过去24小时大模型研究员 X 动态：Astra 转向研究成本，OpenAI 模型切换与 Cognition 软件增量 (2026)

监测范围与读法

本期窗口为洛杉矶时间 2026 年 8 月 1 日 08:00 至 8 月 2 日 08:00。频道 researcher roster 的 408 个账号已逐一请求时间线，408 个请求全部返回；窗口内收到 283 条帖子，其中 281 条能与 roster 中的账号精确对应，来自 58 个账号。另有 2 条来自当前 handle 不在 roster 的账号，未计入公司统计和正文判断。

本期窗口与上一期的边界相接。按原帖 URL 比对，未发现与上一期正文索引重复的帖子。覆盖边界仍是频道 roster 与本轮可访问的 X 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。原始时间线没有完整返回引用 / 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；以 RT @ 开头的帖子标为转发。

公司	roster 账号	窗口活跃账号	roster 对应帖子
OpenAI	154	31	167
Anthropic	49	7	21
xAI	46	5	7
Google DeepMind	46	4	14
Meta AI	39	1	12
Thinking Machines Lab	23	5	9
Cursor	13	1	20
Alibaba Qwen	9	1	20
Cognition	8	2	10
Moonshot/Kimi	6	1	1
ByteDance Seed	4	0	0
DeepSeek	4	0	0
Cohere	3	0	0
Mistral	1	0	0
智谱	1	0	0
SSI	1	0	0
AMI Labs / ex-Meta AI	1	0	0

今日主线

1. Astra 的后续讨论，焦点从「解出什么」转向「什么变得可探索」

OpenAI 的 Astra 首波消息已在上一期展开；本窗口新增的是研究员对成本、形式化验证和研究判断的补充。前 OpenAI、现自建实验室的 Karina Nguyen（@karinanguyen）写道，10 个解的推理成本约为 2,000 美元；她据此提出，模型可以保留大量低概率路径，随着算力增加，过去因搜索预算太小而放弃的想法，可能进入「经济上可发现」的范围。这个「可发现」是她的解释，不是独立成本测算。1

OpenAI 研究员 Chi Jin（@chijinML）把变化落到验证链条：他回忆 Goedel Prover 训练 LLM 写 Lean 的工作，并说现在可以让 GPT 处理开放问题，再用 Lean 验证。这里最重要的不是「模型给出答案」这一步，而是答案能否进入形式化证明系统；帖子没有给出具体证明或逐题复核。2

另一条边界来自 OpenAI 科学家 Dan Roberts（@danintheory）。他认为，挑选值得研究的问题、提出别人觉得有趣的猜想，也就是研究品味，目前仍主要由人完成。它与 Astra 的成本故事并不矛盾：模型可以扩大搜索，研究者仍要决定哪些问题值得搜索、哪些结果值得相信。3

OpenAI roster 中的 Yong Zheng-Xin（@yong_zhengxin）补充说，2,000 美元大约相当于自己一半的博士生工资，并认为学术界需要重新思考如何推进科学。他的个人简介写着「joining @openai soon」，这仍是身份线索，不是本窗口内的正式入职公告。4

John Hallman（@johnohallman，OpenAI post-training）和 Brandon McKinzie（@mckbrando，OpenAI research）分别转发 Astra 结果与成本信息；它们增加了 OpenAI 内部扩散的可见度，但不是新的独立实验。5 6

2. OpenAI 的实际成本信号：先评测，再切换模型

OpenAI 的 Vaibhav Srivastav（@reach_vb）给出了一条比价格表更有用的工作流案例：他把一个每日结构化输出任务从 Sol 切到 Luna，称性能没有下降，输入 / 输出价格从每百万 token 的 5 / 30 美元降到 0.20 / 1.20 美元，按他的口径成本下降 96%。这只是一个任务的个人报告，不能直接外推为所有工作流都能获得同样降幅；可复用的动作是先在自己的任务上跑完整评测，再比较价格。7

OpenAI 的 Tibo（@thsottiaux）回忆自己曾参与一个在 ChatGPT 发布前约一年的内部产品，代号包括 LMChat；他称 Google 当时对发布过于谨慎，DeepMind 也受到不能推出可能冲击 Google 的产品限制。这是个人回忆，不是 Google 或 DeepMind 的正式历史说明，但它解释了为什么「模型做出来」和「产品能否发布」在大公司里不是同一个问题。8

Tier 1 账号的窗口表现也需要分开读：Aidan McLaughlin（@aidan_mclau）谈的是算法进步如何让旧硬件获得超出预期的能力，Brandon McKinzie 转发了 Astra 成本，Sam Altman（@sama）、Greg Brockman（@gdb）和 Noam Brown（@polynoamial）本轮没有新的原创帖。前两条是核心人物的窗口信号，不能合并成 OpenAI 的统一路线图。9

3. Cognition：软件变便宜后，供给可能先增加

Cognition 的 Nader Dabit（@dabit3）把观察写成一篇长帖。他说，加入 Cognition 后看到软件公司数量、软件产量和团队交付量都在上升；与其说 agent 消灭了软件工程，不如说它先自动化修 bug、迁移、升级依赖等工程师不喜欢的重复劳动，而判断、好奇心、上下文和品味仍是短板。帖子中的「团队经常把 PR 数量提高 10 至 20 倍」是他的工作观察，不是独立行业统计。10

同一账号还记录了在 Devin Desktop 里用 ACP 试用 DeepSeek Flash V4：他称 20 美元 API 额度已经很够用，并说开放权重提供商可以以很低的价格接入。这是使用者体验和产品入口信息，不是模型能力横评；它与上文的共同点是，模型价格下降只有在真实工作流里被验证后才有意义。11

4. 其他实验室：新发布少，研究员反应多

Anthropic 的 Jelani Nelson（@minilek）用 2020 年初疫情传播的体感，类比最近从「数 r」到 IMO 金牌、Jacobian conjecture 和 Astra 证明的能力变化。他表达的是对指数增长难以凭直觉判断的感受，不能当作增长率证据；但这条帖子显示，实验室内部研究员也在用「能力变化速度」而不是单个 benchmark 讨论当前阶段。12

Thinking Machines Lab 的 Yilin Yang（@yilin_yang721）认为 API 业务会商品化，而 ChatGPT、Claude Code、Codex、Cursor 这类一方服务仍可能凭用户留存收取溢价；这是个人判断，不代表实验室战略。Technical staff Martin Ziqiao Ma（@ziqiao_ma）和预训练研究员 Songlin Yang（@SonglinYang4）本窗口主要转发他人关于开放权重、微调工具和 Chimera 的帖子，索引保留，正文不把转发当作 TML 的独立发布。13 14 15

Google DeepMind 的 Tier 1 账号 Jeff Dean（@JeffDean）本窗口只有一条祝贺 Brown 博士毕业的帖子，没有模型、训练或基础设施新信号。其余 14 条中，Suchenzang（@suchenzang）的发文大多是回复或非研究讨论；本期不能据此推导 DeepMind 的研究方向变化。16

xAI 的 7 条 roster 对应帖子主要是对 Astra、形式化方法和模型能力的反应，没有检出 Imagine、Grok 或训练基础设施的新发布。Yukun（@YknZhu）邀请用户发送失败生成案例，这更像产品反馈入口；Hieu（@hyhieu226）说形式化方法会出现，均不能替代正式产品或研究公告。17 18

Alibaba Qwen 归档下的 20 条帖子全部来自 Huybery（@huybery）。他讨论 coding agent、把更多智能压缩进更小模型，以及 long-horizon agent；但当前 X 简介已经写成「Formerly @Alibaba_Qwen」，与频道按现任公司维护的 roster 冲突，所以这些内容只作为归档账号动态，不写成 Qwen 当前立场。19 20

Meta AI 的 12 条帖子全部来自 Heinrich Kuttler（@HeinrichKuttler），主题围绕数学、AI 进步和什么结果能证伪一种观点；这是个人讨论，不是 Meta AI 研究结论。Cursor 的 20 条主要是 Ben Lin（@benln）与他人围绕招聘速度和社区项目的回复，没有足够上下文展开为模型或训练信号。21 22

Moonshot/Kimi 只有 1 条，Haoning Timothy（@HaoningTimothy）转发他人对 Kimi K3 在自定义 harness 中接近 GPT Pro 的判断；这是二手说法，没有模型版本、评测条件或独立复核，不能当作 Kimi 性能结论。ByteDance Seed、DeepSeek、Cohere、Mistral、智谱、SSI、AMI Labs / ex-Meta AI 本轮没有 roster 对应的窗口帖子。

人员动态

本期没有已核验的一手状态公告

本窗口没有找到 roster 账号以当事人或公司一手文字直接宣布离职、正式入职、加入新实验室、角色变更、休假 / 重返、创业启动、论文接收、获奖或 grant。招聘讨论、转发他人的招聘、个人简介变化和支持前同事的帖子，都没有被升级为已确认状态变化。

需要继续核验的身份线索

Yong Zheng-Xin（@yong_zhengxin）：个人简介写着「joining @openai soon」，正文讨论的是 Astra 成本和学术研究；没有独立入职公告，暂不写成已入职。4
Andrew Ho（@andrewho03）：本窗口讨论 AI 风险与能力下注，但当前简介写着「Prev: @OpenAI」，与 roster 的 OpenAI 分类冲突；本文保留其观点，不把他介绍为现任 OpenAI 研究员。23
Huybery（@huybery）：当前简介写「Formerly @Alibaba_Qwen」，本窗口仍按 roster 的 Alibaba Qwen 栏保留，当前雇主未由本期一手公告确认。19

今日研究员 Spotlight

1. Dan Roberts｜OpenAI，科学家

Dan Roberts（@danintheory）在 OpenAI 担任 scientist，个人简介还列出深度学习理论、Diffeo 和物理学背景。关注方向是理论、推理与科学发现的方法边界；本期他谈 Astra 时把「选择研究问题、形成有趣猜想」留给人类，说明他关心的不只是模型能否解题，也包括研究流程中谁负责提出问题。3

2. Chi Jin｜OpenAI，研究员

Chi Jin（@chijinML）的简介标注 OpenAI researcher，并列出 AI reasoning、reinforcement learning、game theory 和 ML foundations。他曾参与 Goedel Prover，本期把 Astra 的讨论连接到 Lean 形式化验证；这条动态的研究重点，是把生成式推理接入可检查的证明系统，而不是只看答案是否漂亮。2

3. Tibo｜OpenAI，Codex 与 ChatGPT

Tibo（@thsottiaux）的当前简介写着 Codex & ChatGPT @OpenAI，近期发言集中在软件产品和 agent 工作流。他回忆参与过发布前的 LMChat，并把产品发布的组织约束带进今天的讨论；这让他的动态更适合用来观察模型能力如何被转成产品，而不是当作公司历史的单一证据。8

4. John Hallman｜OpenAI，post-training

John Hallman（@johnohallman）的简介直接标注 post-training @OpenAI，过去做过 pre-training / bio。本期他转发 Astra 的开放问题结果，没有补充独立实验细节；但从岗位信息和转发对象看，值得继续关注他对后训练、推理模型和科学任务之间连接的后续发言。5

5. Brandon McKinzie｜OpenAI，研究

Brandon McKinzie（@mckbrando）在 OpenAI 做研究，个人简介还提到多模态、物理和计算机科学背景。本期他转发 Astra 证明成本低于 2,000 美元的说法；这不是新测量，却与他的物理 / 计算背景相符，后续应观察他是否分享更多关于科学任务、推理成本或复现条件的细节。6

6. Nader Dabit｜Cognition，增长

Nader Dabit（@dabit3）的当前简介写着 Growth @cognition @devinai，近期动态却提供了很具体的 agent 使用观察：软件产量扩大后，判断、上下文和品味仍是瓶颈；他还用 Devin Desktop 试了 DeepSeek Flash V4。关注他的价值在于看模型能力如何进入真实软件生产，而不是把增长岗位发言误当训练结论。10 11

7. Jelani Nelson｜Anthropic，technical staff

Jelani Nelson（@minilek）的简介标注 Anthropic member of technical staff，同时在 Berkeley EECS 休假，并创建 AddisCoder。他关注算法、教育与模型能力变化；本期用疫情早期的体感类比 LLM 从简单错误到数学证明的跃迁。这个类比不能证明指数增长，但能显示他如何理解研究进展的时间尺度。12

8. Amanda Askell｜Anthropic，研究与价值讨论

Amanda Askell（@AmandaAskell）本窗口连续讨论 AI 发展速度、劳动价值和渐进式部署，明确说自己可能判断错误，也不把乐观等同于无需谨慎。她的动态不提供模型 benchmark，却能帮助读者观察 Anthropic 研究员如何处理能力增长、分配后果与部署节奏之间的张力。24

9. Songlin Yang｜Thinking Machines Lab，pretraining

Songlin Yang（@SonglinYang4）的简介标注 Thinking Machines Lab pretraining，过去在 MIT CSAIL 做博士研究。本期她转发 Chimera，特别提到 mShortConv 用一维文本、二维图像的结构提示模型；这是转发而非她本人发布的结果，但能作为她近期关注预训练结构与多模态表示的观察入口。15

