过去24小时研究员 X 动态：Qwen3.8-Max 预告开放权重，OpenAI 把 Codex 推向业务工作流

监测范围与读法

今日主线

1. Qwen3.8-Max 的信号不只是参数，而是「长时程工作」和开放权重同时推进

2. OpenAI 的 Codex 讨论转向三个闭环：教学工具、客户反馈、企业运营

3. Anthropic 的讨论落在模型专用 harness，以及低成本带来的依赖管理压力

4. 研究员发言继续把「推理时优化」和「可审计性」推到前台

按公司查看

OpenAI

• @giffmana 讨论自己让模型查看实验结果并建议下一步，认为约 10% 的建议与自己的想法一致，其他多数是无效的微调建议；这更像一线研究工作流反馈，不是受控评测。16
• @andrew_n_carr 把 Qwen3.8-Max 公告里的 10 天以上运行和论文复现，联系到「模型压缩 / compaction」是否可能成为通向 AGI 的未知问题之一；这是个人推测，原帖没有给出实验。17
• @gdb、@aidan_mclau、@ryanbrewer 的产品与招聘帖构成了本期 OpenAI 最强的非论文信号；其余帖子主要是对 ChatGPT Work、Codex、代码规范和模型反馈的转发或回复。

Anthropic

Alibaba Qwen

Google DeepMind

AMI Labs / ex-Meta AI

Meta AI

Thinking Machines Lab

Moonshot/Kimi

Cursor

Cognition

ByteDance Seed

xAI

无窗口命中的公司

人员动态

• @huybery 的 profile 写「Formerly @Alibaba_Qwen」，但这不是本窗口内的个人状态 announce；因此继续按频道既定 roster 归类，同时标注当前身份冲突。
• @minilek 说某个 dynamic rank/select 项目是「在 Google 做的最后一些工作之一」，这是对过往项目的回顾，不是本期离开 Anthropic 或加入新团队的公告。27
• @jxnlco 在回复中说自己「四个月前加入」OpenAI，没有岗位或团队信息，也不是窗口内发生的加入事件；记录为低信息量历史状态线索，不列入本期新入职。28

今日研究员 Spotlight

1. Greg Brockman（@gdb，OpenAI）：OpenAI President and Co-Founder，当前公开信号集中在 Codex 的产品化方向。他连续把 Codex 指向交互式教育工具、客户反馈到 roadmap、以及企业运营，说明他关注的不只是代码生成，而是模型进入组织工作流后的闭环。4
2. Andrej Karpathy（@karpathy，Anthropic）：频道 roster 将他归入 Anthropic；其公开简介强调训练大型深度神经网络。近期开源浏览器里的 pelican 生成项目，并讨论极小用户空间下 n-gram、决策树和神经网络的比较，关注模型训练、生成程序和效率边界。11
3. Katelyn Lesse（@katelyn_lesse，Anthropic）：Anthropic Head of Platform，关注 agent 平台和 Claude Managed Agents。她最近把模型专用 harness 与通用多模型 harness 做了对比，强调 harness 能否利用模型家族的独特能力，直接关系到 agent 的实际性能。9
4. Binyuan Hui（@huybery，Alibaba Qwen roster）：研究兴趣落在 LLM 的推理、规划、代码和工具使用评测。他最近提出用构建游戏测试这些能力，并指出社区缺标准化 game-building benchmark。当前 profile 写着 Formerly Qwen，因此这里保留研究信号，不把在职关系写死。19
5. Yilin（@yilin_yang721，Thinking Machines Lab）：MTS，曾参与 GPT Realtime Translate 与 Seamless Translation。其关注点包括实时多模态翻译和产品使用模式，最近讨论 ChatGPT 与 Gemini 的用户规模判断，虽然是个人估计，却能反映他如何把模型能力放回实际产品竞争中。29
6. Marco Mascorro（@Mascobot，Thinking Machines Lab）：研究 LLM，也做机器人，过去参与 Fellow Robots 等项目。他最近从工业空间、能耗、稳定性、末端定位和安全成本出发，解释为什么多数工业应用不需要腿式机器人，关注语言模型与物理系统的落地约束。22
7. Zhiqing Sun（@EdwardSun0909，Meta AI）：Meta MSL TBD Lab 的 agent research lead，研究重点是 agent 与 post-training，曾在 OpenAI 做相关研究。他最近祝贺 Qwen CUA 发布新模型，动态本身没有评测细节，但显示他持续关注 computer-use agent 的同行进展。21
8. Yann LeCun（@ylecun，AMI Labs / ex-Meta AI）：AMI Labs Executive Chairman，研究方向包括 EBM、ODAI、世界模型、机器学习和机器人。他最近解释连续变量推断中的梯度优化，并把世界模型规划作为例子，仍在公开推进推理时优化这条研究路线。13
9. Ben Lang（@benln，Cursor）：Cursor team 成员，也运营 Next Play newsletter，关注开发者工具的产品传播和社区反馈。他最近公布 Cafe Cursor 8 月覆盖 24 个国家，虽然不是模型研究结果，却能帮助读者观察 Cursor 如何把产品社区扩展到开发者现场。25
10. Crystal（@crystalsssup，Moonshot/Kimi）：Kimi Staff，关注 Kimi 产品与用户工作流。她所在的 roster 账号池窗口内出现 Kimi Slides 的完整教程，展示 Kimi K3 如何被包装进研究、结构、图表和可编辑幻灯片流程；这里引用的是官方产品帖，Crystal 本人的窗口内代表帖较少。24

