ypwang61 转发 DeepSeek-V4-Flash 在第三方榜单和 Fireworks 上线的说法，并发出一条关于 35B 模型和 AI4AI 后训练研究的短评；该账号当前简介写 Prev. @xai，应读作 roster 历史分类下的个人动态，不作 xAI 官方结果。

Google DeepMind｜9/46 个账号活跃，23 条

Wenhao Chai 的 token factory 讨论见今日主线。Jaidev Shah（@JaidevShah4）的窗口内帖子回顾了一个 40 分钟 panel，主题包括 agents、memory、continual learning 和 RL，简介标注其负责 memory and personalization。15 其余主要是论文转发、短回复和非研究内容；Jeff Dean 等已知核心账号本窗口没有可展开的研究帖。

Thinking Machines Lab｜10/23 个账号活跃，23 条

Mira Murati 和 Martin Ziqiao Ma 的开放权重 / 安全讨论构成本板块主体。其他命中主要是对 TML 原帖的转发，说明 Inkling 的发布方法引发了 roster 内部的即时扩散，但转发数量不等于独立观点数量；本期不把同一段 TML 原文重复计为多条研究结论。

Cursor｜2/13 个账号活跃，7 条

Sasha Rush（@srush_nlp）展示了 JAX opaque types 与自定义 tangent 的用法，并用可微光栅化例子说明从张量走向更一般的几何对象。16 另一个活跃账号主要是个人休息安排，没有形成模型或训练信号。

Meta AI｜7/39 个账号活跃，8 条

窗口内最明确的技术判断是 @EdwardSun0909 对 Muse Spark 1.1 的一句正面评价，@honglin_c 随后转发；内容太短，无法支撑模型能力或 benchmark 结论。17 其余是论文讨论和简短回复。

Cognition｜1/8 个账号活跃，3 条

仅 1 个账号产生 3 条命中，主要是短回复和转发，没有足够上下文的新产品、agent 工程或人员状态信号。

Moonshot/Kimi｜1/6 个账号活跃，1 条

唯一命中是 @HaoningTimothy 的表情回复，没有可核实的 Kimi 模型、评测或人员动态。与前几期相比，本期不能从频道 roster 时间线推导 Kimi 有或没有产品变化。

其余公司

Alibaba Qwen（9）、ByteDance Seed（4）、DeepSeek（4）、Cohere（3）、Mistral（1）、SSI（1）、智谱（1）和 AMI Labs / ex-Meta AI（1）本窗口均无命中。这里的「0」只描述本轮可访问时间线，不是对公司公开活动的否定。

人员动态

已核验的个人状态变化

休假：Robert Bye（@RobertJBye，Anthropic）：在回复中明确写道自己正在 parental leave。这是当事人原帖，对休假状态有直接说明；原帖没有给出开始或结束日期，不能推导回岗时间。18
新团队招聘：Robert Bye（@RobertJBye，Anthropic）：表示 Claude 移动端和桌面端会增加更多 craft / interaction，并邀请关注设计的工程师私信，称新团队正在招聘。这是招聘公告，不是已入职人员名单。14

需要继续核验的身份线索

Yong Zheng-Xin（@yong_zhengxin）：个人简介写 joining @openai soon，本期窗口内的原帖讨论 self-jailbreaking 研究，但没有一条单独的、带日期的入职公告。暂按 OpenAI roster 归类，不把简介文字写成已完成入职。
Andrew Ho（@andrewho03）：个人简介写 Prev: @OpenAI，与频道 roster 的 OpenAI 标签冲突。本期保留其数据集讨论，但不据此断言当前仍在 OpenAI。
Yiping Wang（@ypwang61）：个人简介写 Prev. @xai，本期仍按 roster 的 xAI 分类保留其 AI4AI / 后训练相关帖子；当前雇主未由本期一手公告确认。

本窗口未确认 roster 账号发布可直接核验的离职、正式入职、创业启动、论文接收 / 获奖或 grant 公告。转发他人的招聘、公司产品宣传和对第三方研究的评论不计入个人状态变化。

今日研究员 Spotlight

1. Noam Brown｜OpenAI，reasoning 研究

Noam Brown（@polynoamial）在 OpenAI 研究 reasoning，简介列出 Libratus、Pluribus、CICERO 和 o-series。关注点是推理模型如何进入科研工作流；本期 Astra 的 10 个开放问题、成本和边界说明，是近两周最有代表性的动态。1

2. Mira Murati｜Thinking Machines Lab，实验室负责人

Mira Murati（@miramurati）目前在建设 Thinking Machines Lab，曾任 OpenAI CTO。她关注模型发布、开放权重和安全之间的组织与工程条件；本期围绕 Inkling 解释分阶段访问、测试和生态防御如何组合，是她近期最明确的团队方向表达。5

3. Martin Ziqiao Ma｜Thinking Machines Lab，technical staff

Martin Ziqiao Ma（@ziqiao_ma）是 Thinking Machines Lab technical staff，同时参与 ACL mentorship。其研究兴趣从个人简介无法完整判断，但近期发言集中在模型安全与开放权重取舍；他本期称团队会因安全顾虑放弃性能更高的 checkpoint，适合作为 TML 安全实践的观察窗口。6

4. Martin Ma｜xAI，视频生成

Martin Ma（@martin_ma_007）的简介直接标注 Video Gen @xAI，工作方向是视频生成产品与模型能力。本期他宣布 omni-reference 上线，强调控制和质量，随后回复用户反馈；近况显示关注点从单次生成扩展到参考图、角色一致性和可用性。8

5. Frank Hu｜xAI，Grok / Imagine

Frank Hu（@hexiang）曾在 Google DeepMind 做 Gemini 与 Imagen，目前简介指向 @SpaceXAI，关注 Grok 的理解与想象能力。他本期把 omni-reference 与 Imagine 1.5 的长故事、角色一致性联系起来，提供了比单纯发布口号更具体的使用方向，但没有给出独立质量数据。9

6. Wenhao Chai｜Google DeepMind，模型经济性观察

Wenhao Chai（@wenhaocha1）是 Google DeepMind 相关研究员、普林斯顿计算机科学博士生，长期关注模型与系统的成本、能力和部署后果。本期他讨论 token factory 的商品化风险，并提出保证 / 保险可能成为可靠性溢价，代表其把推理价格与任务失败成本放在一起思考。11

7. Jiaxin Wen｜Anthropic，AI research

Jiaxin Wen（@jiaxinwen22）的简介写着 research @berkeley_ai @anthropicai，关注研究方法、安全与模型生态。本期他评价 TML 建立安全团队，同时反对用与其他开放权重模型的相对危险性作为充分论据；这条动态的价值在于把开放发布的比较尺度和安全工作的激励问题放在一起。7

8. Jaidev Shah｜Google DeepMind，memory and personalization

Jaidev Shah（@JaidevShah4）负责 Google DeepMind 的 memory and personalization，简介还标注 Columbia。其兴趣落在记忆、个性化、持续学习和 agent；本期他回顾一个同时讨论 agents、memory、continual learning 与 RL 的 panel，说明这些方向在应用研究中仍被放在同一系统问题里讨论。15

9. Andrew Ho｜按 roster 归入 OpenAI，数据与 RL 任务

Andrew Ho（@andrewho03）在频道 roster 中归入 OpenAI，但当前简介已写 Prev: @OpenAI，身份应继续核验。他近期关注高质量数据、长时程统计分析、工具使用和 RL 任务；本期围绕 GDP-Val 讨论 benchmark 的标注成本与歧义判定，代表动态是从模型能力转向训练数据质量。10

10. Robert Bye｜Anthropic，Claude 产品

Robert Bye（@RobertJBye）是 Anthropic 的 Claude 产品经理，曾在 Figma、AllTrails 和 Google 工作。他关注产品交互、应用目的与面向用户的 craft；近况包括 parental leave 自述，以及为新团队招聘重视设计的工程师。前者是个人状态，后者是招聘信号，两者都不能扩写成 Claude 的产品路线图。18

本窗口新增原帖索引

以下索引保留本轮去重后的 381 条新增原帖。内容简短或非研究向的帖子不在正文展开，但保留原作者、时间、类型和原帖 URL，便于读者自行复核；所有日期和时间均为洛杉矶时间。