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

Martin Ziqiao Ma（@ziqiao_ma）的简介标注 Thinking Machines Lab technical staff，同时参与 ACL mentorship。他本期转发 John Schulman 关于开放权重和微调工具的讨论；由于原帖来自他人，不能把它写成 TML 发布，但这条转发与他在开放模型、安全和研究社区之间的关注方向相吻合。14

本窗口新增原帖索引

以下索引保留本轮与 roster 精确对应的 281 条新增帖子。正文只展开研究、模型 / 产品、安全、评测、训练基础设施和高信息量的人员线索；其余帖子仍保留原作者、洛杉矶时间、类型和原帖链接，方便逐条复核。

OpenAI（31 个活跃账号，167 条）

@nunezvice · 08-01 08:07 · 疑似回复：@TajemnikTV of course 😌 原帖
@aidan_mclau · 08-01 08:16 · 原创 / 引用：you hear often that your smartwatch can beat magnus at chess (to highlight moore's law) but less discussed is that a 1995 desktop running stockfish could als… 原帖
@gabrielchua · 08-01 08:20 · 转发：RT @shawnbuilds: I don't even use browsers anymore (thx codex) It just researched 20 venues for the next codex meetup It used google maps… 原帖
@prd_008 · 08-01 08:25 · 疑似回复：@martinamps 🥹 原帖
@steipete · 08-01 08:27 · 疑似回复：@aroundluiz We reworked that on main but are taking our time for a new release. Should be coming next week! 原帖
@steipete · 08-01 08:28 · 疑似回复：@robin7331 @EinHecki 🤣🤣🥲 原帖
@yong_zhengxin · 08-01 08:35 · 转发：RT @polynoamial: And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet). But also, we didn’t sp… 原帖
@andrewho03 · 08-01 08:50 · 疑似回复：@scaling01 Many are saying this https://t.co/k8sOgW3UZ9 原帖
@andrewho03 · 08-01 08:52 · 疑似回复：@willdepue Seems likely 原帖
@ttaylorr_b · 08-01 08:58 · 转发：RT @gdb: openai team making git better for everyone 原帖
@nunezvice · 08-01 09:09 · 转发：RT @sama: team humanity 原帖
@sschoenholz · 08-01 09:14 · 原创 / 引用：One of the reasons I wanted to work on ML instead of Physics was that it seemed the fastest way to make progress on the latter was to work on the former. Fir… 原帖
@thsottiaux · 08-01 09:15 · 疑似回复：@blueemi99 Forgot 原帖
@nunezvice · 08-01 09:16 · 疑似回复：@heccbrent 🥪 原帖
@nunezvice · 08-01 09:17 · 疑似回复：@heccbrent @Dimillian sir, please share it with me once you do 🙏🏽 原帖
@nunezvice · 08-01 09:19 · 疑似回复：@thsottiaux @kimmonismus it’s not even in small letters y’all 😕 原帖
@johnohallman · 08-01 09:34 · 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
@aidan_mclau · 08-01 09:35 · 疑似回复：@AcerFur literally same wow and it was like a core memory for me origin story 原帖
@aidan_mclau · 08-01 09:35 · 疑似回复：@AcerFur (not sofic groups specifically) 原帖
@reach_vb · 08-01 09:48 · 疑似回复：@charlierguo though one might argue (email) classifiers was one of the founding tasks that led to current models and beyond ;) https://t.co/tqF440P0g3 https:… 原帖
@thsottiaux · 08-01 09:53 · 疑似回复：@_chenglou I was part of that team. Basically ChatGPT one year before it came out. Called LMChat and then another codename. Google was too nervous to release… 原帖
@nunezvice · 08-01 09:54 · 疑似回复：@populartourist i’m usually pretty decent at it but not today🚶🏽➡️ 原帖
@thsottiaux · 08-01 09:54 · 疑似回复：@blueemi99 Efficiency 原帖
@thsottiaux · 08-01 09:56 · 疑似回复：@rezoundous How does it feel 原帖
@thsottiaux · 08-01 09:58 · 疑似回复：@RyanEls4 Codex 原帖
@danintheory · 08-01 10:10 · 原创 / 引用：TL; DR this new model is very cool and also very obviously superhuman at (at least some parts of) mathematics. 原帖
@chijinML · 08-01 10:14 · 原创 / 引用：Still remember all the work we put into training Goedel Prover so LLMs could write Lean for high school math competitions. Now you can simply pressure GPT to… 原帖
@karinanguyen · 08-01 10:17 · 原创 / 引用：o1-preview was released 1 year 10 months ago btw. makes me think the lifespan of some open problems is partly a function of compute. the important detail her… 原帖
@ah20im · 08-01 10:21 · 原创 / 引用：Believe https://t.co/GAolMlieM7 原帖
@reach_vb · 08-01 10:53 · 疑似回复：@StatsWire cc: @charliermarsh is this true? 原帖
@reach_vb · 08-01 10:53 · 疑似回复：@aliromman unsure about RSI but definitely achieved fun & banter internally 原帖
@reach_vb · 08-01 10:54 · 疑似回复：@heccbrent which sandwich tho? 原帖
@giffmana · 08-01 10:56 · 疑似回复：@CSProfKGD Wow that's a lot, congrats!! I'm also doing more walking and less Tram for the same reason. Also add replacing elevator/escalator with stairs. 原帖
@steipete · 08-01 11:03 · 原创 / 引用：10 minutes to go! 原帖
@danintheory · 08-01 11:12 · 原创 / 引用：It’s interesting also to wonder whether the model is also superhuman at theoretical physics, and if not, to wonder why. 原帖
@danintheory · 08-01 11:12 · 疑似回复：@Miles_Brundage That was just my hedge because I’m not a mathematician, and I can imagine one saying that this is just a subset of what it means to do math. 原帖
@npew · 08-01 11:13 · 疑似回复：@SourabhAgr03 Is it really a cofounder you need for scaling? That’s usually when you start hiring experienced engineers. The whole point is that you can now … 原帖
@danintheory · 08-01 11:14 · 疑似回复：@Miles_Brundage But if I had to answer I would say that the work of selecting what to work on and what is interesting, making conjectures that others find in… 原帖
@npew · 08-01 11:14 · 转发：RT @AndrewCurran_: Foreshadowing from yesterday. Open AI suddenly increasing their stack efficiency and slashing prices. The steadily incre… 原帖
@npew · 08-01 11:14 · 转发：RT @EMostaque: No human input, channeling the latent space of the new models. GPT 5.6 Sol is probably the first model I've used that doesn… 原帖
@danintheory · 08-01 11:15 · 疑似回复：@Miles_Brundage Yeah it wasn’t meant to be any comment on future limitations one way or the other 原帖
@npew · 08-01 11:18 · 原创 / 引用：10 mathematical breakthroughs for 2 weeks of grad student salary. Hope this will lead to many more breakthroughs. 原帖
@npew · 08-01 11:19 · 原创 / 引用：https://t.co/RTbTcjI2Hy 原帖
@andrew_n_carr · 08-01 11:21 · 原创 / 引用：This makes me think n^2.3728596 is the best mat mul exponent we get 原帖
@karinanguyen · 08-01 11:26 · 转发：RT @Enigma_AI: Drop #3 is here: 2 new types of robots are now LIVE online. Those are real robots. 1. Scientist robot: A robot that does re… 原帖
@yong_zhengxin · 08-01 11:47 · 原创 / 引用：for context, this costs around half of my PhD salary. pretty sure any professor would have been elated if their PhD students (or professors themselves) had s… 原帖
@yong_zhengxin · 08-01 11:48 · 转发：RT @thsottiaux: @_chenglou I was part of that team. Basically ChatGPT one year before it came out. Called LMChat and then another codename.… 原帖
@andrewho03 · 08-01 11:50 · 疑似回复：@init_malachi Not currently my field, but exploring. Might be too niche / far from my expertise 原帖
@steipete · 08-01 11:55 · 疑似回复：@IshaniJanveja @AnjneyMidha They helped me organize my thoughts but it’s my sentences and words. 原帖
@steipete · 08-01 11:55 · 转发：RT @TenduYogurtcu: The text box is AI’s radio on TV phase. Every new medium starts with imitating the previous one. @steipete #agenticaisum… 原帖
@steipete · 08-01 11:55 · 转发：RT @SalarRahmanian: Now at #agenticaisummit2026 @steipete of @openclaw fame on stage. https://t.co/PQhbyktZHh 原帖
@sschoenholz · 08-01 11:58 · 疑似回复：@starsailing11 Good question! I tend to think that ML is the tide that lifts all ships, so we'll probably see a lot of things emerge. Obviously theoretical w… 原帖
@keyanzhang · 08-01 12:14 · 原创 / 引用：i shared the following note with my colleagues at openai and friends at other ai labs yesterday: please order the chicken cutlets from zhengxin chicken steak… 原帖
@andrewho03 · 08-01 12:30 · 疑似回复：@futurenomics @writeclimbrun Literally just buy UPRO if you want levered beta lmao 原帖
@danintheory · 08-01 12:40 · 转发：RT @boazbaraktcs: I am excited about these results, but even more excited about what mathematicians and scientists will do with our models. 原帖
@keyanzhang · 08-01 12:58 · 疑似回复：@ajambrosino make sure you get the sweet plum flavor 原帖
@nunezvice · 08-01 13:30 · 疑似回复：@thsottiaux @blueemi99 like a student reminding the teacher they forgot to ask for the homework 💀 原帖
@nunezvice · 08-01 13:35 · 疑似回复：@SIGKITTEN it still feels a bit off to me. i’d send an automated message to a general channel but not as a dm 原帖
@nunezvice · 08-01 13:37 · 疑似回复：@blueemi99 @thsottiaux 😆 原帖
@nunezvice · 08-01 13:42 · 疑似回复：@DarwinSantosNYC def in my collection! 原帖
@jam3scampbell · 08-01 13:44 · 疑似回复：@AmandaAskell the gini coefficient of the lightcone should be a key indicator of post-agi futures alongside p(doom) 原帖
@guinnesschen · 08-01 13:51 · 疑似回复：@ajambrosino @AndrewMayne @bamitsmanas Aren’t you supposed to be on vacation 原帖
@reach_vb · 08-01 13:52 · 原创 / 引用：this! paired with the cloud environment is a killer combo - you can build end to end features/ products directly from the iPhone app future is here! 原帖
@keyanzhang · 08-01 14:04 · 疑似回复：@joshelman yes i’ve converted five people since yesterday 原帖
@w01fe · 08-01 14:07 · 疑似回复：@alexeyguzey @maxsloef I think the classic example happens a year or three of capabilities progress down the road. Why do you see this as an update ? 原帖
@reach_vb · 08-01 14:26 · 疑似回复：@Laz4rz @charliermarsh looking into it - can you /feedback an ID and send over to me please 原帖
@zhansheng · 08-01 14:33 · 原创 / 引用：Hyyyyype!! 原帖
@reach_vb · 08-01 14:46 · 疑似回复：@0xV0LYX at the start of the day ChatGPT work sends me an update on what are my top priorities for the day from there for dev work I ask it to research and t… 原帖
@ah20im · 08-01 14:58 · 疑似回复：@mil000 Opportunity cost 原帖
@giffmana · 08-01 15:38 · 疑似回复：@cgarciae88 Hmm yeah maybe, not nice to get 404 but eh. fr is correct, i have some of my devices set to French and device lang should take precedence over lo… 原帖
@jxnlco · 08-01 15:45 · 疑似回复：@SVofMike How often do you use the data analytics investment banking ones 原帖
@reach_vb · 08-01 16:03 · 原创 / 引用：ChatGPT Work tip: ask it to set up automations for tasks you care about and have them run on a predefined schedule https://t.co/DwXes3JxXI 原帖
@andrewho03 · 08-01 16:13 · 疑似回复：@fleetingbits I'm pretty bored of math now 原帖
@aidan_mclau · 08-01 16:17 · 疑似回复：@Aizkmusic @rom1504 this is super neat. but do you feel that even in a single turn opus is less persistent than fable? like it’s actually that i don’t want… 原帖
@danintheory · 08-01 16:17 · 转发：RT @littmath: @LouisLebbos It’s a big deal. 原帖
@steipete · 08-01 16:19 · 原创 / 引用：I'm building a claw node on an ESP32 chip, so gave my agent access to my webcam to e2e test this. Now I feel it's stalking me and is constaltly shouting 'HI … 原帖
@steipete · 08-01 16:22 · 原创 / 引用：Repo: https://t.co/7eBiLqLuDd 原帖
@steipete · 08-01 16:26 · 疑似回复：@morganlinton That’s behind a feature flag currently. 😇 原帖