完整原帖索引

OpenAI（40 个活跃账号，182 条）

• 8 月 2 日 08:03 @_samirism（原创）：finally watched the odyssey - found some decent last minute tickets using the chatgpt work cloud browser movie a little over-hyped imo https://t.co/JAfUrgnfId 原帖
• 8 月 2 日 08:13 @gdb（原创）：chatgpt for building interactive educational tools 原帖
• 8 月 2 日 08:19 @_samirism（疑似回复）：@ChristianInProd ensure its Work mode 原帖
• 8 月 2 日 08:19 @sherwinwu（疑似回复）：@charliermarsh imagine thinking that nonsofic groups don’t exist 原帖
• 8 月 2 日 08:27 @charliermarsh（疑似回复）：@HotAisle Maybe "increasingly concerned" 原帖
• 8 月 2 日 08:29 @_samirism（疑似回复）：@zachwaugh use work - easiest way right now cloud computer, synced files, memory and history all from phone - with no setup needed 原帖
• 8 月 2 日 08:42 @charliermarsh（疑似回复）：@foges3 @bodonoghue85 @thsottiaux I'm looking into both of these 原帖
• 8 月 2 日 08:42 @nunezvice（转发）：RT @ajambrosino: the Codex app launched 6 months ago today 原帖
• 8 月 2 日 08:43 @nunezvice（疑似回复）：@a7mddra @ajambrosino @thsottiaux ain’t no way 😂 原帖
• 8 月 2 日 08:45 @nunezvice（疑似回复）：@jxnlco why are we giving them ideas to catch up? 😂 原帖
• 8 月 2 日 08:47 @nunezvice（疑似回复）：@JoanneShang don’t yell at me, joanne! 😂 原帖
• 8 月 2 日 08:52 @jxnlco（疑似回复）：@simonw like in this view right? https://t.co/EXzu4Q0l91 原帖
• 8 月 2 日 08:57 @jxnlco（疑似回复）：@loosenedspirit @simonw yup understood 原帖
• 8 月 2 日 09:07 @jxnlco（原创）：i apologize to all my coworkers i've been slacking since 6am, i thought it was monday... 原帖
• 8 月 2 日 09:08 @npew（原创）：Our 9-year old is obsessed with Rubik’s Cube and is building a website for anyone who wants to learn to solve it. Check it out and give him feedback in the replies: https://t.co/Is 原帖
• 8 月 2 日 09:08 @ajambrosino（疑似回复）：@sound4movement @ChatGPT glad we have subtitles here 原帖
• 8 月 2 日 09:10 @ajambrosino（疑似回复）：@SantanuDatta94 🔜 原帖
• 8 月 2 日 09:10 @ajambrosino（疑似回复）：@identity_matrix 🔜 原帖
• 8 月 2 日 09:10 @ajambrosino（疑似回复）：@ForwardEditor 👍 原帖
• 8 月 2 日 09:10 @jxnlco（疑似回复）：@fuckpoasting love the rick 原帖
• 8 月 2 日 09:15 @jxnlco（疑似回复）：@nunezvice i just feel it in my bones 原帖
• 8 月 2 日 09:18 @jxnlco（疑似回复）：@illscience @JonathanAbrams how can i help you make this happen 原帖
• 8 月 2 日 09:33 @jxnlco（原创）：I’ve literally just been like tagging people and threats on the weekend. Oh my fucking God. 原帖
• 8 月 2 日 09:34 @jxnlco（原创）：https://t.co/2lOnP4APTg 原帖
• 8 月 2 日 09:40 @yong_zhengxin（原创）：was seeing how someone said that an undergrad is not worthy working on Erdos-level problems: 1/ as a PhD, I have met so many undergrads or high school dropouts smarter than me. Cre 原帖
• 8 月 2 日 09:53 @ryanbrewer（疑似回复）：@chrisbedoya_ This was yesterday 原帖
• 8 月 2 日 10:02 @andrewho03（疑似回复）：@garrytan 💯 原帖
• 8 月 2 日 10:04 @npew（疑似回复）：@twominutepapers It’s a fun challenge! 原帖
• 8 月 2 日 10:06 @aidan_mclau（疑似回复）：@segyges she’s amazing 原帖
• 8 月 2 日 10:14 @ryanbrewer（疑似回复）：@splitbycomma Yesterday btw 原帖
• 8 月 2 日 10:15 @nunezvice（疑似回复）：@jxnlco this whole year has been a blur, i feel you 原帖
• 8 月 2 日 10:19 @ajambrosino（疑似回复）：@_heyglassy will do 原帖
• 8 月 2 日 10:28 @ryanbrewer（疑似回复）：@aiissent Thanks John 原帖
• 8 月 2 日 10:31 @charliermarsh（疑似回复）：@dexhorthy @charlespacker @dillon_mulroy What’s a rig. Do I need it 原帖
• 8 月 2 日 10:32 @ajambrosino（疑似回复）：@raffichill cool IDE 原帖
• 8 月 2 日 10:35 @ryanbrewer（转发）：RT @LongFormMath: 1. This is incredible. 2. For the first time, AI solved a problem that I have spent time working on (multicolor Ramsey n… 原帖
• 8 月 2 日 10:36 @TheRohanVarma（原创）：We make music too apparently 👌🏾 https://t.co/1jsG4TFM2o 原帖
• 8 月 2 日 10:48 @ryanbrewer（疑似回复）：@JohnAltman71213 @splitbycomma Lol I’m sure you can do all these in 5 minutes John 原帖
• 8 月 2 日 10:55 @charliermarsh（疑似回复）：@HSVSphere Computer is the best 原帖
• 8 月 2 日 10:56 @jxnlco（疑似回复）：@KyleSPatel @loosenedspirit @simonw Oh I never knew I could click it. Almost feels like codex view should show. And work view should close. 原帖
• 8 月 2 日 10:57 @AriX（疑似回复）：@ChrisGPT Surprisingly, it barely makes Computer Use faster at all. Computer Use doesn't produce many output tokens; the bottlenecks are more on tool call performance, prefill (inp 原帖
• 8 月 2 日 10:58 @ryanbrewer（疑似回复）：@JohnAltman71213 @splitbycomma Thanks John 原帖
• 8 月 2 日 10:59 @jxnlco（转发）：RT @_samirism: finally watched the odyssey - found some decent last minute tickets using the chatgpt work cloud browser movie a little ove… 原帖
• 8 月 2 日 11:00 @giffmana（原创）：I've been using the models alongside my experimenting, I let them look at the results etc. out of curiosity i often ask them about next steps. 10% of the time they suggest exactly 原帖
• 8 月 2 日 11:00 @steipete（原创）：That's a fairly new kind of spam. https://t.co/nwdRM5e7Hv https://t.co/7hXqYZnanw 原帖
• 8 月 2 日 11:03 @steipete（疑似回复）：@natbusa @mrza_ch Dark theme doesn't darken emails. 原帖
• 8 月 2 日 11:03 @nickbaumann_（疑似回复）：@jxnlco Probably should 原帖
• 8 月 2 日 11:03 @steipete（疑似回复）：@Sawtaytoes @raphh doesn't darken emails. 原帖
• 8 月 2 日 11:20 @steipete（转发）：RT @davis7: Add this to ur Codex config, makes it so that Codex will use the multiple choice tool outside plan mode. Makes it feel 100x b… 原帖
• 8 月 2 日 11:22 @giffmana（原创）：Damn this is way better than it should be!! Looks like the guy is @jeffgwoah and has a recorded version on youtube too, love it: https://t.co/FBuXvPMqb5 And of course from Waterloo 原帖
• 8 月 2 日 11:31 @markchen90（转发）：RT @ChrisPeikert: WOW!! 🤯 Among many jaw-dropping results, this proves NP-hardness of the Closest Vector and Nearest Codeword Problems for… 原帖
• 8 月 2 日 11:36 @nunezvice（疑似回复）：@robinebers it’s not just the resets keeping everyone around, i’d bet most users don’t even know about them. 原帖
• 8 月 2 日 11:40 @_samirism（疑似回复）：@RyanMorey lifes too short to spell things correctly for chatgpt 原帖
• 8 月 2 日 11:42 @giffmana（疑似回复）：@desert_mouse lol 原帖
• 8 月 2 日 11:43 @giffmana（疑似回复）：@cgarciae88 It's not bo10! I mean in 9 instances (scenarios) out of 10 all of the suggestions are just hyper-detail crap. 原帖
• 8 月 2 日 11:44 @giffmana（疑似回复）：@hassonofer_ It's been a long time we haven't seen @david_picard here! 原帖
• 8 月 2 日 11:51 @npew（转发）：RT @thebuggeddev: Only if education could be this interactive ❤️🔥 I've had a looong wish to build something genuinely useful through vibe… 原帖
• 8 月 2 日 11:51 @ryanbrewer（原创）：Come join the team! Really cool stuff coming up from us 原帖
• 8 月 2 日 11:53 @npew（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
• 8 月 2 日 11:58 @trekedge（疑似回复）：@rileybrown That would be lovely 原帖
• 8 月 2 日 12:11 @jxnlco（转发）：RT @sharifshameem: I just built a functioning React app by describing what I wanted to GPT-3. I'm still in awe. https://t.co/UUKSYz2NJO 原帖
• 8 月 2 日 12:13 @giffmana（疑似回复）：@iScienceLuvr For me, it's an exciting event when i overhear some ai discussion here :) Last month in a coffee i got all excited because I overheard one guy explaining how transfor 原帖
• 8 月 2 日 12:13 @andrewho03（疑似回复）：@ianchanning Certainly not without risk 原帖
• 8 月 2 日 12:18 @trekedge（原创）：I’ll take my refund in cash seeing as I paid the bill. 原帖
• 8 月 2 日 12:18 @giffmana（原创）：Damn. I guess it's starting now. Time to max out our weekly limits with Ultras and Fables to harden the things we care about, folks... (and hope no stupid filter will block us) 原帖
• 8 月 2 日 12:18 @andrewho03（疑似回复）：@johnloeber sobering 原帖
• 8 月 2 日 12:22 @giffmana（疑似回复）：@nathanbenaich mmmm you around? 原帖