Anthropic

07-31 08:30 @minilek · 疑似回复：@thegautamkamath The second one seems to harm students who come from fields where adding advisors as co-authors isn't the norm (e.g. theory). 原帖
07-31 09:42 @trq212 · 疑似回复：@RobertHaisfield @nisten @Lon Yeah basically 原帖
07-31 09:44 @katelyn_lesse · 疑似回复：@TobinSouth @elliot_garreffa decided to ship it this week after all huh 😜 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @TheRealDanSaedi: By 2030 when inference for superintelligence is super constrained everyone on here making fun of Leopold will be makin… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @bayeslord: Sincerely hope everyone at situational awareness limited partnership is doing ok. Seems like an unbelievably stressful week.… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @saranormous: yes, markets/risk management lessons to be learned, but many of the folks dunking on Leopold have never realized a dollar… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @APompliano: I see a lot of people dunking on @leopoldasch, which feels like a classic internet mob looking for attention. Leo\"s story… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @alexeyguzey: People keep making fun of Leopold on the TL but 1. everyone needs to get margin called once in their life 2. Leo still ma… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @buccocapital: It\"s more impactful if you do it like this https://t.co/3kB0vv7ii2 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @zebulgar: Silicon Valley somehow unaware That basically all of the GOATs on Wall Street / Hedge Fund-land Have had some sort of blow-… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @lulumeservey: Funny how much expectations shape perceptions Imagine hearing: - a 24 year old is running one of the world\"s largest he… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @tbpn: Here is the full letter Leopold sent to his LPs last night. Rumors of his demise are greatly exaggerated. https://t.co/5yIzR0DKIp 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @nic_carter: this is what cracks me up when people construct these elaborate bear cases based on AI being "too expensive". ok just wait… 原帖
07-31 10:06 @_sholtodouglas · 转发：RT @thecaptain_nemo: leopold disclosing he\"s still up 80% ytd after everyone dunked on him and left him for dead https://t.co/fpjWLcKNG2 原帖
07-31 10:22 @_sholtodouglas · 原创/引用：Prediction: SALP will be bigger than Citadel by the end of the decade. Leopold has predicted the last two years better than anyone else - now that he can combine that with very expensive lessons in risk he will be unstop 原帖
07-31 10:27 @minilek · 疑似回复：@thegautamkamath @AlexGDimakis Would be good to see who the authors are causing the uptick. Then we could try to craft a classifier that limits them without blocking legit grad student work. 原帖
07-31 10:41 @_sholtodouglas · 疑似回复：@AgustinLebron3 sure, 1k for fun? (anything more is better put in the fund at even odds! 😅) 原帖
07-31 10:42 @_sholtodouglas · 疑似回复：@AgustinLebron3 Deal 🤝 原帖
07-31 10:43 @jiaxinwen22 · 转发：RT @idavidrein: https://t.co/9qIpePNAgk 原帖
07-31 10:52 @jonchu · 疑似回复：@_sholtodouglas @AgustinLebron3 Can I get in on this? Would love to take this bet too 🙂. No shade on Leopold, but mostly a hat tip to Ken. 原帖
07-31 10:55 @_sholtodouglas · 疑似回复：@jonchu @AgustinLebron3 Ken is utterly incredible for sure. Taken one bet - will put other $ into thesis because returns are better that way :) 原帖
07-31 10:57 @jonchu · 疑似回复：@_sholtodouglas @AgustinLebron3 Fair point. You'll likely do better as an LP in either of those funds 原帖
07-31 10:58 @_sholtodouglas · 转发：RT @gabriel1: hating on leopold is so weird. he predicted so many things right, put out a timeline thesis, bet all in on it, and priced in… 原帖
07-31 11:51 @minilek · 疑似回复：@thegautamkamath @AlexGDimakis What kinds of affiliations do they have? 原帖
07-31 12:00 @_sholtodouglas · 转发：RT @dylan522p: Love that all FinTwit people making fun of Leopold still get mogged by his returns this year. Leopold is a great guy and he'… 原帖
07-31 13:35 @RobertJBye · 疑似回复：@clairevo Getting married and starting and family is the probably the most cracked thing anyone can ever do. 原帖
07-31 13:37 @RobertJBye · 原创/引用：Did @googlecalendar just drop a little Liquid Glass update, or have I been slow to notice this?! https://t.co/t83Jn8dyhB 原帖
07-31 13:38 @RobertJBye · 原创/引用：Also TBT to the 3 games a day World Cup era 🥺 原帖
07-31 13:53 @minilek · 转发：RT @AddisCoder: Congrats to the Ethiopian teams at the Pan African Informatics Olympiad (Rwanda) and Pan African Math Olympiad (Ivory Coast… 原帖
07-31 14:04 @minilek · 原创/引用："Mission, part of the San Francisco Unified School District, has for four consecutive years had the highest percentage of students accepted to UC Berkeley of any California public high school with more than 50 applicants 原帖
07-31 14:15 @RobertJBye · 疑似回复：@itsnoahd @googlecalendar Better late than never! 原帖
07-31 14:36 @iamsashasax · 疑似回复：@AlexiGlad Very nice insight! As we increase k modes, flops/step increases but so does data efficiency. So there is some optimal compute-efficient k, and does that change with scale? I haven\"t read the full paper yet an 原帖
07-31 14:39 @RobertJBye · 疑似回复：@novikoff WAIT WHAT?!?!?!?! Yoooooooooof all I've ever wanted is to be able to make my iPhone photos look like my fuji ones for consistency 原帖
07-31 15:17 @jiaxinwen22 · 转发：RT @AndrewCurran_: The chairmen of two House Select Committees (CNBC does not name them but I'm assuming this is Reps. John Moolenaar and A… 原帖
07-31 15:22 @RobertJBye · 疑似回复：@triketora I now want a physical switchboard to all my agents https://t.co/QA4876tzUD 原帖
07-31 15:43 @trq212 · 疑似回复：@badlogicgames I don't understand this actually. The asymmetry is in our favor (you can have perfect software that is not vulnerable), but don't defenders need more time to harden before attackers? Isn't the ideal model 原帖
07-31 15:52 @trq212 · 疑似回复：@badlogicgames yeah not good, probes are hard but I think will improve fast agree that perfect (non-vulnerable) software is maybe like a frictionless surface, but hopefully the asymmetry is still in our favor 原帖
07-31 16:04 @RobertJBye · 原创/引用：Coffee nerds! I\"m making a tier list of all the best decaf coffees — ideally Swiss water process & organic. I\"m about a dozen in, and this is the best so far. I\"ve tried sightglass, counter culture, groundwork, red ro 原帖
07-31 16:04 @RobertJBye · 疑似回复：@joshuatopolsky Imagine this plus Claude though 原帖
07-31 16:59 @RobertJBye · 疑似回复：@novikoff Got to have some small wins during parental leave 😅 原帖
07-31 17:00 @RobertJBye · 疑似回复：@_jamesmck Oooh nice, we\"ll try it out! 原帖
07-31 17:01 @RobertJBye · 疑似回复：@CristianRus4 @claudeai Shoutout to @daniel__designs! This was the first thing he worked on when he joined a few months ago. 原帖
07-31 17:01 @RobertJBye · 转发：RT @CristianRus4: ok @claudeai, I was not aware of your game... https://t.co/1CzOwvzi1O 原帖
07-31 17:05 @RobertJBye · 疑似回复：@SambitBiswas @googlecalendar Nice! It looks way better 原帖
07-31 17:54 @jiaxinwen22 · 原创/引用：great to see TML starts a safety team. This is rare in the open weight ecosystem. Nit: we should avoid arguments like "releasing our model is fine because it's no more dangerous than Grok or other open-weight models" 原帖
07-31 18:13 @AlecTPhD · 原创/引用：this has that big moment feel 原帖
07-31 18:16 @RobertJBye · 疑似回复：@Aurelien_Gz This is so cool! At college I designed a hardware version of this idea. https://t.co/u17ZbiwSgr 原帖
07-31 18:24 @jiaxinwen22 · 疑似回复：@PradyuPrasad it can incentivize everyone to skip any safety work 原帖
07-31 18:37 @_sholtodouglas · 疑似回复：@mpauldaniels @AgustinLebron3 Specifically the hedge fund! 原帖
07-31 18:39 @jiaxinwen22 · 疑似回复：@ziqiao_ma 🫡 原帖
07-31 19:04 @RobertJBye · 疑似回复：@madsmccaus Espresso! Flat white or macchiato mainly 原帖
07-31 19:04 @RobertJBye · 原创/引用：Camera reviewers rarely talk about photography Phone reviewers rarely talk about apps Headphone reviewers rarely talk about music Car reviewers rarely talk about the best drives Benchmarks rarely talk about what they mak 原帖
07-31 19:15 @RobertJBye · 疑似回复：@MKelner So true. I guess reviewers and doers are very different skills. 原帖
07-31 19:17 @RobertJBye · 疑似回复：@mattcassinelli I I feel in each of those markets above there must be room to use the things in the way they were intended, and to show the output of them. There are a few very talented Photographer is out there who also 原帖
07-31 20:07 @RobertJBye · 疑似回复：@heyderekj This is a legit review ha ha 原帖
07-31 20:55 @_sholtodouglas · 疑似回复：@krishnanrohit AUM in hedge fund 原帖
07-31 21:08 @angjiang · 转发：RT @badlogicgames: kv cache strikes again. https://t.co/UBLZrXIXlP 原帖
08-01 00:13 @angjiang · 疑似回复：@miramurati This is awesome! 原帖
08-01 00:33 @angjiang · 疑似回复：@suchenzang :( 原帖
08-01 02:59 @katelyn_lesse · 疑似回复：@devagrawal09 interesting! curious how you think about the \"harness\" definition - sounds like system prompt, tools, etc are all considered part of the harness but skills are not? seems like a somewhat arbitrary line. w 原帖
08-01 07:53 @RobertJBye · 原创/引用：Lots more craft and interaction moments like this will be coming to our mobile and desktop apps soon! And you\"re an engineer who cares about design, DM me, we\"re hiring on a new team… 原帖