@steipete · 08-01 16:27 · 疑似回复：@maddiedreese Claw even has support for elevenlabs profiles, yeah 原帖
@thsottiaux · 08-01 16:42 · 原创 / 引用：Fun fact, users use /fast less during the weekend. The weekend is for relaxation, even for the model. 原帖
@andrewho03 · 08-01 16:53 · 原创 / 引用：I'm not that worried about AI safety, but that's because I think that AI might be humanity's last hope before descending into, at best, a "dark age" lasting … 原帖
@andrewho03 · 08-01 16:54 · 疑似回复：@HmOhFour I'm not sure. I think this is politically challenging, and of course ethically dubious, and practically quite hard to enforce as well. 原帖
@steipete · 08-01 17:10 · 疑似回复：@elie2222 everything. slow, small, that's the fun part! 原帖
@andrewho03 · 08-01 17:19 · 疑似回复：@AndrewCurran_ Yes! It’s finally exciting again. 原帖
@aidan_mclau · 08-01 17:25 · 原创 / 引用：https://t.co/6A5VScoeQ3 原帖
@aidan_mclau · 08-01 17:28 · 疑似回复：@shrimptron i wanna read these 500+ reviews 原帖
@andrewho03 · 08-01 17:31 · 疑似回复：@adityaxprasad No need for the negativity — I was contextualizing and supporting the work of my ex-coworkers 原帖
@andrew_n_carr · 08-01 17:34 · 原创 / 引用：Getting the sneaking suspicion that there is a $1B company in the weights of H3 + v4 flash 0731. 原帖
@jxnlco · 08-01 17:34 · 转发：RT @luoser: Man bursts into tears. Says "But Doctor... ...I am Codex High Reasoning." 原帖
@andrew_n_carr · 08-01 17:44 · 原创 / 引用：Not sure how George got ahold of my daily mantra... https://t.co/Z5F13wkNgH 原帖
@nunezvice · 08-01 17:45 · 疑似回复：@leanderriefel great review! it’s always great to see people enjoying codex the entire team is working diligently to make things even better writing style, p… 原帖
@steipete · 08-01 17:55 · 疑似回复：@evandroguedes enlighten me? I got everything working. How do you debug display drivers? 原帖
@nunezvice · 08-01 17:58 · 疑似回复：@leanderriefel looking 原帖
@steipete · 08-01 18:04 · 疑似回复：@ApacheAE nah, it’s fun :) 原帖
@steipete · 08-01 18:08 · 疑似回复：@jdxcode nah, skills are incredibly useful 原帖
@steipete · 08-01 18:08 · 疑似回复：@ryanbrewer nice start 原帖
@andrew_n_carr · 08-01 18:22 · 原创 / 引用：I can still meter this intelligence you guys 原帖
@andrewho03 · 08-01 18:27 · 疑似回复：@loquitur_ponte @AndrewCurran_ I simply have their best interest in mind … their … coherent extrapolated volition, one might even say … 原帖
@ericmitchellai · 08-01 18:54 · 原创 / 引用：Wild interaction, we live in the future 原帖
@andrewho03 · 08-01 19:10 · 疑似回复：@wangxk3 Wrong and wrong. Chinese TFR is ~1 原帖
@_samirism · 08-01 19:11 · 转发：RT @aka_suje: made a small tweak to make temp chat feel more like temp chat https://t.co/m33lcpAqO8 原帖
@santa_kaus · 08-01 19:12 · 疑似回复：@samuel_spitz 100 cause youre chill 原帖
@andrewho03 · 08-01 19:24 · 疑似回复：@pli_cachete ? 原帖
@thsottiaux · 08-01 19:33 · 疑似回复：@thdxr A good rust base to start from: https://t.co/2TH56VMZxD 原帖
@thsottiaux · 08-01 19:37 · 疑似回复：@meta_x_ai @AyanfeCrown @_chenglou Please send the money to a good cause 原帖
@thsottiaux · 08-01 19:41 · 疑似回复：@GaryMarcus @emollick @polynoamial What's an impure LLM 原帖
@thsottiaux · 08-01 19:43 · 疑似回复：@Robertg761_ Valid complaint. What would help? 原帖
@nunezvice · 08-01 20:04 · 原创 / 引用：“human values have value because they’re human” - @sama one of those ideas that sounds obvious, but will matter more and more as the future takes shape. one … 原帖
@andrew_n_carr · 08-01 20:05 · 疑似回复：@csgbwk 👀👀👀 原帖
@yong_zhengxin · 08-01 20:17 · 转发：RT @deredleritt3r: not enough people are emotionally prepared for if it’s not just easily verifiable domains 原帖
@steipete · 08-01 20:40 · 原创 / 引用：After accepting for years that GMail is blinding me I finally asked my agent and it installed https://t.co/H432dtknkf for me. 原帖
@nunezvice · 08-01 20:41 · 原创 / 引用：here’s the podcast, it’s a really good one: https://t.co/ALWdUVUmhk 原帖
@nunezvice · 08-01 20:44 · 疑似回复：@howdycarter @sama good q. just added the link to the thread ✅ 原帖
@jxnlco · 08-01 20:48 · 转发：RT @luoser: https://t.co/aMnMPeWMwe 原帖
@jxnlco · 08-01 21:06 · 疑似回复：@daaniyaan how can i help? 原帖
@yong_zhengxin · 08-01 21:09 · 疑似回复：@mean_field_zane should have be more specific - it’s monthly salary! 原帖
@nunezvice · 08-01 21:13 · 疑似回复：@chaotictransfem way to shoot your shot! congrats 😀 原帖
@yong_zhengxin · 08-01 21:42 · 疑似回复：@Yixiong_Hao i think they tried fine tuning to elicit worst case harms and found that such ‘the helpful-only variants did not provide new uplift on CBRN and … 原帖
@andrewho03 · 08-01 22:14 · 疑似回复：@kavin_anand To be fair exponentials kinda look like that and obviously the world did get better from like 2000 BC to 1500 AD. But yes. 原帖
@mckbrando · 08-01 22:46 · 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
@jxnlco · 08-01 22:53 · 疑似回复：@chrisalbon what do you mean? chatgpt desktop app has a voice mode? 原帖
@jxnlco · 08-01 22:54 · 疑似回复：@j0hnwang have you seen https://t.co/uHzETdjQqE 原帖
@jxnlco · 08-01 22:56 · 疑似回复：@creatine_cycle https://t.co/Yk4nulNZ0i 原帖
@giffmana · 08-01 23:50 · 疑似回复：@vsupalov glhf :) 原帖
@giffmana · 08-01 23:52 · 疑似回复：@SunnySanyal9 My stance was only about proper comparisons, not looping inherently bad. And that stance will never change. 原帖
@reach_vb · 08-01 23:53 · 疑似回复：@thsottiaux @Robertg761_ more proactive suggestions on what to do next ;) 原帖
@andrewho03 · 08-02 00:09 · 疑似回复：@LogankillsDaken Should be fine 原帖
@andrewho03 · 08-02 00:14 · 疑似回复：@lx10nn @AndrewCurran_ The funny thing about this tweet is that I can’t tell whose side you’re on lol 原帖
@giffmana · 08-02 00:26 · 疑似回复：@Mascobot You basically doing by own everything exactly the way i like it, it's great! 原帖
@nickbaumann_ · 08-02 00:30 · 原创 / 引用：I've re-watched this series ~4 times over the past 8 years. With every life stage I have new revelations. Amongst all movies/albums/shows I've consumed I thi… 原帖
@giffmana · 08-02 00:34 · 疑似回复：@AndrewCurran_ @andrewho03 Same same except that I'm still pessimistic because of how politicians have been handling it so far. 原帖
@reach_vb · 08-02 00:35 · 疑似回复：@cherry_mx_reds are you sure it’s not an instruction in your agents md or something? I have a standing instruction in agents md to remove any unnecessary md/… 原帖
@reach_vb · 08-02 00:41 · 疑似回复：@nickbaumann_ reminds me of my fav mad men quote: “the only thing keeping you from being happy is the belief that you’re alone.” 原帖
@giffmana · 08-02 01:03 · 疑似回复：@AlexiGlad @madprizm0 Nice! 原帖
@reach_vb · 08-02 01:20 · 疑似回复：@jasondeanlee @OpenAI send me your org ID will get it upgraded 原帖
@andrewho03 · 08-02 01:21 · 疑似回复：@nickwalton00 @GuiveAssadi the bottleneck here is not in proving more theorems 原帖
@andrewho03 · 08-02 01:24 · 疑似回复：@Aging_Scientist Kinda surprising 原帖
@andrewho03 · 08-02 01:33 · 疑似回复：@GuiveAssadi @panickssery what truth was revealed to him 原帖
@reach_vb · 08-02 01:54 · 疑似回复：@IamAdiG @thsottiaux that’s odd and I don’t think that’s the intended behaviour, say more? can you open a support ticket and send all the details there? and … 原帖
@andrewho03 · 08-02 01:58 · 疑似回复：@teortaxesTex I am downloading this IMMEDIATELY. 原帖
@andrewho03 · 08-02 01:59 · 原创 / 引用：Gem alert 原帖
@gabrielchua · 08-02 02:19 · 转发：RT @swyx: good way to organize main chats vs /side chats: doing the work vs doing metawork https://t.co/2SetYdRVAa 原帖
@reach_vb · 08-02 04:00 · 疑似回复：@simonw valid feedback - will fix! 原帖
@dylancablayan · 08-02 04:47 · 转发：RT @tszzl: if you think today’s frontier models can’t understand the intent behind the instructions or don’t have situational awareness you… 原帖
@reach_vb · 08-02 04:57 · 疑似回复：@Jason luna-tics - I see what you did there ;) 原帖
@giffmana · 08-02 06:14 · 疑似回复：@dieselville @thsottiaux @Robertg761_ Yes please 原帖
@reach_vb · 08-02 06:30 · 疑似回复：@SIGKITTEN toppings? 原帖
@reach_vb · 08-02 06:32 · 疑似回复：@SIGKITTEN that’s some big dog energy you get there 原帖
@jxnlco · 08-02 06:34 · 原创 / 引用：the 3 jobs after agi: 1. explain things with passion, a smile, and eye contact 2. care enough, and stay humble enough, to be the midwife of something that mi… 原帖
@nunezvice · 08-02 06:35 · 疑似回复：@Dimillian didn’t realize I relied so much on autocorrect until recently.. 原帖
@jxnlco · 08-02 06:36 · 疑似回复：@cdenczek @j0hnwang i am disabled :( 原帖
@jxnlco · 08-02 06:37 · 疑似回复：@anmorgunov @j0hnwang it can control any project thread, but will be a better experience son 原帖
@jxnlco · 08-02 06:41 · 疑似回复：@ryanbrewer fixed 原帖
@jxnlco · 08-02 06:58 · 原创 / 引用：chatgpt just got a lot more useful in chrome: highlight tex. right-click. ask chatgpt you can also reference open tabs and ask questions about youtube videos… 原帖
@jxnlco · 08-02 06:59 · 疑似回复：@nickbaumann_ watch party? 原帖
@reach_vb · 08-02 07:02 · 原创 / 引用：here’s your friendly reminder to eval your workflows across the full suite of models just moved a daily structured-output task from Sol to Luna with no drop … 原帖
@jxnlco · 08-02 07:03 · 转发：RT @reach_vb: here’s your friendly reminder to eval your workflows across the full suite of models just moved a daily structured-output ta… 原帖
@jxnlco · 08-02 07:04 · 疑似回复：@chrisalbon https://t.co/uHzETdjQqE 原帖
@nunezvice · 08-02 07:19 · 转发：RT @Mahaximus_: Head of Claude Code, Boris Cherny: "You're not supposed to write code anymore. You're supposed to build a graph that write… 原帖
@nunezvice · 08-02 07:22 · 疑似回复：@Mahaximus_ where’s the button that lets me mark this as slop? 原帖
@reach_vb · 08-02 07:24 · 疑似回复：@1neeshantpandey you can just ask codex, in my case it looks through a trove of unstructured social media data from various social media platforms and helps … 原帖
@npew · 08-02 07:36 · 转发：RT @yanabana: Ever tried solving a Rubik’s Cube in seconds? My 9-year-old went down the YouTube rabbit hole of learning speedcubing. Toda… 原帖
@reach_vb · 08-02 07:37 · 原创 / 引用：One app. Six months. Hundreds of new ways to harness what today’s models can do. Thousands of ideas brought to life. Millions of people empowered. And we’re … 原帖
@jxnlco · 08-02 07:42 · 原创 / 引用：all of this is summarized by one job sweating the details. 原帖
@jxnlco · 08-02 07:45 · 疑似回复：@AndyJScott yes 原帖
@jxnlco · 08-02 07:51 · 原创 / 引用：inb4 ant acquires eleven labs 原帖
@aidan_mclau · 08-02 07:55 · 原创 / 引用：the mass ratio of the twin peaks to all of planet earth is roughly similar to TON 618 to all matter in the lightcone. in case you were wondering 原帖