• 8 月 2 日 12:23 @giffmana（疑似回复）：@SIGKITTEN Yooo are you Kylie's cat?? 原帖
• 8 月 2 日 12:25 @giffmana（疑似回复）：@pfau I remember Sberbank had a neurips booth and was doing llm. They gave out socks, i may still have them lol 原帖
• 8 月 2 日 12:27 @trekedge（疑似回复）：@briansowards @rileybrown Have you tried work mode? 原帖
• 8 月 2 日 12:28 @trekedge（疑似回复）：@alan_flexer @rileybrown Codex is essentially developer mode. It’s also unique to the desktop app for now. Work mode using codex under the hood but exists in the web and mobile. It 原帖
• 8 月 2 日 12:30 @giffmana（疑似回复）：@TheMingjie Of course 👀 https://t.co/UgS28ByNgi 原帖
• 8 月 2 日 12:31 @giffmana（疑似回复）：@sorianmaran @jeffgwoah Relevant: 原帖
• 8 月 2 日 12:34 @ryanbrewer（转发）：RT @ryanbrewer: My team at @OpenAI is hiring an investment banker! Join us to build the next generation of financial intelligence https://… 原帖
• 8 月 2 日 12:40 @andrew_n_carr（原创）：Friendly reminder for all the DeepSeek v4 flash 0731 lovers: serve the model in the same precision it was trained. Do not fall for the half bit quant dream 原帖
• 8 月 2 日 12:45 @giffmana（疑似回复）：@mandylu Is that so https://t.co/usAfwdi1Nn 原帖
• 8 月 2 日 12:48 @SebastienBubeck（转发）：RT @ChrisPeikert: 1/ Initial reactions after some hours with this groundbreaking result proving the NP-hardness of poly-approx CVP/NCP: It… 原帖
• 8 月 2 日 13:05 @ah20im（疑似回复）：@wonforall @charlierguo Believe in Believe 原帖
• 8 月 2 日 13:19 @ah20im（疑似回复）：@wonforall @charlierguo I believe in you 原帖
• 8 月 2 日 13:24 @giffmana（原创）：Did you know there's a github project called MSLK, Meta Superintelligence Labs Kernels? I didn't, lol! But it is undocumented, so I had my agents create a one-pager, originally jus 原帖
• 8 月 2 日 13:26 @giffmana（疑似回复）：@tszzl math-y people were mean to arts-y people when it was arts-y people's time. Now it is math-y people's time. 原帖
• 8 月 2 日 13:27 @santa_kaus（疑似回复）：@liangsays @knowerofmarkets @MTSlive @nejatian @Opendoor monitor 原帖
• 8 月 2 日 13:31 @ericmitchellai（疑似回复）：@KarelDoostrlnck https://t.co/rViee2i6qZ 原帖
• 8 月 2 日 13:33 @ryanbrewer（疑似回复）：@JVanden wanna bet 原帖
• 8 月 2 日 13:42 @charliermarsh（原创）：TIL that #[serde(flatten)] can have meaningful overhead, since all fields not explicitly declared on the outer struct are buffered https://t.co/uP4hwJO4P7 原帖
• 8 月 2 日 13:44 @giffmana（疑似回复）：@deliprao I don't know if they or me want to maintain this. It's more like a snapshot in time, plus an example of missing docs not even mattering anymore! 原帖
• 8 月 2 日 13:45 @jxnlco（转发）：RT @davis7: Add this to ur Codex config, makes it so that Codex will use the multiple choice tool outside plan mode. Makes it feel 100x b… 原帖
• 8 月 2 日 13:47 @yong_zhengxin（转发）：RT @MedusaOnchain: UPDATE: to anyone saying claude code scoured the whole internet for discussions by humans and then added it to their kno… 原帖
• 8 月 2 日 13:54 @giffmana（疑似回复）：@suryaasub definitely! Like, I'm in that org for a year and only randomly noticed it this week lol 原帖
• 8 月 2 日 13:57 @santa_kaus（疑似回复）：@deedydas post-economic bench 原帖
• 8 月 2 日 13:59 @santa_kaus（原创）：best spot to buy books in bulk? 原帖
• 8 月 2 日 14:01 @santa_kaus（原创）：(home library not data i promise) 原帖
• 8 月 2 日 14:08 @santa_kaus（疑似回复）：@nikesharora @StephenCurry30 @justinbieber mujhe bhi le jao uncle 原帖
• 8 月 2 日 14:13 @TairanHe99（原创）：Everyone will have their Lee Sedol moment, just at a different time. 原帖
• 8 月 2 日 14:18 @jxnlco（疑似回复）：@jefftangx @sharifshameem Can you try now 原帖
• 8 月 2 日 14:19 @guinnesschen（疑似回复）：@eidzoku Yeah, we currently use Terra-low as the backing model for voice chats. We tested a bunch of setups, and it ended up striking the best balance between capability and latenc 原帖
• 8 月 2 日 14:26 @reach_vb（疑似回复）：@shuv1337 @kcosr same, you can also add an instruction in your agents md to not block on long running goals 原帖
• 8 月 2 日 14:27 @reach_vb（疑似回复）：@OanaGoge still experimental and looking at feedback from the community 原帖
• 8 月 2 日 14:41 @ajambrosino（疑似回复）：@jp54362 true 原帖
• 8 月 2 日 15:00 @yong_zhengxin（转发）：RT @tenobrus: imo most of the "why are you mathematicians crying is the point of math to figure out true results or a jobs program for u gu… 原帖
• 8 月 2 日 15:03 @mmeigooni（疑似回复）：@_lopopolo Skills can call scripts, why not have a script that creates a specifically scoped subagent? 原帖
• 8 月 2 日 15:08 @giffmana（原创）：Just learned about a cool project in Switzerland! They're building a huge 1.2GW/2.1GWh battery, right where the CH, DE, and FR grids connect. And on top (literally) of it they buil 原帖
• 8 月 2 日 15:11 @reach_vb（疑似回复）：@nico_laqua It’s a good model sir! https://t.co/Pn8b1RXpkv 原帖
• 8 月 2 日 15:25 @giffmana（疑似回复）：@SIGKITTEN @iScienceLuvr shhh 原帖
• 8 月 2 日 15:31 @yong_zhengxin（转发）：RT @stuhlmueller: few people have had more foresight than wei dai: 1. he's been writing about the singularity since the 90s, back then on e… 原帖
• 8 月 2 日 15:38 @jxnlco（疑似回复）：@zoidz00 I joined 4 months ago. I don’t get that much equity. 原帖
• 8 月 2 日 16:00 @ajambrosino（疑似回复）：@LLMJunky oops yes 原帖
• 8 月 2 日 16:04 @santa_kaus（疑似回复）：@samuel_spitz no i just need my stack hitting the ceiling man https://t.co/mjcmlwbljR 原帖
• 8 月 2 日 16:13 @andrewho03（疑似回复）：@AISafetyMemes Wrong. You are making stuff up — seriously embarrassing yourself 原帖
• 8 月 2 日 16:18 @jxnlco（转发）：RT @athyuttamre: what is the #1 thing we should fix or add to ChatGPT Voice? 原帖
• 8 月 2 日 17:22 @reach_vb（原创）：SF, I’m in you! 🇺🇸 https://t.co/RY3z8anhbw 原帖
• 8 月 2 日 17:28 @reach_vb（原创）：man you can't make this shit up, I get in my cab and my driver is telling me about Astra, the model OpenAI is about to release solved 10 math problems !! 原帖
• 8 月 2 日 17:29 @gdb（原创）：codex for customer feedback -> roadmap 原帖
• 8 月 2 日 17:31 @reach_vb（疑似回复）：@heccbrent bro I’m having profound conversations about gpt-live rn 原帖
• 8 月 2 日 17:37 @reach_vb（疑似回复）：@romainhuet excited to meet you IRL, I think this will be the 4th country I’ll catch up with you in lol 原帖
• 8 月 2 日 17:38 @reach_vb（疑似回复）：@0xV0LYX detailed write up here: 原帖
• 8 月 2 日 17:39 @jam3scampbell（原创）：it’s a generally good heuristic to act as if you have way less time than you think you have 原帖
• 8 月 2 日 17:40 @jxnlco（疑似回复）：@Ren_Simmons @thsottiaux @ChatGPT @OpenAIDevs can you /feedback and give me id 原帖
• 8 月 2 日 17:40 @ryanbrewer（转发）：RT @JoeMcCormick01: I don’t post here often, but joining OpenAI last year has been one of the best decisions I’ve made. Proud to be workin… 原帖
• 8 月 2 日 17:46 @MostafaRohani（转发）：RT @patrickc: Someone asked me recently why I’ve become interested in aesthetics after having spent most of my life more interested in STEM… 原帖
• 8 月 2 日 17:54 @MostafaRohani（疑似回复）：@patrickc I would be so delighted to have @patrickc talk with @ezraklein and @DKThomp on a podcast. I am very abundance/YIMBY-pilled, but I also find myself repulsed by most modern 原帖
• 8 月 2 日 17:55 @nunezvice（疑似回复）：@romainhuet @reach_vb see you guys on tuesday 🤩 原帖
• 8 月 2 日 17:59 @npew（转发）：RT @LLMJunky: GPT 5.6 Sol is 10 miles ahead of every Anthropic model when it comes to computer use. They seemed to have resolved every pr… 原帖
• 8 月 2 日 18:16 @KarelDoostrlnck（原创）：this is where I codex from https://t.co/eAJB0SRfTd 原帖
• 8 月 2 日 18:25 @reach_vb（疑似回复）：@BubbleBrain nah I don’t tell I where I work and where I live for obvious reasons - which makes this even funnier 原帖