Cognition

07-31 11:28 @jhanikhil · 转发：RT @jhanikhil: even better, 4 unique vowels (linguistically) when pronounced in general american english u ɑ ʊ ɔ 原帖
07-31 12:38 @jhanikhil · 疑似回复：@repligate wow all quite reasonable choices creatively / stylistically except maybe the cut (in both video and the abrupt super upbeat instrument) at ~4s 原帖
07-31 12:39 @jhanikhil · 疑似回复：@repligate insanely impressive especially considering this model can't hear 原帖

Cursor

07-31 09:03 @benln · 原创/引用：Signing off for the next 25 hours. No electricity, no internet, no driving, no tech, no money, no music, no work, no cooking. Just rest, family, prayer, friends, and food. I do this every week for Shabbat, no exceptions. 原帖
07-31 11:23 @srush_nlp · 原创/引用：Jax added support for opaque types with custom tangents. One cool use is proper handling of batched geometric objects. Here's a short demo of a differentiable rasterization, i.e. min L2 w/ 1500 triangles to a photo. Inte 原帖
07-31 11:26 @srush_nlp · 原创/引用：Jax docs is the best webpage. 原帖
07-31 11:27 @srush_nlp · 原创/引用：Here's a cow as well. https://t.co/qX7qjKdguL 原帖
07-31 11:38 @srush_nlp · 疑似回复：@alexUnder_sky Oh that library is cool, I hadn't see it. The 2D physics thing is interesting as well. (My example is just writing out through cairo) 原帖
07-31 11:56 @srush_nlp · 转发：RT @cornell_tech: Cornell Tech has appointed two longtime faculty leaders to new academic leadership roles, naming Mert Sabuncu as Associat… 原帖
07-31 13:09 @srush_nlp · 疑似回复：@TheyCallMeMr_ yeah, its a great name. also "its values lower to ordinary (\"lojax\") arrays," 原帖

Google DeepMind

07-31 09:34 @suchenzang · 原创/引用：> Research-wise, Simile in 2026 for simulation is like OpenAI in 2019 for AI assistants - signs of life, a path to scaling, yet still many open research questions. interesting... was that really how things looked in 2 原帖
07-31 09:58 @suchenzang · 疑似回复：@shaped 2019 had more than that, but it sure wasn't rolling out AI assistants (anthropic people were still around then, no way would they insult the singularity this way) 原帖
07-31 11:50 @suchenzang · 原创/引用：the real magic of the current zeitgeist is finding out just how many fat babies there are everywhere with candy for the taking 原帖
07-31 11:56 @suchenzang · 疑似回复：@Rafa_Schwinger false. true. false. 原帖
07-31 11:57 @JaidevShah4 · 原创/引用：Had a fun time on the @joinaicircle panel this Wednesday! agents, memory, continual learning and RL all in 40 mins :) https://t.co/exQARUnN1g 原帖
07-31 12:02 @TianfuF · 疑似回复：@LiTianleli Congrats Tim 原帖
07-31 12:30 @suchenzang · 原创/引用：YC, the frontier applied ai team, with rag co in '23, wrapper in '24, agent in '25, and data in '26 原帖
07-31 12:35 @suchenzang · 原创/引用：would've ripped so hard in the right order at the right time 原帖
07-31 13:20 @wenhaocha1 · 转发：RT @huskydogewoof: 🔥 𝐍𝐞𝐰 𝐛𝐥𝐨𝐠: 𝐓𝐨𝐰𝐚𝐫𝐝𝐬 𝐋𝐨𝐨𝐩𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 𝐃𝐨𝐧𝐞 𝐑𝐢𝐠𝐡𝐭 — 𝐏𝐚𝐫𝐭 𝐈 Looped models reuse the same weights across depth, promising a be… 原帖
07-31 14:50 @suchenzang · 原创/引用：do you think The Information has alpha? 原帖
07-31 15:01 @TobyPhln · 疑似回复：@novikoff Agreed 原帖
07-31 15:28 @letian_fu · 转发：RT @anniechen: Under the hood, this requires: - Action model that coordinates whole-body movement and fine motor skills simultaneously, w… 原帖
07-31 15:28 @suchenzang · 疑似回复：@creatine_cycle as the youths might say: "trvke" 原帖
07-31 16:38 @divy93t · 疑似回复：@BlancheMinerva @roydanroy pinged him, ty! 原帖
07-31 17:41 @suchenzang · 疑似回复：@abdallah197_ for good reason 原帖
07-31 18:52 @JaidevShah4 · 疑似回复：@ashpreetbedi why is this surprising ? 原帖
07-31 19:45 @ekindogus · 原创/引用：Physics is a never-ending quest, and the ultimate proving ground for AI and RSI. 原帖
07-31 22:58 @suchenzang · 原创/引用：just had someone on a bike grab my bag off my shoulder and attempt to ride off. i held on to my bag, yelled at them, and they let go and biked off. valencia and 18th street, sf. what a dump. 原帖
08-01 01:35 @suchenzang · 原创/引用：finally called the cops an hour later after seeing this twat circle the same block again looking for a new victim. also calmed down a bit and realize how stupid it was for me to yell at this thief. bag wasn't even valuab 原帖
08-01 01:51 @suchenzang · 疑似回复：@perryfromsoma i used to carry that when i lived in new york if there's wind you get screwed 原帖
08-01 03:21 @wenhaocha1 · 转发：RT @ducx_du: Introducing AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents. Long-horizon coding trajectories oft… 原帖
08-01 04:08 @wenhaocha1 · 原创/引用：Something I've been thinking about a lot lately. My worry about token factories is pricing power. People will keep buying tokens, that part is fine. The question is whether anyone can charge extra. Open models keep catch 原帖
08-01 07:24 @JoeFenton · 原创/引用：…or just talk with your kids like a human being 原帖

Meta AI

07-31 08:45 @zhuokaiz · 疑似回复：@jaxkoh_ sounds like you'd be the best person to create it 👍 原帖
07-31 11:16 @sainingxie · 疑似回复：@tommiekerssies @AlexiGlad yeah, i immediately thought of deltaworld. both are great papers! 原帖
07-31 11:28 @sainingxie · 疑似回复：@AlexiGlad 🔥🍿 mode forcing sounds really exciting too! wish i could binge the whole series :) 原帖
07-31 11:58 @EdwardSun0909 · 原创/引用：muse spark 1.1 is a good model 原帖
07-31 12:43 @maddiehfaulkner · 疑似回复：@alvinsng Wow that\"s epic 原帖
07-31 13:13 @honglin_c · 转发：RT @EdwardSun0909: muse spark 1.1 is a good model 原帖
07-31 15:51 @JIACHENLIU8 · 疑似回复：@AnneliesGamble @AdaFang_ Glad to see the community is converging to make science verifiable ! 原帖
07-31 23:41 @WenhuChen · 疑似回复：@jasondeanlee Well, that\"s much harder. 原帖