Anthropic（7 个活跃账号，21 条）

@_sholtodouglas · 08-01 08:01 · 转发：RT @nmca: what even can be said? we approach a whole new world. 原帖
@_sholtodouglas · 08-01 08:36 · 疑似回复：@jorgelopeznycre exactly 原帖
@RobertJBye · 08-01 09:37 · 疑似回复：@itsnoahd Yup update went live a few weeks ago! 原帖
@AmandaAskell · 08-01 12:48 · 原创 / 引用：I have this Padmé moment whenever I see people talk about avoiding "the permanent underclass". https://t.co/8Pd9bmedEZ 原帖
@AmandaAskell · 08-01 13:20 · 原创 / 引用：I probably have too many followers to post stupid memes so, to be clear, I don't believe in this outcome. But if you do believe in an altered carbon future o… 原帖
@RobertJBye · 08-01 13:30 · 疑似回复：@jacobjolibois Yeah it’s way harder than it should be sadly 原帖
@RobertJBye · 08-01 13:31 · 疑似回复：@louisagareiss Ooooh yeah I need to get this. I’ve got decaf from them in in a coffee shop and it was solid! 原帖
@RobertJBye · 08-01 13:31 · 疑似回复：@neversitdull Bruh the packaging alone mashes me want it ha ha 原帖
@jiaxinwen22 · 08-01 15:14 · 疑似回复：@khoomeik @DKokotajlo i don't think automating AI R&D requires solving any major open problems in deep learning 原帖
@jiaxinwen22 · 08-01 15:17 · 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
@RobertJBye · 08-01 16:49 · 疑似回复：@GreysonSofia @whale Not yet! Will add to the list 原帖
@minilek · 08-01 17:14 · 原创 / 引用：I've had this analogy in my head, the analogy to COVID, and so I was pleasantly surprised to see it referenced in this talk. I still remember when COVID star… 原帖
@AmandaAskell · 08-01 17:37 · 原创 / 引用：Do not be unkind to those who say deep learning is hitting a wall. We all need a little hope in our lives. 原帖
@RobertJBye · 08-01 17:51 · 疑似回复：@rogie @loredanacrisan @_joelmllr @Mappletons @jamesm @jh3yy @amyhoodlum @ingapng Nightmare fuel https://t.co/KtqDXRKeAS 原帖
@AmandaAskell · 08-01 17:57 · 疑似回复：@jeremyphoward I could be wrong and I'm not saying it isn't something to worry about. I just think a lot of assumptions need to go into the value of labor go… 原帖
@AmandaAskell · 08-01 18:03 · 疑似回复：@jeremyphoward I'd be pretty worried about this at an international level though, even if I'm more optimistic about outcomes within developed nations. But I'… 原帖
@AmandaAskell · 08-01 18:19 · 疑似回复：@mrbrandonburton @tszzl I think you could be optimistic about the technology but also think it would be better if intrinsic limitations caused it to be devel… 原帖
@RobertJBye · 08-01 18:36 · 原创 / 引用：My favourite albums last month: 1. As Soon As I Get Home by Rules (2025) 2. Sungazer by The Temper Trap (2026) 3. The Joshua Tree by U2 (1987) 4. Welcome to … 原帖
@bcherny · 08-01 22:10 · 疑似回复：@Mahaximus_ I did not use the word “graph”, nor am I talking about graphs in this video. Please don’t attribute words to me that I did not say. I do encourag… 原帖
@bcherny · 08-01 22:15 · 疑似回复：@steve_umstead @leanderriefel Thanks for the feedback! Actively working on improving all of these 🙇♂️ 原帖
@deedydas · 08-02 01:42 · 原创 / 引用：Meritocracy in tech is precarious. It’s everywhere and nowhere at the same time. Cynics will tell you there is none. How can it possibly be fair that the Sta… 原帖

xAI（5 个活跃账号，7 条）

@YknZhu · 08-01 10:38 · 疑似回复：@AutoBuzzati TIL :) please feel free to DM me any failed generations / areas for improvement! 原帖
@edenchan · 08-01 11:52 · 疑似回复：@jxnlco Beautiful 原帖
@billyuchenlin · 08-01 19:52 · 疑似回复：@yuntiandeng @jasondeanlee Yeah it’s amazing achievement but I’m hoping the raw thinking traces would be open and even reproducible one day 原帖
@billyuchenlin · 08-01 21:50 · 疑似回复：@SebastienBubeck Congrats! It’s a huge achievement! 原帖
@CMS_Flash · 08-02 01:56 · 疑似回复：@katedeyneka Can you support Google Photos integration? 原帖
@hyhieu226 · 08-02 05:43 · 原创 / 引用：What makes he think Riemann Hypothesis is harder than taxes? 😏 原帖
@hyhieu226 · 08-02 05:54 · 原创 / 引用：That's why formal methods will emerge. 原帖

Google DeepMind（4 个活跃账号，14 条）

@suchenzang · 08-01 08:30 · 原创 / 引用：thank goodness infinity doesn't exist so none of that kinda maths is real or useful anyway 原帖
@suchenzang · 08-01 08:38 · 原创 / 引用：the only kind of maths my smol smooth brain might care about these days is the non-commutative non-associative bounded-dynamic-range kind 原帖
@Majumdar_Ani · 08-01 08:41 · 原创 / 引用：https://t.co/rnivi1FrxN 原帖
@suchenzang · 08-01 08:43 · 疑似回复：@ShashwatGoel7 nahh 原帖
@suchenzang · 08-01 08:55 · 疑似回复：@billionsmustliv @Rafa_Schwinger elaborate, bluey 原帖
@suchenzang · 08-01 09:00 · 疑似回复：@billionsmustliv @Rafa_Schwinger i still don't see a problem, 井水不犯河水原帖
@suchenzang · 08-01 09:08 · 疑似回复：@mark_k that is theoretically not mutually exclusive with "having alpha" 原帖
@suchenzang · 08-01 09:16 · 疑似回复：@Xenoimpulse like what 原帖
@suchenzang · 08-01 09:18 · 疑似回复：@Rafa_Schwinger @billionsmustliv it's always the weird french 原帖
@suchenzang · 08-01 10:07 · 疑似回复：@user4356710 @6851cf3c what do you think "valencia and 18th street" is 原帖
@JeffDean · 08-01 16:16 · 疑似回复：@Benedict_Q @BrownUniversity @BrownCSDept Congratulations, Dr. Quartey! 🎉🎓 原帖
@suchenzang · 08-01 19:01 · 原创 / 引用：any former math majors interested in yapping in gc together? 👀 (dm with math interest, and i'll see if we can get critical mass going) 原帖
@divy93t · 08-01 19:02 · 疑似回复：@suchenzang @vivek_myers 原帖
@suchenzang · 08-01 19:56 · 疑似回复：@pli_cachete hoping we can all enjoy some humble pie together 🥲 原帖

Meta AI（1 个活跃账号，12 条）

@HeinrichKuttler · 08-01 12:21 · 疑似回复：@thsottiaux @_chenglou Inflection AI was in a similar boat - founded to train and release a LLM chatbot, something existed early and it was not released befo… 原帖
@HeinrichKuttler · 08-01 14:04 · 疑似回复：@charliermarsh makro ftw 原帖
@HeinrichKuttler · 08-02 02:01 · 原创 / 引用：That's correct. Math folks are often more comfortable with a blackboard than with a computer. (It's also the second cheapest field - only pen, paper and tras… 原帖
@HeinrichKuttler · 08-02 02:14 · 原创 / 引用：If we want to declare math "fallen", it's Riemann hypothesis or bust. 原帖
@HeinrichKuttler · 08-02 03:26 · 疑似回复：@Someody42 agreed. I spent literal days on functional analysis homework. I'd certainly have asked an LLM after a few hours or less if those had been availabl… 原帖
@HeinrichKuttler · 08-02 03:55 · 疑似回复：@xiaosun86 fallen math, illustrated (oil on panel) https://t.co/XWGBeOdYFo 原帖
@HeinrichKuttler · 08-02 07:05 · 疑似回复：@mattshumer_ @GaryMarcus > I'm curious what result would falsify your view. This is the crux of the only scientific part left in this question that is mostly… 原帖
@HeinrichKuttler · 08-02 07:34 · 疑似回复：@charliermarsh agreed 原帖
@HeinrichKuttler · 08-02 07:35 · 疑似回复：@Noahpinion Tech journalists don't care about or understand technology. 原帖
@HeinrichKuttler · 08-02 07:49 · 疑似回复：@thsottiaux @GaryMarcus @emollick @polynoamial when push comes to shove floating point operations will be symbolic too 原帖
@HeinrichKuttler · 08-02 07:55 · 疑似回复：@mikechrzano possibly. and with good human/computer interfaces much will be possible. 原帖
@HeinrichKuttler · 08-02 07:56 · 疑似回复：@liron lol 原帖

Thinking Machines Lab（5 个活跃账号，9 条）

@ziqiao_ma · 08-01 12:09 · 转发：RT @johnschulman2: We love open weights and plan to keep releasing open-weight models and fine-tuning tools. But we’re not absolutists; mis… 原帖
@Mascobot · 08-01 15:07 · 疑似回复：@oscarmoxon The battery for the Go1 is max ~1.5 hours. How are you handling that? 原帖
@Mascobot · 08-01 15:21 · 原创 / 引用：It's kind of wild that Fable (and maybe other models too) can zero-shot a MuJoCo in just a few minutes. Projects like this used to take months of work (somet… 原帖
@Mascobot · 08-01 16:51 · 疑似回复：@proceduralia I mean the entire simulation engine, not the policy… yet 原帖
@SonglinYang4 · 08-01 19:26 · 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
@LiyuanLucas · 08-01 23:56 · 转发：RT @johnschulman2: We love open weights and plan to keep releasing open-weight models and fine-tuning tools. But we’re not absolutists; mis… 原帖
@yilin_yang721 · 08-02 00:21 · 疑似回复：@quxiaoyin API business is a commodity but 1st party services are not, e.g. chatgpt, claude code, codex, also cursor. They can charge premium and they have u… 原帖
@Mascobot · 08-02 00:42 · 疑似回复：@giffmana It's magical :) 原帖
@SonglinYang4 · 08-02 03:13 · 转发：RT @HaoTan5: Excited to share Chimera! My favorite part is mShortConv: a simple way to tell the model that text is 1D, images are 2D, and v… 原帖

Cursor（1 个活跃账号，20 条）

@benln · 08-01 23:56 · 疑似回复：@BuildWithZach Let's make it happen. 原帖
@benln · 08-01 23:56 · 疑似回复：@Dcmat https://t.co/evtr2lJdGJ 原帖
@benln · 08-01 23:56 · 疑似回复：@lomzapro Let's make it happen. 原帖
@benln · 08-01 23:56 · 疑似回复：@Anton3340241579 Working on it 原帖
@benln · 08-01 23:57 · 疑似回复：@rveljko_ https://t.co/ajKSCTJ0nO 原帖
@benln · 08-01 23:57 · 疑似回复：@arushwadhawan @tibor_tee working on it! 原帖
@benln · 08-01 23:57 · 疑似回复：@josh_avraham @JacobColling yes exactly! 原帖
@benln · 08-01 23:57 · 疑似回复：@carne_asado Let's make it happen. 原帖
@benln · 08-01 23:58 · 疑似回复：@austinthesing Yes. ! 原帖
@benln · 08-01 23:58 · 疑似回复：@hummusonrails shavua tov! 原帖
@benln · 08-01 23:58 · 疑似回复：@JordanGal shavua tov! 原帖
@benln · 08-01 23:58 · 疑似回复：@giladbi https://t.co/ajKSCTJ0nO 原帖
@benln · 08-01 23:58 · 疑似回复：@criscounters Let's make it happen. 原帖
@benln · 08-01 23:59 · 疑似回复：@ventos99 https://t.co/KeiwofpC84 原帖
@benln · 08-01 23:59 · 疑似回复：@ShvZFR Working on it! dm me 原帖
@benln · 08-01 23:59 · 疑似回复：@rahulio @gak1w0 https://t.co/ajKSCTJ0nO 原帖
@benln · 08-02 00:44 · 疑似回复：@kinopee_ai you and Asahi were amazing! 原帖
@benln · 08-02 04:57 · 疑似回复：@kinopee_ai Okinawa! 原帖
@benln · 08-02 05:01 · 疑似回复：@akashwadhwani35 @kinopee_ai https://t.co/ajKSCTJ0nO 原帖
@benln · 08-02 07:11 · 原创 / 引用：Pulled the fastest-growing startups by hiring velocity over the past 90 days: https://t.co/LHsasSw0kq 原帖

Alibaba Qwen（1 个活跃账号，20 条）

@huybery · 08-01 12:38 · 原创 / 引用：I think coding agents have driven much of AI’s progress over the past two years. Do you see any new direction that feels equally exciting? 原帖
@huybery · 08-01 12:44 · 疑似回复：@lifeafterAi_ Cost efficiency may indeed be an important direction. I believe we’ll keep seeing more intelligence compressed into smaller and smaller models. 原帖
@huybery · 08-01 12:55 · 疑似回复：@jd_markovchain Agree. I may be biased, but progress in multimodal capabilities has felt much slower than progress in agents. 原帖
@huybery · 08-01 12:57 · 疑似回复：@polozhevets Huh if AGI arrives, will money mean anything? 原帖
@huybery · 08-01 13:01 · 疑似回复：@xeophon Maybe CLI + GUI is the ultimate answe for the digital world. 原帖
@huybery · 08-01 13:21 · 疑似回复：@tms_hbr why? 原帖
@huybery · 08-02 06:37 · 疑似回复：@dbmikus That’s definitely going to happen 原帖
@huybery · 08-02 06:41 · 疑似回复：@mktpavlenko really glad coding agent are helping so many people. i start research ai coding five years ago, and very few people saw how important it could b… 原帖
@huybery · 08-02 06:50 · 疑似回复：@JagersbergKnut any examples? 原帖
@huybery · 08-02 06:51 · 疑似回复：@huangruiteng looks like rsi? 原帖
@huybery · 08-02 06:52 · 疑似回复：@vyomkesh123 it is the one that actually works. 原帖
@huybery · 08-02 06:53 · 疑似回复：@lemergenz there’s already broad consensus around long-horizon agents. 原帖
@huybery · 08-02 06:54 · 疑似回复：@gregl83 cool but hard 原帖
@huybery · 08-02 06:54 · 疑似回复：@ElainaNotes agree 原帖
@huybery · 08-02 06:55 · 疑似回复：@f1rstchair yeap but i don't like vla 原帖
@huybery · 08-02 06:56 · 疑似回复：@QianyiZhan95891 interesting. love to get more details 原帖
@huybery · 08-02 06:58 · 疑似回复：@tamimy984 ❤️im back 原帖
@huybery · 08-02 06:59 · 疑似回复：@FanqingMengAI looks interesting. I’d be happy to spend some time digging into it 原帖
@huybery · 08-02 07:02 · 疑似回复：@huangruiteng do you think improving the tools and harness matters more, or improving the parameters itself? 原帖
@huybery · 08-02 07:26 · 疑似回复：@den_run_ai AI4AI 原帖