• 8 月 2 日 18:26 @reach_vb（疑似回复）：@mweinbach @Dimillian yeah gotta plan what to have for dinner 原帖
• 8 月 2 日 18:32 @w01fe（转发）：RT @nickcammarata: a lot of the anti-slowdown people also think ai just won’t go that fast so a high speed limit is probably better, above… 原帖
• 8 月 2 日 18:34 @reach_vb（疑似回复）：@heccbrent @mweinbach @Dimillian https://t.co/mPOWB3Uiv7 原帖
• 8 月 2 日 18:37 @reach_vb（疑似回复）：@SIGKITTEN @Dimillian thomas 原帖
• 8 月 2 日 18:43 @reach_vb（疑似回复）：@SIGKITTEN @Dimillian yeah ofc - goes without saying 原帖
• 8 月 2 日 18:54 @reach_vb（疑似回复）：@grabbou yess! 原帖
• 8 月 2 日 18:59 @ryanbrewer（疑似回复）：@SmallParcelGuy Haha yes it is, more reach always better though can’t hurt 原帖
• 8 月 2 日 19:09 @andrew_n_carr（疑似回复）：@garvvee We're in NYC, what sort of things are you looking to work on? 原帖
• 8 月 2 日 19:23 @aidan_mclau（疑似回复）：@KarelDoostrlnck he’s graduated from touching grass 原帖
• 8 月 2 日 19:23 @aidan_mclau（转发）：RT @goodside: ChatGPT Work (5.6 Sol Max) installs Blender and models a chess set with the board position of Byrne vs. Fischer (1956) immedi… 原帖
• 8 月 2 日 19:26 @KarelDoostrlnck（疑似回复）：@aidan_mclau In my touching water era now 原帖
• 8 月 2 日 19:28 @KarelDoostrlnck（疑似回复）：@madelinelawren @hadesdevs I bumped into an aikido one on my walk to work this week! Long way from Belgium 原帖
• 8 月 2 日 19:29 @gdb（原创）：Codex for helping operate your business 原帖
• 8 月 2 日 19:30 @KarelDoostrlnck（疑似回复）：@madelinelawren @hadesdevs * an normal advertisement not a billboard 原帖
• 8 月 2 日 19:31 @andrew_n_carr（原创）：This model just broke the harnessmaxximg approach of the closed labs. By mixing in harnesses as part of the unified reward suite, the new 2T+ Qwen model generalizes across harnesse 原帖
• 8 月 2 日 19:31 @aidan_mclau（转发）：RT @goodside: ChatGPT Work (5.6 Sol Light) installs Blender and renders a 5s video fly-through of a lvl. 5 Menger sponge, returning a video… 原帖
• 8 月 2 日 19:50 @prd_008（转发）：RT @athyuttamre: what is the #1 thing we should fix or add to ChatGPT Voice? 原帖
• 8 月 2 日 19:59 @KarelDoostrlnck（疑似回复）：@madelinelawren @hadesdevs Wholesome 🫶 原帖
• 8 月 2 日 20:14 @ajambrosino（疑似回复）：@rithulbm and that version is still somehow live, quite unchanged 原帖
• 8 月 2 日 20:15 @ajambrosino（疑似回复）：@TheAgileMaker somewhere between 10 and a few thousand 原帖
• 8 月 2 日 20:15 @ajambrosino（疑似回复）：@SIGKITTEN yep 原帖
• 8 月 2 日 20:15 @ajambrosino（疑似回复）：@soumitrashukla9 I do 原帖
• 8 月 2 日 20:36 @aidan_mclau（转发）：RT @goodside: ChatGPT Work (5.6 Sol High) installs Blender and models a chess set with the board position of Byrne vs. Fischer (1956) follo… 原帖
• 8 月 2 日 20:37 @reach_vb（疑似回复）：@dkundel @heccbrent sorry sensei - will be more American next time 原帖
• 8 月 2 日 20:38 @JoanneShang（原创）：TIL: “ack” in Slack does not mean ACK 😱 原帖
• 8 月 2 日 21:28 @garymlin（原创）：TIL anthropic's github org username is "anthropics" and "anthropic" is owned by an Aussie, Marcel J Bennett https://t.co/uvh9yxO0jh 原帖
• 8 月 2 日 21:33 @andrew_n_carr（原创）：The Qwen 3.8 max post hinted at compaction working well in their 10+ day run and their paper reimplementation. Although it didn't seem like they targeted this specifically? In any 原帖
• 8 月 2 日 21:34 @jxnlco（转发）：RT @JoeMcCormick01: I don’t post here often, but joining OpenAI last year has been one of the best decisions I’ve made. Proud to be workin… 原帖
• 8 月 2 日 21:40 @andrewho03（转发）：RT @andrewho03: I'm not that worried about AI safety, but that's because I think that AI might be humanity's last hope before descending in… 原帖
• 8 月 2 日 21:41 @jxnlco（转发）：RT @BHolmesDev: I just tried the new ChatGPT voice mode. I'm... having a moment right now. It let me monologue for 10 minutes, took notes… 原帖
• 8 月 2 日 22:15 @andrewho03（转发）：RT @thewaifuai: pretty much yea 原帖
• 8 月 2 日 22:38 @andrewho03（转发）：RT @cremieuxrecueil: I know at least two dozen people who pursued careers in AI because they see it as the only route to overcome their pes… 原帖
• 8 月 2 日 23:02 @jxnlco（转发）：RT @hannahkoch: @jxnlco is so talented and a great reminder you can just do things. like make an OpenAI mission video. Thank you for maki… 原帖
• 8 月 2 日 23:51 @andrewho03（疑似回复）：@airkatakana Don't know how I put up with this shit as a high school EOP 原帖
• 8 月 3 日 00:00 @dylancablayan（疑似回复）：@creatine_cycle I think someone paid an Etsy witch or something, I got hit with a meniscus tear a week ago 原帖
• 8 月 3 日 01:01 @JustinBleuel（疑似回复）：@argofowl @cgmelamed great minds 原帖
• 8 月 3 日 01:04 @yong_zhengxin（转发）：RT @Jacob_Tsimerman: A couple of websites I want to share a couple websites aimed specifically at mathematicians regarding AI-safety and A… 原帖
• 8 月 3 日 01:25 @andrewho03（疑似回复）：@shaurizard @catboosted Many truth nukes in this thread. 原帖
• 8 月 3 日 01:37 @thsottiaux（原创）：Crazy thing about OpenAI is you just open your laptop, codex a PR into existence and ship an improvement to 1B users. Just like that. 原帖
• 8 月 3 日 03:26 @reach_vb（原创）：gm chat, the jet lag is upon us https://t.co/jDGVb5X8k3 原帖
• 8 月 3 日 03:36 @reach_vb（疑似回复）：@threepointone oh that’s a brill shout 原帖
• 8 月 3 日 04:03 @giffmana（疑似回复）：@DimitrisPapail 3 is exactly the same in coding btw. All my agentsmd files contain instructions to (1) not introduce one/two-time use variables and prefer inlining instead. And ( 原帖
• 8 月 3 日 04:03 @reach_vb（疑似回复）：@swyx @akshaynathan_ @AriX thumbnail looks like you’re both angry at codex 原帖
• 8 月 3 日 04:55 @reach_vb（疑似回复）：@pvncher to be fair, I’m the same 原帖
• 8 月 3 日 06:30 @gabrielchua（转发）：RT @goodside: ChatGPT Work (5.6 Sol Max) installs Blender and models a chess set with the board position of Byrne vs. Fischer (1956) immedi… 原帖
• 8 月 3 日 06:31 @JustinBleuel（疑似回复）：@itsnotbernhard @villi is likely there 原帖
• 8 月 3 日 06:57 @gabrielchua（疑似回复）：@Dimillian smh 原帖
• 8 月 3 日 06:57 @gabrielchua（转发）：RT @JoeMcCormick01: I don’t post here often, but joining OpenAI last year has been one of the best decisions I’ve made. Proud to be workin… 原帖
• 8 月 3 日 07:13 @andrew_n_carr（疑似回复）：@garvvee Super cool! I haven't played in a bit, trying to look in, but I did 40 line sprints and battle :) We don't have much learning theory work - I know AMI does though, and the 原帖
• 8 月 3 日 07:23 @gabrielchua（转发）：RT @goodside: ChatGPT Work (5.6 Sol Light) installs Blender and renders a 5s video fly-through of a lvl. 5 Menger sponge, returning a video… 原帖
• 8 月 3 日 07:23 @charliermarsh（原创）：A great side-effect of Codex is that all my work becomes legible... I can ask Codex how many times we've run X or Y command over the past week, or how much time we've lost on A or 原帖
• 8 月 3 日 07:31 @reach_vb（疑似回复）：@threepointone holy shitt! that’s pretty! is today your birthday? if so, happy birthdayyyyy 原帖
• 8 月 3 日 07:41 @nunezvice（疑似回复）：@derrickcchoi 🥲 原帖
• 8 月 3 日 07:45 @jxnlco（原创）：you watch youtube videos instead of reaching source books from the library that were hand written by a monk 原帖
• 8 月 3 日 07:58 @charliermarsh（疑似回复）：@lcm_in_ai @jbrown_acuity @kettlecorn Doesn't this mean all projects on your machine share the same build lock? 原帖
• 8 月 3 日 07:59 @j_asminewang（转发）：RT @Jacob_Tsimerman: A couple of websites I want to share a couple websites aimed specifically at mathematicians regarding AI-safety and A… 原帖
• 8 月 3 日 07:59 @andrew_n_carr（疑似回复）：@garvvee Insanity! I was, many moons ago, ranked - I hit a couple of 19s but usually sat in the mid to high twenties depending 原帖