Moonshot/Kimi

08-01 05:04 @HaoningTimothy · 原创/引用：🐋 原帖

OpenAI

07-31 08:06 @keyanzhang · 原创/引用：so um what exactly is an Engineering Leader on linkedin 原帖
07-31 08:11 @gabrielchua · 转发：RT @JangLawrenceK: Luna is one of my favorite models out right now, it also does really well on the computer-use benchmarks we tested On O… 原帖
07-31 08:44 @andrew_n_carr · 原创/引用：Really terrible week for manifesto writers everywhere 原帖
07-31 08:59 @ajambrosino · 疑似回复：@rennsong still better than a slop doc 原帖
07-31 09:01 @sama · 原创/引用：cool use case of chatgpt work i heard last night: connect your family calendars and explain your kids' interests. every morning for the drive to school, have it make a podcast that talks about one kid's soccer game that 原帖
07-31 09:10 @sandersted · 疑似回复：@DineshDataAI yep, already getting these settings for free 原帖
07-31 09:13 @steipete · 疑似回复：@raphh @faborio No, maybe @mstroeck does? 原帖
07-31 09:28 @ajambrosino · 疑似回复：@Dimillian @kaiyuhsu @tarekoyou 原帖
07-31 10:03 @ajambrosino · 疑似回复：@iansilber all hand coded I assume 原帖
07-31 10:03 @ajambrosino · 疑似回复：@ejstembler yeah yeah yeah 原帖
07-31 10:09 @trekedge · 原创/引用：I just invested at a 2b dollar valuation 原帖
07-31 10:13 @andrew_n_carr · 原创/引用：Lots of great tidbits in the H3 blog post. The one I like most is that instead of tacking on a Super Resolution model, like we have seen all over, they bake it into the model and generate low to high. This feels akin to 原帖
07-31 10:14 @yong_zhengxin · 疑似回复：@BronsonSchoen thanks for referencing our self-jailbreaking work!! Funny thing is that during our study, we thought this wouldn't happen for closed models because we found safety post-training mitigated it. guess we're w 原帖
07-31 10:15 @yong_zhengxin · 转发：RT @BronsonSchoen: This seems extremely clearly motivated reasoning IMO and I\"m surprised the incident report is so credulous of Claude\"s r… 原帖
07-31 10:23 @ajambrosino · 原创/引用：https://t.co/T8eVc49bDI 原帖
07-31 10:37 @ajambrosino · 原创/引用：which? 原帖
07-31 10:52 @KarelDoostrlnck · 疑似回复：@morganlinton It was a raffle! But I like to think I got it for spending a lot of tokens instead 原帖
07-31 11:00 @nickbaumann_ · 原创/引用：https://t.co/DLEYnc5nmu 原帖
07-31 11:08 @JustinBleuel · 疑似回复：@EverydayAI_ @Conor_D_Dart At your service 🫡 原帖
07-31 11:08 @ajambrosino · 疑似回复：@mrlesk I use steer all the time anyway, I don't need the raw cot 原帖
07-31 11:09 @ajambrosino · 疑似回复：@SIGKITTEN interrupt 原帖
07-31 11:21 @nickbaumann_ · 原创/引用：Tl;dr I gave ChatGPT Work (from the mobile app) a 30 second sample of @jxnlco talking, a link to his essays, and then it tuned a voice model and made those essays into podcast episodes 原帖
07-31 11:38 @steipete · 转发：RT @tenderizzation: two schools of thought https://t.co/M7XjnTfIUW 原帖
07-31 11:40 @guinnesschen · 疑似回复：@youphilthis Working on it! 原帖
07-31 11:53 @ajambrosino · 疑似回复：@acrognali true 原帖
07-31 11:55 @giffmana · 疑似回复：@tozhou I did not know what it was, thank you, makes sense. https://t.co/d1UinYPbf9 原帖
07-31 11:57 @reach_vb · 原创/引用：check it out here: https://t.co/cS517cR9EI 原帖
07-31 12:01 @reach_vb · 疑似回复：@heccbrent thought you\"d never ask 💕 原帖
07-31 12:07 @giffmana · 疑似回复：@Test20061722 What for? Not a fan of the recent wave of startups just clearly shooting for acquisition. (Nice pic though! Maybe i should do more grooming) 原帖
07-31 12:13 @giffmana · 疑似回复：@Stone_Tao @physical_int Quickest intern to fulltime conversion, like, ever? 原帖
07-31 12:14 @giffmana · 疑似回复：@johnamonaco @growing_daniel https://t.co/dYCOtI5Dya 原帖
07-31 12:15 @giffmana · 原创/引用：lol https://t.co/67fkaI9ahm 原帖
07-31 12:18 @KarelDoostrlnck · 疑似回复：@cooperjsaye @OpenAI Welcome ! 原帖
07-31 12:18 @giffmana · 疑似回复：@aidan_clark The one thing that llama4 was thiiiis 🤏 close to getting right: Scout, Maverick, ... Uhm! 原帖
07-31 12:28 @danintheory · 疑似回复：@deanwball @tszzl Amusingly, for a startup in the early 2010s it used to be important to brand as \"machine learning\" and not say \"AI\"… 原帖
07-31 12:28 @johnohallman · 转发：RT @signulll: dunking on ppl who have tried extraordinary things & then had a set back is lame. 原帖
07-31 12:44 @TheRohanVarma · 转发：RT @daniel_mac8: Codex users: do this right now. Max reasoning effort is off by default. Luna at Max reasoning is ~ Sol Medium / Opus 5… 原帖
07-31 13:04 @giffmana · 疑似回复：@JustinSmethers Fast takeover 原帖
07-31 13:07 @reach_vb · 疑似回复：@ashwinexe @devfolio congrats on such an epic run! 原帖
07-31 13:09 @steipete · 转发：RT @nunezvice: let\"s put this 80% price drop for 5.6 🌙 in perspective three weeks ago, gpt-5.5 xhigh was our frontier. it completed 67% of… 原帖
07-31 13:10 @giffmana · 疑似回复：@gabriel1 So jealous. When i had the time to go, i was too poor. When i stopped being too poor, i stopped having the time 😢 原帖
07-31 13:11 @giffmana · 疑似回复：@Test20061722 That's a good reason tbh 😁 原帖
07-31 13:14 @reach_vb · 疑似回复：@mihaimaruseac wow only 6 months? congrats on shipping privacy filter and everything else that came with it! 🐐 原帖
07-31 13:39 @zhansheng · 疑似回复：@idavidrein Well, where\"s the LinkedIn post sir 原帖
07-31 13:42 @zhansheng · 原创/引用：\"This wedding will have everything\" https://t.co/Y1MuSh4oMI 原帖
07-31 13:44 @neil_projects · 疑似回复：@signulll I wouldn\"t bet against @demishassabis 原帖
07-31 13:47 @prd_008 · 转发：RT @OpenAIDevs: See what needs you next. The new Activity view in the ChatGPT desktop app brings together conversations that need your att… 原帖
07-31 13:53 @nickbaumann_ · 疑似回复：@pashmerepat @jxnlco https://t.co/SMSvZmvrxF 原帖
07-31 14:01 @reach_vb · 疑似回复：@ChatGPT cc: @altryne 原帖
07-31 14:04 @KarelDoostrlnck · 疑似回复：@kdaigle If you\"re not coding on vinyl you\"re ngmi 原帖
07-31 14:06 @ajambrosino · 疑似回复：@thenanyu @joshpuckett https://t.co/a0JPne6u2y 原帖
07-31 14:31 @reach_vb · 原创/引用：You should be Luna-maxxing: > Luna (Max) matches GPT‑5.5 xHigh for roughly 12× less > Terra Max scores 2.6 points higher while costing 45% less https://t.co/TlzNBoXT99 原帖
07-31 14:35 @reach_vb · 原创/引用：Made w/ ChatGPT Work on my iPhone: https://t.co/IXiUF3pBBj 原帖
07-31 14:37 @andrew_n_carr · 疑似回复：@eladgil @leopoldasch 👀👀👀 原帖
07-31 14:38 @johnohallman · 原创/引用：Don't be a doomer. It is possible to get AI alignment right. It is possible to build a society where everyone benefits from this technology, one without disempowerment and upsetting inequality. And the future this would 原帖
07-31 14:47 @andrew_n_carr · 原创/引用：My guess is that the models are getting harder to understand because they're being optimized for sub agent spawning and communication. We are not the intended audience 原帖