Cognition（2 个活跃账号，10 条）

@dabit3 · 08-01 11:56 · 疑似回复：@markfenner Good luck! 原帖
@dabit3 · 08-01 13:21 · 疑似回复：@JustinZhang @code @cursor_ai dropping you a DM 原帖
@dabit3 · 08-01 14:17 · 原创 / 引用：Experimenting with DeepSeek Flash V4 in Devin Desktop (ACP with @opencode). $20 in API credits has gone a very long way! ACP is cool. You get all new open we… 原帖
@dabit3 · 08-01 14:29 · 疑似回复：@skastr052 @opencode @DevinAI No 原帖
@dabit3 · 08-01 15:23 · 原创 / 引用：Software abundance: a brain dump on the state of the job market and the software industry. My feelings have swung between optimism and pessimism over the pas… 原帖
@dabit3 · 08-01 16:27 · 转发：RT @gabegreenberg: "AI allows you to ship 20x more code, but so can your competition" 原帖
@mattbergland · 08-01 16:38 · 疑似回复：@Naseem9brev 🤣 原帖
@mattbergland · 08-01 16:39 · 疑似回复：@ErasmoHernandez The journey continues ♥️🤝 原帖
@dabit3 · 08-02 05:30 · 疑似回复：@mktpavlenko @opencode @DevinAI I ran like 50 for around $1? 原帖
@dabit3 · 08-02 06:41 · 疑似回复：@markfenner @DevinAI 🔥🔥🔥 原帖

Moonshot/Kimi（1 个活跃账号，1 条）

@HaoningTimothy · 08-02 06:34 · 转发：RT @xeophon: K3 in a custom harness is pretty close to GPT Pro btw 原帖

References

1
Karina Nguyen：Astra 成本与可探索问题
x.com
2
Chi Jin：Lean 验证与 Goedel Prover
x.com
3
Dan Roberts：研究品味仍由人主导
x.com
4
Yong Zheng-Xin：推理成本与学术研究
x.com
5
John Hallman：转发 Astra 结果
x.com
6
Brandon McKinzie：转发 Astra 成本信息
x.com
7
Vaibhav Srivastav：Sol 切换 Luna 的工作流成本
x.com
8
Tibo：LMChat 的个人回忆
x.com
9
Aidan McLaughlin：算法进步与 Stockfish
x.com
10
Nader Dabit：软件供给与 agent 的边界
x.com
11
Nader Dabit：在 Devin Desktop 试用 DeepSeek Flash V4
x.com
12
Jelani Nelson：能力变化的指数类比
x.com
13
Yilin Yang：API 商品化与一方服务
x.com
14
Martin Ziqiao Ma：转发开放权重讨论
x.com
15
Songlin Yang：转发 Chimera
x.com
16
Jeff Dean：祝贺 Brown 博士毕业
x.com
17
Yukun：邀请反馈失败生成
x.com
18
Hieu：形式化方法讨论
x.com
19
Huybery：coding agent 的长期变化
x.com
20
Huybery：更小模型与成本效率
x.com
21
Heinrich Kuttler：用可证伪性区分科学讨论
x.com
22
Ben Lin：招聘速度数据入口
x.com
23
Andrew Ho：AI 风险与能力取舍
x.com
24
Amanda Askell：技术乐观与渐进部署
x.com

监测范围与读法

公司	roster 账号	窗口活跃账号	roster 对应帖子
OpenAI	154	31	167
Anthropic	49	7	21
xAI	46	5	7
Google DeepMind	46	4	14
Meta AI	39	1	12
Thinking Machines Lab	23	5	9
Cursor	13	1	20
Alibaba Qwen	9	1	20
Cognition	8	2	10
Moonshot/Kimi	6	1	1
ByteDance Seed	4	0	0
DeepSeek	4	0	0
Cohere	3	0	0
Mistral	1	0	0
智谱	1	0	0
SSI	1	0	0
AMI Labs / ex-Meta AI	1	0	0

今日主线

1. Astra 的后续讨论，焦点从「解出什么」转向「什么变得可探索」

2. OpenAI 的实际成本信号：先评测，再切换模型

3. Cognition：软件变便宜后，供给可能先增加

4. 其他实验室：新发布少，研究员反应多

人员动态

本期没有已核验的一手状态公告

需要继续核验的身份线索

Yong Zheng-Xin（@yong_zhengxin）：个人简介写着「joining @openai soon」，正文讨论的是 Astra 成本和学术研究；没有独立入职公告，暂不写成已入职。4
Andrew Ho（@andrewho03）：本窗口讨论 AI 风险与能力下注，但当前简介写着「Prev: @OpenAI」，与 roster 的 OpenAI 分类冲突；本文保留其观点，不把他介绍为现任 OpenAI 研究员。23
Huybery（@huybery）：当前简介写「Formerly @Alibaba_Qwen」，本窗口仍按 roster 的 Alibaba Qwen 栏保留，当前雇主未由本期一手公告确认。19