Anthropic（10 个活跃账号，29 条）

• 8 月 2 日 08:46 @karpathy（疑似回复）：@Izkimar love it! :) 原帖
• 8 月 2 日 09:02 @minilek（疑似回复）：@LocasaleLab This shows that those places wanted MIT PhDs. It doesn’t show that most MIT PhDs wanted them back. I remember when KAUST was founded — they tried to entice MIT undergr 原帖
• 8 月 2 日 09:10 @karpathy（原创）：More on the pelican on the bicycle test from @simonw: https://t.co/OXmtODyTKj I uploaded the source here so it's playable in the browser, forkable etc. https://t.co/w3Nctc888d Look 原帖
• 8 月 2 日 09:19 @deedydas（疑似回复）：@drBalls0 I could give you more modern examples, beyond the ones included (Dylan / Dwarkesh). Younger people tend to be less ubiquitously well known but plenty examples in engineer 原帖
• 8 月 2 日 10:17 @katelyn_lesse（疑似回复）：@thdxr agreed, appreciate your take. jess is a PM on our Platform team who shipped Claude Managed Agents (an agent api suite). her point here is that a harness+model pair where the 原帖
• 8 月 2 日 10:31 @minilek（疑似回复）：@LocasaleLab As far as missing the point, my point is you haven't demonstrated any convincing evidence supporting your point. I also retract my previous statement "This shows tha 原帖
• 8 月 2 日 11:06 @trq212（原创）：you can already see Jevons paradox at work in mathematics there is more happening, it is easier to understand and mathematicians have more time to discuss it with us at higher abst 原帖
• 8 月 2 日 11:07 @trq212（原创）：i think there are lots of parallels to what happened with chess 原帖
• 8 月 2 日 11:37 @deedydas（原创）：I love dropping floor plans of houses in the Bay Area and asking LLMs to redesign in a different aesthetic, Kyoto in this case. Models have gotten phenomenal at 3D. https://t.co/NU 原帖
• 8 月 2 日 11:45 @chrisrohlf（转发）：RT @defendtheworld: It's BlackHat/DEF CON week so I'd like to interrupt your regularly scheduled chest drumming feed! Let's talk about a cr… 原帖
• 8 月 2 日 11:48 @deedydas（疑似回复）：@hmsmotapa Statistics can lie too. I don’t have that numbers but I reckon if you look at Stanford graduate outcomes in the Bay Area vs any other school the former will do much bett 原帖
• 8 月 2 日 11:48 @deedydas（疑似回复）：@_RohitSurana This was Fable 原帖
• 8 月 2 日 12:07 @itsclivetime（转发）：RT @calxoxo: tag your vc 原帖
• 8 月 2 日 12:47 @deedydas（原创）：Here’s a link if you want to play with it https://t.co/uHj1Ty0uMp 原帖
• 8 月 2 日 13:45 @karpathy（疑似回复）：@rainisto @DavidmComfort agree!! i quite like the idea of procedural code for storyboarding and control, and then video to video models for texturing and looksmaxxing. 原帖
• 8 月 2 日 14:16 @AlecTPhD（疑似回复）：@alxfazio Studying them is a good starting point for sure, but the truly best way to learn is to build your own. 原帖
• 8 月 2 日 15:49 @jonchu（疑似回复）：@coolcoder56 This one is easy. Leetcode gives you the answer key to the test ahead of time so passing it isn't that high a bar anymore. Imagine never practicing a single coding or 原帖
• 8 月 2 日 15:53 @jonchu（疑似回复）：@patrickc I think you'd enjoy talking to @armon who's heading to architecture school after Hashicorp 🙂 原帖
• 8 月 2 日 15:55 @jonchu（疑似回复）：@ericbahn I think it was a little post 2012 iirc. CS really took off when Google FB did their talent war and pushed salaries up to 300-400k for senior folks. That was around 2013-2 原帖
• 8 月 2 日 16:20 @karpathy（疑似回复）：@MattBeton so fun! :) at some point i wonder if ngram (tables) or even something like decision trees start to give superior log probs, and at much smaller program lengths overall ( 原帖
• 8 月 2 日 17:54 @karpathy（疑似回复）：@threejs I was curious and gave your feedback to Opus. TIL! https://t.co/t5HvyZsBlh 原帖
• 8 月 2 日 20:32 @EthanJPerez（转发）：RT @Jacob_Tsimerman: A couple of websites I want to share a couple websites aimed specifically at mathematicians regarding AI-safety and A… 原帖
• 8 月 2 日 21:28 @deedydas（转发）：RT @deedydas: I love dropping floor plans of houses in the Bay Area and asking LLMs to redesign in a different aesthetic, Kyoto in this cas… 原帖
• 8 月 2 日 22:53 @deedydas（原创）：8 of the 50 most congested cities in the world are in India. Bangalore, at 3:37min/km is only 1.4x faster than running. Poland (!), Brazil and Columbia are #2-4. Commuters lose 39m 原帖
• 8 月 2 日 22:53 @deedydas（原创）：Source: https://t.co/zQcpd3b1zQ 原帖
• 8 月 3 日 06:16 @chrisrohlf（原创）：Your contract with a software vendor is one of the highest security leverage points you have. Demand they use the latest frontier models to continually improve the security of that 原帖
• 8 月 3 日 06:52 @minilek（原创）：This was a fun project and one of the last things I worked on at Google. dynamic rank/select is a basic data structural problem: maintain a database of items in {1,...,N} subject t 原帖
• 8 月 3 日 07:21 @chrisrohlf（原创）："Rewrite vs Import" is the new "Build vs Buy". The OSS footprint of your average enterprise is going to shrink over time. What remains will represent a calculated decision where th 原帖
• 8 月 3 日 07:21 @chrisrohlf（原创）：But at the same time token costs are coming down the cost of dependency management, including vulnerability management, is arguably rising, at least in the short to medium term. Ev 原帖