07-31 14:51 @reach_vb · 疑似回复：@OpenAIDevs thank you 5.4 for everything! 🫡 time to move to terra & luna: 原帖
07-31 14:56 @jeffreygwang · 转发：RT @johnohallman: Don't be a doomer. It is possible to get AI alignment right. It is possible to build a society where everyone benefits fr… 原帖
07-31 15:09 @trekedge · 疑似回复：@ajambrosino Steer is the default queue is for power users 原帖
07-31 15:18 @prd_008 · 转发：RT @ChatGPT: Work with just your voice, hands free. See—and hear— ChatGPT Voice in action, now on macOS and Windows. https://t.co/D3HrcfN8… 原帖
07-31 15:25 @reach_vb · 原创/引用：I had a sandwich today 原帖
07-31 15:45 @nickbaumann_ · 转发：RT @ChatGPT: Work with just your voice, hands free. See—and hear— ChatGPT Voice in action, now on macOS and Windows. https://t.co/D3HrcfN8… 原帖
07-31 15:46 @dylancablayan · 转发：RT @Dexerto: YouTuber Hank Green has paused uploading on multiple channels after admitting he relies "too heavily" on AI "I need to come t… 原帖
07-31 15:56 @nickbaumann_ · 疑似回复：@jxnlco https://t.co/1R6nf4vfoO 原帖
07-31 16:01 @steipete · 转发：RT @vltansky: This is so neat. @cursor_ai @ClaudeDevs what happened to you guys? Codex been here just few months and you already seems behi… 原帖
07-31 16:01 @steipete · 转发：RT @threepointone: https://t.co/1fFlz1P0K1 原帖
07-31 16:02 @steipete · 转发：RT @mamagnus00: What is openAI doing?? Just winning in accuracy, speed and cost? 原帖
07-31 16:05 @steipete · 原创/引用：will we have to wear radiation shielded underwear now? 原帖
07-31 16:09 @prd_008 · 原创/引用：There's good writing days and bad writing days, and no matter how much you practice, you just have to eventually accept that you have no control over what kind of day its going to be. 原帖
07-31 16:13 @steipete · 疑似回复：@truongphillipt2 growth hacking? 原帖
07-31 16:18 @thsottiaux · 疑似回复：@ForwardEditor What's next in the pipeline? 原帖
07-31 16:19 @thsottiaux · 疑似回复：@theo They shipped it one day early. Thanks Theo. Let's keep learning together. https://t.co/eecKQYRRv7 原帖
07-31 16:20 @thsottiaux · 疑似回复：@viewsfrom02108 Let's circle back soon 原帖
07-31 16:21 @thsottiaux · 疑似回复：@theo openai = everyone Flattered. But leave some for the others in the space. 原帖
07-31 16:48 @andrew_n_carr · 原创/引用：One of my best contributions was emojis in oai slack that corresponded to the neuron count of our models (e.g., fruit flies and meal worms) 原帖
07-31 16:50 @thsottiaux · 疑似回复：@theo Wait. You can get paid on this app? 原帖
07-31 16:52 @yong_zhengxin · 转发：RT @thinkymachines: Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a… 原帖
07-31 17:04 @dylancablayan · 转发：RT @reach_vb: I had a sandwich today 原帖
07-31 17:06 @jxnlco · 转发：RT @heyjgold: if you\"re in tech, your taste and vision has never mattered more. AI can make anything now. which means knowing what\"s wort… 原帖
07-31 17:07 @jxnlco · 转发：RT @steipete: Jason is doing such great work. Love this. 原帖
07-31 17:07 @jxnlco · 转发：RT @DeryaTR_: I can confirm this firsthand. I\"ve interacted with quite a few people working at @OpenAI over the past 2-years & they were al… 原帖
07-31 17:07 @jxnlco · 转发：RT @nickbaumann_: OpenAI is composed of so many wonderful people like @jxnlco who are kind, creative, & extremely ambitious And the cultur… 原帖
07-31 17:07 @jxnlco · 转发：RT @reach_vb: One of the things I love most about this video is how well it captures the ethos of OpenAI and what drives the people behind… 原帖
07-31 17:11 @jxnlco · 转发：RT @derrickcchoi: We're lucky to have this legend on the team 原帖
07-31 17:11 @jxnlco · 转发：RT @Dimillian: This is such an amazing video, and every time I watch this, I get chills. I'm so lucky to work and be surrounded by such pas… 原帖
07-31 17:11 @jxnlco · 转发：RT @anshitasaini_: wow this is amazing, really resonates 原帖
07-31 17:12 @jxnlco · 转发：RT @prd_008: There\"s hope in trusting in humanity 原帖
07-31 17:12 @jxnlco · 转发：RT @TechFollowrazzi: 🗞️ @sama followed @jxnlco @jxnlco is the \"vibes @openai\" Codex DX guy who built Instructor and is currently the one… 原帖
07-31 17:12 @jxnlco · 转发：RT @coreyching: The future is bright. Checkout @jxnlco's video and if you haven't already follow him for increasing the vibes of the intern… 原帖
07-31 17:12 @jxnlco · 转发：RT @FieroTy: "Everyone deserves the power and tools to be curious, build things, and shape their own futures" I love working at @OpenAI ❤️ 原帖
07-31 17:12 @jxnlco · 转发：RT @dl_insider: How OpenAI turned the situation around is something that should be studied in marketing and business schools. It is true t… 原帖
07-31 17:19 @trekedge · 疑似回复：@nunezvice Rest when? 原帖
07-31 17:43 @guinnesschen · 疑似回复：@jxnlco @arthaud_ > to switch between text & voice mid chat Coming soon! 原帖
07-31 17:44 @guinnesschen · 转发：RT @alliekmiller: It is nearly 2am and the only reason I just finished all of my urgent work is Codex Voice Mode. Good god. How is ever… 原帖
07-31 17:48 @andrew_n_carr · 原创/引用：What if you were once good at Tetris? 原帖
07-31 18:14 @KarelDoostrlnck · 疑似回复：@RylanSchaeffer Email ? 原帖
07-31 18:24 @nickbaumann_ · 转发：RT @reach_vb: I had a sandwich today 原帖
07-31 18:39 @andrewho03 · 疑似回复：@teortaxesTex Seriously impressive 原帖
07-31 18:43 @andrewho03 · 疑似回复：@_sholtodouglas Will take this bet. What odds and with what size? 原帖
07-31 18:43 @andrewho03 · 疑似回复：@alexeyguzey Good point 原帖
07-31 18:47 @steipete · 转发：RT @aarondfrancis: On a bit of a whim, I let Codex and Claude talk to each other about whatever they wanted and told them they didn't need… 原帖
07-31 18:48 @guinnesschen · 疑似回复：@arthaud_ @jxnlco Wdym by that? 原帖
07-31 18:49 @andrewho03 · 原创/引用：My dermatologist agreed, btw. 原帖
07-31 18:51 @steipete · 原创/引用：Queue was the way but with 5.5 the model doesn\"t get confused anymore, you can just throw stuff at it while it works and it will diligently process whatever you throw at it. 原帖
07-31 18:51 @yong_zhengxin · 转发：RT @ziqiao_ma: @jiaxinwen22 We\"ve always had a safety team, and we\"ve discarded higher-performing checkpoints because of safety concerns. 😂 原帖
07-31 19:00 @steipete · 转发：RT @sarah_edo: We (Chrome) published a big paper about security work that's important to share. It shouldn\"t surprise anyone that open sour… 原帖
07-31 19:05 @thsottiaux · 原创/引用：I don't come often to GMail (OpenAI is a Slack company), but I swear, every time I do open it, there is a new button somewhere around the top right corner. https://t.co/ZpasAaoaFl 原帖