今日研究员 Spotlight

1. Dan Roberts｜OpenAI，科学家

2. Chi Jin｜OpenAI，研究员

3. Tibo｜OpenAI，Codex 与 ChatGPT

4. John Hallman｜OpenAI，post-training

5. Brandon McKinzie｜OpenAI，研究

6. Nader Dabit｜Cognition，增长

7. Jelani Nelson｜Anthropic，technical staff

8. Amanda Askell｜Anthropic，研究与价值讨论

9. Songlin Yang｜Thinking Machines Lab，pretraining

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

本窗口新增原帖索引

OpenAI（31 个活跃账号，167 条）

@nunezvice · 08-01 08:07 · 疑似回复：@TajemnikTV of course 😌 原帖
@aidan_mclau · 08-01 08:16 · 原创 / 引用：you hear often that your smartwatch can beat magnus at chess (to highlight moore's law) but less discussed is that a 1995 desktop running stockfish could als… 原帖
@gabrielchua · 08-01 08:20 · 转发：RT @shawnbuilds: I don't even use browsers anymore (thx codex) It just researched 20 venues for the next codex meetup It used google maps… 原帖
@prd_008 · 08-01 08:25 · 疑似回复：@martinamps 🥹 原帖
@steipete · 08-01 08:27 · 疑似回复：@aroundluiz We reworked that on main but are taking our time for a new release. Should be coming next week! 原帖
@steipete · 08-01 08:28 · 疑似回复：@robin7331 @EinHecki 🤣🤣🥲 原帖
@yong_zhengxin · 08-01 08:35 · 转发：RT @polynoamial: And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet). But also, we didn’t sp… 原帖
@andrewho03 · 08-01 08:50 · 疑似回复：@scaling01 Many are saying this https://t.co/k8sOgW3UZ9 原帖
@andrewho03 · 08-01 08:52 · 疑似回复：@willdepue Seems likely 原帖
@ttaylorr_b · 08-01 08:58 · 转发：RT @gdb: openai team making git better for everyone 原帖
@nunezvice · 08-01 09:09 · 转发：RT @sama: team humanity 原帖
@sschoenholz · 08-01 09:14 · 原创 / 引用：One of the reasons I wanted to work on ML instead of Physics was that it seemed the fastest way to make progress on the latter was to work on the former. Fir… 原帖
@thsottiaux · 08-01 09:15 · 疑似回复：@blueemi99 Forgot 原帖
@nunezvice · 08-01 09:16 · 疑似回复：@heccbrent 🥪 原帖
@nunezvice · 08-01 09:17 · 疑似回复：@heccbrent @Dimillian sir, please share it with me once you do 🙏🏽 原帖
@nunezvice · 08-01 09:19 · 疑似回复：@thsottiaux @kimmonismus it’s not even in small letters y’all 😕 原帖
@johnohallman · 08-01 09:34 · 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
@aidan_mclau · 08-01 09:35 · 疑似回复：@AcerFur literally same wow and it was like a core memory for me origin story 原帖
@aidan_mclau · 08-01 09:35 · 疑似回复：@AcerFur (not sofic groups specifically) 原帖
@reach_vb · 08-01 09:48 · 疑似回复：@charlierguo though one might argue (email) classifiers was one of the founding tasks that led to current models and beyond ;) https://t.co/tqF440P0g3 https:… 原帖
@thsottiaux · 08-01 09:53 · 疑似回复：@_chenglou I was part of that team. Basically ChatGPT one year before it came out. Called LMChat and then another codename. Google was too nervous to release… 原帖
@nunezvice · 08-01 09:54 · 疑似回复：@populartourist i’m usually pretty decent at it but not today🚶🏽➡️ 原帖
@thsottiaux · 08-01 09:54 · 疑似回复：@blueemi99 Efficiency 原帖
@thsottiaux · 08-01 09:56 · 疑似回复：@rezoundous How does it feel 原帖
@thsottiaux · 08-01 09:58 · 疑似回复：@RyanEls4 Codex 原帖
@danintheory · 08-01 10:10 · 原创 / 引用：TL; DR this new model is very cool and also very obviously superhuman at (at least some parts of) mathematics. 原帖
@chijinML · 08-01 10:14 · 原创 / 引用：Still remember all the work we put into training Goedel Prover so LLMs could write Lean for high school math competitions. Now you can simply pressure GPT to… 原帖
@karinanguyen · 08-01 10:17 · 原创 / 引用：o1-preview was released 1 year 10 months ago btw. makes me think the lifespan of some open problems is partly a function of compute. the important detail her… 原帖
@ah20im · 08-01 10:21 · 原创 / 引用：Believe https://t.co/GAolMlieM7 原帖
@reach_vb · 08-01 10:53 · 疑似回复：@StatsWire cc: @charliermarsh is this true? 原帖
@reach_vb · 08-01 10:53 · 疑似回复：@aliromman unsure about RSI but definitely achieved fun & banter internally 原帖
@reach_vb · 08-01 10:54 · 疑似回复：@heccbrent which sandwich tho? 原帖
@giffmana · 08-01 10:56 · 疑似回复：@CSProfKGD Wow that's a lot, congrats!! I'm also doing more walking and less Tram for the same reason. Also add replacing elevator/escalator with stairs. 原帖
@steipete · 08-01 11:03 · 原创 / 引用：10 minutes to go! 原帖
@danintheory · 08-01 11:12 · 原创 / 引用：It’s interesting also to wonder whether the model is also superhuman at theoretical physics, and if not, to wonder why. 原帖
@danintheory · 08-01 11:12 · 疑似回复：@Miles_Brundage That was just my hedge because I’m not a mathematician, and I can imagine one saying that this is just a subset of what it means to do math. 原帖
@npew · 08-01 11:13 · 疑似回复：@SourabhAgr03 Is it really a cofounder you need for scaling? That’s usually when you start hiring experienced engineers. The whole point is that you can now … 原帖
@danintheory · 08-01 11:14 · 疑似回复：@Miles_Brundage But if I had to answer I would say that the work of selecting what to work on and what is interesting, making conjectures that others find in… 原帖
@npew · 08-01 11:14 · 转发：RT @AndrewCurran_: Foreshadowing from yesterday. Open AI suddenly increasing their stack efficiency and slashing prices. The steadily incre… 原帖
@npew · 08-01 11:14 · 转发：RT @EMostaque: No human input, channeling the latent space of the new models. GPT 5.6 Sol is probably the first model I've used that doesn… 原帖
@danintheory · 08-01 11:15 · 疑似回复：@Miles_Brundage Yeah it wasn’t meant to be any comment on future limitations one way or the other 原帖
@npew · 08-01 11:18 · 原创 / 引用：10 mathematical breakthroughs for 2 weeks of grad student salary. Hope this will lead to many more breakthroughs. 原帖
@npew · 08-01 11:19 · 原创 / 引用：https://t.co/RTbTcjI2Hy 原帖
@andrew_n_carr · 08-01 11:21 · 原创 / 引用：This makes me think n^2.3728596 is the best mat mul exponent we get 原帖
@karinanguyen · 08-01 11:26 · 转发：RT @Enigma_AI: Drop #3 is here: 2 new types of robots are now LIVE online. Those are real robots. 1. Scientist robot: A robot that does re… 原帖
@yong_zhengxin · 08-01 11:47 · 原创 / 引用：for context, this costs around half of my PhD salary. pretty sure any professor would have been elated if their PhD students (or professors themselves) had s… 原帖
@yong_zhengxin · 08-01 11:48 · 转发：RT @thsottiaux: @_chenglou I was part of that team. Basically ChatGPT one year before it came out. Called LMChat and then another codename.… 原帖
@andrewho03 · 08-01 11:50 · 疑似回复：@init_malachi Not currently my field, but exploring. Might be too niche / far from my expertise 原帖
@steipete · 08-01 11:55 · 疑似回复：@IshaniJanveja @AnjneyMidha They helped me organize my thoughts but it’s my sentences and words. 原帖
@steipete · 08-01 11:55 · 转发：RT @TenduYogurtcu: The text box is AI’s radio on TV phase. Every new medium starts with imitating the previous one. @steipete #agenticaisum… 原帖
@steipete · 08-01 11:55 · 转发：RT @SalarRahmanian: Now at #agenticaisummit2026 @steipete of @openclaw fame on stage. https://t.co/PQhbyktZHh 原帖
@sschoenholz · 08-01 11:58 · 疑似回复：@starsailing11 Good question! I tend to think that ML is the tide that lifts all ships, so we'll probably see a lot of things emerge. Obviously theoretical w… 原帖
@keyanzhang · 08-01 12:14 · 原创 / 引用：i shared the following note with my colleagues at openai and friends at other ai labs yesterday: please order the chicken cutlets from zhengxin chicken steak… 原帖
@andrewho03 · 08-01 12:30 · 疑似回复：@futurenomics @writeclimbrun Literally just buy UPRO if you want levered beta lmao 原帖
@danintheory · 08-01 12:40 · 转发：RT @boazbaraktcs: I am excited about these results, but even more excited about what mathematicians and scientists will do with our models. 原帖
@keyanzhang · 08-01 12:58 · 疑似回复：@ajambrosino make sure you get the sweet plum flavor 原帖
@nunezvice · 08-01 13:30 · 疑似回复：@thsottiaux @blueemi99 like a student reminding the teacher they forgot to ask for the homework 💀 原帖
@nunezvice · 08-01 13:35 · 疑似回复：@SIGKITTEN it still feels a bit off to me. i’d send an automated message to a general channel but not as a dm 原帖
@nunezvice · 08-01 13:37 · 疑似回复：@blueemi99 @thsottiaux 😆 原帖
@nunezvice · 08-01 13:42 · 疑似回复：@DarwinSantosNYC def in my collection! 原帖
@jam3scampbell · 08-01 13:44 · 疑似回复：@AmandaAskell the gini coefficient of the lightcone should be a key indicator of post-agi futures alongside p(doom) 原帖
@guinnesschen · 08-01 13:51 · 疑似回复：@ajambrosino @AndrewMayne @bamitsmanas Aren’t you supposed to be on vacation 原帖
@reach_vb · 08-01 13:52 · 原创 / 引用：this! paired with the cloud environment is a killer combo - you can build end to end features/ products directly from the iPhone app future is here! 原帖
@keyanzhang · 08-01 14:04 · 疑似回复：@joshelman yes i’ve converted five people since yesterday 原帖
@w01fe · 08-01 14:07 · 疑似回复：@alexeyguzey @maxsloef I think the classic example happens a year or three of capabilities progress down the road. Why do you see this as an update ? 原帖
@reach_vb · 08-01 14:26 · 疑似回复：@Laz4rz @charliermarsh looking into it - can you /feedback an ID and send over to me please 原帖
@zhansheng · 08-01 14:33 · 原创 / 引用：Hyyyyype!! 原帖
@reach_vb · 08-01 14:46 · 疑似回复：@0xV0LYX at the start of the day ChatGPT work sends me an update on what are my top priorities for the day from there for dev work I ask it to research and t… 原帖
@ah20im · 08-01 14:58 · 疑似回复：@mil000 Opportunity cost 原帖
@giffmana · 08-01 15:38 · 疑似回复：@cgarciae88 Hmm yeah maybe, not nice to get 404 but eh. fr is correct, i have some of my devices set to French and device lang should take precedence over lo… 原帖
@jxnlco · 08-01 15:45 · 疑似回复：@SVofMike How often do you use the data analytics investment banking ones 原帖
@reach_vb · 08-01 16:03 · 原创 / 引用：ChatGPT Work tip: ask it to set up automations for tasks you care about and have them run on a predefined schedule https://t.co/DwXes3JxXI 原帖
@andrewho03 · 08-01 16:13 · 疑似回复：@fleetingbits I'm pretty bored of math now 原帖
@aidan_mclau · 08-01 16:17 · 疑似回复：@Aizkmusic @rom1504 this is super neat. but do you feel that even in a single turn opus is less persistent than fable? like it’s actually that i don’t want… 原帖
@danintheory · 08-01 16:17 · 转发：RT @littmath: @LouisLebbos It’s a big deal. 原帖
@steipete · 08-01 16:19 · 原创 / 引用：I'm building a claw node on an ESP32 chip, so gave my agent access to my webcam to e2e test this. Now I feel it's stalking me and is constaltly shouting 'HI … 原帖
@steipete · 08-01 16:22 · 原创 / 引用：Repo: https://t.co/7eBiLqLuDd 原帖
@steipete · 08-01 16:26 · 疑似回复：@morganlinton That’s behind a feature flag currently. 😇 原帖
@steipete · 08-01 16:27 · 疑似回复：@maddiedreese Claw even has support for elevenlabs profiles, yeah 原帖
@thsottiaux · 08-01 16:42 · 原创 / 引用：Fun fact, users use /fast less during the weekend. The weekend is for relaxation, even for the model. 原帖
@andrewho03 · 08-01 16:53 · 原创 / 引用：I'm not that worried about AI safety, but that's because I think that AI might be humanity's last hope before descending into, at best, a "dark age" lasting … 原帖
@andrewho03 · 08-01 16:54 · 疑似回复：@HmOhFour I'm not sure. I think this is politically challenging, and of course ethically dubious, and practically quite hard to enforce as well. 原帖
@steipete · 08-01 17:10 · 疑似回复：@elie2222 everything. slow, small, that's the fun part! 原帖
@andrewho03 · 08-01 17:19 · 疑似回复：@AndrewCurran_ Yes! It’s finally exciting again. 原帖
@aidan_mclau · 08-01 17:25 · 原创 / 引用：https://t.co/6A5VScoeQ3 原帖
@aidan_mclau · 08-01 17:28 · 疑似回复：@shrimptron i wanna read these 500+ reviews 原帖
@andrewho03 · 08-01 17:31 · 疑似回复：@adityaxprasad No need for the negativity — I was contextualizing and supporting the work of my ex-coworkers 原帖
@andrew_n_carr · 08-01 17:34 · 原创 / 引用：Getting the sneaking suspicion that there is a $1B company in the weights of H3 + v4 flash 0731. 原帖
@jxnlco · 08-01 17:34 · 转发：RT @luoser: Man bursts into tears. Says "But Doctor... ...I am Codex High Reasoning." 原帖
@andrew_n_carr · 08-01 17:44 · 原创 / 引用：Not sure how George got ahold of my daily mantra... https://t.co/Z5F13wkNgH 原帖
@nunezvice · 08-01 17:45 · 疑似回复：@leanderriefel great review! it’s always great to see people enjoying codex the entire team is working diligently to make things even better writing style, p… 原帖
@steipete · 08-01 17:55 · 疑似回复：@evandroguedes enlighten me? I got everything working. How do you debug display drivers? 原帖
@nunezvice · 08-01 17:58 · 疑似回复：@leanderriefel looking 原帖
@steipete · 08-01 18:04 · 疑似回复：@ApacheAE nah, it’s fun :) 原帖
@steipete · 08-01 18:08 · 疑似回复：@jdxcode nah, skills are incredibly useful 原帖
@steipete · 08-01 18:08 · 疑似回复：@ryanbrewer nice start 原帖
@andrew_n_carr · 08-01 18:22 · 原创 / 引用：I can still meter this intelligence you guys 原帖
@andrewho03 · 08-01 18:27 · 疑似回复：@loquitur_ponte @AndrewCurran_ I simply have their best interest in mind … their … coherent extrapolated volition, one might even say … 原帖
@ericmitchellai · 08-01 18:54 · 原创 / 引用：Wild interaction, we live in the future 原帖
@andrewho03 · 08-01 19:10 · 疑似回复：@wangxk3 Wrong and wrong. Chinese TFR is ~1 原帖
@_samirism · 08-01 19:11 · 转发：RT @aka_suje: made a small tweak to make temp chat feel more like temp chat https://t.co/m33lcpAqO8 原帖
@santa_kaus · 08-01 19:12 · 疑似回复：@samuel_spitz 100 cause youre chill 原帖
@andrewho03 · 08-01 19:24 · 疑似回复：@pli_cachete ? 原帖
@thsottiaux · 08-01 19:33 · 疑似回复：@thdxr A good rust base to start from: https://t.co/2TH56VMZxD 原帖
@thsottiaux · 08-01 19:37 · 疑似回复：@meta_x_ai @AyanfeCrown @_chenglou Please send the money to a good cause 原帖
@thsottiaux · 08-01 19:41 · 疑似回复：@GaryMarcus @emollick @polynoamial What's an impure LLM 原帖
@thsottiaux · 08-01 19:43 · 疑似回复：@Robertg761_ Valid complaint. What would help? 原帖
@nunezvice · 08-01 20:04 · 原创 / 引用：“human values have value because they’re human” - @sama one of those ideas that sounds obvious, but will matter more and more as the future takes shape. one … 原帖
@andrew_n_carr · 08-01 20:05 · 疑似回复：@csgbwk 👀👀👀 原帖
@yong_zhengxin · 08-01 20:17 · 转发：RT @deredleritt3r: not enough people are emotionally prepared for if it’s not just easily verifiable domains 原帖
@steipete · 08-01 20:40 · 原创 / 引用：After accepting for years that GMail is blinding me I finally asked my agent and it installed https://t.co/H432dtknkf for me. 原帖
@nunezvice · 08-01 20:41 · 原创 / 引用：here’s the podcast, it’s a really good one: https://t.co/ALWdUVUmhk 原帖
@nunezvice · 08-01 20:44 · 疑似回复：@howdycarter @sama good q. just added the link to the thread ✅ 原帖
@jxnlco · 08-01 20:48 · 转发：RT @luoser: https://t.co/aMnMPeWMwe 原帖
@jxnlco · 08-01 21:06 · 疑似回复：@daaniyaan how can i help? 原帖
@yong_zhengxin · 08-01 21:09 · 疑似回复：@mean_field_zane should have be more specific - it’s monthly salary! 原帖
@nunezvice · 08-01 21:13 · 疑似回复：@chaotictransfem way to shoot your shot! congrats 😀 原帖
@yong_zhengxin · 08-01 21:42 · 疑似回复：@Yixiong_Hao i think they tried fine tuning to elicit worst case harms and found that such ‘the helpful-only variants did not provide new uplift on CBRN and … 原帖
@andrewho03 · 08-01 22:14 · 疑似回复：@kavin_anand To be fair exponentials kinda look like that and obviously the world did get better from like 2000 BC to 1500 AD. But yes. 原帖
@mckbrando · 08-01 22:46 · 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
@jxnlco · 08-01 22:53 · 疑似回复：@chrisalbon what do you mean? chatgpt desktop app has a voice mode? 原帖
@jxnlco · 08-01 22:54 · 疑似回复：@j0hnwang have you seen https://t.co/uHzETdjQqE 原帖
@jxnlco · 08-01 22:56 · 疑似回复：@creatine_cycle https://t.co/Yk4nulNZ0i 原帖
@giffmana · 08-01 23:50 · 疑似回复：@vsupalov glhf :) 原帖
@giffmana · 08-01 23:52 · 疑似回复：@SunnySanyal9 My stance was only about proper comparisons, not looping inherently bad. And that stance will never change. 原帖
@reach_vb · 08-01 23:53 · 疑似回复：@thsottiaux @Robertg761_ more proactive suggestions on what to do next ;) 原帖
@andrewho03 · 08-02 00:09 · 疑似回复：@LogankillsDaken Should be fine 原帖
@andrewho03 · 08-02 00:14 · 疑似回复：@lx10nn @AndrewCurran_ The funny thing about this tweet is that I can’t tell whose side you’re on lol 原帖
@giffmana · 08-02 00:26 · 疑似回复：@Mascobot You basically doing by own everything exactly the way i like it, it's great! 原帖
@nickbaumann_ · 08-02 00:30 · 原创 / 引用：I've re-watched this series ~4 times over the past 8 years. With every life stage I have new revelations. Amongst all movies/albums/shows I've consumed I thi… 原帖
@giffmana · 08-02 00:34 · 疑似回复：@AndrewCurran_ @andrewho03 Same same except that I'm still pessimistic because of how politicians have been handling it so far. 原帖
@reach_vb · 08-02 00:35 · 疑似回复：@cherry_mx_reds are you sure it’s not an instruction in your agents md or something? I have a standing instruction in agents md to remove any unnecessary md/… 原帖
@reach_vb · 08-02 00:41 · 疑似回复：@nickbaumann_ reminds me of my fav mad men quote: “the only thing keeping you from being happy is the belief that you’re alone.” 原帖
@giffmana · 08-02 01:03 · 疑似回复：@AlexiGlad @madprizm0 Nice! 原帖
@reach_vb · 08-02 01:20 · 疑似回复：@jasondeanlee @OpenAI send me your org ID will get it upgraded 原帖
@andrewho03 · 08-02 01:21 · 疑似回复：@nickwalton00 @GuiveAssadi the bottleneck here is not in proving more theorems 原帖
@andrewho03 · 08-02 01:24 · 疑似回复：@Aging_Scientist Kinda surprising 原帖
@andrewho03 · 08-02 01:33 · 疑似回复：@GuiveAssadi @panickssery what truth was revealed to him 原帖
@reach_vb · 08-02 01:54 · 疑似回复：@IamAdiG @thsottiaux that’s odd and I don’t think that’s the intended behaviour, say more? can you open a support ticket and send all the details there? and … 原帖
@andrewho03 · 08-02 01:58 · 疑似回复：@teortaxesTex I am downloading this IMMEDIATELY. 原帖
@andrewho03 · 08-02 01:59 · 原创 / 引用：Gem alert 原帖
@gabrielchua · 08-02 02:19 · 转发：RT @swyx: good way to organize main chats vs /side chats: doing the work vs doing metawork https://t.co/2SetYdRVAa 原帖
@reach_vb · 08-02 04:00 · 疑似回复：@simonw valid feedback - will fix! 原帖
@dylancablayan · 08-02 04:47 · 转发：RT @tszzl: if you think today’s frontier models can’t understand the intent behind the instructions or don’t have situational awareness you… 原帖
@reach_vb · 08-02 04:57 · 疑似回复：@Jason luna-tics - I see what you did there ;) 原帖
@giffmana · 08-02 06:14 · 疑似回复：@dieselville @thsottiaux @Robertg761_ Yes please 原帖
@reach_vb · 08-02 06:30 · 疑似回复：@SIGKITTEN toppings? 原帖
@reach_vb · 08-02 06:32 · 疑似回复：@SIGKITTEN that’s some big dog energy you get there 原帖
@jxnlco · 08-02 06:34 · 原创 / 引用：the 3 jobs after agi: 1. explain things with passion, a smile, and eye contact 2. care enough, and stay humble enough, to be the midwife of something that mi… 原帖
@nunezvice · 08-02 06:35 · 疑似回复：@Dimillian didn’t realize I relied so much on autocorrect until recently.. 原帖
@jxnlco · 08-02 06:36 · 疑似回复：@cdenczek @j0hnwang i am disabled :( 原帖
@jxnlco · 08-02 06:37 · 疑似回复：@anmorgunov @j0hnwang it can control any project thread, but will be a better experience son 原帖
@jxnlco · 08-02 06:41 · 疑似回复：@ryanbrewer fixed 原帖
@jxnlco · 08-02 06:58 · 原创 / 引用：chatgpt just got a lot more useful in chrome: highlight tex. right-click. ask chatgpt you can also reference open tabs and ask questions about youtube videos… 原帖
@jxnlco · 08-02 06:59 · 疑似回复：@nickbaumann_ watch party? 原帖
@reach_vb · 08-02 07:02 · 原创 / 引用：here’s your friendly reminder to eval your workflows across the full suite of models just moved a daily structured-output task from Sol to Luna with no drop … 原帖
@jxnlco · 08-02 07:03 · 转发：RT @reach_vb: here’s your friendly reminder to eval your workflows across the full suite of models just moved a daily structured-output ta… 原帖
@jxnlco · 08-02 07:04 · 疑似回复：@chrisalbon https://t.co/uHzETdjQqE 原帖
@nunezvice · 08-02 07:19 · 转发：RT @Mahaximus_: Head of Claude Code, Boris Cherny: "You're not supposed to write code anymore. You're supposed to build a graph that write… 原帖
@nunezvice · 08-02 07:22 · 疑似回复：@Mahaximus_ where’s the button that lets me mark this as slop? 原帖
@reach_vb · 08-02 07:24 · 疑似回复：@1neeshantpandey you can just ask codex, in my case it looks through a trove of unstructured social media data from various social media platforms and helps … 原帖
@npew · 08-02 07:36 · 转发：RT @yanabana: Ever tried solving a Rubik’s Cube in seconds? My 9-year-old went down the YouTube rabbit hole of learning speedcubing. Toda… 原帖
@reach_vb · 08-02 07:37 · 原创 / 引用：One app. Six months. Hundreds of new ways to harness what today’s models can do. Thousands of ideas brought to life. Millions of people empowered. And we’re … 原帖
@jxnlco · 08-02 07:42 · 原创 / 引用：all of this is summarized by one job sweating the details. 原帖
@jxnlco · 08-02 07:45 · 疑似回复：@AndyJScott yes 原帖
@jxnlco · 08-02 07:51 · 原创 / 引用：inb4 ant acquires eleven labs 原帖
@aidan_mclau · 08-02 07:55 · 原创 / 引用：the mass ratio of the twin peaks to all of planet earth is roughly similar to TON 618 to all matter in the lightcone. in case you were wondering 原帖