Google DeepMind（8 个活跃账号，30 条）

• 8 月 2 日 08:48 @vitaliychiley（原创）：pretty soon, this graph will need ax.set_yscale('log') 原帖
• 8 月 2 日 11:07 @Majumdar_Ani（原创）：Real-time uncertainty quantification from Gemini Robotics 2 原帖
• 8 月 2 日 14:55 @ekindogus（转发）：RT @sschoenholz: @starsailing11 Good question! I tend to think that ML is the tide that lifts all ships, so we'll probably see a lot of thi… 原帖
• 8 月 2 日 17:40 @wenhaocha1（转发）：RT @machinestein: Recently we found a similar algorithm in the RL for LLMs domain. Updating on max(R) leads to better exploration under bes… 原帖
• 8 月 2 日 17:57 @wenhaocha1（转发）：RT @KL_Div: Was just as excited as everyone else to read this cool new paper, and it felt like a trip down memory lane! XM seems to be the… 原帖
• 8 月 2 日 20:11 @suchenzang（原创）：probed it with a math question: seems to have the formatting style of sol and the reasoning laziness of opus, but the weights are dropping next week so i'm definitely not complaini 原帖
• 8 月 2 日 20:12 @suchenzang（原创）：also muting everyone who QT'd with zero additional information-add improve your cogsec, kids 原帖
• 8 月 2 日 21:18 @suchenzang（疑似回复）：@Xxi5olc are you complaining or gloating or ratioing or 原帖
• 8 月 2 日 21:52 @suchenzang（原创）：so... the explorative-modeling-third-pretraining-axis thing was effectively plagiarism? 👀 原帖
• 8 月 2 日 22:03 @suchenzang（原创）：incredible, they even plagiarized the "mode forcing" motivation, citing a 34-page preprint paper (https://t.co/JlbC72KeQX) that reads like pure claude slop, instead of the 2023 pap 原帖
• 8 月 2 日 22:04 @suchenzang（疑似回复）：@du_yilun are you aware of this ^ ? 原帖
• 8 月 2 日 22:06 @suchenzang（疑似回复）：@ErrkC did you read the thread? 原帖
• 8 月 2 日 22:31 @suchenzang（疑似回复）：@du_yilun yeah... not a great look 原帖
• 8 月 2 日 23:31 @suchenzang（疑似回复）：@ErrkC oh no 原帖
• 8 月 2 日 23:54 @suchenzang（疑似回复）：@bozavlado the claude slop citation violence is even more egregious, i think 原帖
• 8 月 3 日 00:06 @suchenzang（疑似回复）：@eric_alcaide not quite 原帖
• 8 月 3 日 00:07 @suchenzang（疑似回复）：@YouJiacheng that's not quite the worst thing 原帖
• 8 月 3 日 00:14 @suchenzang（疑似回复）：@YouJiacheng so you wouldn't be writing a 34 page paper describing this trivial motivation then? 原帖
• 8 月 3 日 00:20 @suchenzang（疑似回复）：@eric_alcaide yeah, they hallucinated a whole separate paper, crazy! https://t.co/78dOPcuJoy 原帖
• 8 月 3 日 00:30 @suchenzang（原创）：didn't even notice at first, but apparently this is what the fellows at flappy airplanes get up to - creating slop papers to cite in more slop work seems like that neolabs' fellows 原帖
• 8 月 3 日 00:32 @suchenzang（疑似回复）：@AlexiGlad rip https://t.co/LsTQYWFniM 原帖
• 8 月 3 日 01:11 @xiao_ted（疑似回复）：@porterstansb @pangram 原帖
• 8 月 3 日 04:37 @_philschmid（疑似回复）：@PaulPlayStudio Hey 👋 Part of gemini team, would love to help you stay successful. Feel free to reach out via DM. 原帖
• 8 月 3 日 06:53 @vitaliychiley（转发）：RT @OfficialLoganK: Reminder: you are probably not model pilled enough. However much you are betting on the models, it’s likely under the… 原帖
• 8 月 3 日 06:54 @vitaliychiley（疑似回复）：@sonofalli https://t.co/ckviAuDayV 原帖
• 8 月 3 日 07:16 @_philschmid（疑似回复）：@PaulPlayStudio Did you already try 3.5 flash lite on your setup? What are the results? 原帖
• 8 月 3 日 07:26 @suchenzang（原创）：incredible. not only does this "mode-forcing draft paper" (cited in the potentially mischaracterized-as-plagiarism work) cite Yet Another Unpublished Draft Paper, it is So Thick Wi 原帖
• 8 月 3 日 07:33 @JaidevShah4（原创）：Yup It’s also a lot more fun to operate with this assumption 原帖
• 8 月 3 日 07:46 @suchenzang（疑似回复）：@doomslide @tekbog rip, still there's some nonzero entertainment value to seeing how a cult classic opines, like, artistically 原帖
• 8 月 3 日 07:52 @suchenzang（原创）：i've been informed of the possibility that this crew could be pioneering an AI slop citation scheme, and i wouldn't be any the wiser because i have honestly not considered such thi 原帖