07-31 19:13 @gabrielchua · 转发：RT @GregKamradt: I had a gnarly research project I needed done - 78K different classifications I let GPT-5.6 Luna run over night, batch mo… 原帖
07-31 19:26 @thsottiaux · 原创/引用：And yet, I don't see a reset button there. 原帖
07-31 19:32 @andrewho03 · 原创/引用：The ability for DeepSeek to be price competitive with GPT-5.6 Luna is genuinely incredible. The inference team at OpenAI was brilliant from what I heard. No clue how DeepSeek does it with their resource constraints. Marg 原帖
07-31 19:42 @gabrielchua · 转发：RT @sinasanm: i just tried gpt 5.6 luna and yo....at this rate i will not be using open source models anymore! i perceived open source mode… 原帖
07-31 19:43 @andrewho03 · 疑似回复：@quantyboi @obsolete26 Not explicitly but it\"s pretty general and should be applicable 原帖
07-31 19:56 @andrewho03 · 疑似回复：@alexframegreen Newport Beach guy too so you know he\"s good. 原帖
07-31 20:01 @jam3scampbell · 原创/引用：everyone wants to have a take on leopold, very few are counting the OOMs 原帖
07-31 20:01 @thsottiaux · 原创/引用：ChatGPT Work counted https://t.co/XlnkkYvx1s 原帖
07-31 20:04 @ah20im · 原创/引用：I had a sandwich today 原帖
07-31 20:06 @jam3scampbell · 疑似回复：@theojaffee who would\"ve thought the eschaton would be monitored 原帖
07-31 20:32 @thsottiaux · 原创/引用：To celebrate a week of efficiency and let you run 100'000 Luna threads this weekend... that's right... wait for it... I have reset usage limits for Codex and ChatGPT Work. Enjoy. 原帖
07-31 20:33 @thsottiaux · 疑似回复：@maria_rcks Fixed it 原帖
07-31 20:36 @andrewho03 · 疑似回复：@sensho Wish I understood the financial numbers here — feels a bit like a fool\"s errand but who knows 原帖
07-31 20:44 @andrewho03 · 疑似回复：@sensho indeed. I suppose maybe we will see when IPO time comes 原帖
07-31 20:47 @ajambrosino · 疑似回复：@BennettBuhner https://t.co/312ZxSp5yP 原帖
07-31 20:55 @dylancablayan · 转发：RT @angelbrodin: As an original vlogbrother fan/nerdfighter from the brotherhood 2.0 days, I\"m not surprised there\"s a good population of H… 原帖
07-31 21:08 @andrewho03 · 原创/引用：Constructing high-quality datasets at scale is really hard. Incredibly so. No AI lab researcher is an expert in every field, let alone necessarily more than one; to produce even a simple benchmark means continually stari 原帖
07-31 21:09 @jxnlco · 转发：RT @dyett: This is the best expression I\"ve seen of what makes OpenAI\"s mission and values are distinctive in the industry And a single… 原帖
07-31 21:09 @jxnlco · 疑似回复：@DoctorYev At the office maybe twice a week. 原帖
07-31 21:11 @ajambrosino · 转发：RT @danshipper: pretty cool to have the kicker in this @WSJ piece on OpenAI vs Anthropic! i stand behind this…has been pretty clear since… 原帖
07-31 21:14 @gabrielchua · 转发：RT @thsottiaux: To celebrate a week of efficiency and let you run 100'000 Luna threads this weekend... that's right... wait for it... I hav… 原帖
07-31 21:34 @andrewho03 · 疑似回复：@jklpuzo1923 problem here is making sure your grading functions are good 原帖
07-31 21:44 @andrewho03 · 疑似回复：@0xf3mi Privileged access to infra, accumulated know-how. 原帖
07-31 22:05 @ajambrosino · 疑似回复：@bryceoflife That\"s what steer is 原帖
07-31 22:12 @jxnlco · 原创/引用：Touching grass this weekend. https://t.co/eCXVIjzE5V 原帖
07-31 22:14 @andrewho03 · 疑似回复：@santiaghini ? 原帖
07-31 22:18 @thsottiaux · 疑似回复：@jxnlco Where are you haha 原帖
07-31 22:18 @thsottiaux · 疑似回复：@maria_rcks Liquidated 原帖
07-31 22:28 @guinnesschen · 疑似回复：@xGoatJames @jxnlco Try asking it to set a heartbeat that runs every 2 min 原帖
07-31 22:34 @thsottiaux · 疑似回复：@jaredpalmer https://t.co/NHfYSOfuGD 原帖
07-31 22:40 @thsottiaux · 原创/引用：Optimize for curiosity 原帖
07-31 22:42 @andrewho03 · 原创/引用：Looking for connects to anyone at Mistral interested in these datasets: long-horizon statistical analysis, CAD environments (CUA or tool use), PDF parsing (finance & other white collar work), biology in general, chem 原帖
07-31 22:45 @jxnlco · 转发：RT @dhh: Omarchy Quattro will automatically detect if you're using Claude or Codex subscriptions, and then make an agent panel available to… 原帖
07-31 22:46 @jxnlco · 疑似回复：@thsottiaux 1) was talking at an OpenAI influencer event upstate New York 2) i was in toronto for a codex marketing production touching the fake grass. 原帖
07-31 22:55 @andrewho03 · 疑似回复：@theasarkar working exceptionally well! 原帖
07-31 23:26 @reach_vb · 疑似回复：@swyx ngl timing matched perfectly!!! 原帖
07-31 23:27 @reach_vb · 原创/引用：time to kick off your agent runs before you sleep! 原帖
07-31 23:32 @reach_vb · 疑似回复：@swyx OpenAIs mission is to ensure AGI benefits humanity (through resets) 原帖
07-31 23:35 @reach_vb · 疑似回复：@dkundel I\"m glad you asked: A toasted baguette w/ fresh pesto Lots of arugula Topped up w/ Prosciutto Thick layer of fresh buffala Chilli oil and balsamic glaze 原帖
07-31 23:48 @reach_vb · 原创/引用：I love ChatGPT Work 原帖
07-31 23:50 @steipete · 疑似回复：@ayaangazali 🙌 原帖
07-31 23:59 @reach_vb · 原创/引用：Usage limits reset across Codex and ChatGPT Work! This weekend let your curiosity run wild & free, think of an idea offload to Chat Work or Codex! 原帖
07-31 23:59 @BarrAlexandra · 疑似回复：@meshtimes_ @jxnlco @OpenAI dang how do i get an invite to my own corp events :\") 原帖
08-01 00:00 @andrewho03 · 疑似回复：@init_malachi Approach in what sense? 原帖
08-01 00:03 @reach_vb · 疑似回复：@Dimillian yes, if you want ideas: 原帖
08-01 00:05 @BarrAlexandra · 原创/引用：when u say tautological instead of redundant >:( 原帖
08-01 00:08 @giffmana · 疑似回复：@tkukurin Sometimes he's kid1 and sometimes he's kid2 原帖
08-01 00:08 @giffmana · 疑似回复：@Stone_Tao @physical_int Well congrats and enjoy! 原帖
08-01 00:08 @reach_vb · 疑似回复：@lukeknight Lifemaxxing! Congrats on a stellar week! 原帖
08-01 00:09 @BarrAlexandra · 疑似回复：@cooperjsaye @OpenAI welcome!! you\"re joining a great and important team :) 原帖
08-01 00:30 @andrewho03 · 原创/引用：https://t.co/KAGMPzVFep 原帖
08-01 00:34 @giffmana · 疑似回复：@suchenzang Come to Zurich (Sorry to hear, it sucks, well done holding onto it) 原帖
08-01 00:37 @andrewho03 · 疑似回复：@GuiveAssadi @papasmurfffs Oops. Disagree 原帖
08-01 00:40 @giffmana · 疑似回复：@cgarciae88 Maybe this to begin with, what are they even doing... https://t.co/JclcnrBoCL 原帖
08-01 00:41 @Yuf_Zh · 原创/引用：3 years ago 原帖
08-01 00:42 @alexwei_ · 转发：RT @SebastienBubeck: yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model.… 原帖
08-01 00:42 @andrewho03 · 疑似回复：@fleetingbits I think the biggest factor is that, in general, at almost all companies, it's hard to get rewarded for infra work. It is ESPECIALLY hard to get rewarded for infra work that makes the lives of people who are 原帖