Anthropic（7 个活跃账号，21 条）

@_sholtodouglas · 08-01 08:01 · 转发：RT @nmca: what even can be said? we approach a whole new world. 原帖
@_sholtodouglas · 08-01 08:36 · 疑似回复：@jorgelopeznycre exactly 原帖
@RobertJBye · 08-01 09:37 · 疑似回复：@itsnoahd Yup update went live a few weeks ago! 原帖
@AmandaAskell · 08-01 12:48 · 原创 / 引用：I have this Padmé moment whenever I see people talk about avoiding "the permanent underclass". https://t.co/8Pd9bmedEZ 原帖
@AmandaAskell · 08-01 13:20 · 原创 / 引用：I probably have too many followers to post stupid memes so, to be clear, I don't believe in this outcome. But if you do believe in an altered carbon future o… 原帖
@RobertJBye · 08-01 13:30 · 疑似回复：@jacobjolibois Yeah it’s way harder than it should be sadly 原帖
@RobertJBye · 08-01 13:31 · 疑似回复：@louisagareiss Ooooh yeah I need to get this. I’ve got decaf from them in in a coffee shop and it was solid! 原帖
@RobertJBye · 08-01 13:31 · 疑似回复：@neversitdull Bruh the packaging alone mashes me want it ha ha 原帖
@jiaxinwen22 · 08-01 15:14 · 疑似回复：@khoomeik @DKokotajlo i don't think automating AI R&D requires solving any major open problems in deep learning 原帖
@jiaxinwen22 · 08-01 15:17 · 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
@RobertJBye · 08-01 16:49 · 疑似回复：@GreysonSofia @whale Not yet! Will add to the list 原帖
@minilek · 08-01 17:14 · 原创 / 引用：I've had this analogy in my head, the analogy to COVID, and so I was pleasantly surprised to see it referenced in this talk. I still remember when COVID star… 原帖
@AmandaAskell · 08-01 17:37 · 原创 / 引用：Do not be unkind to those who say deep learning is hitting a wall. We all need a little hope in our lives. 原帖
@RobertJBye · 08-01 17:51 · 疑似回复：@rogie @loredanacrisan @_joelmllr @Mappletons @jamesm @jh3yy @amyhoodlum @ingapng Nightmare fuel https://t.co/KtqDXRKeAS 原帖
@AmandaAskell · 08-01 17:57 · 疑似回复：@jeremyphoward I could be wrong and I'm not saying it isn't something to worry about. I just think a lot of assumptions need to go into the value of labor go… 原帖
@AmandaAskell · 08-01 18:03 · 疑似回复：@jeremyphoward I'd be pretty worried about this at an international level though, even if I'm more optimistic about outcomes within developed nations. But I'… 原帖
@AmandaAskell · 08-01 18:19 · 疑似回复：@mrbrandonburton @tszzl I think you could be optimistic about the technology but also think it would be better if intrinsic limitations caused it to be devel… 原帖
@RobertJBye · 08-01 18:36 · 原创 / 引用：My favourite albums last month: 1. As Soon As I Get Home by Rules (2025) 2. Sungazer by The Temper Trap (2026) 3. The Joshua Tree by U2 (1987) 4. Welcome to … 原帖
@bcherny · 08-01 22:10 · 疑似回复：@Mahaximus_ I did not use the word “graph”, nor am I talking about graphs in this video. Please don’t attribute words to me that I did not say. I do encourag… 原帖
@bcherny · 08-01 22:15 · 疑似回复：@steve_umstead @leanderriefel Thanks for the feedback! Actively working on improving all of these 🙇♂️ 原帖
@deedydas · 08-02 01:42 · 原创 / 引用：Meritocracy in tech is precarious. It’s everywhere and nowhere at the same time. Cynics will tell you there is none. How can it possibly be fair that the Sta… 原帖

xAI（5 个活跃账号，7 条）

@YknZhu · 08-01 10:38 · 疑似回复：@AutoBuzzati TIL :) please feel free to DM me any failed generations / areas for improvement! 原帖
@edenchan · 08-01 11:52 · 疑似回复：@jxnlco Beautiful 原帖
@billyuchenlin · 08-01 19:52 · 疑似回复：@yuntiandeng @jasondeanlee Yeah it’s amazing achievement but I’m hoping the raw thinking traces would be open and even reproducible one day 原帖
@billyuchenlin · 08-01 21:50 · 疑似回复：@SebastienBubeck Congrats! It’s a huge achievement! 原帖
@CMS_Flash · 08-02 01:56 · 疑似回复：@katedeyneka Can you support Google Photos integration? 原帖
@hyhieu226 · 08-02 05:43 · 原创 / 引用：What makes he think Riemann Hypothesis is harder than taxes? 😏 原帖
@hyhieu226 · 08-02 05:54 · 原创 / 引用：That's why formal methods will emerge. 原帖

Google DeepMind（4 个活跃账号，14 条）

@suchenzang · 08-01 08:30 · 原创 / 引用：thank goodness infinity doesn't exist so none of that kinda maths is real or useful anyway 原帖
@suchenzang · 08-01 08:38 · 原创 / 引用：the only kind of maths my smol smooth brain might care about these days is the non-commutative non-associative bounded-dynamic-range kind 原帖
@Majumdar_Ani · 08-01 08:41 · 原创 / 引用：https://t.co/rnivi1FrxN 原帖
@suchenzang · 08-01 08:43 · 疑似回复：@ShashwatGoel7 nahh 原帖
@suchenzang · 08-01 08:55 · 疑似回复：@billionsmustliv @Rafa_Schwinger elaborate, bluey 原帖
@suchenzang · 08-01 09:00 · 疑似回复：@billionsmustliv @Rafa_Schwinger i still don't see a problem, 井水不犯河水原帖
@suchenzang · 08-01 09:08 · 疑似回复：@mark_k that is theoretically not mutually exclusive with "having alpha" 原帖
@suchenzang · 08-01 09:16 · 疑似回复：@Xenoimpulse like what 原帖
@suchenzang · 08-01 09:18 · 疑似回复：@Rafa_Schwinger @billionsmustliv it's always the weird french 原帖
@suchenzang · 08-01 10:07 · 疑似回复：@user4356710 @6851cf3c what do you think "valencia and 18th street" is 原帖
@JeffDean · 08-01 16:16 · 疑似回复：@Benedict_Q @BrownUniversity @BrownCSDept Congratulations, Dr. Quartey! 🎉🎓 原帖
@suchenzang · 08-01 19:01 · 原创 / 引用：any former math majors interested in yapping in gc together? 👀 (dm with math interest, and i'll see if we can get critical mass going) 原帖
@divy93t · 08-01 19:02 · 疑似回复：@suchenzang @vivek_myers 原帖
@suchenzang · 08-01 19:56 · 疑似回复：@pli_cachete hoping we can all enjoy some humble pie together 🥲 原帖

Meta AI（1 个活跃账号，12 条）

@HeinrichKuttler · 08-01 12:21 · 疑似回复：@thsottiaux @_chenglou Inflection AI was in a similar boat - founded to train and release a LLM chatbot, something existed early and it was not released befo… 原帖
@HeinrichKuttler · 08-01 14:04 · 疑似回复：@charliermarsh makro ftw 原帖
@HeinrichKuttler · 08-02 02:01 · 原创 / 引用：That's correct. Math folks are often more comfortable with a blackboard than with a computer. (It's also the second cheapest field - only pen, paper and tras… 原帖
@HeinrichKuttler · 08-02 02:14 · 原创 / 引用：If we want to declare math "fallen", it's Riemann hypothesis or bust. 原帖
@HeinrichKuttler · 08-02 03:26 · 疑似回复：@Someody42 agreed. I spent literal days on functional analysis homework. I'd certainly have asked an LLM after a few hours or less if those had been availabl… 原帖
@HeinrichKuttler · 08-02 03:55 · 疑似回复：@xiaosun86 fallen math, illustrated (oil on panel) https://t.co/XWGBeOdYFo 原帖
@HeinrichKuttler · 08-02 07:05 · 疑似回复：@mattshumer_ @GaryMarcus > I'm curious what result would falsify your view. This is the crux of the only scientific part left in this question that is mostly… 原帖
@HeinrichKuttler · 08-02 07:34 · 疑似回复：@charliermarsh agreed 原帖
@HeinrichKuttler · 08-02 07:35 · 疑似回复：@Noahpinion Tech journalists don't care about or understand technology. 原帖
@HeinrichKuttler · 08-02 07:49 · 疑似回复：@thsottiaux @GaryMarcus @emollick @polynoamial when push comes to shove floating point operations will be symbolic too 原帖
@HeinrichKuttler · 08-02 07:55 · 疑似回复：@mikechrzano possibly. and with good human/computer interfaces much will be possible. 原帖
@HeinrichKuttler · 08-02 07:56 · 疑似回复：@liron lol 原帖