Meta AI（4 个活跃账号，17 条）

• 8 月 2 日 08:41 @HeinrichKuttler（疑似回复）：@Noahpinion and the context was this, arguably unrelated, tweet 原帖
• 8 月 2 日 08:42 @HeinrichKuttler（疑似回复）：@ziv_ravid it seems to influence my feed 原帖
• 8 月 2 日 08:58 @HeinrichKuttler（疑似回复）：@SunnyDfan4eva "If I were to awaken after having slept for a hundred days, my first question would be: has the Riemann hypothesis been proven?" - Hilbert, allegedly 原帖
• 8 月 2 日 09:04 @HeinrichKuttler（疑似回复）：@liberzonn decade? https://t.co/d9Gd1M4oRg 原帖
• 8 月 2 日 10:10 @HeinrichKuttler（疑似回复）：@kai_fell literally nothing is more complex than C++ 原帖
• 8 月 2 日 10:18 @HeinrichKuttler（疑似回复）：@QualiaQuanta crank bs 原帖
• 8 月 2 日 10:25 @HeinrichKuttler（疑似回复）：@SunnyDfan4eva incidentally apocryphal, with a Barbarossa reference (if only!) https://t.co/Mj9lLx0Aub 原帖
• 8 月 2 日 10:30 @HeinrichKuttler（疑似回复）：@gauravisnotme I'm not convinced we still need either. Let the AI write your kernels. 原帖
• 8 月 2 日 11:15 @HeinrichKuttler（疑似回复）：@QualiaQuanta after considering the responses, let me correct myself: definitely crank bs 原帖
• 8 月 2 日 13:36 @maddiehfaulkner（疑似回复）：@benln I’m confused why factory isn’t on this — we 2x’d in size in the last 3 mos 原帖
• 8 月 2 日 14:06 @HeinrichKuttler（疑似回复）：@finbarrtimbers @thsottiaux @_chenglou it was eventually released but the trade off between readiness and speed wasn't right in retrospect 原帖
• 8 月 2 日 14:28 @HeinrichKuttler（原创）：where is the bbc? where is German public tv (budget a slim $10B annually!)? random media startups report from Ceuta. 原帖
• 8 月 2 日 14:49 @HeinrichKuttler（疑似回复）：@abeirami banger 原帖
• 8 月 2 日 17:19 @honglin_c（转发）：RT @xf1280: Testing LLM by asking it to create your favorite game that you feel nostalgic about. Everything here is procedurally generated;… 原帖
• 8 月 2 日 22:29 @EdwardSun0909（原创）：Congrats to the Qwen CUA team on releasing a sota model! 原帖
• 8 月 3 日 00:28 @HeinrichKuttler（疑似回复）：@Ilya_Kuprov real life impact from alphafold so far has been underwhelming 原帖
• 8 月 3 日 02:58 @HeinrichKuttler（疑似回复）：@geoffreyirving @davidad Agreed. 原帖

Thinking Machines Lab（4 个活跃账号，12 条）

• 8 月 2 日 11:27 @yilin_yang721（疑似回复）：@extrasalt1 @quxiaoyin Well chatgpt isn’t usage based, and it has 10b active user. People say gemini is catching up but I doubt it 原帖
• 8 月 2 日 12:23 @yilin_yang721（疑似回复）：@extrasalt1 @quxiaoyin I think openai’s product design a moat, both chatgpt and codex feels like an apple product 原帖
• 8 月 2 日 12:49 @yilin_yang721（疑似回复）：@Sentdex Idk why a simple question prompt can run two hours with Opus5 原帖
• 8 月 2 日 14:05 @Mascobot（疑似回复）：@berman66 👀 原帖
• 8 月 2 日 14:15 @Mascobot（疑似回复）：@tunguz 😆 原帖
• 8 月 2 日 14:34 @ziqiao_ma（疑似回复）：@anpaure What benchmark did you use? I use SimpleQA with low reasoning efforts as my personal estimate metric.👀 原帖
• 8 月 2 日 14:37 @ziqiao_ma（疑似回复）：@anpaure Interesting let me read 🙏 原帖
• 8 月 2 日 17:41 @Mascobot（原创）：For most industrial applications, robots don’t need legs. The reality is that most industrial spaces are already designed and optimized for wheeled equipment (carts, dollies, forkl 原帖
• 8 月 2 日 17:48 @Mascobot（疑似回复）：@fabrisera2000 Wait are you a DJ? 原帖
• 8 月 2 日 18:18 @x_hanyu（原创）：how to get one 原帖
• 8 月 2 日 18:21 @Mascobot（疑似回复）：@JasonMa2020 Yeah for anything outside of industrial stuff, they have their place for sure 原帖
• 8 月 3 日 02:30 @Mascobot（原创）：💯 And how fast this is about to happen 原帖

Cursor（1 个活跃账号，10 条）

• 8 月 2 日 20:53 @benln（疑似回复）：@TedMerz https://t.co/ajeueQ5gR4 原帖
• 8 月 2 日 20:53 @benln（疑似回复）：@axiospoint headcount growth %, pulled from https://t.co/ajeueQ5gR4 原帖
• 8 月 2 日 22:48 @benln（疑似回复）：@nelopuchades just trying to be helpful to folks looking for their next thing! 原帖
• 8 月 3 日 02:38 @benln（疑似回复）：@k1euu exciting! 原帖
• 8 月 3 日 06:32 @benln（原创）：Cafe Cursor takeovers coming to 24 countries in August 🇨🇦🇮🇩🇦🇷🇹🇭🇲🇦🇮🇳🇺🇸🇿🇲🇰🇪🇿🇦🇦🇴🇪🇨🇧🇷🇻🇳🇪🇹🇸🇻🇨🇷🇨🇱🇷🇼🇵🇱🇦🇹🇬🇹🇵🇦🇨🇭 原帖
• 8 月 3 日 07:29 @benln（疑似回复）：@iamfnote https://t.co/5M0QKNUjF9 原帖
• 8 月 3 日 07:29 @benln（疑似回复）：@mmmikhaeel https://t.co/5M0QKNUjF9 原帖
• 8 月 3 日 07:29 @benln（疑似回复）：@eve_bouff yesterday! https://t.co/7mfsCvJgk4 原帖
• 8 月 3 日 07:38 @benln（疑似回复）：@marco_dewey https://t.co/5M0QKNUjF9 原帖
• 8 月 3 日 07:39 @benln（疑似回复）：@akshit_mokani Pune: https://t.co/MdpqMpSMFd Bangalore: https://t.co/IcrCZqhnj2 原帖

Alibaba Qwen（4 个活跃账号，25 条）

• 8 月 2 日 10:30 @BowenWangNLP（原创）：Check out what we have been building for over 1 year, kudos to @DunjieLu1219 原帖
• 8 月 2 日 11:15 @TianbaoX（原创）：Finally comes out after nearly 1 year! Congrats the team and who remains 🍾 原帖
• 8 月 2 日 11:26 @TianbaoX（疑似回复）：@DunjieLu1219 Hope/(and not hope) you are using the infra I built. 原帖
• 8 月 2 日 18:44 @huybery（原创）：Cool!! building games is a great way to test LLM. It covers reasoning, planning, coding, tool (lib) use, multimdal understanding, and more. But it seems like the community still la 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：📢 Meet Qwen3.8-Max — our most capable model to date. Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉 Qwen3. 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：LM Performance https://t.co/zMUpEbIupH 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：VL Performance https://t.co/gsuM0XTYTY 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：16 days autonomous coding https://t.co/OX0RN7itxP 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：Cowork: Any role, build beyond https://t.co/Hggkj3DFSs 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：Dynamic workflows to quant strategies https://t.co/cGL1EeJdeK 原帖
• 8 月 2 日 19:15 @Alibaba_Qwen（原创）：Visual agentic intelligence https://t.co/bUhsuYAqzb 原帖
• 8 月 2 日 19:27 @Alibaba_Qwen（原创）：Text Arena: https://t.co/FyzZab0cCK https://t.co/0NGPCr5Dkh 原帖
• 8 月 2 日 19:57 @Alibaba_Qwen（原创）：🏆 Alibaba hits #2 on the Text Arena. https://t.co/s3YQ70T7zi 原帖
• 8 月 2 日 20:07 @Alibaba_Qwen（原创）：LM Performance - English Version https://t.co/Bi9WefHHLB 原帖
• 8 月 2 日 20:09 @Alibaba_Qwen（原创）：VL Performance - English Version https://t.co/95NsmZI1hz 原帖
• 8 月 2 日 20:25 @huybery（疑似回复）：@0xkato thanks! 原帖
• 8 月 2 日 20:56 @Alibaba_Qwen（疑似回复）：@cline Appreciate the shoutout! Can't wait to see Qwen3.8 in Cline soon. 🚀 原帖
• 8 月 2 日 21:55 @Alibaba_Qwen（疑似回复）：@GenAI_is_real Appreciate your support!❤️ 原帖
• 8 月 2 日 23:15 @Alibaba_Qwen（原创）：Thanks for the recognition. We'll keep building! 🚀 原帖
• 8 月 2 日 23:32 @Alibaba_Qwen（原创）：Appreciate it! Now let's understand the world through the eyes of Qwen3.8. 🥳 原帖
• 8 月 3 日 01:05 @Alibaba_Qwen（疑似回复）：@chamath We appreciate it! Stay tuned for what's next. 原帖
• 8 月 3 日 02:30 @Alibaba_Qwen（原创）：Let's Go!🚀 原帖
• 8 月 3 日 02:31 @Alibaba_Qwen（原创）：Live in Command Code!🎉 原帖
• 8 月 3 日 03:50 @Alibaba_Qwen（疑似回复）：@Teknium Thanks! Looking forward to seeing Qwen3.8-Max on Hermes Agent! 原帖
• 8 月 3 日 07:09 @Alibaba_Qwen（原创）：Ask anything, anonymously. Qwen3.8-Max has landed on Venice. Give it a try! 原帖