08-01 00:45 @yubai01 · 原创/引用：My jaw dropped 10 times 😅 But really this is going to be an avalanche. Been throwing a few open questions of mine at Astra too, boy is it strong. 原帖
08-01 00:48 @HongxunWu · 原创/引用：https://t.co/63ZEIiPuuL 原帖
08-01 00:48 @Yuf_Zh · 原创/引用：Astra 🚀 excited to see these amazing results coming out from the new frontier! 原帖
08-01 00:50 @zhansheng · 原创/引用：My team works on some really cool stuff! 原帖
08-01 01:17 @polynoamial · 原创/引用：An internal version of Astra, @OpenAI\"s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step for scientific reas 原帖
08-01 01:29 @giffmana · 疑似回复：@Lon Can't believe I defend anthro on this, but you need to up your standards for what you call "definite proof". Code for something existing doesn't mean it actually happening. There's dead code everywhere in the indust 原帖
08-01 01:32 @polynoamial · 原创/引用：The cost of generating the proofs for all 10 of these breakthroughs combined was under $2,000 at Sol API prices. We\"re excited to see what scientists and researchers are able to create with our upcoming Astra models! 原帖
08-01 01:33 @jxnlco · 疑似回复：@JasonBotterill @apples_jimmy @coreyching what happened to the list 原帖
08-01 01:36 @xiangyuqi_pton · 转发：RT @wjmzbmr1: 10 proofs from our next major model Astra on long-standing open problems in mathematics and theoretical computer science (als… 原帖
08-01 01:37 @alexwei_ · 原创/引用：It is remarkable—and takes a moment to process—how quickly the next generation of models will accelerate our research, and breathe new life into old problems. This was an incredible project done by an incredible team (an 原帖
08-01 01:46 @reach_vb · 原创/引用：An internal version of Astra, our next major model, found new results across 10 long-standing open problems in math and theoretical computer science. The total token cost to find all 10 solutions? Roughly $2,000 at Sol A 原帖
08-01 01:48 @reach_vb · 原创/引用：Check out the proofs here: https://t.co/PCMtQfVakl 原帖
08-01 01:48 @polynoamial · 转发：RT @alexwei_: It is remarkable—and takes a moment to process—how quickly the next generation of models will accelerate our research, and br… 原帖
08-01 01:56 @polynoamial · 疑似回复：@iamgroguu @DaveShapi We still haven\"t solved math. Astra isn\"t building new branches of mathematics, or posing interesting new conjectures. Though I admit it\"s hard to believe that tweet was only a year ago. A lot ha 原帖
08-01 02:01 @polynoamial · 原创/引用：And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet). But also, we didn\"t spend a lot on each problem. It\"s possible to push test-time compute much further. 原帖
08-01 02:10 @reach_vb · 原创/引用：ngl, Astra is a pretty cool name 原帖
08-01 02:10 @gabrielchua · 原创/引用：Astra ! 原帖
08-01 02:19 @polynoamial · 疑似回复：@full_kelly_ We believe the fastest way to solve a Millennium Prize problem is to have that large team help build the next generation of general-purpose models. 原帖
08-01 05:13 @JustinBleuel · 原创/引用：Thank you @ChatGPT for answering all my questions big and small https://t.co/xZbPgvZ5m2 原帖
08-01 05:17 @JustinBleuel · 疑似回复：@maria_rcks @thsottiaux It\"d be more believable if the stray jacket had a hood 原帖
08-01 05:19 @gabrielchua · 原创/引用：ChatGPT works for you! https://t.co/fxLFAo1Ayk 原帖
08-01 05:20 @gabrielchua · 转发：RT @polynoamial: The cost of generating the proofs for all 10 of these breakthroughs combined was under $2,000 at Sol API prices. We\"re exc… 原帖
08-01 06:13 @giffmana · 疑似回复：@DimitrisPapail Phew 原帖
08-01 06:17 @giffmana · 疑似回复：@francoisfleuret /btw 原帖
08-01 06:22 @giffmana · 原创/引用：And that's why i recently made a google account for my son, so that he can book slots on my calendar. I love technology! 原帖
08-01 06:57 @sama · 转发：RT @SebastienBubeck: yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model.… 原帖
08-01 06:58 @aidan_clark · 原创/引用：https://t.co/6B7HpgF6uD 原帖
08-01 06:59 @sama · 转发：RT @polynoamial: An internal version of Astra, @OpenAI\"s next major model family, solved 10 major open problems in mathematics, quantum com… 原帖
08-01 07:00 @gabrielchua · 疑似回复：@nathanclark_ @ChatGPT awesome! what are you gonna cook on your 160th say! 原帖
08-01 07:06 @gabrielchua · 转发：RT @dkundel: @dl_insider That business school class would be relatively short: hire people who care, let them bring their authentic selfs,… 原帖
08-01 07:13 @thsottiaux · 原创/引用：The week was for efficiency. The weekend is for 10 major breakthroughs in science. There will be signs. 原帖
08-01 07:14 @thsottiaux · 疑似回复：@kimmonismus \"select customers\" 原帖
08-01 07:14 @ajambrosino · 原创/引用：our new album \"there will be signs\" 原帖
08-01 07:15 @thsottiaux · 疑似回复：@Robertg761_ Too cheap to meter 原帖
08-01 07:18 @gabrielchua · 转发：RT @gregpr07: GPT‑5.6 Luna xhigh is near Opus 5 performance being 17x cheaper 🤯 The new Browser Use pareto frontier is wild 👀 Luna was alre… 原帖
08-01 07:20 @thsottiaux · 疑似回复：@loss_gobbler @JuniperViews You can\"t make this up. Hahahaha 原帖
08-01 07:22 @yong_zhengxin · 转发：RT @polynoamial: An internal version of Astra, @OpenAI\"s next major model family, solved 10 major open problems in mathematics, quantum com… 原帖
08-01 07:23 @gabrielchua · 原创/引用：Cached inputs for 5.6 Luna are now $0.01/M tokens 1 cent! 🫨🫨🫨 原帖
08-01 07:26 @gabrielchua · 原创/引用：Thanks, Luna 🌙 https://t.co/JPMdbZphrE 原帖
08-01 07:29 @sama · 原创/引用：team humanity 原帖
08-01 07:34 @thsottiaux · 疑似回复：@jayair Tell us about Luna going to the moon. The people want to know 原帖
08-01 07:35 @jxnlco · 转发：RT @sama: team humanity 原帖
08-01 07:37 @giffmana · 疑似回复：@adnancagri he needs to book a meeting for that 原帖
08-01 07:37 @giffmana · 疑似回复：@SimonGoodman_ or a calendly link that hacks his twitter, to teach a lesson? 原帖
08-01 07:38 @gabrielchua · 原创/引用：$0.01/M tokens for cached inputs 1 cent! wild!!! https://t.co/CiURwdDv8o 原帖
08-01 07:41 @thsottiaux · 疑似回复：@NanneWielinga @gdb All of them actually. The subscription is very generous in tokens. 原帖
08-01 07:43 @dylancablayan · 原创/引用：Ad Astra 🚀 原帖
08-01 07:43 @dylancablayan · 转发：RT @nunezvice: ad astra per aspera 💫 \"to the stars through hardships\" 原帖
08-01 07:47 @giffmana · 疑似回复：@Lon > You also stated I didn't watch server traffic or server replies. And this is not accurate at all impossible to know from your post as it's never mentioned. But that would be more interesting. Basically everything 原帖
08-01 07:51 @andrewho03 · 疑似回复：@bleysg Damn that\"s impressive 原帖