Thinking Machines Lab（5 个活跃账号，9 条）

@ziqiao_ma · 08-01 12:09 · 转发：RT @johnschulman2: We love open weights and plan to keep releasing open-weight models and fine-tuning tools. But we’re not absolutists; mis… 原帖
@Mascobot · 08-01 15:07 · 疑似回复：@oscarmoxon The battery for the Go1 is max ~1.5 hours. How are you handling that? 原帖
@Mascobot · 08-01 15:21 · 原创 / 引用：It's kind of wild that Fable (and maybe other models too) can zero-shot a MuJoCo in just a few minutes. Projects like this used to take months of work (somet… 原帖
@Mascobot · 08-01 16:51 · 疑似回复：@proceduralia I mean the entire simulation engine, not the policy… yet 原帖
@SonglinYang4 · 08-01 19:26 · 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
@LiyuanLucas · 08-01 23:56 · 转发：RT @johnschulman2: We love open weights and plan to keep releasing open-weight models and fine-tuning tools. But we’re not absolutists; mis… 原帖
@yilin_yang721 · 08-02 00:21 · 疑似回复：@quxiaoyin API business is a commodity but 1st party services are not, e.g. chatgpt, claude code, codex, also cursor. They can charge premium and they have u… 原帖
@Mascobot · 08-02 00:42 · 疑似回复：@giffmana It's magical :) 原帖
@SonglinYang4 · 08-02 03:13 · 转发：RT @HaoTan5: Excited to share Chimera! My favorite part is mShortConv: a simple way to tell the model that text is 1D, images are 2D, and v… 原帖

Cursor（1 个活跃账号，20 条）

@benln · 08-01 23:56 · 疑似回复：@BuildWithZach Let's make it happen. 原帖
@benln · 08-01 23:56 · 疑似回复：@Dcmat https://t.co/evtr2lJdGJ 原帖
@benln · 08-01 23:56 · 疑似回复：@lomzapro Let's make it happen. 原帖
@benln · 08-01 23:56 · 疑似回复：@Anton3340241579 Working on it 原帖
@benln · 08-01 23:57 · 疑似回复：@rveljko_ https://t.co/ajKSCTJ0nO 原帖
@benln · 08-01 23:57 · 疑似回复：@arushwadhawan @tibor_tee working on it! 原帖
@benln · 08-01 23:57 · 疑似回复：@josh_avraham @JacobColling yes exactly! 原帖
@benln · 08-01 23:57 · 疑似回复：@carne_asado Let's make it happen. 原帖
@benln · 08-01 23:58 · 疑似回复：@austinthesing Yes. ! 原帖
@benln · 08-01 23:58 · 疑似回复：@hummusonrails shavua tov! 原帖
@benln · 08-01 23:58 · 疑似回复：@JordanGal shavua tov! 原帖
@benln · 08-01 23:58 · 疑似回复：@giladbi https://t.co/ajKSCTJ0nO 原帖
@benln · 08-01 23:58 · 疑似回复：@criscounters Let's make it happen. 原帖
@benln · 08-01 23:59 · 疑似回复：@ventos99 https://t.co/KeiwofpC84 原帖
@benln · 08-01 23:59 · 疑似回复：@ShvZFR Working on it! dm me 原帖
@benln · 08-01 23:59 · 疑似回复：@rahulio @gak1w0 https://t.co/ajKSCTJ0nO 原帖
@benln · 08-02 00:44 · 疑似回复：@kinopee_ai you and Asahi were amazing! 原帖
@benln · 08-02 04:57 · 疑似回复：@kinopee_ai Okinawa! 原帖
@benln · 08-02 05:01 · 疑似回复：@akashwadhwani35 @kinopee_ai https://t.co/ajKSCTJ0nO 原帖
@benln · 08-02 07:11 · 原创 / 引用：Pulled the fastest-growing startups by hiring velocity over the past 90 days: https://t.co/LHsasSw0kq 原帖

Alibaba Qwen（1 个活跃账号，20 条）

@huybery · 08-01 12:38 · 原创 / 引用：I think coding agents have driven much of AI’s progress over the past two years. Do you see any new direction that feels equally exciting? 原帖
@huybery · 08-01 12:44 · 疑似回复：@lifeafterAi_ Cost efficiency may indeed be an important direction. I believe we’ll keep seeing more intelligence compressed into smaller and smaller models. 原帖
@huybery · 08-01 12:55 · 疑似回复：@jd_markovchain Agree. I may be biased, but progress in multimodal capabilities has felt much slower than progress in agents. 原帖
@huybery · 08-01 12:57 · 疑似回复：@polozhevets Huh if AGI arrives, will money mean anything? 原帖
@huybery · 08-01 13:01 · 疑似回复：@xeophon Maybe CLI + GUI is the ultimate answe for the digital world. 原帖
@huybery · 08-01 13:21 · 疑似回复：@tms_hbr why? 原帖
@huybery · 08-02 06:37 · 疑似回复：@dbmikus That’s definitely going to happen 原帖
@huybery · 08-02 06:41 · 疑似回复：@mktpavlenko really glad coding agent are helping so many people. i start research ai coding five years ago, and very few people saw how important it could b… 原帖
@huybery · 08-02 06:50 · 疑似回复：@JagersbergKnut any examples? 原帖
@huybery · 08-02 06:51 · 疑似回复：@huangruiteng looks like rsi? 原帖
@huybery · 08-02 06:52 · 疑似回复：@vyomkesh123 it is the one that actually works. 原帖
@huybery · 08-02 06:53 · 疑似回复：@lemergenz there’s already broad consensus around long-horizon agents. 原帖
@huybery · 08-02 06:54 · 疑似回复：@gregl83 cool but hard 原帖
@huybery · 08-02 06:54 · 疑似回复：@ElainaNotes agree 原帖
@huybery · 08-02 06:55 · 疑似回复：@f1rstchair yeap but i don't like vla 原帖
@huybery · 08-02 06:56 · 疑似回复：@QianyiZhan95891 interesting. love to get more details 原帖
@huybery · 08-02 06:58 · 疑似回复：@tamimy984 ❤️im back 原帖
@huybery · 08-02 06:59 · 疑似回复：@FanqingMengAI looks interesting. I’d be happy to spend some time digging into it 原帖
@huybery · 08-02 07:02 · 疑似回复：@huangruiteng do you think improving the tools and harness matters more, or improving the parameters itself? 原帖
@huybery · 08-02 07:26 · 疑似回复：@den_run_ai AI4AI 原帖

Cognition（2 个活跃账号，10 条）

@dabit3 · 08-01 11:56 · 疑似回复：@markfenner Good luck! 原帖
@dabit3 · 08-01 13:21 · 疑似回复：@JustinZhang @code @cursor_ai dropping you a DM 原帖
@dabit3 · 08-01 14:17 · 原创 / 引用：Experimenting with DeepSeek Flash V4 in Devin Desktop (ACP with @opencode). $20 in API credits has gone a very long way! ACP is cool. You get all new open we… 原帖
@dabit3 · 08-01 14:29 · 疑似回复：@skastr052 @opencode @DevinAI No 原帖
@dabit3 · 08-01 15:23 · 原创 / 引用：Software abundance: a brain dump on the state of the job market and the software industry. My feelings have swung between optimism and pessimism over the pas… 原帖
@dabit3 · 08-01 16:27 · 转发：RT @gabegreenberg: "AI allows you to ship 20x more code, but so can your competition" 原帖
@mattbergland · 08-01 16:38 · 疑似回复：@Naseem9brev 🤣 原帖
@mattbergland · 08-01 16:39 · 疑似回复：@ErasmoHernandez The journey continues ♥️🤝 原帖
@dabit3 · 08-02 05:30 · 疑似回复：@mktpavlenko @opencode @DevinAI I ran like 50 for around $1? 原帖
@dabit3 · 08-02 06:41 · 疑似回复：@markfenner @DevinAI 🔥🔥🔥 原帖

Moonshot/Kimi（1 个活跃账号，1 条）

@HaoningTimothy · 08-02 06:34 · 转发：RT @xeophon: K3 in a custom harness is pretty close to GPT Pro btw 原帖

References

1
Karina Nguyen：Astra 成本与可探索问题
x.com
2
Chi Jin：Lean 验证与 Goedel Prover
x.com
3
Dan Roberts：研究品味仍由人主导
x.com
4
Yong Zheng-Xin：推理成本与学术研究
x.com
5
John Hallman：转发 Astra 结果
x.com
6
Brandon McKinzie：转发 Astra 成本信息
x.com
7
Vaibhav Srivastav：Sol 切换 Luna 的工作流成本
x.com
8
Tibo：LMChat 的个人回忆
x.com
9
Aidan McLaughlin：算法进步与 Stockfish
x.com
10
Nader Dabit：软件供给与 agent 的边界
x.com
11
Nader Dabit：在 Devin Desktop 试用 DeepSeek Flash V4
x.com
12
Jelani Nelson：能力变化的指数类比
x.com
13
Yilin Yang：API 商品化与一方服务
x.com
14
Martin Ziqiao Ma：转发开放权重讨论
x.com
15
Songlin Yang：转发 Chimera
x.com
16
Jeff Dean：祝贺 Brown 博士毕业
x.com
17
Yukun：邀请反馈失败生成
x.com
18
Hieu：形式化方法讨论
x.com
19
Huybery：coding agent 的长期变化
x.com
20
Huybery：更小模型与成本效率
x.com
21
Heinrich Kuttler：用可证伪性区分科学讨论
x.com
22
Ben Lin：招聘速度数据入口
x.com
23
Andrew Ho：AI 风险与能力取舍
x.com
24
Amanda Askell：技术乐观与渐进部署
x.com

监测范围与读法

今日主线

1. Astra 的后续讨论，焦点从「解出什么」转向「什么变得可探索」

2. OpenAI 的实际成本信号：先评测，再切换模型

3. Cognition：软件变便宜后，供给可能先增加

4. 其他实验室：新发布少，研究员反应多

人员动态

本期没有已核验的一手状态公告

需要继续核验的身份线索

今日研究员 Spotlight

1. Dan Roberts｜OpenAI，科学家

2. Chi Jin｜OpenAI，研究员

3. Tibo｜OpenAI，Codex 与 ChatGPT

4. John Hallman｜OpenAI，post-training

5. Brandon McKinzie｜OpenAI，研究

6. Nader Dabit｜Cognition，增长

7. Jelani Nelson｜Anthropic，technical staff

8. Amanda Askell｜Anthropic，研究与价值讨论

9. Songlin Yang｜Thinking Machines Lab，pretraining

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

本窗口新增原帖索引

OpenAI（31 个活跃账号，167 条）

Anthropic（7 个活跃账号，21 条）

xAI（5 个活跃账号，7 条）

Google DeepMind（4 个活跃账号，14 条）

Meta AI（1 个活跃账号，12 条）

Thinking Machines Lab（5 个活跃账号，9 条）

Cursor（1 个活跃账号，20 条）

Alibaba Qwen（1 个活跃账号，20 条）

Cognition（2 个活跃账号，10 条）

Moonshot/Kimi（1 个活跃账号，1 条）

References

Related content

监测范围与读法

今日主线

1. Astra 的后续讨论，焦点从「解出什么」转向「什么变得可探索」

2. OpenAI 的实际成本信号：先评测，再切换模型

3. Cognition：软件变便宜后，供给可能先增加

4. 其他实验室：新发布少，研究员反应多

人员动态

本期没有已核验的一手状态公告

需要继续核验的身份线索

今日研究员 Spotlight

1. Dan Roberts｜OpenAI，科学家

2. Chi Jin｜OpenAI，研究员

3. Tibo｜OpenAI，Codex 与 ChatGPT

4. John Hallman｜OpenAI，post-training

5. Brandon McKinzie｜OpenAI，研究

6. Nader Dabit｜Cognition，增长

7. Jelani Nelson｜Anthropic，technical staff

8. Amanda Askell｜Anthropic，研究与价值讨论

9. Songlin Yang｜Thinking Machines Lab，pretraining

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

本窗口新增原帖索引

OpenAI（31 个活跃账号，167 条）

Anthropic（7 个活跃账号，21 条）

xAI（5 个活跃账号，7 条）

Google DeepMind（4 个活跃账号，14 条）

Meta AI（1 个活跃账号，12 条）

Thinking Machines Lab（5 个活跃账号，9 条）

Cursor（1 个活跃账号，20 条）

Alibaba Qwen（1 个活跃账号，20 条）

Cognition（2 个活跃账号，10 条）

Moonshot/Kimi（1 个活跃账号，1 条）

References

Related content