Cognition（3 个活跃账号，6 条）

• 8 月 2 日 09:23 @dabit3（转发）：RT @rohanpaul_ai: Shenzhen’s DAMODA has a mobile drone system that charges, deploys, and launches light shows directly from the vehicle. At… 原帖
• 8 月 2 日 09:31 @jeffwang（原创）：Thanks again for the hospitality in Singapore! First Sanjeev Tripathi and I did a chat at @Infosys with many financial service leaders about unlocking the value of AI. The biggest 原帖
• 8 月 2 日 15:45 @dabit3（疑似回复）：@superligious @jamonholmgren https://t.co/VFg5CCqpG3 原帖
• 8 月 2 日 18:07 @jhanikhil（疑似回复）：@SIGKITTEN hmm wait i use this one all the time and it works fine, maybe the real q is what plane / airline 原帖
• 8 月 2 日 18:45 @dabit3（疑似回复）：@ben_ai_eng looks awesome 原帖
• 8 月 3 日 07:06 @dabit3（转发）：RT @jonathan_wilke: Over the weekend I tried @DevinAI for some supastarter tasks and man, this thing has become reaaaaally good. Did you… 原帖

Moonshot/Kimi（3 个活跃账号，4 条）

• 8 月 2 日 08:50 @ShengjieWa34067（转发）：RT @huybery: I think coding agents have driven much of AI’s progress over the past two years. Do you see any new direction that feels equal… 原帖
• 8 月 2 日 18:47 @ShengjieWa34067（转发）：RT @xf1280: Testing LLM by asking it to create your favorite game that you feel nostalgic about. Everything here is procedurally generated;… 原帖
• 8 月 3 日 04:51 @Kimi_Moonshot（原创）：Build Slides with Kimi Work - Tutorial #1. Kimi Slides handles the entire slide-building process: - Clear structure and research, powered by Kimi K3 - Cohesive design, including po 原帖
• 8 月 3 日 04:52 @crystalsssup（转发）：RT @Kimi_Moonshot: Build Slides with Kimi Work - Tutorial #1. Kimi Slides handles the entire slide-building process: - Clear structure and… 原帖

ByteDance Seed（1 个活跃账号，4 条）

• 8 月 2 日 21:50 @GenAI_is_real（原创）：Congrats to the Qwen team! The recent wave of model releases across the ecosystem has been genuinely impressive, but what stands out to me here is the continued commitment to open 原帖
• 8 月 2 日 22:00 @GenAI_is_real（原创）：Ha! Thanks! my whole local pipeline is Qwen: ASR to transcribe, then LLM to process, all running locally. this stack handles a real chunk of my work and life now - i even drafted t 原帖
• 8 月 2 日 22:03 @GenAI_is_real（原创）：If a chicken can't make it into KFC or McDonald's, can't get into Chick-fil-A, can't even land at a local fried chicken joint — does that chicken's life have any meaning at all? — 原帖
• 8 月 2 日 22:03 @GenAI_is_real（原创）：如果一只鸡进不了肯德基、麦当劳这种大企业，进不了 Chick-fil-A，甚至连正新鸡排都进不了，这样的鸡生还有任何意义么？ ——读某些 linkedin 帖子有感 原帖

xAI（2 个活跃账号，2 条）

• 8 月 2 日 13:24 @JasonBud（疑似回复）：@bradshannon @Baconbrix is working on adding a project view. analytics is also a great idea. which metrics would be most useful for you? 原帖
• 8 月 2 日 16:46 @hyhieu226（原创）：The beginning of singularity. 原帖

AMI Labs / ex-Meta AI（1 个活跃账号，15 条）

• 8 月 2 日 13:01 @ylecun（转发）：RT @alex_peys: maybe i am lucky but across my work (ai+bio research) and relevant hobbies (music, motorsports) there have been 0 model rele… 原帖
• 8 月 2 日 13:26 @ylecun（疑似回复）：@KellyScaletta 😂😂😂 原帖
• 8 月 2 日 13:30 @ylecun（转发）：RT @ClementDelangue: It's not time to slow down but to accelerate! The recent AI-powered cyberattacks have everyone talking about the risk… 原帖
• 8 月 2 日 23:44 @ylecun（疑似回复）：@MRatable 😂 原帖
• 8 月 3 日 00:25 @ylecun（疑似回复）：@martin_casado Mathematics and computing have had a very tight relationship in some corners of academia. Richard Courant proposed methods to solve PDEs numerically in 1928. Von Neu 原帖
• 8 月 3 日 00:38 @ylecun（疑似回复）：@willdepue Using optimization at inference time is a foundational concept of Energy-Based Models (EBM) and Objective-Driven AI architectures (ODAI). When the variables to be inferr 原帖
• 8 月 3 日 00:38 @ylecun（转发）：RT @ylecun: @willdepue Using optimization at inference time is a foundational concept of Energy-Based Models (EBM) and Objective-Driven AI… 原帖
• 8 月 3 日 00:39 @ylecun（转发）：RT @Dan_Jeffries1: All Chinese AI company media: AI will be a wondrous and wonderful and give you back precious time in your life to do thi… 原帖
• 8 月 3 日 00:41 @ylecun（疑似回复）：@pmddomingos 😂 原帖
• 8 月 3 日 01:02 @ylecun（疑似回复）：@olivier_evan @martin_casado Peut-être ne savez-vous pas que j'ai travaillé sur l'OCR dans les années 80 et 90 ? Les techniques que j'ai proposées à l'époque sont à la base des mét 原帖
• 8 月 3 日 01:06 @ylecun（疑似回复）：@ByShovelINC @martin_casado False. In many universities, computer science was a spinoff of electrical engineering, and is still part of the School of Engineering. In some universit 原帖
• 8 月 3 日 02:14 @ylecun（疑似回复）：@Andrea__M 😂 Saved by the Moravec paradox ! 原帖
• 8 月 3 日 02:33 @ylecun（疑似回复）：@ramez SV-shaped bubble? 原帖
• 8 月 3 日 03:31 @ylecun（疑似回复）：@nxthompson 😱 原帖
• 8 月 3 日 03:32 @ylecun（疑似回复）：@halsinger 😱😱😱 原帖