Thinking Machines Lab

07-31 09:40 @SonglinYang4 · 转发：RT @jun_song: This is just SO CRAZY. I can\"t sleep anymore. 原帖
07-31 11:27 @ziqiao_ma · 疑似回复：@LigengZhu 不如异国他湘原帖
07-31 11:53 @Mascobot · 转发：RT @_sholtodouglas: Prediction: SALP will be bigger than Citadel by the end of the decade. Leopold has predicted the last two years better… 原帖
07-31 12:06 @Mascobot · 原创/引用：💯 原帖
07-31 12:35 @Mascobot · 转发：RT @dylan522p: Love that all FinTwit people making fun of Leopold still get mogged by his returns this year. Leopold is a great guy and he'… 原帖
07-31 12:48 @Mascobot · 原创/引用：More weddings should have this. 原帖
07-31 14:17 @Mascobot · 转发：RT @eladgil: I just asked to invest in @leopoldasch fund for the first time :) 原帖
07-31 16:02 @Ber18791531 · 转发：RT @junlin45300: 🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement… 原帖
07-31 16:11 @ziqiao_ma · 转发：RT @ValsAI: Inkling Small dropped two weeks after Thinking Machines shipped its debut model, Inkling. It is roughly a quarter the size,… 原帖
07-31 16:48 @miramurati · 原创/引用：We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosystem it enters, and how testing, staged access, and stronger defenses can create a path toward gr 原帖
07-31 16:48 @ZhongRuiqi · 转发：RT @thinkymachines: Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a… 原帖
07-31 16:49 @ZhongRuiqi · 转发：RT @miramurati: We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosy… 原帖
07-31 16:55 @Guangxuan_Xiao · 转发：RT @thinkymachines: Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a… 原帖
07-31 16:57 @rown · 转发：RT @thinkymachines: Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a… 原帖
07-31 16:57 @YueYangAI · 转发：RT @miramurati: We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosy… 原帖
07-31 17:01 @shizhediao · 原创/引用：Toward a Future of Open, Secure AI 🚀 原帖
07-31 17:06 @ziqiao_ma · 转发：RT @thinkymachines: Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a… 原帖
07-31 17:08 @miramurati · 转发：RT @clarejtbirch: Alex is good to work with, can kick a ball and captain a football team pretty bloody well, and has often-good thoughts on… 原帖
07-31 17:13 @SonglinYang4 · 转发：RT @miramurati: We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosy… 原帖
07-31 17:16 @SonglinYang4 · 转发：RT @huskydogewoof: 🔥 𝐍𝐞𝐰 𝐛𝐥𝐨𝐠: 𝐓𝐨𝐰𝐚𝐫𝐝𝐬 𝐋𝐨𝐨𝐩𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 𝐃𝐨𝐧𝐞 𝐑𝐢𝐠𝐡𝐭 — 𝐏𝐚𝐫𝐭 𝐈 Looped models reuse the same weights across depth, promising a be… 原帖
07-31 17:19 @ZhongRuiqi · 转发：RT @soumithchintala: Our plan keeps open weights and safety compatible with each other. We hope that this will help resolve debates and p… 原帖
07-31 18:32 @ziqiao_ma · 疑似回复：@jiaxinwen22 We\"ve always had a safety team, and we\"ve discarded higher-performing checkpoints because of safety concerns. 😂 原帖
08-01 00:41 @Ber18791531 · 转发：RT @huskydogewoof: 🔥 𝐍𝐞𝐰 𝐛𝐥𝐨𝐠: 𝐓𝐨𝐰𝐚𝐫𝐝𝐬 𝐋𝐨𝐨𝐩𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 𝐃𝐨𝐧𝐞 𝐑𝐢𝐠𝐡𝐭 — 𝐏𝐚𝐫𝐭 𝐈 Looped models reuse the same weights across depth, promising a be… 原帖

xAI

07-31 08:17 @milichab · 转发：RT @jasonkneen: Honestly and this is not sponsored, @grok and @SpaceXAI grok build are just astonishing. I just paid for the promo for grok… 原帖
07-31 08:17 @milichab · 疑似回复：@jasonkneen @grok @SpaceXAI Glad you liked it, please send over some feedback! 原帖
07-31 08:22 @sivil_taram · 疑似回复：@AndrewZeng17 Nice model! Congrats! 原帖
07-31 08:36 @ypwang61 · 转发：RT @james_y_zou: Auto-research without a verifier 🚀 #SqueezeEvolve is accepted to #COLM2026 and now works in Claude Code! It uses model co… 原帖
07-31 08:42 @sivil_taram · 疑似回复：@WeiLiu99 Dr. Wei 😎 原帖
07-31 08:43 @sivil_taram · 转发：RT @ducx_du: Introducing AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents. Long-horizon coding trajectories oft… 原帖
07-31 08:45 @sivil_taram · 疑似回复：@ducx_du Nice work and impressive homepage for the work! Q: How is the frequency of AutoCompact specified? I'm asking this because it might corrupt the prefix, making it hard to reuse cached prefix, right? 原帖
07-31 08:46 @sivil_taram · 疑似回复：@yuzhenh17 🐐 of agent 原帖
07-31 09:09 @ypwang61 · 转发：RT @yuzhenh17: Come join us ! 原帖
07-31 09:27 @JasonBud · 疑似回复：@guillaumemeyer @milichab yeah makes total sense. @kevngg is on it. 原帖
07-31 10:06 @forwarddeploy · 转发：RT @kamath_sutra: We have passed the Turing test for Voice AI and raised $13M in Series A funding to bring the next generation of Voice to… 原帖
07-31 11:45 @edenchan · 疑似回复：@ryanbrewer Ya he\"s cool 原帖
07-31 14:32 @mycharmspace · 疑似回复：@elonmusk @BasedMikeLee Based 原帖
07-31 15:58 @hexiang · 原创/引用：One last (but nice) bit for July 🌞 https://t.co/7sBnvXVMBA 原帖
07-31 17:32 @hyhieu226 · 转发：RT @AnushElangovan: Power of OPEN: Mi455/CDNA5 ISA: https://t.co/xFsF8Cytom 原帖
07-31 17:49 @milichab · 转发：RT @elonmusk: SpaceX is hiring exceptional engineering & skilled trades talent to build & operate the most powerful AI supercomputer cluste… 原帖
07-31 17:52 @milichab · 原创/引用：Try Imagine Video 1.5 with new image + voice references 原帖
07-31 17:56 @hexiang · 原创/引用：With omni-reference, now imagine 1.5 is much easier to use for making long story with consistent characters 原帖
07-31 18:06 @hexiang · 疑似回复：@IamEmily2050 It is closer to the API model (not exactly the same as it supports more form of inputs); We are always cooking 原帖
07-31 18:11 @XiangLi5217365 · 原创/引用：Try it! 原帖
07-31 18:13 @zeliu_ · 转发：RT @grok: Imagine Video 1.5 launched last month as our best video model yet, with more lifelike motion and sound. Today it goes even furthe… 原帖
07-31 18:13 @forwarddeploy · 转发：RT @elonmusk: SpaceX is hiring exceptional engineering & skilled trades talent to build & operate the most powerful AI supercomputer cluste… 原帖
07-31 18:16 @YknZhu · 原创/引用：Have fun :) 原帖
07-31 18:20 @JasonBud · 转发：RT @grok: Imagine Video 1.5 launched last month as our best video model yet, with more lifelike motion and sound. Today it goes even furthe… 原帖
07-31 18:27 @jefffhj · 原创/引用：Try Grok Imagine Video 1.5 with omni-reference 原帖
07-31 18:45 @ypwang61 · 转发：RT @arena: DeepSeek-V4-Flash-High by @deepseek_ai is #7 overall in the Frontend Code Arena with 1,586 pts! It\"s #3 among open. In categori… 原帖
07-31 18:57 @martin_ma_007 · 原创/引用：Launching our omni-reference model 🚀 More controls. Higher quality. We are cooking to enable more functionalities. Have fun and let us know what you\"d like to see improved!! 原帖
07-31 19:43 @ypwang61 · 疑似回复：@arena @deepseek_ai are there results for max effort? 原帖
07-31 19:59 @hyhieu226 · 疑似回复：@hsu_byron @sgl_project @lmsysorg FP8 KVCache: is there a magic number? iykyk. 原帖
07-31 20:05 @ypwang61 · 转发：RT @FireworksAI_HQ: DeepSeek-V4-Flash-0731 is live on Fireworks, day-zero. DeepSeek reports it beats V4 Pro across all 9 agentic evals, in… 原帖
07-31 20:24 @milichab · 转发：RT @chaitu: Imagine Video 1.5 Upgrades are now live on our API and consumer products. Try them out! 原帖
07-31 20:45 @ypwang61 · 原创/引用：Very impressive results with 35B model! Here are some systematic study of AI4AI researches covering post-training and evolving system 原帖
07-31 20:50 @forwarddeploy · 转发：RT @ishand: See you all next Saturday at the @SpaceXAI Grokathon! https://t.co/KDJjlj5BE5 原帖
07-31 21:30 @milichab · 转发：RT @elonmusk: Grok Build can do almost anything you can think of https://t.co/lOJK57wDP5 原帖
08-01 00:44 @martin_ma_007 · 疑似回复：@heavypulp thanks for being our first users and precious feedbacks 🙌 原帖
08-01 00:45 @martin_ma_007 · 疑似回复：@armandokirwin noted, will pass on this feedback to our product team, stay tuned :) 原帖
08-01 00:46 @martin_ma_007 · 疑似回复：@MaxPazow updating the video extension feature is on our roadmap! 原帖
08-01 00:47 @martin_ma_007 · 疑似回复：@fQQKJmwURyekDcf noted, thanks! 原帖
08-01 00:47 @martin_ma_007 · 疑似回复：@JayRetro85 glad that you are having fun with it! 原帖
08-01 00:48 @martin_ma_007 · 疑似回复：@Tim_Daust you got it, it's on our roadmap, stay tuned! 原帖
08-01 06:34 @ypwang61 · 转发：RT @Zaddyzaddy: The new deepseek-v4-flash is extremely cracked at vulnerability detection, this was a quick internal benchmark done on the… 原帖